Inaccuracy of the International Classification of Diseases (ICD-9-CM) in identifying the diagnosis of ischemic cerebrovascular disease

Risk of Postpartum Hemorrhage in Hypertensive Disorders of Pregnancy: Stratified by Severity

OBJECTIVE

 We aimed to determine the composite maternal hemorrhagic outcome (CMHO) among individuals with and without hypertensive disorders of pregnancy (HDP), stratified by disease severity. Additionally, we investigated the composite neonatal adverse outcome (CNAO) among individuals with HDP who had postpartum hemorrhage (PPH) versus did not have PPH.

STUDY DESIGN

 Our retrospective cohort study included all singletons who delivered at a Level IV center over two consecutive years. The primary outcome was the rate of CMHO, defined as blood loss ≥1,000 mL, use of uterotonics, mechanical tamponade, surgical techniques for atony, transfusion, venous thromboembolism, intensive care unit admission, hysterectomy, or maternal death. A subgroup analysis was performed to investigate the primary outcome stratified by (1) chronic hypertension, (2) gestational hypertension and preeclampsia without severe features, and (3) preeclampsia with severe features. A multivariable regression analysis was performed to investigate the association of HDP with and without PPH on a CNAO which included APGAR <7 at 5 minutes, bronchopulmonary dysplasia, intraventricular hemorrhage, necrotizing enterocolitis, seizures, neonatal sepsis, meconium aspiration syndrome, ventilation >6 hours, hypoxic-ischemic encephalopathy, or neonatal death.

RESULTS

 Of 8,357 singletons, 2,827 (34%) had HDP. Preterm delivery <37 weeks, induction of labor, prolonged oxytocin use, and magnesium sulfate usage were more common in those with versus without HDP (p < 0.001). CMHO was higher among individuals with HDP than those without HDP (26 vs. 19%; adjusted relative risk [aRR] = 1.11, 95% CI: 1.01-1.22). In the subgroup analysis, only individuals with preeclampsia with severe features were associated with higher CMHO (n = 802; aRR = 1.52, 95% CI: 1.32-1.75). There was a higher likelihood of CNAO in individuals with both HDP and PPH compared to those with HDP without PPH (aRR = 1.49, 95% CI: 1.06-2.09).

CONCLUSION

 CMHO was higher among those with HDP. After stratification, only those with preeclampsia with severe features had an increased risk of CMHO. Among individuals with HDP, those who also had a PPH had worse neonatal outcomes than those without hemorrhage.

KEY POINTS

· Individuals with HDP had an 11% higher likelihood of CMHO.. · After stratification, increased CMHO was limited to those with preeclampsia with severe features.. · There was a higher likelihood of CNAO in those with both HDP and PPH compared to HDP without PPH..

Can researchers trust ICD-10 coding of medical comorbidities in orthopaedic trauma patients?

Objectives

The 10th revision of the International Classification of Diseases (ICD-10) coding system may prove useful to orthopaedic trauma researchers to identify and document populations based on comorbidities. However, its use for research first necessitates determination of its reliability. The purpose of this study was to assess the reliability of electronic medical record (EMR) ICD-10 coding of nonorthopaedic diagnoses in orthopaedic trauma patients relative to the gold standard of prospective data collection.

Design

Nonexperimental cross-sectional study.

Setting

Level 1 Trauma Center.

Patients/Participants

Two hundred sixty-three orthopaedic trauma patients from 2 prior prospective studies from September 2018 to April 2022.

Intervention

Prospectively collected data were compared with EMR ICD-10 code abstraction for components of the Charlson Comorbidity Index (CCI), obesity, alcohol abuse, and tobacco use (retrospective data).

Main Outcome Measurements

Percent agreement and Cohen's kappa reliability.

Results

Percent agreement ranged from 86.7% to 96.9% for all CCI diagnoses and was as low as 72.6% for the diagnosis "overweight." Only 2 diagnoses, diabetes without end-organ damage (kappa = 0.794) and AIDS (kappa = 0.798) demonstrated Cohen's kappa values to indicate substantial agreement.

Conclusion

EMR diagnostic coding for medical comorbidities in orthopaedic trauma patients demonstrated variable reliability. Researchers may be able to rely on EMR coding to identify patients with diabetes without complications or AIDS. Chart review may still be necessary to confirm diagnoses. Low prevalence of most comorbidities led to high percentage agreement with low reliability.

Level of Evidence

Level 1 diagnostic.

Developing a national surveillance system for stroke and acute myocardial infarction using claims data in the Republic of Korea: a retrospective study

BACKGROUND

Limited information is available concerning the epidemiology of stroke and acute myocardial infarction (AMI) in the Republic of Korea. This study aimed to develop a national surveillance system to monitor the incidence of stroke and AMI using national claims data.

METHODS

We developed and validated identification algorithms for stroke and AMI using claims data. This validation involved a 2-stage stratified sampling method with a review of medical records for sampled cases. The weighted positive predictive value (PPV) and negative predictive value (NPV) were calculated based on the sampling structure and the corresponding sampling rates. Incident cases and the incidence rates of stroke and AMI in the Republic of Korea were estimated by applying the algorithms and weighted PPV and NPV to the 2018 National Health Insurance Service claims data.

RESULTS

In total, 2,200 cases (1,086 stroke cases and 1,114 AMI cases) were sampled from the 2018 claims database. The sensitivity and specificity of the algorithms were 94.3% and 88.6% for stroke and 97.9% and 90.1% for AMI, respectively. The estimated number of cases, including recurrent events, was 150,837 for stroke and 40,529 for AMI in 2018. The age- and sex-standardized incidence rate for stroke and AMI was 180.2 and 46.1 cases per 100,000 person-years, respectively, in 2018.

CONCLUSION

This study demonstrates the feasibility of developing a national surveillance system based on claims data and identification algorithms for stroke and AMI to monitor their incidence rates.

Contrastive diagnostic embedding (CDE) model for automated coding - A case study using emergency department encounters

BACKGROUND

Billing codes are utilized for medical reimbursement, clinical quality metric valuation and for epidemiologic purposes to report and follow disease trends and outcomes. The current paradigm of manual coding can be expensive, time-consuming, and subject to human error. Though automation of the billing codes has been widely reported in the literature via rule-based and supervised approaches, existing strategies lack generalizability and robustness towards large and constantly changing ICD hierarchical structure.

METHOD

We propose a weakly supervised training strategy by leveraging contrastive learning, contrastive diagnosis embedding (CDE) to capture the fine semantic variations between the diagnosis codes. The approach consists of a two-phase contrastive training for generating the semantic embedding space adapted to incorporate hierarchical information of ICD-10 vocabulary and a weakly supervised retrieval scheme. Core strength of the proposed method is that it puts no limit on the 70 K ICD-10 codes set and can handle all rare codes for coding the diagnosis.

RESULTS

Our CDE model outperformed string-based partial matching and ClinicalBERT embedding on three test cases (a retrospective testset, a prospective testset, and external testset) and produced an accurate prediction of rare and newly introduced diagnosis codes. A detailed ablation study showed the importance of each phase of the proposed multi-phase training. Each successive phase of training - ICD-10 group sensitive training (phase 1.1), ICD-10 subgroup sensitive training (phase 1.2), free-text diagnosis description-based training (phase 2) - improved performance beyond the previous phase of training. The model also outperformed existing supervised models like CAML and PLM-ICD and produced satisfactory performance on the rare codes.

CONCLUSION

Compared to the existing rule-based and supervised models, the proposed weakly supervised contrastive learning overcomes the limitations in terms of generalization capability and increases the robustness of the automated billing. Such a model will allow flexibility through accurate billing code automation for practice convergence and gains efficiencies in a value-based care payment environment.

Weakly Semi-supervised phenotyping using Electronic Health records

OBJECTIVE

Electronic Health Record (EHR) based phenotyping is a crucial yet challenging problem in the biomedical field. Though clinicians typically determine patient-level diagnoses via manual chart review, the sheer volume and heterogeneity of EHR data renders such tasks challenging, time-consuming, and prohibitively expensive, thus leading to a scarcity of clinical annotations in EHRs. Weakly supervised learning algorithms have been successfully applied to various EHR phenotyping problems, due to their ability to leverage information from large quantities of unlabeled samples to better inform predictions based on a far smaller number of patients. However, most weakly supervised methods are subject to the challenge to choose the right cutoff value to generate an optimal classifier. Furthermore, since they only utilize the most informative features (i.e., main ICD and NLP counts) they may fail for episodic phenotypes that cannot be consistently detected via ICD and NLP data. In this paper, we propose a label-efficient, weakly semi-supervised deep learning algorithm for EHR phenotyping (WSS-DL), which overcomes the limitations above.

MATERIALS AND METHODS

WSS-DL classifies patient-level disease status through a series of learning stages: 1) generating silver standard labels, 2) deriving enhanced-silver-standard labels by fitting a weakly supervised deep learning model to data with silver standard labels as outcomes and high dimensional EHR features as input, and 3) obtaining the final prediction score and classifier by fitting a supervised learning model to data with a minimal number of gold standard labels as the outcome, and the enhanced-silver-standard labels and a minimal set of most informative EHR features as input. To assess the generalizability of WSS-DL across different phenotypes and medical institutions, we apply WSS-DL to classify a total of 17 diseases, including both acute and chronic conditions, using EHR data from three healthcare systems. Additionally, we determine the minimum quantity of training labels required by WSS-DL to outperform existing supervised and semi-supervised phenotyping methods.

RESULTS

The proposed method, in combining the strengths of deep learning and weakly semi-supervised learning, successfully leverages the crucial phenotyping information contained in EHR features from unlabeled samples. Indeed, the deep learning model's ability to handle high-dimensional EHR features allows it to generate strong phenotype status predictions from silver standard labels. These predictions, in turn, provide highly effective features in the final logistic regression stage, leading to high phenotyping accuracy in notably small subsets of labeled data (e.g. n = 40 labeled samples).

CONCLUSION

Our method's high performance in EHR datasets with very small numbers of labels indicates its potential value in aiding doctors to diagnose rare diseases as well as conditions susceptible to misdiagnosis.

Analyzing real world data of blood transfusion adverse events: Opportunities and challenges

BACKGROUND

Blood transfusions are a vital component of modern healthcare, yet adverse reactions to blood product transfusions can cause morbidity, and rarely result in mortality. Therefore, accurate reporting of transfusion related adverse events (TRAEs) is paramount to improved transfusion practice. This study aims to investigate real-world data (RWD) on TRAEs by evaluating differences between ICD 9/10-based electronic health records (EHR) and blood bank-specific reporting.

STUDY DESIGN AND METHODS

TRAE data were retrospectively collected from a blood bank-specific database between Jan 2015 and June 2019 as the reference data source and compared it to ICD 9/10 diagnostic codes corresponding to various TRAEs. Seven reactions that have corresponding ICD 9/10 diagnostic codes were evaluated: Transfusion related circulatory overload (TACO), transfusion related acute lung injury (TRALI), febrile non-hemolytic reaction (FNHTR), transfusion-related anaphylactic reaction (TRA), acute hemolytic transfusion reaction (AHTR), delayed hemolytic transfusion reaction (DHTR), and delayed serologic reaction (DSTR). These accounted for 33% of the TRAEs at an academic institution during the study period.

RESULTS

Among 18637 adult blood transfusion recipients, there were 229 unique patients with 263 TRAE related ICD codes in the EHR, while there were 191 unique patients with 287 TRAEs identified in the blood bank database. None of the categories of reaction we investigated had perfect alignment between ICD 9/10 codes and blood bank specific diagnoses.

DISCUSSION

Multiple systemic challenges were identified that hinder effective reporting of TRAEs. Identifying factors causing inconsistent reporting between blood banks and EHRs is paramount to developing effective workability between these electronic systems, as well as across clinical and laboratory teams.

Toward attention-based learning to predict the risk of brain degeneration with multimodal medical data

Introduction

Brain degeneration is commonly caused by some chronic diseases, such as Alzheimer's disease (AD) and diabetes mellitus (DM). The risk prediction of brain degeneration aims to forecast the situation of disease progression of patients in the near future based on their historical health records. It is beneficial for patients to make an accurate clinical diagnosis and early prevention of disease. Current risk predictions of brain degeneration mainly rely on single-modality medical data, such as Electronic Health Records (EHR) or magnetic resonance imaging (MRI). However, only leveraging EHR or MRI data for the pertinent and accurate prediction is insufficient because of single-modality information (e.g., pixel or volume information of image data or clinical context information of non-image data).

Methods

Several deep learning-based methods have used multimodal data to predict the risks of specified diseases. However, most of them simply integrate different modalities in an early, intermediate, or late fusion structure and do not care about the intra-modal and intermodal dependencies. A lack of these dependencies would lead to sub-optimal prediction performance. Thus, we propose an encoder-decoder framework for better risk prediction of brain degeneration by using MRI and EHR. An encoder module is one of the key components and mainly focuses on feature extraction of input data. Specifically, we introduce an encoder module, which integrates intra-modal and inter-modal dependencies with the spatial-temporal attention and cross-attention mechanism. The corresponding decoder module is another key component and mainly parses the features from the encoder. In the decoder module, a disease-oriented module is used to extract the most relevant disease representation features. We take advantage of a multi-head attention module followed by a fully connected layer to produce the predicted results.

Results

As different types of AD and DM influence the nature and severity of brain degeneration, we evaluate the proposed method for three-class prediction of AD and three-class prediction of DM. Our results show that the proposed method with integrated MRI and EHR data achieves an accuracy of 0.859 and 0.899 for the risk prediction of AD and DM, respectively.

Discussion

The prediction performance is significantly better than the benchmarks, including MRI-only, EHR-only, and state-of-the-art multimodal fusion methods.

Association between Air Pollutants and Initial Hospital Admission for Ischemic Stroke in Korea from 2002 to 2013

OBJECTIVES

There is limited information regarding the association between air pollution exposure and stroke incidence. Therefore, this study aimed to assess the associations between short-term exposure to ambient air pollutants and initial hospital admission for ischemic stroke.

MATERIALS AND METHODS

From the Korea National Health Insurance Service-National Sample Cohort 2002-2013 database in South Korea, 55,852 first hospital admissions for ischemic stroke were identified. A generalized additive Poisson model was used to explore the association between air pollutants, including particulate matter, sulfur dioxide, nitrogen dioxide, and carbon monoxide and admissions for ischemic stroke.

RESULTS

All air pollutant models showed significant associations with ischemic stroke in the single lag model. In all air pollutant models excluding particulate matter 10 μm, a significant association was found between nitrogen dioxide exposure and initial admission for ischemic stroke after adjusting for other pollutants. An increment of 10 μg/m³ in nitrogen dioxide concentration at lag 0 and 14 days corresponded to a 0.259% (95% confidence interval, 0.231-0.287%) and 0.110% (95% confidence interval, 0.097-0.124) increase in initial admission for ischemic stroke, respectively.

CONCLUSIONS

The exposure-response relationship between nitrogen dioxide and initial admissions for ischemic stroke was approximately linear, with a sharper response at higher concentrations. Short-term exposure to nitrogen dioxide was positively associated with initial hospital admission for ischemic stroke.

Accelerometer-derived physical activity and risk of atrial fibrillation

AIMS

Physical activity may be an important modifiable risk factor for atrial fibrillation (AF), but associations have been variable and generally based on self-reported activity.

METHODS AND RESULTS

We analysed 93 669 participants of the UK Biobank prospective cohort study without prevalent AF who wore a wrist-based accelerometer for 1 week. We categorized whether measured activity met the standard recommendations of the European Society of Cardiology, American Heart Association, and World Health Organization [moderate-to-vigorous physical activity (MVPA) ≥150 min/week]. We tested associations between guideline-adherent activity and incident AF (primary) and stroke (secondary) using Cox proportional hazards models adjusted for age, sex, and each component of the Cohorts for Heart and Aging Research in Genomic Epidemiology AF (CHARGE-AF) risk score. We also assessed correlation between accelerometer-derived and self-reported activity. The mean age was 62 ± 8 years and 57% were women. Over a median of 5.2 years, 2338 incident AF events occurred. In multivariable adjusted models, guideline-adherent activity was associated with lower risks of AF [hazard ratio (HR) 0.82, 95% confidence interval (CI) 0.75-0.89; incidence 3.5/1000 person-years, 95% CI 3.3-3.8 vs. 6.5/1000 person-years, 95% CI 6.1-6.8] and stroke (HR 0.76, 95% CI 0.64-0.90; incidence 1.0/1000 person-years, 95% CI 0.9-1.1 vs. 1.8/1000 person-years, 95% CI 1.6-2.0). Correlation between accelerometer-derived and self-reported MVPA was weak (Spearman r = 0.16, 95% CI 0.16-0.17). Self-reported activity was not associated with incident AF or stroke.

CONCLUSIONS

Greater accelerometer-derived physical activity is associated with lower risks of AF and stroke. Future preventive efforts to reduce AF risk may be most effective when targeting adherence to objective activity thresholds.

Systematic Review and Meta-Analysis of Diagnostic Agreement in Suspected TIA

Objective

To determine the interrater variability for TIA diagnostic agreement among expert clinicians (neurologists/stroke physicians), administrative data, and nonspecialists.

Methods

We performed a meta-analysis of studies from January 1984 to January 2019 using MEDLINE, EMBASE, and PubMed. Two reviewers independently screened for eligible studies and extracted interrater variability measurements using Cohen's kappa scores to assess diagnostic agreement.

Results

Nineteen original studies consisting of 19,421 patients were included. Expert clinicians demonstrate good agreement for TIA diagnosis (κ = 0.71, 95% confidence interval [CI] = 0.62-0.81). Interrater variability between clinicians' TIA diagnosis and administrative data also demonstrated good agreement (κ = 0.68, 95% CI = 0.62-0.74). There was moderate agreement (κ = 0.41, 95% CI = 0.22-0.61) between referring clinicians and clinicians at TIA clinics receiving the referrals. Sixty percent of 748 patient referrals to TIA clinics were TIA mimics.

Conclusions

Overall agreement between expert clinicians was good for TIA diagnosis, although variation still existed for a sizeable proportion of cases. Diagnostic agreement for TIA decreased among nonspecialists. The substantial number of patients being referred to TIA clinics with other (often neurologic) diagnoses was large, suggesting that clinicians, who are proficient in managing TIAs and their mimics, should run TIA clinics.

Implementing the US Department of Health and Human Services definition of multimorbidity: a comparison between billing codes and medical record review in a population-based sample of persons 4084 years old

OBJECTIVE

To assess the validity of the US Department of Health and Human Services (DHHS) definition of multimorbidity using International Classification of Diseases, ninth edition (ICD-9) codes from administrative data.

DESIGN

Cross-sectional comparison of two ICD-9 billing code algorithms to data abstracted from medical records.

SETTING

Olmsted County, Minnesota, USA.

PARTICIPANTS

An age-stratified and sex-stratified random sample of 1509 persons ages 40-84 years old residing in Olmsted County on 31 December 2010.

STUDY MEASURES

Seventeen chronic conditions identified by the US DHHS as important in studies of multimorbidity were identified through medical record review of each participant between 2006 and 2010. ICD-9 administrative billing codes corresponding to the 17 conditions were extracted using the Rochester Epidemiology Project records-linkage system. Persons were classified as having each condition using two algorithms: at least one code or at least two codes separated by more than 30 days. We compared the ICD-9 code algorithms with the diagnoses obtained through medical record review to identify persons with multimorbidity (defined as ≥2, ≥3 or ≥4 chronic conditions).

RESULTS

Use of a single code to define each of the 17 chronic conditions resulted in sensitivity and positive predictive values (PPV) ≥70%, and in specificity and negative predictive values (NPV) ≥70% for identifying multimorbidity in the overall study population. PPV and sensitivity were highest in persons 65-84 years of age, whereas NPV and specificity were highest in persons 40-64 years. The results varied by condition, and by age and sex. The use of at least two codes reduced sensitivity, but increased specificity.

CONCLUSIONS

The use of a single code to identify each of the 17 chronic conditions may be a simple and valid method to identify persons who meet the DHHS definition of multimorbidity in populations with similar demographic, socioeconomic, and health care characteristics.

High-throughput multimodal automated phenotyping (MAP) with application to PheWAS

OBJECTIVE

Electronic health records linked with biorepositories are a powerful platform for translational studies. A major bottleneck exists in the ability to phenotype patients accurately and efficiently. The objective of this study was to develop an automated high-throughput phenotyping method integrating International Classification of Diseases (ICD) codes and narrative data extracted using natural language processing (NLP).

MATERIALS AND METHODS

We developed a mapping method for automatically identifying relevant ICD and NLP concepts for a specific phenotype leveraging the Unified Medical Language System. Along with health care utilization, aggregated ICD and NLP counts were jointly analyzed by fitting an ensemble of latent mixture models. The multimodal automated phenotyping (MAP) algorithm yields a predicted probability of phenotype for each patient and a threshold for classifying participants with phenotype yes/no. The algorithm was validated using labeled data for 16 phenotypes from a biorepository and further tested in an independent cohort phenome-wide association studies (PheWAS) for 2 single nucleotide polymorphisms with known associations.

RESULTS

The MAP algorithm achieved higher or similar AUC and F-scores compared to the ICD code across all 16 phenotypes. The features assembled via the automated approach had comparable accuracy to those assembled via manual curation (AUCMAP 0.943, AUCmanual 0.941). The PheWAS results suggest that the MAP approach detected previously validated associations with higher power when compared to the standard PheWAS method based on ICD codes.

CONCLUSION

The MAP approach increased the accuracy of phenotype definition while maintaining scalability, thereby facilitating use in studies requiring large-scale phenotyping, such as PheWAS.

Trends in endovascular interventions for pediatric ischemic stroke at the national level: data from 2000 to 2009

PURPOSE

Present knowledge is limited with regard to endovascular and interventional management of pediatric acute ischemic stroke (AIS). The current practice of neurointerventions in this population was analyzed via a national database.

METHODS

The Kids' Inpatient Database for years 2000, 2003, 2006, and 2009 was examined for patients aged < 18 years discharged with a primary diagnosis of AIS and identified according to ICD-9 codes. Descriptive statistics were tabulated on each of the subcohorts.

RESULTS

There were 3467 patients identified; 920 (26.5%) underwent angiograms, 51 (1.5%) angiogram + thrombolysis, and 18 (0.5%) received angiogram + endovascular recanalization. The angiogram only subcohort was significantly younger compared with thrombolysis and endovascular procedure subcohorts (9.8 vs. 12.2 vs. 14.9 years, P < 0.001). Mortality was 4.3%, significantly lower for angiogram only than for thrombolysis (1.1% vs. 18.2%, P < 0.0001). Thrombolysis also had significantly higher hospital charges ($149,045 vs. $64,826, P < 0.0001). While not many differences in outcomes between angiogram only versus endovascular procedures, the latter had higher financial burden ($122,482 vs. $64,826, P < 0.0001).

CONCLUSIONS

This national study suggests that children receiving neurointerventions tend to be older (> 12 years) and heart and valvular defects are their most likely comorbidities. There was a lower mortality and fewer complications with endovascular procedures when compared with intravenous/intraarterial thrombolysis alone. Thrombolysis was also associated with more non-routine discharges and lengthier stay.

Comparative analysis, applications, and interpretation of electronic health record-based stroke phenotyping methods

BACKGROUND

Accurate identification of acute ischemic stroke (AIS) patient cohorts is essential for a wide range of clinical investigations. Automated phenotyping methods that leverage electronic health records (EHRs) represent a fundamentally new approach cohort identification without current laborious and ungeneralizable generation of phenotyping algorithms. We systematically compared and evaluated the ability of machine learning algorithms and case-control combinations to phenotype acute ischemic stroke patients using data from an EHR.

MATERIALS AND METHODS

Using structured patient data from the EHR at a tertiary-care hospital system, we built and evaluated machine learning models to identify patients with AIS based on 75 different case-control and classifier combinations. We then estimated the prevalence of AIS patients across the EHR. Finally, we externally validated the ability of the models to detect AIS patients without AIS diagnosis codes using the UK Biobank.

RESULTS

Across all models, we found that the mean AUROC for detecting AIS was 0.963 ± 0.0520 and average precision score 0.790 ± 0.196 with minimal feature processing. Classifiers trained with cases with AIS diagnosis codes and controls with no cerebrovascular disease codes had the best average F1 score (0.832 ± 0.0383). In the external validation, we found that the top probabilities from a model-predicted AIS cohort were significantly enriched for AIS patients without AIS diagnosis codes (60-150 fold over expected).

CONCLUSIONS

Our findings support machine learning algorithms as a generalizable way to accurately identify AIS patients without using process-intensive manual feature curation. When a set of AIS patients is unavailable, diagnosis codes may be used to train classifier models.

Identification of Patients with Nontraumatic Intracranial Hemorrhage Using Administrative Claims Data

INTRODUCTION

Nontraumatic intracranial hemorrhage (ICH) is a neurological emergency of research interest; however, unlike ischemic stroke, has not been well studied in large datasets due to the lack of an established administrative claims-based definition. We aimed to evaluate both explicit diagnosis codes and machine learning methods to create a claims-based definition for this clinical phenotype.

METHODS

We examined all patients admitted to our tertiary medical center with a primary or secondary International Classification of Disease version 9 (ICD-9) or 10 (ICD-10) code for ICH in claims from any portion of the hospitalization in 2014-2015. As a gold standard, we defined the nontraumatic ICH phenotype based on manual chart review. We tested explicit definitions based on ICD-9 and ICD-10 that had been previously published in the literature as well as four machine learning classifiers including support vector machine (SVM), logistic regression with LASSO, random forest and xgboost. We report five standard measures of model performance for each approach.

RESULTS

A total of 1830 patients with 2145 unique ICD-10 codes were included in the initial dataset, of which 437 (24%) were true positive based on manual review. The explicit ICD-10 definition performed best (Sensitivity = 0.89 (95% CI 0.85-0.92), Specificity = 0.83 (0.81-0.85), F-score = 0.73 (0.69-0.77)) and improves on an explicit ICD-9 definition (Sensitivity = 0.87 (0.83-0.90), Specificity = 0.77 (0.74-0.79), F-score = 0.67 (0.63-0.71). Among machine learning classifiers, SVM performed best (Sensitivity = 0.78 (0.75-0.82), Specificity = 0.84 (0.81-0.87), AUC = 0.89 (0.87-0.92), F-score = 0.66 (0.62-0.69)).

CONCLUSIONS

An explicit ICD-10 definition can be used to accurately identify patients with a nontraumatic ICH phenotype with substantially better performance than ICD-9. An explicit ICD-10 based definition is easier to implement and quantitatively not appreciably improved with the additional application of machine learning classifiers. Future research utilizing large datasets should utilize this definition to address important research gaps.

Reduction in Stroke After Transient Ischemic Attack in a Province-Wide Cohort Between 2003 and 2015

BACKGROUND

Improvements in management of transient ischemic attack (TIA) have decreased stroke and mortality post-TIA. Studies examining trends over time on a provincial level are limited. We analyzed whether efforts to improve management have decreased the rate of stroke and mortality after TIA from 2003 to 2015 across an entire province.

METHODS

Using administrative data from the Canadian Institute for Health Information's (CIHI) databases from 2003 to 2015, we identified a cohort of patients with a diagnosis of TIA upon discharge from the emergency department (ED). We examined stroke rates at Day 1, 2, 7, 30, 90, 180, and 365 post-TIA and 1-year mortality rates and compared trends over time between 2003 and 2015.

RESULTS

From 2003 to 2015 in Ontario, there were 61,710 patients with an ED diagnosis of TIA. Linear regressions of stroke after the index TIA showed a significant decline between 2003 and 2015, decreasing by 25% at Day 180 and 32% at 1 year (p < 0.01). The 1-year stroke rate decreased from 6.0% in 2003 to 3.4% in 2015. Early (within 48 h) stroke after TIA continued to represent approximately half of the 1-year event rates. The 1-year mortality rate after ED discharge following a TIA decreased from 1.3% in 2003 to 0.3% in 2015 (p < 0.001).

INTERPRETATION

At a province-wide level, 1-year rates of stroke and mortality after TIA have declined significantly between 2003 and 2015, suggesting that efforts to improve management may have contributed toward the decline in long-term risk of stroke and mortality. Continued efforts are needed to further reduce the immediate risk of stroke following a TIA.

Discharge against medical advice among neurological patients: Characteristics and outcomes

OBJECTIVE

To study characteristics and outcomes of patients with stroke, traumatic brain injury (TBI), and epilepsy with discharge against medical advice (DAMA).

DATA SOURCES/STUDY SETTING

Retrospective analysis of the 2013 Nationwide Readmissions Database, a nationally representative inpatient administrative dataset.

STUDY DESIGN

Associations between predictors and DAMA at index admission were analyzed using adjusted logistic models. We examined 30-day all-cause readmissions.

DATA COLLECTION METHODS

Patients aged ≥18 years at index admission for International Classification of Diseases-9 diagnosis code of epilepsy, TBI, or stroke were included.

PRINCIPAL FINDINGS

Discharge against medical advice occurred in 1998/58278 patients (3.43 percent) in the epilepsy group, 1762/211 213 (0.83 percent) in the stroke group, and 1289/74 652 (1.73 percent) in the TBI group. Factors consistently associated with increased likelihood of DAMA included lower age, male sex, non-Medicare and nonprivate insurance, lower socioeconomic status, and behavioral risk factors (smoking history, alcohol history, and drug use). The crude 30-day all-cause readmission rate for those with DAMA from their index admission was 16.4 percent for the stroke cohort, 13.9 percent for epilepsy, and 13.4 percent for TBI. DAMA at index admission was significantly associated with increased risk of 30-day all-cause readmission among all groups (adjusted odds ratio 1.79, 95% CI: 1.65-1.94, P < .0001).

CONCLUSIONS

Age, sex, insurance status, socioeconomic status, and behavioral factors were associated with DAMA in neurological patients. Further research is needed to develop interventions to reduce DAMA in high-risk groups.

FasTag: Automatic text classification of unstructured medical narratives

Unstructured clinical narratives are continuously being recorded as part of delivery of care in electronic health records, and dedicated tagging staff spend considerable effort manually assigning clinical codes for billing purposes. Despite these efforts, however, label availability and accuracy are both suboptimal. In this retrospective study, we aimed to automate the assignment of top-level International Classification of Diseases version 9 (ICD-9) codes to clinical records from human and veterinary data stores using minimal manual labor and feature curation. Automating top-level annotations could in turn enable rapid cohort identification, especially in a veterinary setting. To this end, we trained long short-term memory (LSTM) recurrent neural networks (RNNs) on 52,722 human and 89,591 veterinary records. We investigated the accuracy of both separate-domain and combined-domain models and probed model portability. We established relevant baseline classification performances by training Decision Trees (DT) and Random Forests (RF). We also investigated whether transforming the data using MetaMap Lite, a clinical natural language processing tool, affected classification performance. We showed that the LSTM-RNNs accurately classify veterinary and human text narratives into top-level categories with an average weighted macro F1 score of 0.74 and 0.68 respectively. In the "neoplasia" category, the model trained on veterinary data had a high validation accuracy in veterinary data and moderate accuracy in human data, with F1 scores of 0.91 and 0.70 respectively. Our LSTM method scored slightly higher than that of the DT and RF models. The use of LSTM-RNN models represents a scalable structure that could prove useful in cohort identification for comparative oncology studies. Digitization of human and veterinary health information will continue to be a reality, particularly in the form of unstructured narratives. Our approach is a step forward for these two domains to learn from and inform one another.

Secondary Stroke Prevention: A Population-Based Cohort Study on Anticoagulation and Antiplatelet Treatments, and the Risk of Death or Recurrence

Using claims databases of a public healthcare program (Quebec) for the years 2010-2013, we conducted a cohort study of patients with acute ischemic stroke (AIS) to describe secondary prevention treatments and determine how they stood against practice guidelines. We compared the risk of death or AIS recurrence over 1 year in patients treated with anticoagulants, antiplatelets, and/or other cardiovascular drugs. In the month after discharge, 44.3% of the patients did not receive the recommended treatment and > 20% did not have any treatment. Untreated patients were younger, had less comorbidities, and a more severe AIS. Anticoagulants and antiplatelets were associated with a reduced risk of death or recurrence (hazard ratio (HR) 0.27; 95% confidence interval (CI) 0.20-0.36 and HR 0.25; 95% CI 0.16-0.38, respectively) compared with the untreated group. Effect size was similar for the other treatments. Findings confirm treatment benefits shown in clinical trials and emphasize the importance of AIS secondary prevention.

Reliability of International Classification of Disease-9 Versus International Classification of Disease-10 Coding for Proximal Femur Fractures at a Level 1 Trauma Center

INTRODUCTION

The Centers for Medicare & Medicaid services proposed that transitioning from the 9th to the 10th revision of the International Classification of Disease (ICD) would provide better data for research. This study sought to determine the reliability of ICD-10 compared with ICD-9 for proximal femur fractures.

METHODS

Available imaging studies from 196 consecutively treated proximal femur fractures were retrospectively reviewed and assigned ICD codes by three physicians. Intercoder reliability (ICR) was calculated. Collectively, the physicians agreed on what should be the correct codes for each fracture, and this was compared with coding found in the medical and billing records.

RESULTS

No significant difference was observed in ICR for both ICD-9 and ICD-10 exact coding, which were both unreliable. Less specific coding improved ICR. ICD-9 general coding was better than ICD-10. Electronic medical record coding was unreliable. Billing codes were also unreliable, yet ICD-10 was better than ICD-9.

DISCUSSION

ICD-9 and ICD-10 lack reliability in coding proximal femur fractures. ICD-10 results in data that are no more reliable than those found with ICD-9.

LEVEL OF EVIDENCE

Level I diagnostic.

Human gene and disease associations for clinical-genomics and precision medicine research

We are entering the era of personalized medicine in which an individual's genetic makeup will eventually determine how a doctor can tailor his or her therapy. Therefore, it is becoming critical to understand the genetic basis of common diseases, for example, which genes predispose and rare genetic variants contribute to diseases, and so on. Our study focuses on helping researchers, medical practitioners, and pharmacists in having a broad view of genetic variants that may be implicated in the likelihood of developing certain diseases. Our focus here is to create a comprehensive database with mobile access to all available, authentic and actionable genes, SNPs, and classified diseases and drugs collected from different clinical and genomics databases worldwide, including Ensembl, GenCode, ClinVar, GeneCards, DISEASES, HGMD, OMIM, GTR, CNVD, Novoseek, Swiss-Prot, LncRNADisease, Orphanet, GWAS Catalog, SwissVar, COSMIC, WHO, and FDA. We present a new cutting-edge gene-SNP-disease-drug mobile database with a smart phone application, integrating information about classified diseases and related genes, germline and somatic mutations, and drugs. Its database includes over 59 000 protein-coding and noncoding genes; over 67 000 germline SNPs and over a million somatic mutations reported for over 19 000 protein-coding genes located in over 1000 regions, published with over 3000 articles in over 415 journals available at the PUBMED; over 80 000 ICDs; over 123 000 NDCs; and over 100 000 classified gene-SNP-disease associations. We present an application that can provide new insights into the information about genetic basis of human complex diseases and contribute to assimilating genomic with phenotypic data for the availability of gene-based designer drugs, precise targeting of molecular fingerprints for tumor, appropriate drug therapy, predicting individual susceptibility to disease, diagnosis, and treatment of rare illnesses are all a few of the many transformations expected in the decade to come.

Capturing anaphylaxis through medical records: Are ICD and CPT codes sufficient?

OBJECTIVE

The identification of anaphylaxis cases is imperative for optimal clinicalprovider knowledge deficiencies in diagnosis and treatment and the efficacy of reimbursement codes, such as International Classification of Diseases (ICD) and current procedural terminology (CPT) codes, in detecting anaphylaxis.

DATA SOURCES

Pubmed.

STUDY SELECTIONS

Recent and clinically relevant literature on anaphylaxis and provider knowledge, ICD, CPT, Healthcare Common Procedural Coding System (HCPCS), and E-codes were selected and reviewed.

RESULTS

Reimbursement codes are used to detect anaphylaxis in administrative claims databases. Inaccurate recognition of the diagnosis by providers, underreporting, and cause identification are challenges faced by health researchers using reimbursement codes for anaphylaxis case identification. Anaphylactic shock-specific ICD codes were noted to have a positive predictive value (PPV) of 52% to 53% of anaphylaxis events compared with physician chart review, which was improved to 63% to 67.3% when used in conjunction with anaphylaxis symptom-specific ICD, CPT, HCPCS, and E-codes 31, 34, and 35.

CONCLUSION

Education of providers to properly diagnose and treat anaphylaxis requires systematic and educational investments. The ICD codes specific to anaphylactic shock have suboptimal PPV to identify anaphylaxis in administrative claims databases. Use of algorithms incorporating other reimbursement codes improve the PPV, but they are limited by inaccurate diagnoses and underreporting of anaphylaxis. Future ICD-11 reclassification may improve anaphylaxis detection by reimbursement codes.

Impact of Personality Disorder Cluster on Depression Outcomes Within Collaborative Care Management Model of Care

BACKGROUND

Previous studies have suggested that having a comorbid personality disorder (PD) along with major depression is associated with poorer depression outcomes relative to those without comorbid PD. However, few studies have examined the influence of specific PD cluster types. The purpose of the current study is to compare depression outcomes between cluster A, cluster B, and cluster C PD patients treated within a collaborative care management (CCM), relative to CCM patients without a PD diagnosis. The overarching goal was to identify cluster types that might confer a worse clinical prognosis.

METHODS

This retrospective chart review study examined 2826 adult patients with depression enrolled in CCM. The cohort was divided into 4 groups based on the presence of a comorbid PD diagnosis (cluster A/nonspecified, cluster B, cluster C, or no PD). Baseline clinical and demographic variables, along with 6-month follow-up Patient Health Questionnaire-9 (PHQ-9) scores were obtained for all groups. Depression remission was defined as a PHQ-9 score <5 at 6 months, and persistent depressive symptoms (PDS) was defined as a PHQ-9 score ≥10 at 6 months. Adjusted odds ratios (AORs) were determined for both remission and PDS using logistic regression modeling for the 6-month PHQ-9 outcome, while retaining all study variables.

RESULTS

A total of 59 patients (2.1%) had a cluster A or nonspecified PD diagnosis, 122 patients (4.3%) had a cluster B diagnosis, 35 patients (1.2%) had a cluster C diagnosis, and 2610 patients (92.4%) did not have any PD diagnosis. The presence of a cluster A/nonspecified PD diagnosis was associated with a 62% lower likelihood of remission at 6 months (AOR = 0.38; 95% CI 0.20-0.70). The presence of a cluster B PD diagnosis was associated with a 71% lower likelihood of remission at 6 months (AOR = 0.29; 95% CI 0.18-0.47). Conversely, having a cluster C diagnosis was not associated with a significantly lower likelihood of remission at 6 months (AOR = 0.83; 95% CI 0.42-1.65). Increased odds of having PDS at 6-month follow-up were seen with cluster A/nonspecified PD patients (AOR = 3.35; 95% CI 1.92-5.84) as well as cluster B patients (AOR = 3.66; 95% CI 2.45-5.47). However, cluster C patents did not have significantly increased odds of experiencing persistent depressive symptoms at 6-month follow-up (AOR = 0.95; 95% CI 0.45-2.00).

CONCLUSIONS

Out of the 3 clusters, the presence of a cluster B PD diagnosis was most significantly associated with poorer depression outcomes at 6-month follow-up, including reduced remission rates and increased risk for PDS. The cluster A/nonspecified PD group also showed poor outcomes; however, the heterogeneity of this subgroup with regard to PD features must be noted. The development of novel targeted interventions for at-risk clusters may be warranted in order to improve outcomes of these patients within the CCM model of care.

Alcohol-Related Hospital Encounters Trigger Thrombotic and Hemorrhagic Vascular Events

BACKGROUND AND PURPOSE

We investigated the associations between alcohol-related emergency department visits and hospitalizations and vascular events including acute ischemic stroke, intracerebral hemorrhage (ICH), and subarachnoid hemorrhage.

METHODS

The New York State Inpatient and Emergency Department Databases were examined (2006-2013). Validated International Classification of Diseases 9th edition definitions identified index vascular hospitalizations and alcohol abuse encounters. We used case cross-over analysis with conditional logistic regression to estimate odds ratios (OR) for the association between alcohol-related encounters during 6 case periods (7, 14, 30, 60, 90, and 120 days before index event) compared to control periods (1 year before). Multivariate logistic regression was used to examine the association between an alcohol-related encounter within 6 months before index admission and 30-day readmission after discharge.

RESULTS

An alcohol encounter before index admission was associated with acute ischemic stroke (OR = 1.765 within 60 days, 1.418 within 90 days, and 1.287 within 120 days) and subarachnoid hemorrhage (OR = 2.375 within 90 days), but not ICH. Alcohol-related encounters within 6 months before index vascular events increased the likelihood of 30-day readmission after index admission.

CONCLUSION

We found that a recent alcohol-related counter was associated with occurrence of vascular events, but not ICH, as well as worse outcomes after index admission.

Text preprocessing for improving hypoglycemia detection from clinical notes - A case study of patients with diabetes

BACKGROUND AND OBJECTIVE

Hypoglycemia is a common safety event when attempting to optimize glycemic control in diabetes (DM). While electronic medical records provide a natural ground for detecting and analyzing hypoglycemia, ICD codes used in the databases may be invalid, insensitive or non-specific in detecting new hypoglycemic events. We developed text preprocessing methods to improve automatic detection of hypoglycemia from analysis of clinical encounter text notes.

METHODS

We set out to improve hypoglycemia detection from clinical notes by introducing three preprocessing methods: stop word filtering, medication signaling, and ICD narrative enrichment. To test the proposed methods, we selected clinical notes from VA Maryland Healthcare System, based on various combinations of three criteria that are suggestive of hypoglycemia, including ICD-9 code of diabetes and hypoglycemia, laboratory glucose values < 70 md/dL, and text reference to a proximate hypoglycemia event. In addition, we constructed one dataset of 395 clinical notes from year 2009 and another of 460 notes from year 2014 to test the generality of the proposed methods. For each of the datasets, two physician judges manually reviewed individual clinical notes to determine whether hypoglycemia was present or absent. A third physician judge served as a final adjudicator for disagreements.

RESULTS

Each of the proposed preprocessing methods contributed to the performance of hypoglycemia detection by significantly increasing the F1 score in the range of 5.3∼7.4% on one dataset (p < .01). Among the methods, stop word filtering contributed most to the performance improvement (7.4%). Combining all the preprocessing methods led to greater performance gain (p < .001) compared with using each method individually. Similar patterns were observed for the other dataset with the F1 score being increased in the range of 7.7%∼9.4% by individual methods (p < .001). Nevertheless, combining the three methods did not yield additional performance gain.

CONCLUSION

The proposed text preprocessing methods improved the performance of hypoglycemia detection from clinical text notes. Stop word filtering achieved the most performance improvement. ICD narrative enrichment boosted the recall of detection. Combining the three preprocessing methods led to additional performance gains.

Economic analysis of rapid multiplex polymerase chain reaction testing for meningitis/encephalitis in adult patients

PURPOSE

Many patients with suspected meningitis do not require hospitalization yet are admitted, often resulting in unnecessary care and additional cost. We assessed the possible economic impact of a rapid multiplex test for suspected adult community-acquired meningitis/encephalitis.

METHODS

A model simulated diagnosis, clinical decisions, resource use/costs of standard of care (SOC) and two cerebrospinal fluid (CSF) testing strategies using the FDA-cleared BioFire® FilmArray® System (FA) which provides results in approximately one hour.

RESULTS

Pathogens detected by FA caused approximately 74% of cases, 97% of which would be accurately diagnosed with FA. False positives and false negatives more often led to extended/unnecessary admission than inappropriate discharge/missed admission. Mean cost per case ranged from 16829 to 20791. A strategy of testing all suspected cases yielded greater savings (2213/case) than testing only those with abnormal CSF (812/case) and both were less expensive than SOC.

CONCLUSION

This economic analysis demonstrates that FA can inform more appropriate clinician decisions resulting in cost savings with greater economic benefits achievable with syndromic testing of all cases, rather than SOC or targeted syndromic testing.

Feature extraction for phenotyping from semantic and knowledge resources

OBJECTIVE

Phenotyping algorithms can efficiently and accurately identify patients with a specific disease phenotype and construct electronic health records (EHR)-based cohorts for subsequent clinical or genomic studies. Previous studies have introduced unsupervised EHR-based feature selection methods that yielded algorithms with high accuracy. However, those selection methods still require expert intervention to tweak the parameter settings according to the EHR data distribution for each phenotype. To further accelerate the development of phenotyping algorithms, we propose a fully automated and robust unsupervised feature selection method that leverages only publicly available medical knowledge sources, instead of EHR data.

METHODS

SEmantics-Driven Feature Extraction (SEDFE) collects medical concepts from online knowledge sources as candidate features and gives them vector-form distributional semantic representations derived with neural word embedding and the Unified Medical Language System Metathesaurus. A number of features that are semantically closest and that sufficiently characterize the target phenotype are determined by a linear decomposition criterion and are selected for the final classification algorithm.

RESULTS

SEDFE was compared with the EHR-based SAFE algorithm and domain experts on feature selection for the classification of five phenotypes including coronary artery disease, rheumatoid arthritis, Crohn's disease, ulcerative colitis, and pediatric pulmonary arterial hypertension using both supervised and unsupervised approaches. Algorithms yielded by SEDFE achieved comparable accuracy to those yielded by SAFE and expert-curated features. SEDFE is also robust to the input semantic vectors.

CONCLUSION

SEDFE attains satisfying performance in unsupervised feature selection for EHR phenotyping. Both fully automated and EHR-independent, this method promises efficiency and accuracy in developing algorithms for high-throughput phenotyping.

ICD-10-PCS extension with ICD-9 procedure codes to support integrated access to clinical legacy data

Since the creation of The International Classification of Diseases (ICD), new versions have been released to keep updated with the current medical knowledge. Migrations of Electronic Health Records (EHR) from ICD-9 to ICD-10-PCS as clinical procedure codification system, has been a significant challenge and involved large resources. In addition, it created new barriers for integrated access to legacy medical procedure data (frequently ICD-9 coded) with current data (frequently ICD-10-PCS coded). This work proposes a solution based on extending ICD-10-PCS with a subgroup of ICD-9-CM concepts to facilitate such integrated access. The General Equivalence Mappings (GEMs) has been used as foundation to set the terminology relations of these inserted concepts in ICD-10-PCS hierarchy, but due to the existence of 1-to-many mappings, advanced rules are required to seamlessly integrate both terminologies. With the generation of rules based on GEMs relationships, 2014 ICD-9 concepts were included within the ICD-10-PCS hierarchy. For the rest of the concepts, a new method is also proposed to increase 1-to-1 mappings. As results, with the suggested approach, the percentage of ICD-9-CM procedure concepts that can be mapped accurately (avoiding mappings to a large number of concepts) rise from 11.56% to 69.01% of ICD-9-Proc, through the extended ICD-10-PCS hierarchy.

Validation of ICD-10 codes shows intracranial venous thrombosis incidence to be higher than previously reported

BACKGROUND

Intracranial venous thrombosis (ICVT) accounts for around 0.5% of all stroke cases. There have been no previously published studies of the International Classification of Diseases, Tenth Edition (ICD-10) validation for the identification of ICVT admissions in adults.

OBJECTIVE

The aims of this study were to validate and quantify the performance of the ICD-10 coding system for identifying cases of ICVT in adults and to derive an estimate of incidence.

METHOD

Administrative data were collected for all patients admitted to a regional neurosciences centre over a 5-year period. We searched for the following ICD-10 codes at any position: G08.X (intracranial and intraspinal phlebitis and thrombophlebitis), I67.6 (non-pyogenic thrombosis of intracranial venous system), I63.6 (cerebral infarction due to cerebral venous thrombosis, non-pyogenic), O22.5 (cerebral venous thrombosis in pregnancy) and O87.3 (cerebral venous thrombosis in the puerperium).

RESULTS

Sixty-five admissions were identified by at least one of the relevant ICD-10 codes. The overall positive predictive value (PPV) for confirmed ICVT from all of the admissions combined was 92.3% (60 out of 65) with the results for each code as follows: G08.X 91.5% (54 of 59), O22.5 100% (4 of 4), I67.6 100% (1 of 1), I63.6 100% (1 of 1) and O87.3 100% (1 of 1). There were 40 unique cases of ICVT over a 5-year period giving an annual incidence of ICVT of 5 per million.

CONCLUSIONS

All codes gave a high PPV.

IMPLICATIONS FOR PRACTICE

As demonstrated in previous studies, the incidence of ICVT may be higher than previously thought.

Towards automated clinical coding

BACKGROUND

Patients' encounters with healthcare services must undergo clinical coding. These codes are typically derived from free-text notes. Manual clinical coding is expensive, time-consuming and prone to error. Automated clinical coding systems have great potential to save resources, and realtime availability of codes would improve oversight of patient care and accelerate research. Automated coding is made challenging by the idiosyncrasies of clinical text, the large number of disease codes and their unbalanced distribution.

METHODS

We explore methods for representing clinical text and the labels in hierarchical clinical coding ontologies. Text is represented as term frequency-inverse document frequency counts and then as word embeddings, which we use as input to recurrent neural networks. Labels are represented atomically, and then by learning representations of each node in a coding ontology and composing a representation for each label from its respective node path. We consider different strategies for initialisation of the node representations. We evaluate our methods using the publicly-available Medical Information Mart for Intensive Care III dataset: we extract the history of presenting illness section from each discharge summary in the dataset, then predicting the International Classification of Diseases, ninth revision, Clinical Modification codes associated with these.

RESULTS

Composing the label representations from the clinical-coding-ontology nodes increased weighted F1 for prediction of the 17,561 disease labels to 0.264-0.281 from 0.232-0.249 for atomic representations. Recurrent neural network text representation improved weighted F1 for prediction of the 19 disease-category labels to 0.682-0.701 from 0.662-0.682 using term frequency-inverse document frequency. However, term frequency-inverse document frequency outperformed recurrent neural networks for prediction of the 17,561 disease labels.

CONCLUSIONS

This study demonstrates that hierarchically-structured medical knowledge can be incorporated into statistical models, and produces improved performance during automated clinical coding. This performance improvement results primarily from improved representation of rarer diseases. We also show that recurrent neural networks improve representation of medical text in some settings. Learning good representations of the very rare diseases in clinical coding ontologies from data alone remains challenging, and alternative means of representing these diseases will form a major focus of future work on automated clinical coding.

De facto diagnosis specialties: Recognition and discovery

A medical specialty indicates the skills needed by health care providers to conduct key procedures or make critical judgments. However, documentation about specialties may be lacking or inaccurately specified in a health care institution. Thus, we propose to leverage diagnosis histories to recognize medical specialties that exist in practice. Such specialties that are highly recognizable through diagnosis histories are de facto diagnosis specialties. We aim to recognize de facto diagnosis specialties that are listed in the Health Care Provider Taxonomy Code Set (HPTCS) and discover those that are unlisted. First, to recognize the former, we use similarity and supervised learning models. Next, to discover de facto diagnosis specialties unlisted in the HPTCS, we introduce a general discovery-evaluation framework. In this framework, we use a semi-supervised learning model and an unsupervised learning model, from which the discovered specialties are subsequently evaluated by the similarity and supervised learning models used in recognition. To illustrate the potential for these approaches, we collect 2 data sets of 1 year of diagnosis histories from a large academic medical center: One is a subset of the other except for additional information useful for network analysis. The results indicate that 12 core de facto diagnosis specialties listed in the HPTCS are highly recognizable. Additionally, the semi-supervised learning model discovers a specialty for breast cancer on the smaller data set based on network analysis, while the unsupervised learning model confirms this discovery and suggests an additional specialty for Obesity on the larger data set. The potential correctness of these 2 specialties is reinforced by the evaluation results that they are highly recognizable by similarity and supervised learning models in comparison with 12 core de facto diagnosis specialties listed in the HPTCS.

Automated versus Manual Data Extraction of the Padua Prediction Score for Venous Thromboembolism Risk in Hospitalized Older Adults

OBJECTIVE

Venous thromboembolism (VTE) prophylaxis is an important consideration for hospitalized older adults, and the Padua Prediction Score (PPS) is a risk prediction tool used to prioritize patient selection. We developed an automated PPS (APPS) algorithm using electronic health record (EHR) data. This study examines the accuracy of APPS and its individual components versus manual data extraction.

METHODS

This is a retrospective cohort study of hospitalized general internal medicine patients, aged 70 and over. Fourteen clinical variables were collected to determine their PPS; APPS used EHR data exports from health system databases, and a trained abstractor performed manual chart abstractions. We calculated sensitivity and specificity of the APPS, using manual PPS as the gold standard for classifying risk category (low vs. high). We also examined performance characteristics of the APPS for individual variables.

RESULTS

PPS was calculated by both methods on 311 individuals. The mean PPS was 3.6 (standard deviation, 1.8) for manual abstraction and 2.8 (1.4) for APPS. In detecting patients at high risk for VTE, the sensitivity and specificity of the APPS algorithm were 46 and 94%, respectively. The sensitivity for APPS was poor (range: 6-34%) for detecting acute conditions (i.e., acute myocardial infarction), moderate (range: 52-74%) for chronic conditions (i.e., heart failure), and excellent (range: 94-98%) for conditions of obesity and restricted mobility. Specificity of the automated extraction method for each PPS variable was > 87%.

CONCLUSION

APPS as a stand-alone tool was suboptimal for classifying risk of VTE occurrence. The APPS accurately identified high risk patients (true positives), but lower scores were considered indeterminate.

Administrative data underestimate acute ischemic stroke events and thrombolysis treatments: Data from a multicenter validation survey in Italy

Background

Informing health systems and monitoring hospital performances using administrative data sets, mainly hospital discharge data coded according to International-Classification-Diseases-9edition-Clinical-Modifiers (ICD9-CM), is now commonplace in several countries, but the reliability of diagnostic coding of acute ischemic stroke in the routine practice is uncertain. This study aimed at estimating accuracy of ICD9-CM codes for the identification of acute ischemic stroke and the use of thrombolysis treatment comparing hospital discharge data with medical record review in all the six hospitals of the Florence Area, Italy, through 2015.

Methods

We reviewed the medical records of all the 3915 potential acute stroke events during 2015 across the six hospitals of the Florence Area, Italy. We then estimated sensitivity and Positive Predictive Value of ICD9-CM code-groups 433*1, 434*1 and thrombolysis code 99.10 against medical record review with clinical adjudication. For each false-positive case we obtained the actual diagnosis. For each false-negative case we obtained the primary and secondary ICD9-CM diagnoses.

Results

The medical record review identified 1273 acute ischemic stroke events. The hospital discharge records identified 898 among those (true-positive cases),but missed 375 events (false-negative cases), and identified 104 events that were not eventually confirmed as acute ischemic events (false-positive cases). Code-group specific Positive Predictive Value was 85.7% (95%CI,74.6–93.3) for 433*1 and 89.9% (95%CI, 87.8–91.7) for 434*1 codes. Thrombolysis treatment, as identified by ICD9-CM code 99.10, was only documented in 6.0% of acute ischemic stroke events, but was 13.6% in medical record review.

Conclusions

Hospital discharge data were found to be fairly specific but insensitive in the reporting of acute ischemic stroke and thrombolysis, providing misleading indications about both quantity and quality of acute ischemic stroke hospital care. Efforts to improve coding accuracy should precede the use of hospital discharge data to measure hospital performances in acute ischemic stroke care.

Comparative cardiovascular safety of nonsteroidal anti-inflammatory drugs in patients with hypertension: a population-based cohort study

AIMS

Previous studies have suggested that nonsteroidal anti-inflammatory drugs (NSAIDs) may be associated with higher cardiovascular risks. However, few have been active comparison studies that directly assessed the potential differential cardiovascular risk between NSAID classes or across individual NSAIDs. We compared the risk of major cardiovascular events between cyclooxygenase 2 (COX-2)-selective and nonselective NSAIDs in patients with hypertension.

METHODS

We conducted a cohort study of patients with hypertension who initiated COX-2-selective or nonselective NSAIDs in a population-based Taiwanese database. The outcomes included hospitalization for the following major cardiovascular events: ischaemic stroke, acute myocardial infarction, congestive heart failure, transient ischaemic attack, unstable angina or coronary revascularization. We followed patients for up to 4 weeks, based on the as-treated principle. We used inverse probability weighting to control for baseline and time-varying covariates, and estimated the on-treatment hazard ratios (HRs) and 95% conservative confidence interval (CIs).

RESULTS

We identified 2749 eligible COX-2-selective NSAID users and 52 880 eligible nonselective NSAID users. The HR of major cardiovascular events comparing COX-2-selective with nonselective NSAIDs after adjusting for baseline and time-varying covariates was 1.07 (95% CI 0.65, 1.74). We did not observe a differential risk when comparing celecoxib to diclofenac (HR 1.17; 95% CI 0.61, 2.25), ibuprofen (HR 1.36; 95% CI 0.58, 3.18) or naproxen (HR 0.75; 95% CI 0.23, 2.44). There was an increased risk with COX-2-selective NSAIDs, however, when comparing COX-2-selective NSAIDs with mefenamic acid (HR 2.11; 95% CI 1.09, 4.09).

CONCLUSIONS

Our results provide important information about the comparative cardiovascular safety of NSAIDs in patients with hypertension.

Diagnostic accuracy of ICD-9 code 780.2 for the identification of patients with syncope in the emergency department

PURPOSE

Syncope is a common condition that affects individuals of all ages and is responsible for 1-3% of all emergency department (ED) visits. Prospective studies on syncope are often limited by the exiguous number of subjects enrolled. A possible alternative approach would be to use of hospital discharge diagnoses from administrative databases to identify syncope subjects in epidemiological observational studies. We assessed the accuracy of the International Classification of Diseases, Ninth Revision (ICD-9) code 780.2 "syncope and collapse" to identify patients with syncope.

METHODS

Patients in two teaching hospitals in Milan, Italy with a triage assessment for ED access that was possibly related to syncope were recruited in this study. We considered the index test to be the attribution of the ICD-9 code 780.2 at ED discharge and the reference standard to be the diagnosis of syncope by the ED physician.

RESULTS

The sensitivity, specificity, positive and negative predictive values of the ICD-9 code 780.2 to identify patients with syncope were 0.63 (95% confidence interval [CI] 0.58-0.67), 0.98 (95% CI 0.98-0.99), 0.83 (95% CI 0.79-0.87) and 0.95 (95% CI 0.94-0.95), respectively.

CONCLUSIONS

The moderate sensitivity of ICD-9 code 780.2 should be considered when the code is used to identify patients with syncope through administrative databases.

Surrogate-assisted feature extraction for high-throughput phenotyping

Objective

Phenotyping algorithms are capable of accurately identifying patients with specific phenotypes from within electronic medical records systems. However, developing phenotyping algorithms in a scalable way remains a challenge due to the extensive human resources required. This paper introduces a high-throughput unsupervised feature selection method, which improves the robustness and scalability of electronic medical record phenotyping without compromising its accuracy.

Methods

The proposed Surrogate-Assisted Feature Extraction (SAFE) method selects candidate features from a pool of comprehensive medical concepts found in publicly available knowledge sources. The target phenotype's International Classification of Diseases, Ninth Revision and natural language processing counts, acting as noisy surrogates to the gold-standard labels, are used to create silver-standard labels. Candidate features highly predictive of the silver-standard labels are selected as the final features.

Results

Algorithms were trained to identify patients with coronary artery disease, rheumatoid arthritis, Crohn's disease, and ulcerative colitis using various numbers of labels to compare the performance of features selected by SAFE, a previously published automated feature extraction for phenotyping procedure, and domain experts. The out-of-sample area under the receiver operating characteristic curve and F -score from SAFE algorithms were remarkably higher than those from the other two, especially at small label sizes.

Conclusion

SAFE advances high-throughput phenotyping methods by automatically selecting a succinct set of informative features for algorithm training, which in turn reduces overfitting and the needed number of gold-standard labels. SAFE also potentially identifies important features missed by automated feature extraction for phenotyping or experts.

The obesity paradox and incident cardiovascular disease: A population-based study

Background

Prior work suggests that obesity may confer a survival advantage among persons with cardiovascular disease (CVD). This obesity “paradox” is frequently studied in the context of prevalent disease, a stage in the disease process when confounding from illness-related weight loss and selective survival are especially problematic. Our objective was to examine the association of obesity with mortality among persons with incident CVD, where biases are potentially reduced, and to compare these findings with those based on prevalent disease.

Methods

We used data from the Health and Retirement Study, an ongoing, nationally representative longitudinal survey of U.S. adults age 50 years and older initiated in 1992 and linked to Medicare claims. Cox proportional hazard models were used to estimate the association between weight status and mortality among persons with specific CVD diagnoses. CVD diagnoses were established by self-reported survey data as well as Medicare claims. Prevalent disease models used concurrent weight status, and incident disease models used pre-diagnosis weight status.

Results

We examined myocardial infarction, congestive heart failure, stroke, and ischemic heart disease. A strong and significant obesity paradox was consistently observed in prevalent disease models (hazard of death 18–36% lower for obese class I relative to normal weight), replicating prior findings. However, in incident disease models of the same conditions in the same dataset, there was no evidence of this survival benefit. Findings from models using survey- vs. claims-based diagnoses were largely consistent.

Conclusion

We observed an obesity paradox in prevalent CVD, replicating prior findings in a population-based sample with longer-term follow-up. In incident CVD, however, we did not find evidence of a survival advantage for obesity. Our findings do not offer support for reevaluating clinical and public health guidelines in pursuit of a potential obesity paradox.

Population-based study of ischemic stroke risk after trauma in children and young adults

OBJECTIVE

To quantify the incidence, timing, and risk of ischemic stroke after trauma in a population-based young cohort.

METHODS

We electronically identified trauma patients (<50 years old) from a population enrolled in a Northern Californian integrated health care delivery system (1997-2011). Within this cohort, we identified cases of arterial ischemic stroke within 4 weeks of trauma and 3 controls per case. A physician panel reviewed medical records, confirmed cases, and adjudicated whether the stroke was related to trauma. We calculated the 4-week stroke incidence and estimated stroke odds ratios (OR) by injury location using logistic regression.

RESULTS

From 1,308,009 trauma encounters, we confirmed 52 trauma-related ischemic strokes. The 4-week stroke incidence was 4.0 per 100,000 encounters (95% confidence interval [CI] 3.0-5.2). Trauma was multisystem in 26 (50%). In 19 (37%), the stroke occurred on the day of trauma, and all occurred within 15 days. In 7/28 cases with cerebrovascular angiography at the time of trauma, no abnormalities were detected. In unadjusted analyses, head, neck, chest, back, and abdominal injuries increased stroke risk. Only head (OR 4.1, CI 1.1-14.9) and neck (OR 5.6, CI 1.03-30.9) injuries remained associated with stroke after adjusting for demographics and trauma severity markers (multisystem trauma, motor vehicle collision, arrival by ambulance, intubation).

CONCLUSIONS

Stroke risk is elevated for 2 weeks after trauma. Onset is frequently delayed, providing an opportunity for stroke prevention during this period. However, in one-quarter of stroke cases with cerebrovascular angiography at the time of trauma, no vascular abnormality was detected.

Chart validation of inpatient ICD-9-CM administrative diagnosis codes for ischemic stroke among IGIV users in the Sentinel Distributed Database

The Sentinel Distributed Database (SDD) is a large database of patient-level medical and prescription records, primarily derived from insurance claims and electronic health records, and is sponsored by the U.S. Food and Drug Administration for drug safety assessments. In this chart validation study, we report on the positive predictive value (PPV) of inpatient ICD-9-CM acute ischemic stroke (AIS) administrative diagnosis codes (433.x1, 434.xx, and 436) in the SDD.As part of an assessment of the risk of thromboembolic adverse events following treatment with intravenous immune globulin (IGIV), charts were obtained for 131 potential post-IGIV AIS cases. Charts were abstracted by trained nurses and then adjudicated by stroke experts using pre-specified diagnostic criteria.Case status could be determined for 128 potential AIS cases, of which 34 were confirmed. The PPVs for the inpatient AIS diagnoses recorded in the SDD were 27% overall [95% confidence interval (95% CI): 19-35], 60% (95% CI: 32-84) for principal-position diagnoses, 42% (95% CI: 28-57) for secondary diagnoses, and 6% (95% CI: 2-15) for position-unspecified diagnoses (which in the SDD generally originate from separate physician claims associated with an inpatient stay).Position-unspecified diagnoses were unlikely to represent true AIS cases. PPVs for principal and secondary inpatient diagnosis codes were higher, but still meaningfully lower than estimates from prior chart validation studies. The low PPVs may be specific to the IGIV user study population. Additional research is needed to assess the validity of AIS administrative diagnosis codes in other study populations within the SDD.

Development of Query Strategies to Identify a Histologic Lymphoma Subtype in a Large Linked Database System

Background

Large linked databases (LLDB) represent a novel resource for cancer outcomes research. However, accurate means of identifying a patient population of interest within these LLDBs can be challenging. Our research group developed a fully integrated platform that provides a means of combining independent legacy databases into a single cancer-focused LLDB system. We compared the sensitivity and specificity of several SQL-based query strategies for identifying a histologic lymphoma subtype in this LLDB to determine the most accurate legacy data source for identifying a specific cancer patient population.

Methods

Query strategies were developed to identify patients with follicular lymphoma from a LLDB of cancer registry data, electronic medical records (EMR), laboratory, administrative, pharmacy, and other clinical data. Queries were performed using common diagnostic codes (ICD-9), cancer registry histology codes (ICD-O), and text searches of EMRs. We reviewed medical records and pathology reports to confirm each diagnosis and calculated the sensitivity and specificity for each query strategy.

Results

Together the queries identified 1538 potential cases of follicular lymphoma. Review of pathology and other medical reports confirmed 415 cases of follicular lymphoma, 300 pathology-verified and 115 verified from other medical reports. The query using ICD-O codes was highly specific (96%). Queries using text strings varied in sensitivity (range 7–92%) and specificity (range 86–99%). Queries using ICD-9 codes were both less sensitive (34–44%) and specific (35–87%).

Conclusions

Queries of linked-cancer databases that include cancer registry data should utilize ICD-O codes or employ structured free-text searches to identify patient populations with a precise histologic diagnosis.

Intermediate risk of cardiac events and recurrent stroke after stroke admission in young adults

Background

In older adults with stroke, there is an increased risk of cardiovascular events in the intermediate period, up to one year after stroke. The risk of cardiovascular events in this period in young adults after stroke has not been studied. We hypothesized that in the intermediate risk period, young adults with ischemic stroke have an increased risk of recurrent stroke and a smaller increase of cardiac events.

Methods

Using the National Readmissions Database during the year 2013, we identified ischemic stroke admissions among those aged 18–45 years using International Classification of Disease, Ninth Revision, Clinical Modification codes to identify index vascular events and risk factors. Primary outcomes were readmission for cardiac events and stroke. Multivariable Cox proportional hazard models and Kaplan–Meier analysis were used to estimate risk of primary outcomes.

Results

We identified 12,392 young adults with index stroke. The readmission rate due to recurrent stroke was higher than for cardiac events (2913.3.1 vs. 1132.4 per 100,000 index hospitalizations at 90 days). There was a higher cumulative risk of both cardiac events and recurrent stroke in the presence of baseline diabetes and hypercholesterolemia.

Conclusion

In a large, nationally representative database, the intermediate risk of recurrent stroke after index stroke in young adults was higher than the risk of cardiac events. The presence of vascular risk factors augmented this risk but did not entirely account for it. The aggressive control of hypercholesterolemia and diabetes may play an important role in secondary prevention in young adults with stroke.

Moderate sensitivity and high specificity of emergency department administrative data for transient ischemic attacks

Background

Validation of administrative data case definitions is key for accurate passive surveillance of disease. Transient ischemic attack (TIA) is a condition primarily managed in the emergency department. However, prior validation studies have focused on data after inpatient hospitalization. We aimed to determine the validity of the Canadian 10th International Classification of Diseases (ICD-10-CA) codes for TIA in the national ambulatory administrative database.

Methods

We performed a diagnostic accuracy study of four ICD-10-CA case definition algorithms for TIA in the emergency department setting. The study population was obtained from two ongoing studies on the diagnosis of TIA and minor stroke versus stroke mimic using serum biomarkers and neuroimaging. Two reference standards were used 1) the emergency department clinical diagnosis determined by chart abstractors and 2) the 90-day final diagnosis, both obtained by stroke neurologists, to calculate the sensitivity, specificity, positive and negative predictive values (PPV and NPV) of the ICD-10-CA algorithms for TIA.

Results

Among 417 patients, emergency department adjudication showed 163 (39.1%) TIA, 155 (37.2%) ischemic strokes, and 99 (23.7%) stroke mimics. The most restrictive algorithm, defined as a TIA code in the main position had the lowest sensitivity (36.8%), but highest specificity (92.5%) and PPV (76.0%). The most inclusive algorithm, defined as a TIA code in any position with and without query prefix had the highest sensitivity (63.8%), but lowest specificity (81.5%) and PPV (68.9%). Sensitivity, specificity, PPV, and NPV were overall lower when using the 90-day diagnosis as reference standard.

Conclusions

Emergency department administrative data reflect diagnosis of suspected TIA with high specificity, but underestimate the burden of disease. Future studies are necessary to understand the reasons for the low to moderate sensitivity.

Electronic supplementary material

The online version of this article (10.1186/s12913-017-2612-6) contains supplementary material, which is available to authorized users.