A drug prescription recommendation system based on novel DIAKID ontology and extensive semantic rules

According to the World Health Organization (WHO) data from 2000 to 2019, the number of people living with Diabetes Mellitus and Chronic Kidney Disease (CKD) is increasing rapidly. It is observed that Diabetes Mellitus increased by 70% and ranked in the top 10 among all causes of death, while the rate of those who died from CKD increased by 63% and rose from the 13th place to the 10th place. In this work, we combined the drug dose prediction model, drug-drug interaction warnings and drugs that potassium raising (K-raising) warnings to create a novel and effective ontology-based assistive prescription recommendation system for patients having both Type-2 Diabetes Mellitus (T2DM) and CKD. Although there are several computational solutions that use ontology-based systems for treatment plans for these type of diseases, none of them combine information analysis and treatment plans prediction for T2DM and CKD. The proposed method is novel: (1) We develop a new drug-drug interaction model and drug dose ontology called DIAKID (for drugs of T2DM and CKD). (2) Using comprehensive Semantic Web Rule Language (SWRL) rules, we automatically extract the correct drug dose, K-raising drugs, and drug-drug interaction warnings based on the Glomerular Filtration Rate (GFR) value of T2DM and CKD patients. The proposed work achieves very competitive results, and this is the first time such a study conducted on both diseases. The proposed system will guide clinicians in preparing prescriptions by giving necessary warnings about drug-drug interactions and doses.

Variations in Electronic Health Record-Based Definitions of Diabetic Retinopathy Cohorts: A Literature Review and Quantitative Analysis

Purpose

Use of the electronic health record (EHR) has motivated the need for data standardization. A gap in knowledge exists regarding variations in existing terminologies for defining diabetic retinopathy (DR) cohorts. This study aimed to review the literature and analyze variations regarding codified definitions of DR.

Design

Literature review and quantitative analysis.

Subjects

Published manuscripts.

Methods

Four graders reviewed PubMed and Google Scholar for peer-reviewed studies. Studies were included if they used codified definitions of DR (e.g., billing codes). Data elements such as author names, publication year, purpose, data set type, and DR definitions were manually extracted. Each study was reviewed by ≥ 2 authors to validate inclusion eligibility. Quantitative analyses of the codified definitions were then performed to characterize the variation between DR cohort definitions.

Main Outcome Measures

Number of studies included and numeric counts of billing codes used to define codified cohorts.

Results

In total, 43 studies met the inclusion criteria. Half of the included studies used datasets based on structured EHR data (i.e., data registries, institutional EHR review), and half used claims data. All but 1 of the studies used billing codes such as the International Classification of Diseases 9th or 10th edition (ICD-9 or ICD-10), either alone or in addition to another terminology for defining disease. Of the 27 included studies that used ICD-9 and the 20 studies that used ICD-10 codes, the most common codes used pertained to the full spectrum of DR severity. Diabetic retinopathy complications (e.g., vitreous hemorrhage) were also used to define some DR cohorts.

Conclusions

Substantial variations exist among codified definitions for DR cohorts within retrospective studies. Variable definitions may limit generalizability and reproducibility of retrospective studies. More work is needed to standardize disease cohorts.

Financial Disclosures

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Using machine learning to link electronic health records in cancer registries: On the tradeoff between linkage quality and manual effort

BACKGROUND

Cancer registries link a large number of electronic health records reported by medical institutions to already registered records of the matching individual and tumor. Records are automatically linked using deterministic and probabilistic approaches; machine learning is rarely used. Records that cannot be matched automatically with sufficient accuracy are typically processed manually. For application, it is important to know how well record linkage approaches match real-world records and how much manual effort is required to achieve the desired linkage quality. We study the task of linking reported records to the matching registered tumor in cancer registries.

METHODS

We compare the tradeoff between linkage quality and manual effort of five machine learning methods (logistic regression, random forest, gradient boosting, neural network, and a stacked method) to a deterministic baseline. The record linkage methods are compared in a two-class setting (no-match/ match) and a three-class setting (no-match/ undecided/ match). A cancer registry collected and linked the dataset consisting of categorical variables matching 145,755 reported records with 33,289 registered tumors.

RESULTS

In the two-class setting, the gradient boosting, neural network, and stacked models have higher accuracy and F1 score (accuracy: 0.968-0.978, F1 score: 0.983-0.988) than the deterministic baseline (accuracy: 0.964, F1 score: 0.980) when the same records are manually processed (0.89% of all records). In the three-class setting, these three machine learning methods can automatically process all reported records and still have higher accuracy and F1 score than the deterministic baseline. The linkage quality of the machine learning methods studied, except for the neural network, increase as the number of manually processed records increases.

CONCLUSION

Machine learning methods can significantly improve linkage quality and reduce the manual effort required by medical coders to match tumor records in cancer registries compared to a deterministic baseline. Our results help cancer registries estimate how linkage quality increases as more records are manually processed.

Integrating Electronic Health Records and Polygenic Risk to Identify Genetically Unrelated Comorbidities of Schizophrenia That May Be Modifiable

Background

Patients with schizophrenia have substantial comorbidity that contributes to reduced life expectancy of 10 to 20 years. Identifying modifiable comorbidities could improve rates of premature mortality. Conditions that frequently co-occur but lack shared genetic risk with schizophrenia are more likely to be products of treatment, behavior, or environmental factors and therefore are enriched for potentially modifiable associations.

Methods

Phenome-wide comorbidity was calculated from electronic health records of 250,000 patients across 2 independent health care institutions (Vanderbilt University Medical Center and Mass General Brigham); associations with schizophrenia polygenic risk scores were calculated across the same phenotypes in linked biobanks.

Results

Schizophrenia comorbidity was significantly correlated across institutions (r = 0.85), and the 77 identified comorbidities were consistent with prior literature. Overall, comorbidity and polygenic risk score associations were significantly correlated (r = 0.55, p = 1.29 × 10-118). However, directly testing for the absence of genetic effects identified 36 comorbidities that had significantly equivalent schizophrenia polygenic risk score distributions between cases and controls. This set included phenotypes known to be consequences of antipsychotic medications (e.g., movement disorders) or of the disease such as reduced hygiene (e.g., diseases of the nail), thereby validating the approach. It also highlighted phenotypes with less clear causal relationships and minimal genetic effects such as tobacco use disorder and diabetes.

Conclusions

This work demonstrates the consistency and robustness of electronic health record-based schizophrenia comorbidities across independent institutions and with the existing literature. It identifies known and novel comorbidities with an absence of shared genetic risk, indicating other causes that may be modifiable and where further study of causal pathways could improve outcomes for patients.

Identifying predictors of the tooth loss phenotype in a large periodontitis patient cohort using a machine learning approach

OBJECTIVES

This study aimed to identify predictors associated with the tooth loss phenotype in a large periodontitis patient cohort in the university setting.

METHODS

Information on periodontitis patients and nineteen factors identified at the initial visit was extracted from electronic health records. The primary outcome is tooth loss phenotype (presence or absence of tooth loss). Prediction models were built on significant factors (single or combinatory) selected by the RuleFit algorithm, and these factors were further adopted by regression models. Model performance was evaluated by Area Under the Receiver Operating Characteristic Curve (AUROC) and Area Under the Precision-Recall Curve (AUPRC). Associations between predictors and the tooth loss phenotype were also evaluated by classical statistical approaches to validate the performance of machine learning models.

RESULTS

In total, 7840 patients were included. The machine learning model predicting the tooth loss phenotype achieved AUROC of 0.71 and AUPRC of 0.66. Age, periodontal diagnosis, number of missing teeth at baseline, furcation involvement, and tooth mobility were associated with the tooth loss phenotype in both machine learning and classical statistical models.

CONCLUSIONS

The rule-based machine learning approach improves model explainability compared to classical statistical methods. However, the model's generalizability needs to be further validated by external datasets.

CLINICAL SIGNIFICANCE

Predictors identified by the current machine learning approach using the RuleFit algorithm had clinically relevant thresholds in predicting the tooth loss phenotype in a large and diverse periodontitis patient cohort. The results of this study will assist clinicians in performing risk assessment for periodontitis at the initial visit.

Building large-scale registries from unstructured clinical notes using a low-resource natural language processing pipeline

Building clinical registries is an important step in clinical research and improvement of patient care quality. Natural Language Processing (NLP) methods have shown promising results in extracting valuable information from unstructured clinical notes. However, the structure and nature of clinical notes are very different from regular text that state-of-the-art NLP models are trained and tested on, and they have their own set of challenges. In this study, we propose Sentence Extractor with Keywords (SE-K), an efficient and interpretable classification approach for extracting information from clinical notes and show that it outperforms more computationally expensive methods in text classification. Following the Institutional Review Board (IRB) approval, we used SE-K and two embedding based NLP approaches (Sentence Extractor with Embeddings (SE-E) and Bidirectional Encoder Representations from Transformers (BERT)) to develop comprehensive registry of anterior cruciate ligament surgeries from 20 years of unstructured clinical data at a multi-site tertiary-care regional children's hospital. The low-resource approach (SE-K) had better performance (average AUROC of 0.94 ± 0.04) than the embedding-based approaches (SE-E: 0.93 ± 0.04 and BERT: 0.87 ± 0.09) for out of sample validation, in addition to minimum performance drop between test and out-of-sample validation. Moreover, the SE-K approach was at least six times faster (on CPU) than SE-E (on CPU) and BERT (on GPU) and provides interpretability. Our proposed approach, SE-K, can be effectively used to extract relevant variables from clinic notes to build large-scale registries, with consistently better performance compared to the more resource-intensive approaches (e.g., BERT). Such approaches can facilitate information extraction from unstructured notes for registry building, quality improvement and adverse event monitoring.

Factors impacting the adoption of big data in healthcare: A systematic literature review

BACKGROUND

The term "big data" refers to the vast volume, variety, and velocity of data generated from various sources-e.g., sensors, social media, and online platforms. Big data adoption within healthcare poses an intriguing possibility for improving patients' health, increasing operational efficiency, and enabling data-driven decision-making. Despite considerable interest in the adoption of big data in healthcare, empirical research assessing the factors impacting the adoption process is lacking. Therefore, this review aimed to investigate the literature using a systematic approach to explore the factors that affect big data adoption in healthcare.

METHODS

A systematic literature review was conducted. The methodical and thorough process of discovering, assessing, and synthesizing relevant studies provided a full review of the available data. Several databases were used for the information search. Most of the articles retrieved from the search came from popular medical research databases, such as Scopus, Taylor & Francis, ScienceDirect, Emerald Insights, PubMed, Springer, IEEE, MDPI, Google Scholar, ProQuest Central, ProQuest Public Health Database, and MEDLINE.

RESULTS AND CONCLUSION

The results of the systematic literature review indicated that several theoretical frameworks (including the technology acceptance model; the technology, organization, and environment framework; the interactive communication technology adoption model; diffusion of innovation theory; dynamic capabilities theory; and the absorptive capability framework) can be used to analyze and understand technology acceptance in healthcare. It is vital to consider the safety of electronic health records during the use of big data. Furthermore, several elements were found to determine technological acceptance, including environmental, technological, organizational, political, and regulatory factors.

Prevalence and differences in the co-administration of drugs known to interact: an analysis of three distinct and large populations

BACKGROUND

The co-administration of drugs known to interact greatly impacts morbidity, mortality, and health economics. This study aims to examine the drug-drug interaction (DDI) phenomenon with a large-scale longitudinal analysis of age and gender differences found in drug administration data from three distinct healthcare systems.

METHODS

This study analyzes drug administrations from population-wide electronic health records in Blumenau (Brazil; 133 K individuals), Catalonia (Spain; 5.5 M individuals), and Indianapolis (USA; 264 K individuals). The stratified prevalences of DDI for multiple severity levels per patient gender and age at the time of administration are computed, and null models are used to estimate the expected impact of polypharmacy on DDI prevalence. Finally, to study actionable strategies to reduce DDI prevalence, alternative polypharmacy regimens using drugs with fewer known interactions are simulated.

RESULTS

A large prevalence of co-administration of drugs known to interact is found in all populations, affecting 12.51%, 12.12%, and 10.06% of individuals in Blumenau, Indianapolis, and Catalonia, respectively. Despite very different healthcare systems and drug availability, the increasing prevalence of DDI as patients age is very similar across all three populations and is not explained solely by higher co-administration rates in the elderly. In general, the prevalence of DDI is significantly higher in women - with the exception of men over 50 years old in Indianapolis. Finally, we show that using proton pump inhibitor alternatives to omeprazole (the drug involved in more co-administrations in Catalonia and Blumenau), the proportion of patients that are administered known DDI can be reduced by up to 21% in both Blumenau and Catalonia and 2% in Indianapolis.

CONCLUSIONS

DDI administration has a high incidence in society, regardless of geographic, population, and healthcare management differences. Although DDI prevalence increases with age, our analysis points to a complex phenomenon that is much more prevalent than expected, suggesting comorbidities as key drivers of the increase. Furthermore, the gender differences observed in most age groups across populations are concerning in regard to gender equity in healthcare. Finally, our study exemplifies how electronic health records' analysis can lead to actionable interventions that significantly reduce the administration of known DDI and its associated human and economic costs.

Influenza vaccination accuracy among adults: Self-report compared with electronic health record data

OBJECTIVE

To assess the validity of electronic health record (EHR)-based influenza vaccination data among adults in a multistate network.

METHODS

Following the 2018-2019 and 2019-2020 influenza seasons, surveys were conducted among a random sample of adults who did or did not appear influenza-vaccinated (per EHR data) during the influenza season. Participants were asked to report their influenza vaccination status; self-report was treated as the criterion standard. Results were combined across survey years.

RESULTS

Survey response rate was 44.7% (777 of 1740) for the 2018-2019 influenza season and 40.5% (505 of 1246) for the 2019-2020 influenza season. The sensitivity of EHR-based influenza vaccination data was 75.0% (95% confidence interval [CI] 68.1, 81.1), specificity 98.4% (95% CI 92.9, 99.9), and negative predictive value 73.9% (95% CI 68.0, 79.3).

CONCLUSIONS

In a multistate research network across two recent influenza seasons, there was moderate concordance between EHR-based vaccination data and self-report.

The Brain Gene Registry: a data snapshot

Monogenic disorders account for a large proportion of population-attributable risk for neurodevelopmental disabilities. However, the data necessary to infer a causal relationship between a given genetic variant and a particular neurodevelopmental disorder is often lacking. Recognizing this scientific roadblock, 13 Intellectual and Developmental Disabilities Research Centers (IDDRCs) formed a consortium to create the Brain Gene Registry (BGR), a repository pairing clinical genetic data with phenotypic data from participants with variants in putative brain genes. Phenotypic profiles are assembled from the electronic health record (EHR) and a battery of remotely administered standardized assessments collectively referred to as the Rapid Neurobehavioral Assessment Protocol (RNAP), which include cognitive, neurologic, and neuropsychiatric assessments, as well as assessments for attention deficit hyperactivity disorder (ADHD) and autism spectrum disorder (ASD). Co-enrollment of BGR participants in the Clinical Genome Resource's (ClinGen's) GenomeConnect enables display of variant information in ClinVar. The BGR currently contains data on 479 participants who are 55% male, 6% Asian, 6% Black or African American, 76% white, and 12% Hispanic/Latine. Over 200 genes are represented in the BGR, with 12 or more participants harboring variants in each of these genes: CACNA1A, DNMT3A, SLC6A1, SETD5, and MYT1L. More than 30% of variants are de novo and 43% are classified as variants of uncertain significance (VUSs). Mean standard scores on cognitive or developmental screens are below average for the BGR cohort. EHR data reveal developmental delay as the earliest and most common diagnosis in this sample, followed by speech and language disorders, ASD, and ADHD. BGR data has already been used to accelerate gene-disease validity curation of 36 genes evaluated by ClinGen's BGR Intellectual Disability (ID)-Autism (ASD) Gene Curation Expert Panel. In summary, the BGR is a resource for use by stakeholders interested in advancing translational research for brain genes and continues to recruit participants with clinically reported variants to establish a rich and well-characterized national resource to promote research on neurodevelopmental disorders.

MixEHR-SurG: A joint proportional hazard and guided topic model for inferring mortality-associated topics from electronic health records

Survival models can help medical practitioners to evaluate the prognostic importance of clinical variables to patient outcomes such as mortality or hospital readmission and subsequently design personalized treatment regimes. Electronic Health Records (EHRs) hold the promise for large-scale survival analysis based on systematically recorded clinical features for each patient. However, existing survival models either do not scale to high dimensional and multi-modal EHR data or are difficult to interpret. In this study, we present a supervised topic model called MixEHR-SurG to simultaneously integrate heterogeneous EHR data and model survival hazard. Our contributions are three-folds: (1) integrating EHR topic inference with Cox proportional hazards likelihood; (2) integrating patient-specific topic hyperparameters using the PheCode concepts such that each topic can be identified with exactly one PheCode-associated phenotype; (3) multi-modal survival topic inference. This leads to a highly interpretable survival topic model that can infer PheCode-specific phenotype topics associated with patient mortality. We evaluated MixEHR-SurG using a simulated dataset and two real-world EHR datasets: the Quebec Congenital Heart Disease (CHD) data consisting of 8211 subjects with 75,187 outpatient claim records of 1767 unique ICD codes; the MIMIC-III consisting of 1458 subjects with multi-modal EHR records. Compared to the baselines, MixEHR-SurG achieved a superior dynamic AUROC for mortality prediction, with a mean AUROC score of 0.89 in the simulation dataset and a mean AUROC of 0.645 on the CHD dataset. Qualitatively, MixEHR-SurG associates severe cardiac conditions with high mortality risk among the CHD patients after the first heart failure hospitalization and critical brain injuries with increased mortality among the MIMIC-III patients after their ICU discharge. Together, the integration of the Cox proportional hazards model and EHR topic inference in MixEHR-SurG not only leads to competitive mortality prediction but also meaningful phenotype topics for in-depth survival analysis. The software is available at GitHub: https://github.com/li-lab-mcgill/MixEHR-SurG.

Potentially inappropriate prescribing for people with dementia in ambulatory care: a cross-sectional observational study

BACKGROUND

Studies have shown that potentially inappropriate prescribing (PIP) is highly prevalent among people with dementia (PwD) and linked to negative outcomes, such as hospitalisation and mortality. However, there are limited data on prescribing appropriateness for PwD in Saudi Arabia. Therefore, we aimed to estimate the prevalence of PIP and investigate associations between PIP and other patient characteristics among PwD in an ambulatory care setting.

METHODS

A cross-sectional, retrospective analysis was conducted at a tertiary hospital in Saudi Arabia. Patients who were ≥ 65 years old, had dementia, and visited ambulatory care clinics between 01/01/2019 and 31/12/2021 were included. Prescribing appropriateness was evaluated by applying the Screening Tool of Older Persons Potentially Inappropriate Prescriptions (STOPP) criteria. Descriptive analyses were used to describe the study population. Prevalence of PIP and the prevalence per each STOPP criterion were calculated as a percentage of all eligible patients. Logistic regression analysis was used to investigate associations between PIP, polypharmacy, age and sex; odds ratios (ORs) and 95% confidence intervals (CIs) were calculated. Analyses were conducted using SPSS v27.

RESULTS

A total of 287 PwD were identified; 56.0% (n = 161) were female. The mean number of medications prescribed was 9.0 [standard deviation (SD) ± 4.2]. The prevalence of PIP was 61.0% (n = 175). Common instances of PIP were drugs prescribed beyond the recommended duration (n = 90, 31.4%), drugs prescribed without an evidence-based clinical indication (n = 78, 27.2%), proton pump inhibitors (PPIs) for > 8 weeks (n = 75, 26.0%), and acetylcholinesterase inhibitors with concurrent drugs that reduce heart rate (n = 60, 21.0%). Polypharmacy was observed in 82.6% (n = 237) of patients and was strongly associated with PIP (adjusted OR 24.1, 95% CI 9.0-64.5).

CONCLUSIONS

Findings have revealed a high prevalence of PIP among PwD in Saudi Arabia that is strongly associated with polypharmacy. Future research should aim to explore key stakeholders' experiences and perspectives of medicines management to optimise medication use for this vulnerable patient population.

Landscape analysis of available European data sources amenable for machine learning and recommendations on usability for rare diseases screening

BACKGROUND

Patient registries and databases are essential tools for advancing clinical research in the area of rare diseases, as well as for enhancing patient care and healthcare planning. The primary aim of this study is a landscape analysis of available European data sources amenable to machine learning (ML) and their usability for Rare Diseases screening, in terms of findable, accessible, interoperable, reusable(FAIR), legal, and business considerations. Second, recommendations will be proposed to provide a better understanding of the health data ecosystem.

METHODS

In the period of March 2022 to December 2022, a cross-sectional study using a semi-structured questionnaire was conducted among potential respondents, identified as main contact person of a health-related databases. The design of the self-completed questionnaire survey instrument was based on information drawn from relevant scientific publications, quantitative and qualitative research, and scoping review on challenges in mapping European rare disease (RD) databases. To determine database characteristics associated with the adherence to the FAIR principles, legal and business aspects of database management Bayesian models were fitted.

RESULTS

In total, 330 unique replies were processed and analyzed, reflecting the same number of distinct databases (no duplicates included). In terms of geographical scope, we observed 24.2% (n = 80) national, 10.0% (n = 33) regional, 8.8% (n = 29) European, and 5.5% (n = 18) international registries coordinated in Europe. Over 80.0% (n = 269) of the databases were still active, with approximately 60.0% (n = 191) established after the year 2000 and 71.0% last collected new data in 2022. Regarding their geographical scope, European registries were associated with the highest overall FAIR adherence, while registries with regional and "other" geographical scope were ranked at the bottom of the list with the lowest proportion. Responders' willingness to share data as a contribution to the goals of the Screen4Care project was evaluated at the end of the survey. This question was completed by 108 respondents; however, only 18 of them (16.7%) expressed a direct willingness to contribute to the project by sharing their databases. Among them, an equal split between pro-bono and paid services was observed.

CONCLUSIONS

The most important results of our study demonstrate not enough sufficient FAIR principles adherence and low willingness of the EU health databases to share patient information, combined with some legislation incapacities, resulting in barriers to the secondary use of data.

Therapeutic management, adherence, and clinical outcomes of heart failure in Andalucía. ANDALIC Protocol

Heart failure is a prevalent syndrome with high mortality rates, representing a significant economic burden in terms of healthcare. The lack of systematic information about the treatment and adherence of patients with heart failure limits the understanding of these aspects and potentially the improvement of clinical outcomes.

OBJECTIVE

To describe the clinical characteristics, therapeutic management, adherence, persistence and clinical results, as well as the association between these variables, in a cohort of patients with heart failure in Andalusia.

DESIGN

This study will be an observational, population-based, retrospective cohort study. Data of patients discharged from an Andalusian hospital with a diagnosis of heart failure between 2014 and 2023 will be extracted from the Andalusian population health database.

ANALYSIS

The statistical analysis will incorporate the following strategies: 1) Descriptive analysis of the characteristics of the population cohort, adherence measures, and clinical outcomes. 2) Bivariate analyses to study the association of covariates with adherence, persistence and clinical results. 3) Multivariate logistic regression and Cox regression analysis including relevant covariates. 4) To evaluate changes over time, multivariate Poisson regression models will be used. By conducting this comprehensive study, we aim to gain valuable insights into the clinical characteristics, treatment management, and adherence of heart failure patients in Andalusia, as well as to identify factors that may influence clinical outcomes. These findings could be critical both for the development of optimized strategies that improve medical care and quality of life of patients and for mitigating the health burden of HF in the region.

Risk of herpes zoster in adults with SARS-CoV-2 infection in Spain: A population-based, retrospective cohort study

OBJECTIVES

We aimed to compare the risk of herpes zoster (HZ) in adults with and without laboratory-confirmed severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection.

METHODS

This retrospective dynamic cohort study analyzed data from a public healthcare database in Spain between November 2020 and October 2021. The main outcome was incident cases of HZ in individuals ≥18-year-old. Relative risk (RR) of HZ in SARS-CoV-2-confirmed versus SARS-CoV-2-free individuals was estimated by a multivariable negative binomial regression adjusted by age, sex, and comorbidities.

RESULTS

Data from 4,085,590 adults were analyzed. The overall HZ incidence rate in adults was 5.76 (95% confidence interval [CI], 5.66-5.85) cases per 1000 person-years. Individuals ≥18-year-old with SARS-CoV-2-confirmed infection had a 19% higher risk of developing HZ versus SARS-CoV-2-free ≥18-year-olds (adjusted RR = 1.19; 95% CI, 1.09-1.29); this percentage was 16% (adjusted RR = 1.16; 95% CI, 1.05-1.29) in ≥50-year-olds. Severe (hospitalized) cases of SARS-CoV-2 infection had a 64% (if ≥18 years old) or 44% (if ≥50 years old) higher risk of HZ versus nonhospitalized cases.

CONCLUSION

These results support an association between SARS-CoV-2 infection and HZ, with a greater HZ risk in severe cases of SARS-CoV-2 infection.

FedFSA: Hybrid and federated framework for functional status ascertainment across institutions

INTRODUCTION

Patients' functional status assesses their independence in performing activities of daily living, including basic ADLs (bADL), and more complex instrumental activities (iADL). Existing studies have discovered that patients' functional status is a strong predictor of health outcomes, particularly in older adults. Depite their usefulness, much of the functional status information is stored in electronic health records (EHRs) in either semi-structured or free text formats. This indicates the pressing need to leverage computational approaches such as natural language processing (NLP) to accelerate the curation of functional status information. In this study, we introduced FedFSA, a hybrid and federated NLP framework designed to extract functional status information from EHRs across multiple healthcare institutions.

METHODS

FedFSA consists of four major components: 1) individual sites (clients) with their private local data, 2) a rule-based information extraction (IE) framework for ADL extraction, 3) a BERT model for functional status impairment classification, and 4) a concept normalizer. The framework was implemented using the OHNLP Backbone for rule-based IE and open-source Flower and PyTorch library for federated BERT components. For gold standard data generation, we carried out corpus annotation to identify functional status-related expressions based on ICF definitions. Four healthcare institutions were included in the study. To assess FedFSA, we evaluated the performance of category- and institution-specific ADL extraction across different experimental designs.

RESULTS

ADL extraction performance ranges from an F1-score of 0.907 to 0.986 for bADL and 0.825 to 0.951 for iADL across the four healthcare sites. The performance for ADL extraction with impairment ranges from an F1-score of 0.722 to 0.954 for bADL and 0.674 to 0.813 for iADL across four healthcare sites. For category-specific ADL extraction, laundry and transferring yielded relatively high performance, while dressing, medication, bathing, and continence achieved moderate-high performance. Conversely, food preparation and toileting showed low performance.

CONCLUSION

NLP performance varied across ADL categories and healthcare sites. Federated learning using a FedFSA framework performed higher than non-federated learning for impaired ADL extraction at all healthcare sites. Our study demonstrated the potential of the federated learning framework in functional status extraction and impairment classification in EHRs, exemplifying the importance of a large-scale, multi-institutional collaborative development effort.

Predicting adolescent suicidal behavior following inpatient discharge using structured and unstructured data

BACKGROUND

The objective was to develop and assess performance of an algorithm predicting suicide-related ICD codes within three months of psychiatric discharge.

METHODS

This prognostic study used a retrospective cohort of EHR data from 2789 youth (12 to 20 years old) hospitalized in a safety net institution in the Northeastern United States. The dataset combined structured data with unstructured data obtained through natural language processing of clinical notes. Machine learning approaches compared gradient boosting to random forest analyses.

RESULTS

Area under the ROC and precision-recall curve were 0.88 and 0.17, respectively, for the final Gradient Boosting model. The cutoff point of the model-generated predicted probabilities of suicide that optimally classified the individual as high risk or not was 0.009. When applying the chosen cutoff (0.009) to the hold-out testing set, the model correctly identified 8 positive cases out of 10, and 418 negative cases out 548. The corresponding performance metrics showed 80 % sensitivity, 76 % specificity, 6 % PPV, 99 % NPV, F-1 score of 0.11, and an accuracy of 76 %.

LIMITATIONS

The data in this study comes from a single health system, possibly introducing bias in the model's algorithm. Thus, the model may have underestimated the incidence of suicidal behavior in the study population. Further research should include multiple system EHRs.

CONCLUSIONS

These performance metrics suggest a benefit to including both unstructured and structured data in design of predictive algorithms for suicidal behavior, which can be integrated into psychiatric services to help assess risk.

Using natural language processing to analyze unstructured patient-reported outcomes data derived from electronic health records for cancer populations: a systematic review

INTRODUCTION

Patient-reported outcomes (PROs; symptoms, functional status, quality-of-life) expressed in the 'free-text' or 'unstructured' format within clinical notes from electronic health records (EHRs) offer valuable insights beyond biological and clinical data for medical decision-making. However, a comprehensive assessment of utilizing natural language processing (NLP) coupled with machine learning (ML) methods to analyze unstructured PROs and their clinical implementation for individuals affected by cancer remains lacking.

AREAS COVERED

This study aimed to systematically review published studies that used NLP techniques to extract and analyze PROs in clinical narratives from EHRs for cancer populations. We examined the types of NLP (with and without ML) techniques and platforms for data processing, analysis, and clinical applications.

EXPERT OPINION

Utilizing NLP methods offers a valuable approach for processing and analyzing unstructured PROs among cancer patients and survivors. These techniques encompass a broad range of applications, such as extracting or recognizing PROs, categorizing, characterizing, or grouping PROs, predicting or stratifying risk for unfavorable clinical results, and evaluating connections between PROs and adverse clinical outcomes. The employment of NLP techniques is advantageous in converting substantial volumes of unstructured PRO data within EHRs into practical clinical utilities for individuals with cancer.

Documentation of results and medication prescribing after combinatorial psychiatric pharmacogenetic testing: A case for discrete results

PURPOSE

Combinatorial pharmacogenetic (PGx) panels intended to aid psychiatric prescribing are available to clinicians. Here, we evaluated the documentation of PGx panel results and subsequent prescribing patterns within a tertiary health care system.

METHODS

We performed a query of psychiatry service note text in our electronic health record using 71 predefined PGx terms. Patients who underwent combinatorial PGx testing were identified, and documentation of test results was analyzed. Prescription data following testing were examined for the frequency of prescriptions influenced by genes on the panel along with the medical specialties involved.

RESULTS

A total of 341 patients received combinatorial PGx testing, and documentation of results was found to be absent or incomplete for 198 patients (58%). The predominant method of documentation was through portable document formats uploaded to the electronic health record's "Media" section. Among patients with at least 1 year of follow-up, a large majority (194/228, 85%) received orders for medications affected by the tested genes, including 132 of 228 (58%) patients receiving at least 1 non-psychiatric medication influenced by the test results.

CONCLUSION

Results from combinatorial PGx testing were poorly documented. Medications affected by these results were often prescribed after testing, highlighting the need for discrete results and clinical decision support.

Clinical features-based machine learning models to separate sexually transmitted infections from other skin diagnoses

INTRODUCTION

Many sexual health services are overwhelmed and cannot cater for all the individuals who present with sexually transmitted infections (STIs). Digital health software that separates STIs from non-STIs could improve the efficiency of clinical services. We developed and evaluated a machine learning model that predicts whether patients have an STI based on their clinical features.

METHODS

We manually extracted 25 demographic features and clinical features from 1315 clinical records in the electronic health record system at Melbourne Sexual Health Center. We examined 16 machine learning models to predict a binary outcome of an STI or a non-STI diagnosis. We evaluated the models' performance with the area under the ROC curve (AUC), accuracy and F1-scores.

RESULTS

Our study included 1315 consultations, of which 36.8% (484/1315) were diagnosed with STIs and 63.2% (831/1315) had non-STI conditions. The study population predominantly consisted of heterosexual men (49.5%, 651/1315), followed by gay, bisexual and other men who have sex with men (GBMSM) (25.7%), women (21.6%) and unknown gender (3.2%). The median age was 31 years (intra-quartile range (IQR) 26-39). The top 5 performing models were CatBoost (AUC 0.912), Random Forest (AUC 0.917), LightGBM (AUC 0.907), Gradient Boosting (AUC 0.905) and XGBoost (AUC 0.900). The best model, CatBoost, achieved an accuracy of 0.837, sensitivity of 0.776, specificity of 0.831, precision of 0.782 and F1-score of 0.778. The key important features were lesion duration, type of skin lesions, age, gender, history of skin disorders, number of lesions, dysuria duration, anorectal pain and itchiness.

CONCLUSIONS

Our best model demonstrates a reasonable performance in distinguishing STIs from non-STIs. However, to be clinically useful, more detailed information such as clinical images, may be required to reach sufficient accuracy.

Utilizing nursing standards in electronic health records: A descriptive qualitative study

BACKGROUND

The electronic health record (EHR), including standardized structures and languages, represents an important data source for nurses, to continually update their individual and shared perceptual understanding of clinical situations. Registered nurses' utilization of nursing standards, such as standardized nursing care plans and language in EHRs, has received little attention in the literature. Further research is needed to understand nurses' care planning and documentation practice.

AIMS

This study aimed to describe the experiences and perceptions of nurses' EHR documentation practices utilizing standardized nursing care plans including standardized nursing language, in the daily documentation of nursing care for patients living in special dementia-care units in nursing homes in Norway.

METHODS

A descriptive qualitative study was conducted between April and November 2021 among registered nurses working in special dementia care units in Norwegian nursing homes. In-depth interviews were conducted, and data was analyzed utilizing reflexive thematic analysis with a deductive orientation. Findings Four themes were generated from the analysis. First, the knowledge, skills, and attitude of system users were perceived to influence daily documentation practice. Second, management and organization of documentation work, internally and externally, influenced motivation and engagement in daily documentation processes. Third, usability issues of the EHR were perceived to limit the daily workflow and the nurses' information-needs. Last, nursing standards in the EHR were perceived to contribute to the development of documentation practices, supporting and stimulating ethical awareness, cognitive processes, and knowledge development.

CONCLUSION

Nurses and nursing leaders need to be continuously involved and engaged in EHR documentation to safeguard development and implementation of relevant nursing standards.

Developing deep learning-based strategies to predict the risk of hepatocellular carcinoma among patients with nonalcoholic fatty liver disease from electronic health records

OBJECTIVE

The accuracy of deep learning models for many disease prediction problems is affected by time-varying covariates, rare incidence, covariate imbalance and delayed diagnosis when using structured electronic health records data. The situation is further exasperated when predicting the risk of one disease on condition of another disease, such as the hepatocellular carcinoma risk among patients with nonalcoholic fatty liver disease due to slow, chronic progression, the scarce of data with both disease conditions and the sex bias of the diseases. The goal of this study is to investigate the extent to which the aforementioned issues influence deep learning performance, and then devised strategies to tackle these challenges. These strategies were applied to improve hepatocellular carcinoma risk prediction among patients with nonalcoholic fatty liver disease.

METHODS

We evaluated two representative deep learning models in the task of predicting the occurrence of hepatocellular carcinoma in a cohort of patients with nonalcoholic fatty liver disease (n = 220,838) from a national EHR database. The disease prediction task was carefully formulated as a classification problem while taking censorship and the length of follow-up into consideration.

RESULTS

We developed a novel backward masking scheme to deal with the issue of delayed diagnosis which is very common in EHR data analysis and evaluate how the length of longitudinal information after the index date affects disease prediction. We observed that modeling time-varying covariates improved the performance of the algorithms and transfer learning mitigated reduced performance caused by the lack of data. In addition, covariate imbalance, such as sex bias in data impaired performance. Deep learning models trained on one sex and evaluated in the other sex showed reduced performance, indicating the importance of assessing covariate imbalance while preparing data for model training.

CONCLUSIONS

The strategies developed in this work can significantly improve the performance of hepatocellular carcinoma risk prediction among patients with nonalcoholic fatty liver disease. Furthermore, our novel strategies can be generalized to apply to other disease risk predictions using structured electronic health records, especially for disease risks on condition of another disease.

Robust diagnosis recommendation system for Primary Care Telemedicine using long short-term memory multi-class sequence classification

Background

Telemedicine offers opportunity for robust diagnoses recommendations to support healthcare providers intra-consultation in a way that does not limit providers ability to explore diagnostic codes and make the most appropriate selection for each consultation.

Objective

The objective of this work was to develop a recommendation system for ICD-10 coding using multiclass sequence classification and deep learning. The recommendations are intended to support telemedicine clinicians in making timely and appropriate diagnosis selections. The recommendations allow clinicians to find and select the best diagnosis code much quicker and without leaving the telemedicine platform to search codes and code descriptions.

Methods

We developed an LSTM model for multi-class text sequence classification to make diagnosis recommendations. The LSTM recommender used text-based symptoms, complaints, and consultation request reasons as model inputs. Data were extracted from a live telemedicine platform which spans general medicine, dermatology, and mental health clinical specialties. A popularity-based model was used for baseline comparison.

Results

Using over 2.8 MM telemedicine consultations during 2021 and 2022, our LSTM recommender average accuracy was 31.7%. LSTM recommender average coverage in the top 20 recommended diagnoses was 85.8% with an average personalization score of 0.87.

Conclusions

LSTM multi-class sequence classification recommends diagnoses specific to individual consultations, is retrainable on regular intervals, and could improve diagnoses recommendations such that providers require less time and resources searching for diagnosis codes. In addition, the LSTM recommender is robust enough to make recommendations across clinical specialties such as general medicine, dermatology, and mental health.

Comparing the clinical practice and prescribing safety of locum and permanent doctors: observational study of primary care consultations in England

BACKGROUND

Temporary doctors, known as locums, are a key component of the medical workforce in the NHS but evidence on differences in quality and safety between locum and permanent doctors is limited. We aimed to examine differences in the clinical practice, and prescribing safety for locum and permanent doctors working in primary care in England.

METHODS

We accessed electronic health care records (EHRs) for 3.5 million patients from the CPRD GOLD database with linkage to Hospital Episode Statistics from 1st April 2010 to 31st March 2022. We used multi-level mixed effects logistic regression to compare consultations with locum and permanent GPs for several patient outcomes including general practice revisits; prescribing of antibiotics; strong opioids; hypnotics; A&E visits; emergency hospital admissions; admissions for ambulatory care sensitive conditions; test ordering; referrals; and prescribing safety indicators while controlling for patient and practice characteristics.

RESULTS

Consultations with locum GPs were 22% more likely to involve a prescription for an antibiotic (OR = 1.22 (1.21 to 1.22)), 8% more likely to involve a prescription for a strong opioid (OR = 1.08 (1.06 to 1.09)), 4% more likely to be followed by an A&E visit on the same day (OR = 1.04 (1.01 to 1.08)) and 5% more likely to be followed by an A&E visit within 1 to 7 days (OR = 1.05 (1.02 to 1.08)). Consultations with a locum were 12% less likely to lead to a practice revisit within 7 days (OR = 0.88 (0.87 to 0.88)), 4% less likely to involve a prescription for a hypnotic (OR = 0.96 (0.94 to 0.98)), 15% less likely to involve a referral (OR = 0.85 (0.84 to 0.86)) and 19% less likely to involve a test (OR = 0.81 (0.80 to 0.82)). We found no evidence that emergency admissions, ACSC admissions and eight out of the eleven prescribing safety indicators were different if patients were seen by a locum or a permanent GP.

CONCLUSIONS

Despite existing concerns, the clinical practice and performance of locum GPs did not appear to be systematically different from that of permanent GPs. The practice and performance of both locum and permanent GPs is likely shaped by the organisational setting and systems within which they work.

'Kindling the fire' of NHS patient data exploitations: The care.data controversy in news media discourses

This paper explores news media discourse about care.data: an NHS England programme of work for amalgamating and sharing patient data from primary care for planning and research. It was scrapped in 2016 after three years of public outcry, delays and around 1.5 million opt-outs. I examine UK news media coverage of this programme through the 'fire object' metaphor, focusing upon the visions of purpose and value it inspired, the abrupt discontinuities, juxtapositions and transformations it performed, and the matters of concern that went unheeded. Findings suggest that, in care.data's pursuit of a societal consensus on NHS patient data exploitations, various visions for new and fluid data flows brought to presence narratives of transforming the NHS, saving lives, and growing the economy. Other realities and concerns that mattered for certain stakeholders, such as data ownership and commercialisation, public engagement and informed consent, commitment and leadership, operational capabilities, and NHS privatisation agendas, remained absent or unsettled. False dichotomies kept the controversy alive, sealing its fate. I conclude by arguing that such failed programmes can turn into phantom-like objects, haunting future patient data schemes of similar aspirations. The paper highlights the role news media can have in understanding such energetic public controversies.

Clinical spectrum of Transthyretin amyloidogenic mutations among diverse population origins

PURPOSE

Coding mutations in the Transthyretin (TTR) gene cause a hereditary form of amyloidosis characterized by a complex genotype-phenotype correlation with limited information regarding differences among worldwide populations.

METHODS

We compared 676 diverse individuals carrying TTR amyloidogenic mutations (rs138065384, Phe44Leu; rs730881165, Ala81Thr; rs121918074, His90Asn; rs76992529, Val122Ile) to 12,430 non-carriers matched by age, sex, and genetically-inferred ancestry to assess their clinical presentations across 1,693 outcomes derived from electronic health records in UK biobank.

RESULTS

In individuals of African descent (AFR), Val122Ile mutation was linked to multiple outcomes related to the circulatory system (fold-enrichment = 2.96, p = 0.002) with the strongest associations being cardiac congenital anomalies (phecode 747.1, p = 0.003), endocarditis (phecode 420.3, p = 0.006), and cardiomyopathy (phecode 425, p = 0.007). In individuals of Central-South Asian descent (CSA), His90Asn mutation was associated with dermatologic outcomes (fold-enrichment = 28, p = 0.001). The same TTR mutation was linked to neoplasms in European-descent individuals (EUR, fold-enrichment = 3.09, p = 0.003). In EUR, Ala81Thr showed multiple associations with respiratory outcomes related (fold-enrichment = 3.61, p = 0.002), but the strongest association was with atrioventricular block (phecode 426.2, p = 2.81 × 10- 4). Additionally, the same mutation in East Asians (EAS) showed associations with endocrine-metabolic traits (fold-enrichment = 4.47, p = 0.003). In the cross-ancestry meta-analysis, Val122Ile mutation was associated with peripheral nerve disorders (phecode 351, p = 0.004) in addition to cardiac congenital anomalies (fold-enrichment = 6.94, p = 0.003).

CONCLUSIONS

Overall, these findings highlight that TTR amyloidogenic mutations present ancestry-specific and ancestry-convergent associations related to a range of health domains. This supports the need to increase awareness regarding the range of outcomes associated with TTR mutations across worldwide populations to reduce misdiagnosis and delayed diagnosis of TTR-related amyloidosis.

Secure Cloud-Based Electronic Health Records: Cross-Patient Block-Level Deduplication with Blockchain Auditing

In today's data-driven world, the exponential growth of digital information poses significant challenges in data management. In recent years, the adoption of cloud-based Electronic Health Records (EHR) sharing schemes has yielded numerous advantages like improved accessibility, availability, and enhanced interoperability. However, the centralized nature of cloud storage presents challenges in terms of information storage, privacy protection, and security. Despite several approaches that have been presented to ensure secure deduplication of similar EHRs, the validation of data integrity without a third-party auditor (TPA) remains a persistent task. Because involving a TPA raises concerns about the confidentiality and privacy of crucial healthcare information. To tackle this challenge, a novel cloud storage auditing technique is proposed that incorporates cross-patient block-level deduplication while upholding strong privacy protection, ensuring that EHR is not compromised. Here, we introduced blockchain technology to achieve integrity verification, thus eliminating the need for a TPA by providing a decentralized and transparent mechanism. Additionally, an index for all EHRs has been generated to facilitate block-level duplicate checks and employ a novel strategy to prevent adversaries from acquiring original information saved in the cloud storage. The security of the proposed approach is established against factorization attacks and decrypt exponent attacks. The performance evaluation demonstrates the superior efficiency of the proposed scheme in terms of file authenticator generation, challenge creation, and proof verification to other existing client-side deduplication approaches.

Using machine learning models to predict falls in hospitalised adults

BACKGROUND

Identifying patients at high risk of falling is crucial in implementing effective fall prevention programs. While the integration of information systems is becoming more widespread in the healthcare industry, it poses a significant challenge in analysing vast amounts of data to identify factors that could enhance patient safety.

OBJECTIVE

To determine fall-associated factors and develop high-performance prediction tools for at-risk patients in acute and sub-acute care services in Australia.

METHODS

A retrospective study of 672,400 patients admitted to acute and sub-acute care services within a large metropolitan tertiary health service in Victoria, Australia, between January 1, 2019, and December 31, 2021. Data were obtained from four sources: the Department of Health Victorian Admitted Episodes Dataset, RiskManTM, electronic health records, and the health workforce dataset. Machine learning techniques, including Random Forest and Deep Neural Network models, were used to analyse the data, predict patient falls, and identify the most important risk factors for falls in this population. Model performance was evaluated using accuracy, F1-score, precision, recall, specificity, Matthew's correlation coefficient, and the area under the receiver operating characteristic curve (AUC).

RESULTS

The deep neural network and random forest models were highly accurate in predicting hospital patient falls. The deep neural network model achieved an accuracy of 0.988 and a specificity of 0.999, while the RF achieved an accuracy of 0.989 and a specificity of 1.000. The top 20 variables impacting falls were compared across both models, and 12 common factors were identified. These factors can be broadly classified into three categories: patient-related factors, staffing-related factors, and admission-related factors. Although not all factors are modifiable, they must be considered when planning fall prevention interventions.

CONCLUSION

The study demonstrated machine learning's potential to predict falls and identify key risk factors. Further validation across diverse populations and settings is essential for broader applicability.

Implementation and external validation of the Cambridge Multimorbidity Score in the UK Biobank cohort

BACKGROUND

Patients with multiple conditions present a growing challenge for healthcare provision. Measures of multimorbidity may support clinical management, healthcare resource allocation and accounting for the health of participants in purpose-designed cohorts. The recently developed Cambridge Multimorbidity scores (CMS) have the potential to achieve these aims using primary care records, however, they have not yet been validated outside of their development cohort.

METHODS

The CMS, developed in the Clinical Research Practice Dataset (CPRD), were validated in UK Biobank participants whose data is not available in CPRD (the cohort used for CMS development) with available primary care records (n = 111,898). This required mapping of the 37 pre-existing conditions used in the CMS to the coding frameworks used by UK Biobank data providers. We used calibration plots and measures of discrimination to validate the CMS for two of the three outcomes used in the development study (death and primary care consultation rate) and explored variation by age and sex. We also examined the predictive ability of the CMS for the outcome of cancer diagnosis. The results were compared to an unweighted count score of the 37 pre-existing conditions.

RESULTS

For all three outcomes considered, the CMS were poorly calibrated in UK Biobank. We observed a similar discriminative ability for the outcome of primary care consultation rate to that reported in the development study (C-index: 0.67 (95%CI:0.66-0.68) for both, 5-year follow-up); however, we report lower discrimination for the outcome of death than the development study (0.69 (0.68-0.70) and 0.89 (0.88-0.90) respectively). Discrimination for cancer diagnosis was adequate (0.64 (0.63-0.65)). The CMS performs favourably to the unweighted count score for death, but not for the outcomes of primary care consultation rate or cancer diagnosis.

CONCLUSIONS

In the UK Biobank, CMS discriminates reasonably for the outcomes of death, primary care consultation rate and cancer diagnosis and may be a valuable resource for clinicians, public health professionals and data scientists. However, recalibration will be required to make accurate predictions when cohort composition and risk levels differ substantially from the development cohort. The generated resources (including codelists for the conditions and code for CMS implementation in UK Biobank) are available online.

Challenges in and Opportunities for Electronic Health Record-Based Data Analysis and Interpretation

Electronic health records (EHRs) have been increasingly adopted in clinical practices across the United States, providing a primary source of data for clinical research, particularly observational cohort studies. EHRs are a high-yield, low-maintenance source of longitudinal real-world data for large patient populations and provide a wealth of information and clinical contexts that are useful for clinical research and translation into practice. Despite these strengths, it is important to recognize the multiple limitations and challenges related to the use of EHR data in clinical research. Missing data are a major source of error and biases and can affect the representativeness of the cohort of interest, as well as the accuracy of the outcomes and exposures. Here, we aim to provide a critical understanding of the types of data available in EHRs and describe the impact of data heterogeneity, quality, and generalizability, which should be evaluated prior to and during the analysis of EHR data. We also identify challenges pertaining to data quality, including errors and biases, and examine potential sources of such biases and errors. Finally, we discuss approaches to mitigate and remediate these limitations. A proactive approach to addressing these issues can help ensure the integrity and quality of EHR data and the appropriateness of their use in clinical studies.

Use of noisy labels as weak learners to identify incompletely ascertainable outcomes: A Feasibility study with opioid-induced respiratory depression

Objective

Assigning outcome labels to large observational data sets in a timely and accurate manner, particularly when outcomes are rare or not directly ascertainable, remains a significant challenge within biomedical informatics. We examined whether noisy labels generated from subject matter experts' heuristics using heterogenous data types within a data programming paradigm could provide outcomes labels to a large, observational data set. We chose the clinical condition of opioid-induced respiratory depression for our use case because it is rare, has no administrative codes to easily identify the condition, and typically requires at least some unstructured text to ascertain its presence.

Materials and methods

Using de-identified electronic health records of 52,861 post-operative encounters, we applied a data programming paradigm (implemented in the Snorkel software) for the development of a machine learning classifier for opioid-induced respiratory depression. Our approach included subject matter experts creating 14 labeling functions that served as noisy labels for developing a probabilistic Generative model. We used probabilistic labels from the Generative model as outcome labels for training a Discriminative model on the source data. We evaluated performance of the Discriminative model with a hold-out test set of 599 independently-reviewed patient records.

Results

The final Discriminative classification model achieved an accuracy of 0.977, an F1 score of 0.417, a sensitivity of 1.0, and an AUC of 0.988 in the hold-out test set with a prevalence of 0.83% (5/599).

Discussion

All of the confirmed Cases were identified by the classifier. For rare outcomes, this finding is encouraging because it reduces the number of manual reviews needed by excluding visits/patients with low probabilities.

Conclusion

Application of a data programming paradigm with expert-informed labeling functions might have utility for phenotyping clinical phenomena that are not easily ascertainable from highly-structured data.

Multimodal risk prediction with physiological signals, medical images and clinical notes

The broad adoption of electronic health record (EHR) systems brings us a tremendous amount of clinical data and thus provides opportunities to conduct data-based healthcare research to solve various clinical problems in the medical domain. Machine learning and deep learning methods are widely used in the medical informatics and healthcare domain due to their power to mine insights from raw data. When adapting deep learning models for EHR data, it is essential to consider its heterogeneous nature: EHR contains patient records from various sources including medical tests (e.g. blood test, microbiology test), medical imaging, diagnosis, medications, procedures, clinical notes, etc. Those modalities together provide a holistic view of patient health status and complement each other. Therefore, combining data from multiple modalities that are intrinsically different is challenging but intuitively promising in deep learning for EHR. To assess the expectations of multimodal data, we introduce a comprehensive fusion framework designed to integrate temporal variables, medical images, and clinical notes in EHR for enhanced performance in clinical risk prediction. Early, joint, and late fusion strategies are employed to combine data from various modalities effectively. We test the model with three predictive tasks: in-hospital mortality, long length of stay, and 30-day readmission. Experimental results show that multimodal models outperform uni-modal models in the tasks involved. Additionally, by training models with different input modality combinations, we calculate the Shapley value for each modality to quantify their contribution to multimodal performance. It is shown that temporal variables tend to be more helpful than CXR images and clinical notes in the three explored predictive tasks.

Sex matters in the association between cardiovascular health and incident dementia: evidence from real world data

BACKGROUND

Cardiovascular health has been associated with dementia onset, but little is known about the variation of such association by sex and age considering dementia subtypes. We assessed the role of sex and age in the association between cardiovascular risk and the onset of all-cause dementia, Alzheimer's disease, and vascular dementia in people aged 50-74 years.

METHODS

This is a retrospective cohort study covering 922.973 Catalans who attended the primary care services of the Catalan Health Institute (Spain). Data were obtained from the System for the Development of Research in Primary Care (SIDIAP database). Exposure was the cardiovascular risk (CVR) at baseline categorized into four levels of Framingham-REGICOR score (FRS): low (FRS < 5%), low-intermediate (5% ≤ FRS < 7.5%), high-intermediate (7.5% ≤ FRS < 10%), high (FRS ≥ 10%), and one group with previous vascular disease. Cases of all-cause dementia and Alzheimer's disease were identified using validated algorithms, and cases of vascular dementia were identified by diagnostic codes. We fitted stratified Cox models using age parametrized as b-Spline.

RESULTS

A total of 51,454 incident cases of all-cause dementia were recorded over a mean follow-up of 12.7 years. The hazard ratios in the low-intermediate and high FRS groups were 1.12 (95% confidence interval: 1.08-1.15) and 1.55 (1.50-1.60) for all-cause dementia; 1.07 (1.03-1.11) and 1.17 (1.11-1.24) for Alzheimer's disease; and 1.34 (1.21-1.50) and 1.90 (1.67-2.16) for vascular dementia. These associations were stronger in women and in midlife compared to later life in all dementia types. Women with a high Framingham-REGICOR score presented a similar risk of developing dementia - of any type - to women who had previous vascular disease, and at age 50-55, they showed three times higher risk of developing dementia risk compared to the lowest Framingham-REGICOR group.

CONCLUSIONS

We found a dose‒response association between the Framingham-REGICOR score and the onset of all dementia types. Poor cardiovascular health in midlife increased the onset of all dementia types later in life, especially in women.

Natural language processing to identify lupus nephritis phenotype in electronic health records

BACKGROUND

Systemic lupus erythematosus (SLE) is a rare autoimmune disorder characterized by an unpredictable course of flares and remission with diverse manifestations. Lupus nephritis, one of the major disease manifestations of SLE for organ damage and mortality, is a key component of lupus classification criteria. Accurately identifying lupus nephritis in electronic health records (EHRs) would therefore benefit large cohort observational studies and clinical trials where characterization of the patient population is critical for recruitment, study design, and analysis. Lupus nephritis can be recognized through procedure codes and structured data, such as laboratory tests. However, other critical information documenting lupus nephritis, such as histologic reports from kidney biopsies and prior medical history narratives, require sophisticated text processing to mine information from pathology reports and clinical notes. In this study, we developed algorithms to identify lupus nephritis with and without natural language processing (NLP) using EHR data from the Northwestern Medicine Enterprise Data Warehouse (NMEDW).

METHODS

We developed five algorithms: a rule-based algorithm using only structured data (baseline algorithm) and four algorithms using different NLP models. The first NLP model applied simple regular expression for keywords search combined with structured data. The other three NLP models were based on regularized logistic regression and used different sets of features including positive mention of concept unique identifiers (CUIs), number of appearances of CUIs, and a mixture of three components (i.e. a curated list of CUIs, regular expression concepts, structured data) respectively. The baseline algorithm and the best performing NLP algorithm were externally validated on a dataset from Vanderbilt University Medical Center (VUMC).

RESULTS

Our best performing NLP model incorporated features from both structured data, regular expression concepts, and mapped concept unique identifiers (CUIs) and showed improved F measure in both the NMEDW (0.41 vs 0.79) and VUMC (0.52 vs 0.93) datasets compared to the baseline lupus nephritis algorithm.

CONCLUSION

Our NLP MetaMap mixed model improved the F-measure greatly compared to the structured data only algorithm in both internal and external validation datasets. The NLP algorithms can serve as powerful tools to accurately identify lupus nephritis phenotype in EHR for clinical research and better targeted therapies.

Predicting transfers to intensive care in children using CEWT and other early warning systems

Background and Objective

The Children's Early Warning Tool (CEWT), developed in Australia, is widely used in many countries to monitor the risk of deterioration in hospitalized children. Our objective was to compare CEWT prediction performance against a version of the Bedside Pediatric Early Warning Score (Bedside PEWS), Between the Flags (BTF), and the pediatric Calculated Assessment of Risk and Triage (pCART).

Methods

We conducted a retrospective observational study of all patient admissions to the Comer Children's Hospital at the University of Chicago between 2009-2019. We compared performance for predicting the primary outcome of a direct ward-to-intensive care unit (ICU) transfer within the next 12 h using the area under the receiver operating characteristic curve (AUC). Alert rates at various score thresholds were also compared.

Results

Of 50,815 ward admissions, 1,874 (3.7%) experienced the primary outcome. Among patients in Cohort 1 (years 2009-2017, on which the machine learning-based pCART was trained), CEWT performed slightly worse than Bedside PEWS but better than BTF (CEWT AUC 0.74 vs. Bedside PEWS 0.76, P < 0.001; vs. BTF 0.66, P < 0.001), while pCART performed best for patients in Cohort 2 (years 2018-2019, pCART AUC 0.84 vs. CEWT AUC 0.79, P < 0.001; vs. BTF AUC 0.67, P < 0.001; vs. Bedside PEWS 0.80, P < 0.001). Sensitivity, specificity, and positive predictive values varied across all four tools at the examined thresholds for alerts.

Conclusion

CEWT has good discrimination for predicting which patients will likely be transferred to the ICU, while pCART performed the best.

Automated Recognition of Visual Acuity Measurements in Ophthalmology Clinical Notes Using Deep Learning

Purpose

Visual acuity (VA) is a critical component of the eye examination but is often only documented in electronic health records (EHRs) as unstructured free-text notes, making it challenging to use in research. This study aimed to improve on existing rule-based algorithms by developing and evaluating deep learning models to perform named entity recognition of different types of VA measurements and their lateralities from free-text ophthalmology notes: VA for each of the right and left eyes, with and without glasses correction, and with and without pinhole.

Design

Cross-sectional study.

Subjects

A total of 319 756 clinical notes with documented VA measurements from approximately 90 000 patients were included.

Methods

The notes were split into train, validation, and test sets. Bidirectional Encoder Representations from Transformers (BERT) models were fine-tuned to identify VA measurements from the progress notes and included BERT models pretrained on biomedical literature (BioBERT), critical care EHR notes (ClinicalBERT), both (BlueBERT), and a lighter version of BERT with 40% fewer parameters (DistilBERT). A baseline rule-based algorithm was created to recognize the same VA entities to compare against BERT models.

Main Outcome Measures

Model performance was evaluated on a held-out test set using microaveraged precision, recall, and F1 score for all entities.

Results

On the human-annotated subset, BlueBERT achieved the best microaveraged F1 score (F1 = 0.92), followed by ClinicalBERT (F1 = 0.91), DistilBERT (F1 = 0.90), BioBERT (F1 = 0.84), and the baseline model (F1 = 0.83). Common errors included labeling VA in sections outside of the examination portion of the note, difficulties labeling current VA alongside a series of past VAs, and missing nonnumeric VAs.

Conclusions

This study demonstrates that deep learning models are capable of identifying VA measurements from free-text ophthalmology notes with high precision and recall, achieving significant performance improvements over a rule-based algorithm. The ability to recognize VA from free-text notes would enable a more detailed characterization of ophthalmology patient cohorts and enhance the development of models to predict ophthalmology outcomes.

Financial Disclosures

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Graph neural networks for clinical risk prediction based on electronic health records: A survey

OBJECTIVE

This study aims to comprehensively review the use of graph neural networks (GNNs) for clinical risk prediction based on electronic health records (EHRs). The primary goal is to provide an overview of the state-of-the-art of this subject, highlighting ongoing research efforts and identifying existing challenges in developing effective GNNs for improved prediction of clinical risks.

METHODS

A search was conducted in the Scopus, PubMed, ACM Digital Library, and Embase databases to identify relevant English-language papers that used GNNs for clinical risk prediction based on EHR data. The study includes original research papers published between January 2009 and May 2023.

RESULTS

Following the initial screening process, 50 articles were included in the data collection. A significant increase in publications from 2020 was observed, with most selected papers focusing on diagnosis prediction (n = 36). The study revealed that the graph attention network (GAT) (n = 19) was the most prevalent architecture, and MIMIC-III (n = 23) was the most common data resource.

CONCLUSION

GNNs are relevant tools for predicting clinical risk by accounting for the relational aspects among medical events and entities and managing large volumes of EHR data. Future studies in this area may address challenges such as EHR data heterogeneity, multimodality, and model interpretability, aiming to develop more holistic GNN models that can produce more accurate predictions, be effectively implemented in clinical settings, and ultimately improve patient care.

Using Natural Language Processing to Identify Stigmatizing Language in Labor and Birth Clinical Notes

INTRODUCTION

Stigma and bias related to race and other minoritized statuses may underlie disparities in pregnancy and birth outcomes. One emerging method to identify bias is the study of stigmatizing language in the electronic health record. The objective of our study was to develop automated natural language processing (NLP) methods to identify two types of stigmatizing language: marginalizing language and its complement, power/privilege language, accurately and automatically in labor and birth notes.

METHODS

We analyzed notes for all birthing people > 20 weeks' gestation admitted for labor and birth at two hospitals during 2017. We then employed text preprocessing techniques, specifically using TF-IDF values as inputs, and tested machine learning classification algorithms to identify stigmatizing and power/privilege language in clinical notes. The algorithms assessed included Decision Trees, Random Forest, and Support Vector Machines. Additionally, we applied a feature importance evaluation method (InfoGain) to discern words that are highly correlated with these language categories.

RESULTS

For marginalizing language, Decision Trees yielded the best classification with an F-score of 0.73. For power/privilege language, Support Vector Machines performed optimally, achieving an F-score of 0.91. These results demonstrate the effectiveness of the selected machine learning methods in classifying language categories in clinical notes.

CONCLUSION

We identified well-performing machine learning methods to automatically detect stigmatizing language in clinical notes. To our knowledge, this is the first study to use NLP performance metrics to evaluate the performance of machine learning methods in discerning stigmatizing language. Future studies should delve deeper into refining and evaluating NLP methods, incorporating the latest algorithms rooted in deep learning.

Cross-institution natural language processing for reliable clinical association studies: a methodological exploration

OBJECTIVES

Natural language processing (NLP) of clinical notes in electronic medical records is increasingly used to extract otherwise sparsely available patient characteristics, to assess their association with relevant health outcomes. Manual data curation is resource intensive and NLP methods make these studies more feasible. However, the methodology of using NLP methods reliably in clinical research is understudied. The objective of this study is to investigate how NLP models could be used to extract study variables (specifically exposures) to reliably conduct exposure-outcome association studies.

STUDY DESIGN AND SETTING

In a convenience sample of patients admitted to the intensive care unit of a US academic health system, multiple association studies are conducted, comparing the association estimates based on NLP-extracted vs. manually extracted exposure variables. The association studies varied in NLP model architecture (Bidirectional Encoder Decoder from Transformers, Long Short-Term Memory), training paradigm (training a new model, fine-tuning an existing external model), extracted exposures (employment status, living status, and substance use), health outcomes (having a do-not-resuscitate/intubate code, length of stay, and in-hospital mortality), missing data handling (multiple imputation vs. complete case analysis), and the application of measurement error correction (via regression calibration).

RESULTS

The study was conducted on 1,174 participants (median [interquartile range] age, 61 [50, 73] years; 60.6% male). Additionally, up to 500 discharge reports of participants from the same health system and 2,528 reports of participants from an external health system were used to train the NLP models. Substantial differences were found between the associations based on NLP-extracted and manually extracted exposures under all settings. The error in association was only weakly correlated with the overall F1 score of the NLP models.

CONCLUSION

Associations estimated using NLP-extracted exposures should be interpreted with caution. Further research is needed to set conditions for reliable use of NLP in medical association studies.

Contextual Word Embedding for Biomedical Knowledge Extraction: a Rapid Review and Case Study

Recent advancements in natural language processing (NLP), particularly contextual word embedding models, have improved knowledge extraction from biomedical and healthcare texts. However, limited comprehensive research compares these models. This study conducts a scoping review and compares the performance of the major contextual word embedding models for biomedical knowledge extraction. From 26 articles identified from Scopus, PubMed, PubMed Central, and Google Scholar between 2017 and 2021, 18 notable contextual word embedding models were identified. These include ELMo, BERT, BioBERT, BlueBERT, CancerBERT, DDS-BERT, RuBERT, LABSE, EhrBERT, MedBERT, Clinical BERT, Clinical BioBERT, Discharge Summary BERT, Discharge Summary BioBERT, GPT, GPT-2, GPT-3, and GPT2-Bio-Pt. A case study compared the performance of six representative models-ELMo, BERT, BioBERT, BlueBERT, Clinical BioBERT, and GPT-3-across text classification, named entity recognition, and question answering. The evaluation utilized datasets comprising biomedical text from tweets, NCBI, PubMed, and clinical notes sourced from two electronic health record datasets. Performance metrics, including accuracy and F1 score, were used. The results of this case study reveal that BioBERT performs the best in analyzing biomedical text, while Clinical BioBERT excels in analyzing clinical notes. These findings offer crucial insights into word embedding models for researchers, practitioners, and stakeholders utilizing NLP in biomedical and clinical document analysis.

Supplementary Information

The online version contains supplementary material available at 10.1007/s41666-023-00157-y.

The viewpoints of parents of children with mental disorders regarding the confidentiality and security of their children's information in the Iranian national electronic health record system

INTRODUCTION

Electronic health records help collect and communicate patient information among healthcare providers. The confidentiality of information, especially for patients with mental disorders, is paramount due to its profound impacts on individuals' lives' social and personal aspects. This study aimed to investigate the viewpoints and concerns of parents of children with mental disorders regarding the confidentiality and security of their children's information in the Iranian National Electronic Health Record System (IEHRS).

METHODS

This is a survey study on parents or guardians of children with mental disorders who visited Kerman's specialised child psychiatry treatment centres. The data collection tool was a researcher-made questionnaire with 28 questions organised in seven sections, including demographic information of parents, children's medical history, Internet use, knowledge about IEHRS, the necessity of data collection, IEHRS security concerns, and privacy concerns. The data were analysed in SPSS 24 software using descriptive statistics and logistic and ordinal regressions to assess the relationship between parents' demographic characteristics and their viewpoints regarding information security and confidentiality concerns.

RESULTS

The results showed that more than 85 % of the parents believed that the security of their children's information in IEHRS was moderate to high. More than two-thirds (71 %) of the parents also believed that IEHRS should tighten its privacy policies. Most participants (87 %) were concerned about their children's information security in IEHRS. In this study, the parents' concerns about the privacy and security of information in IEHRS were not significantly associated with their age, gender, or knowledge about IEHRS.

CONCLUSIONS

Most parents of children with mental disorders were concerned about the security and confidentiality of their children's information in IEHRS. Thus, health policymakers should maintain a high level of security and establish appropriate privacy and confidentiality rules in IEHRS. In addition, they should be transparent about the system's security mechanisms and confidentiality regulations to win public trust.

Multi-view representation learning for tabular data integration using inter-feature relationships

OBJECTIVE

An applied problem facing all areas of data science is harmonizing data sources. Joining data from multiple origins with unmapped and only partially overlapping features is a prerequisite to developing and testing robust, generalizable algorithms, especially in healthcare. This integrating is usually resolved using meta-data such as feature names, which may be unavailable or ambiguous. Our goal is to design methods that create a mapping between structured tabular datasets derived from electronic health records independent of meta-data.

METHODS

We evaluate methods in the challenging case of numeric features without reliable and distinctive univariate summaries, such as nearly Gaussian and binary features. We assume that a small set of features are a priori mapped between two datasets, which share unknown identical features and possibly many unrelated features. Inter-feature relationships are the main source of identification which we expect. We compare the performance of contrastive learning methods for feature representations, novel partial auto-encoders, mutual-information graph optimizers, and simple statistical baselines on simulated data, public datasets, the MIMIC-III medical-record changeover, and perioperative records from before and after a medical-record system change. Performance was evaluated using both mapping of identical features and reconstruction accuracy of examples in the format of the other dataset.

RESULTS

Contrastive learning-based methods overall performed the best, often substantially beating the literature baseline in matching and reconstruction, especially in the more challenging real data experiments. Partial auto-encoder methods showed on-par matching with contrastive methods in all synthetic and some real datasets, along with good reconstruction. However, the statistical method we created performed reasonably well in many cases, with much less dependence on hyperparameter tuning. When validating feature match output in the EHR dataset we found that some mistakes were actually a surrogate or related feature as reviewed by two subject matter experts.

CONCLUSION

In simulation studies and real-world examples, we find that inter-feature relationships are effective at identifying matching or closely related features across tabular datasets when meta-data is not available. Decoder architectures are also reasonably effective at imputing features without an exact match.

Working with the Electronic Health Record and Laboratory Information System to Maximize Ordering and Reporting of Molecular Microbiology Results

Molecular microbiology assays have a higher cost of testing compared to traditional methods and need to be utilized appropriately. Results from these assays may also require interpretation and appropriate follow-up. Electronic tools available in the electronic health record and laboratory information system can be deployed both preanalytically and postanalytically to influence ordering behaviors and positively impact diagnostic stewardship. Next generation technologies, such as machine learning and artificial intelligence, have the potential to expand upon the capabilities currently available and warrant additional study and development but also require regulation around their use in health care.

People with diabetes who read their clinicians' visit notes: Behaviors and attitudes

OBJECTIVES

To understand behaviors and attitudes of adults with diabetes who read their clinicians' visit notes.

METHODS

By linking a large 2017 patient survey involving three institutions with administrative and portal use data, we identified patients with diabetes mellitus from outpatient records and examined reading behaviors related to eligible notes-initial, follow-up, history and physical, and progress notes. We analyzed patients' perceived benefits of reading notes.

RESULTS

2104 respondents had diagnoses of diabetes mellitus and had read ≥1 note in the 12-month period. Patients had an average of 8.7 eligible notes available and read 59% of them. The strongest predictor of reading more notes was having more notes available; the specialties of the authoring clinicians were not correlated with note reading rates. Patients reported understanding notes by primary care clinicians and specialists equally well; more than 90% of patients reported understanding everything or almost everything in a self-selected note. Across visit types, 73-80% of patients reported that note reading was extremely important for taking care of their health.

DISCUSSION

People with diabetes want to read their clinicians' notes, are accessing them at high rates, and report understanding the notes and benefiting from reading them.

Developing a Collaborative Agenda-Setting Intervention (CASI) to promote patient-centered communication in ovarian cancer care: A design thinking approach

OBJECTIVES

Patient-centered communication (PCC) occurs when clinicians respond to patients' needs, preferences, and concerns. While PCC is associated with better health-related quality of life in patients with cancer, patients with ovarian cancer have reported unmet communication needs. We used design thinking to develop an intervention to promote PCC in ovarian cancer care.

METHODS

Following the steps of design thinking, we empathized with stakeholders by reviewing the literature, then created stakeholder and journey maps to define the design challenge. To ideate solutions, we developed a challenge map. Finally, we developed wireframe prototypes and tested them with stakeholders.

RESULTS

Empathizing revealed that misaligned visit priorities precipitated suboptimal communication. Defining the design challenge and ideating solutions highlighted the need to normalize preference assessments, promote communication self-efficacy, and enhance visit efficiency. The Collaborative Agenda-Setting Intervention (CASI) elicits patients' needs and preferences and delivers communication guidance at the point of care. Stakeholders approved of the prototype.

CONCLUSION

Design thinking provided a systematic approach to empathizing with stakeholders, identifying challenges, and innovating solutions.

PRACTICE IMPLICATIONS

To our knowledge, the CASI is the first intervention to set the visit agenda and support communication from within the electronic health record. Future research will assess its usability and acceptability.

Chronic liver disease-associated severe thrombocytopenia in Spain: Results from a retrospective study using machine learning and natural language processing

BACKGROUND

Patients with chronic liver disease (CLD) often develop thrombocytopenia (TCP) as a complication. Severe TCP (platelet count<50×109/L) can increase morbidity and complicate CLD management, increasing bleeding risk during invasive procedures.

OBJECTIVES

To describe the real-world scenario of CLD-associated severe TCP patients' clinical characteristics. To evaluate the association between invasive procedures, prophylactic treatments, and bleeding events in this group of patients. To describe their need of medical resource use in Spain.

METHODS

This is a retrospective, multicenter study including patients who had confirmed diagnosis of CLD and severe TCP in four hospitals within the Spanish National Healthcare Network from January 2014 to December 2018. We analyzed the free-text information from Electronic Health Records (EHRs) of patients using Natural Language Processing (NLP), machine learning techniques, and SNOMED-CT terminology. Demographics, comorbidities, analytical parameters and characteristics of CLD were extracted at baseline and need for invasive procedures, prophylactic treatments, bleeding events and medical resources used in the follow up period. Frequency tables were generated for categorical variables, whereas continuous variables were described in summary tables as mean (SD) and median (Q1-Q3).

RESULTS

Out of 1,765,675 patients, 1787 had CLD and severe TCP; 65.2% were male with a mean age of 54.7 years old. Cirrhosis was detected in 46% (n=820) of patients and 9.1% (n=163) had hepatocellular carcinoma. Invasive procedures were needed in 85.6% of patients during the follow up period. Patients undergoing procedures compared to those patients without invasive procedures presented higher rates of bleeding events (33% vs 8%, p<0.0001) and higher number of bleedings. While prophylactic platelet transfusions were given to 25.6% of patients undergoing procedures, TPO receptor agonist use was only detected in 3.1% of them. Most patients (60.9%) required at least one hospital admission during the follow up and 14.4% of admissions were due to bleeding events with a hospital length of stay of 6 (3, 9) days.

CONCLUSIONS

NLP and machine learning are useful tools to describe real-world data in patients with CLD and severe TCP in Spain. Bleeding events are frequent in those patients who need invasive procedures, even receiving platelet transfusions as a prophylactic treatment, increasing the further use of medical resources. Because that, new prophylactic treatments that are not yet generalized, are needed.

Temporal self-attention for risk prediction from electronic health records using non-stationary kernel approximation

Effective modeling of patient representation from electronic health records (EHRs) is increasingly becoming a vital research topic. Yet, modeling the non-stationarity in EHR data has received less attention. Most existing studies follow a strong assumption of stationarity in patient representation from EHRs. However, in practice, a patient's visits are irregularly spaced over a relatively long period of time, and disease progression patterns exhibit non-stationarity. Furthermore, the time gaps between patient visits often encapsulate significant domain knowledge, potentially revealing undiscovered patterns that characterize specific medical conditions. To address these challenges, we introduce a new method which combines the self-attention mechanism with non-stationary kernel approximation to capture both contextual information and temporal relationships between patient visits in EHRs. To assess the effectiveness of our proposed approach, we use two real-world EHR datasets, comprising a total of 76,925 patients, for the task of predicting the next diagnosis code for a patient, given their EHR history. The first dataset is a general EHR cohort and consists of 11,451 patients with a total of 3,485 unique diagnosis codes. The second dataset is a disease-specific cohort that includes 65,474 pregnant patients and encompasses a total of 9,782 unique diagnosis codes. Our experimental evaluation involved nine prediction models, categorized into three distinct groups. Group 1 comprises the baselines: original self-attention with positional encoding model, RETAIN model, and LSTM model. Group 2 includes models employing self-attention with stationary kernel approximations, specifically incorporating three variations of Bochner's feature maps. Lastly, Group 3 consists of models utilizing self-attention with non-stationary kernel approximations, including quadratic, cubic, and bi-quadratic polynomials. The experimental results demonstrate that non-stationary kernels significantly outperformed baseline methods for NDCG@10 and Hit@10 metrics in both datasets. The performance boost was more substantial in dataset 1 for the NDCG@10 metric. On the other hand, stationary Kernels showed significant but smaller gains over baselines and were nearly as effective as Non-stationary Kernels for Hit@10 in dataset 2. These findings robustly validate the efficacy of employing non-stationary kernels for temporal modeling of EHR data, and emphasize the importance of modeling non-stationary temporal information in healthcare prediction tasks.

A two-stream deep model for automated ICD-9 code prediction in an intensive care unit

Assigning medical codes for patients is essential for healthcare organizations, not only for billing purposes but also for maintaining accurate records of patients' medical histories and analyzing the outputs of certain procedures. Due to the abundance of disease codes, it can be laborious and time-consuming for medical specialists to manually assign these codes to each procedure. To address this problem, we discuss the automatic prediction of ICD-9 codes, the most popular and widely accepted system of medical coding. We introduce a two-stream deep learning framework specifically designed to analyze multi-modal data. This framework is applied to the extensive and publicly available MIMIC-III dataset, enabling us to leverage both numerical and text-based data for improved ICD-9 code prediction. Our system uses text representation models to understand the text-based medical records; the Gated Recurrent Unit (GRU) to model the numerical health records; and fuses these two streams to automatically predict the ICD-9 codes used in the intensive care unit. We discuss the preprocessing and classification methods and demonstrate that our proposed two-stream model outperforms other state-of-the-art studies in the literature.

Digital stressors and resources perceived by emergency physicians and associations to their digital stress perception, mental health, job satisfaction and work engagement

BACKGROUND

Digital technologies are increasingly being integrated into healthcare settings, including emergency departments, with the potential to improve efficiency and patient care. Although digitalisation promises many benefits, the use of digital technologies can also introduce new stressors and challenges among medical staff, which may result in the development of various negative work and health outcomes. Therefore, this study aims to identify existing digital stressors and resources among emergency physicians, examine associations with various work- and health-related parameters, and finally identify the potential need for preventive measures.

METHODS

In this quantitative cross-sectional study, an online questionnaire was used to examine the relationship between digital stressors (technostress creators), digital resources (technostress inhibitors), technostress perception as well as mental health, job satisfaction and work engagement among 204 physicians working in German emergency medicine departments. Data collection lasted from December 2022 to April 2023. Validated scales were used for the questionnaire (e.g. "Technostress"-scale and the Copenhagen Psychosocial Questionnaire (COPSOQ). Descriptive and multiple regression analyses were run to test explorative assumptions.

RESULTS

The study found medium levels of technostress perception among the participating emergency physicians as well as low levels of persisting technostress inhibitors. The queried physicians on average reported medium levels of exhaustion symptoms, high levels of work engagement and job satisfaction. Significant associations between digital stressors and work- as well as health-related outcomes were analyzed.

CONCLUSION

This study provides a preliminary assessment of the persistence of digital stressors, digital resources and technostress levels, and their potential impact on relevant health and work-related outcomes, among physicians working in German emergency departments. Understanding and mitigating these stressors is essential to promote the well-being of physicians and ensure optimal patient care. As digitisation processes will continue to increase, the need for preventive support measures in dealing with technology stressors is obvious and should be expanded accordingly in the clinics. By integrating such support into everyday hospital life, medical staff in emergency departments can better focus on patient care and mitigate potential stress factors associated with digital technologies.

Large Language Models Facilitate the Generation of Electronic Health Record Phenotyping Algorithms

Objectives

Phenotyping is a core task in observational health research utilizing electronic health records (EHRs). Developing an accurate algorithm demands substantial input from domain experts, involving extensive literature review and evidence synthesis. This burdensome process limits scalability and delays knowledge discovery. We investigate the potential for leveraging large language models (LLMs) to enhance the efficiency of EHR phenotyping by generating high-quality algorithm drafts.

Materials and Methods

We prompted four LLMs-GPT-4 and GPT-3.5 of ChatGPT, Claude 2, and Bard-in October 2023, asking them to generate executable phenotyping algorithms in the form of SQL queries adhering to a common data model (CDM) for three phenotypes (i.e., type 2 diabetes mellitus, dementia, and hypothyroidism). Three phenotyping experts evaluated the returned algorithms across several critical metrics. We further implemented the top-rated algorithms and compared them against clinician-validated phenotyping algorithms from the Electronic Medical Records and Genomics (eMERGE) network.

Results

GPT-4 and GPT-3.5 exhibited significantly higher overall expert evaluation scores in instruction following, algorithmic logic, and SQL executability, when compared to Claude 2 and Bard. Although GPT-4 and GPT-3.5 effectively identified relevant clinical concepts, they exhibited immature capability in organizing phenotyping criteria with the proper logic, leading to phenotyping algorithms that were either excessively restrictive (with low recall) or overly broad (with low positive predictive values).

Conclusion

GPT versions 3.5 and 4 are capable of drafting phenotyping algorithms by identifying relevant clinical criteria aligned with a CDM. However, expertise in informatics and clinical experience is still required to assess and further refine generated algorithms.