Design and implementation of a standardized framework to generate and evaluate patient-level prediction models using observational healthcare data

Enhancing readmission prediction models by integrating insights from home healthcare notes: Retrospective cohort study

BACKGROUND

Hospital readmission is an important indicator of inpatient care quality and a significant driver of increasing medical costs. Therefore, it is important to explore the effects of postdischarge information, particularly from home healthcare notes, on enhancing readmission prediction models. Despite the use of Natural Language Processing (NLP) and machine learning in prediction model development, current studies often overlook insights from home healthcare notes.

OBJECTIVE

This study aimed to develop prediction models for 30-day readmissions using home healthcare notes and structured data. In addition, it explored the development of 14- and 180-day prediction models using variables in the 30-day model.

DESIGN

A retrospective observational cohort study.

SETTING(S)

This study was conducted at Ajou University School of Medicine in South Korea.

PARTICIPANTS

Data from electronic health records, encompassing demographic characteristics of 1819 participants, along with information on conditions, drug, and home healthcare, were utilized.

METHODS

Two distinct models were developed for each prediction window (30-, 14-, 180-day): the traditional model, which utilized structured variables alone, and the common data model (CDM)-NLP model, which incorporated structured and topic variables extracted from home healthcare notes. BERTopic facilitated topic generation and risk probability, representing the likelihood of documents being assigned to specific topics. Feature selection involved experimenting with various algorithms. The best-performing algorithm, determined using the area under the receiver operating characteristic curve (AUROC), was used for model development. Model performance was assessed using various learning metrics including AUROC.

RESULTS

Among 1819 patients, 251 (13.80 %) experienced 30-day readmission. The least absolute shrinkage and selection operator was used for feature extraction and model development. The 15 structured features were used in the traditional model. Moreover, five additional topic variables from the home healthcare notes were applied in the CDM-NLP model. The AUROC of the traditional model was 0.739 (95 % CI: 0.672-0.807). The AUROC of the CDM-NLP model was high at 0.824 (95 % CI: 0.768-0.880), which indicated an outstanding performance. The topics in the CDM-NLP model included emotional distress, daily living functions, nutrition, postoperative status, and cardiorespiratory issues. In extended prediction model development for 14- and 180-day readmissions, the CDM-NLP consistently outperformed the traditional model.

CONCLUSIONS

This study developed effective prediction models using both structured and unstructured data, thereby emphasizing the significance of postdischarge information from home healthcare notes in readmission prediction.

Unbiased identification of risk factors for invasive Escherichia coli disease using machine learning

BACKGROUND

Invasive Escherichia coli disease (IED), also known as invasive extraintestinal pathogenic E. coli disease, is a leading cause of sepsis and bacteremia in older adults that can result in hospitalization and sometimes death and is frequently associated with antimicrobial resistance. Moreover, certain patient characteristics may increase the risk of developing IED. This study aimed to validate a machine learning approach for the unbiased identification of potential risk factors that correlate with an increased risk for IED.

METHODS

Using electronic health records from 6.5 million people, an XGBoost model was trained to predict IED from 663 distinct patient features, and the most predictive features were identified as potential risk factors. Using Shapley Additive predictive values, the specific relationships between features and the outcome of developing IED were characterized.

RESULTS

The model independently predicted that older age, a known risk factor for IED, increased the chance of developing IED. The model also predicted that a history of ≥ 1 urinary tract infection, as well as more frequent and/or more recent urinary tract infections, and ≥ 1 emergency department or inpatient visit increased the risk for IED. Outcomes were used to calculate risk ratios in selected subpopulations, demonstrating the impact of individual or combinations of features on the incidence of IED.

CONCLUSION

This study illustrates the viability and validity of using large electronic health records datasets and machine learning to identify correlating features and potential risk factors for infectious diseases, including IED. The next step is the independent validation of potential risk factors using conventional methods.

Annotation-preserving machine translation of English corpora to validate Dutch clinical concept extraction tools

OBJECTIVE

To explore the feasibility of validating Dutch concept extraction tools using annotated corpora translated from English, focusing on preserving annotations during translation and addressing the scarcity of non-English annotated clinical corpora.

MATERIALS AND METHODS

Three annotated corpora were standardized and translated from English to Dutch using 2 machine translation services, Google Translate and OpenAI GPT-4, with annotations preserved through a proposed method of embedding annotations in the text before translation. The performance of 2 concept extraction tools, MedSpaCy and MedCAT, was assessed across the corpora in both Dutch and English.

RESULTS

The translation process effectively generated Dutch annotated corpora and the concept extraction tools performed similarly in both English and Dutch. Although there were some differences in how annotations were preserved across translations, these did not affect extraction accuracy. Supervised MedCAT models consistently outperformed unsupervised models, whereas MedSpaCy demonstrated high recall but lower precision.

DISCUSSION

Our validation of Dutch concept extraction tools on corpora translated from English was successful, highlighting the efficacy of our annotation preservation method and the potential for efficiently creating multilingual corpora. Further improvements and comparisons of annotation preservation techniques and strategies for corpus synthesis could lead to more efficient development of multilingual corpora and accurate non-English concept extraction tools.

CONCLUSION

This study has demonstrated that translated English corpora can be used to validate non-English concept extraction tools. The annotation preservation method used during translation proved effective, and future research can apply this corpus translation method to additional languages and clinical settings.

Development and validation of a patient-level model to predict dementia across a network of observational databases

BACKGROUND

A prediction model can be a useful tool to quantify the risk of a patient developing dementia in the next years and take risk-factor-targeted intervention. Numerous dementia prediction models have been developed, but few have been externally validated, likely limiting their clinical uptake. In our previous work, we had limited success in externally validating some of these existing models due to inadequate reporting. As a result, we are compelled to develop and externally validate novel models to predict dementia in the general population across a network of observational databases. We assess regularization methods to obtain parsimonious models that are of lower complexity and easier to implement.

METHODS

Logistic regression models were developed across a network of five observational databases with electronic health records (EHRs) and claims data to predict 5-year dementia risk in persons aged 55-84. The regularization methods L1 and Broken Adaptive Ridge (BAR) as well as three candidate predictor sets to optimize prediction performance were assessed. The predictor sets include a baseline set using only age and sex, a full set including all available candidate predictors, and a phenotype set which includes a limited number of clinically relevant predictors.

RESULTS

BAR can be used for variable selection, outperforming L1 when a parsimonious model is desired. Adding candidate predictors for disease diagnosis and drug exposure generally improves the performance of baseline models using only age and sex. While a model trained on German EHR data saw an increase in AUROC from 0.74 to 0.83 with additional predictors, a model trained on US EHR data showed only minimal improvement from 0.79 to 0.81 AUROC. Nevertheless, the latter model developed using BAR regularization on the clinically relevant predictor set was ultimately chosen as best performing model as it demonstrated more consistent external validation performance and improved calibration.

CONCLUSIONS

We developed and externally validated patient-level models to predict dementia. Our results show that although dementia prediction is highly driven by demographic age, adding predictors based on condition diagnoses and drug exposures further improves prediction performance. BAR regularization outperforms L1 regularization to yield the most parsimonious yet still well-performing prediction model for dementia.

A multi-center study on the adaptability of a shared foundation model for electronic health records

Foundation models are transforming artificial intelligence (AI) in healthcare by providing modular components adaptable for various downstream tasks, making AI development more scalable and cost-effective. Foundation models for structured electronic health records (EHR), trained on coded medical records from millions of patients, demonstrated benefits including increased performance with fewer training labels, and improved robustness to distribution shifts. However, questions remain on the feasibility of sharing these models across hospitals and their performance in local tasks. This multi-center study examined the adaptability of a publicly accessible structured EHR foundation model (FMSM), trained on 2.57 M patient records from Stanford Medicine. Experiments used EHR data from The Hospital for Sick Children (SickKids) and Medical Information Mart for Intensive Care (MIMIC-IV). We assessed both adaptability via continued pretraining on local data, and task adaptability compared to baselines of locally training models from scratch, including a local foundation model. Evaluations on 8 clinical prediction tasks showed that adapting the off-the-shelf FMSM matched the performance of gradient boosting machines (GBM) locally trained on all data while providing a 13% improvement in settings with few task-specific training labels. Continued pretraining on local data showed FMSM required fewer than 1% of training examples to match the fully trained GBM's performance, and was 60 to 90% more sample-efficient than training local foundation models from scratch. Our findings demonstrate that adapting EHR foundation models across hospitals provides improved prediction performance at less cost, underscoring the utility of base foundation models as modular components to streamline the development of healthcare AI.

Comparing penalization methods for linear models on large observational health data

OBJECTIVE

This study evaluates regularization variants in logistic regression (L1, L2, ElasticNet, Adaptive L1, Adaptive ElasticNet, Broken adaptive ridge [BAR], and Iterative hard thresholding [IHT]) for discrimination and calibration performance, focusing on both internal and external validation.

MATERIALS AND METHODS

We use data from 5 US claims and electronic health record databases and develop models for various outcomes in a major depressive disorder patient population. We externally validate all models in the other databases. We use a train-test split of 75%/25% and evaluate performance with discrimination and calibration. Statistical analysis for difference in performance uses Friedman's test and critical difference diagrams.

RESULTS

Of the 840 models we develop, L1 and ElasticNet emerge as superior in both internal and external discrimination, with a notable AUC difference. BAR and IHT show the best internal calibration, without a clear external calibration leader. ElasticNet typically has larger model sizes than L1. Methods like IHT and BAR, while slightly less discriminative, significantly reduce model complexity.

CONCLUSION

L1 and ElasticNet offer the best discriminative performance in logistic regression for healthcare predictions, maintaining robustness across validations. For simpler, more interpretable models, L0-based methods (IHT and BAR) are advantageous, providing greater parsimony and calibration with fewer features. This study aids in selecting suitable regularization techniques for healthcare prediction models, balancing performance, complexity, and interpretability.

Frameworks for procurement, integration, monitoring, and evaluation of artificial intelligence tools in clinical settings: A systematic review

Research on the applications of artificial intelligence (AI) tools in medicine has increased exponentially over the last few years but its implementation in clinical practice has not seen a commensurate increase with a lack of consensus on implementing and maintaining such tools. This systematic review aims to summarize frameworks focusing on procuring, implementing, monitoring, and evaluating AI tools in clinical practice. A comprehensive literature search, following PRSIMA guidelines was performed on MEDLINE, Wiley Cochrane, Scopus, and EBSCO databases, to identify and include articles recommending practices, frameworks or guidelines for AI procurement, integration, monitoring, and evaluation. From the included articles, data regarding study aim, use of a framework, rationale of the framework, details regarding AI implementation involving procurement, integration, monitoring, and evaluation were extracted. The extracted details were then mapped on to the Donabedian Plan, Do, Study, Act cycle domains. The search yielded 17,537 unique articles, out of which 47 were evaluated for inclusion based on their full texts and 25 articles were included in the review. Common themes extracted included transparency, feasibility of operation within existing workflows, integrating into existing workflows, validation of the tool using predefined performance indicators and improving the algorithm and/or adjusting the tool to improve performance. Among the four domains (Plan, Do, Study, Act) the most common domain was Plan (84%, n = 21), followed by Study (60%, n = 15), Do (52%, n = 13), & Act (24%, n = 6). Among 172 authors, only 1 (0.6%) was from a low-income country (LIC) and 2 (1.2%) were from lower-middle-income countries (LMICs). Healthcare professionals cite the implementation of AI tools within clinical settings as challenging owing to low levels of evidence focusing on integration in the Do and Act domains. The current healthcare AI landscape calls for increased data sharing and knowledge translation to facilitate common goals and reap maximum clinical benefit.

Towards global model generalizability: independent cross-site feature evaluation for patient-level risk prediction models using the OHDSI network

BACKGROUND

Predictive models show promise in healthcare, but their successful deployment is challenging due to limited generalizability. Current external validation often focuses on model performance with restricted feature use from the original training data, lacking insights into their suitability at external sites. Our study introduces an innovative methodology for evaluating features during both the development phase and the validation, focusing on creating and validating predictive models for post-surgery patient outcomes with improved generalizability.

METHODS

Electronic health records (EHRs) from 4 countries (United States, United Kingdom, Finland, and Korea) were mapped to the OMOP Common Data Model (CDM), 2008-2019. Machine learning (ML) models were developed to predict post-surgery prolonged opioid use (POU) risks using data collected 6 months before surgery. Both local and cross-site feature selection methods were applied in the development and external validation datasets. Models were developed using Observational Health Data Sciences and Informatics (OHDSI) tools and validated on separate patient cohorts.

RESULTS

Model development included 41 929 patients, 14.6% with POU. The external validation included 31 932 (UK), 23 100 (US), 7295 (Korea), and 3934 (Finland) patients with POU of 44.2%, 22.0%, 15.8%, and 21.8%, respectively. The top-performing model, Lasso logistic regression, achieved an area under the receiver operating characteristic curve (AUROC) of 0.75 during local validation and 0.69 (SD = 0.02) (averaged) in external validation. Models trained with cross-site feature selection significantly outperformed those using only features from the development site through external validation (P < .05).

CONCLUSIONS

Using EHRs across four countries mapped to the OMOP CDM, we developed generalizable predictive models for POU. Our approach demonstrates the significant impact of cross-site feature selection in improving model performance, underscoring the importance of incorporating diverse feature sets from various clinical settings to enhance the generalizability and utility of predictive healthcare models.

Characterizing the limitations of using diagnosis codes in the context of machine learning for healthcare

BACKGROUND

Diagnostic codes are commonly used as inputs for clinical prediction models, to create labels for prediction tasks, and to identify cohorts for multicenter network studies. However, the coverage rates of diagnostic codes and their variability across institutions are underexplored. The primary objective was to describe lab- and diagnosis-based labels for 7 selected outcomes at three institutions. Secondary objectives were to describe agreement, sensitivity, and specificity of diagnosis-based labels against lab-based labels.

METHODS

This study included three cohorts: SickKids from The Hospital for Sick Children, and StanfordPeds and StanfordAdults from Stanford Medicine. We included seven clinical outcomes with lab-based definitions: acute kidney injury, hyperkalemia, hypoglycemia, hyponatremia, anemia, neutropenia and thrombocytopenia. For each outcome, we created four lab-based labels (abnormal, mild, moderate and severe) based on test result and one diagnosis-based label. Proportion of admissions with a positive label were presented for each outcome stratified by cohort. Using lab-based labels as the gold standard, agreement using Cohen's Kappa, sensitivity and specificity were calculated for each lab-based severity level.

RESULTS

The number of admissions included were: SickKids (n = 59,298), StanfordPeds (n = 24,639) and StanfordAdults (n = 159,985). The proportion of admissions with a positive diagnosis-based label was significantly higher for StanfordPeds compared to SickKids across all outcomes, with odds ratio (99.9% confidence interval) for abnormal diagnosis-based label ranging from 2.2 (1.7-2.7) for neutropenia to 18.4 (10.1-33.4) for hyperkalemia. Lab-based labels were more similar by institution. When using lab-based labels as the gold standard, Cohen's Kappa and sensitivity were lower at SickKids for all severity levels compared to StanfordPeds.

CONCLUSIONS

Across multiple outcomes, diagnosis codes were consistently different between the two pediatric institutions. This difference was not explained by differences in test results. These results may have implications for machine learning model development and deployment.

A comparative patient-level prediction study in OMOP CDM: applicative potential and insights from synthetic data

The emergence of collaborations, which standardize and combine multiple clinical databases across different regions, provide a wealthy source of data, which is fundamental for clinical prediction models, such as patient-level predictions. With the aid of such large data pools, researchers are able to develop clinical prediction models for improved disease classification, risk assessment, and beyond. To fully utilize this potential, Machine Learning (ML) methods are commonly required to process these large amounts of data on disease-specific patient cohorts. As a consequence, the Observational Health Data Sciences and Informatics (OHDSI) collaborative develops a framework to facilitate the application of ML models for these standardized patient datasets by using the Observational Medical Outcomes Partnership (OMOP) common data model (CDM). In this study, we compare the feasibility of current web-based OHDSI approaches, namely ATLAS and "Patient-level Prediction" (PLP), against a native solution (R based) to conduct such ML-based patient-level prediction analyses in OMOP. This will enable potential users to select the most suitable approach for their investigation. Each of the applied ML solutions was individually utilized to solve the same patient-level prediction task. Both approaches went through an exemplary benchmarking analysis to assess the weaknesses and strengths of the PLP R-Package. In this work, the performance of this package was subsequently compared versus the commonly used native R-package called Machine Learning in R 3 (mlr3), and its sub-packages. The approaches were evaluated on performance, execution time, and ease of model implementation. The results show that the PLP package has shorter execution times, which indicates great scalability, as well as intuitive code implementation, and numerous possibilities for visualization. However, limitations in comparison to native packages were depicted in the implementation of specific ML classifiers (e.g., Lasso), which may result in a decreased performance for real-world prediction problems. The findings here contribute to the overall effort of developing ML-based prediction models on a clinical scale and provide a snapshot for future studies that explicitly aim to develop patient-level prediction models in OMOP CDM.

Health-Analytics Data to Evidence Suite (HADES): Open-Source Software for Observational Research

The Health-Analytics Data to Evidence Suite (HADES) is an open-source software collection developed by Observational Health Data Sciences and Informatics (OHDSI). It executes directly against healthcare data such as electronic health records and administrative claims, that have been converted to the Observational Medical Outcomes Partnership (OMOP) Common Data Model. Using advanced analytics, HADES performs characterization, population-level causal effect estimation, and patient-level prediction, potentially across a federated data network, allowing patient-level data to remain locally while only aggregated statistics are shared. Designed to run across a wide array of technical environments, including different operating systems and database platforms, HADES uses continuous integration with a large set of unit tests to maintain reliability. HADES implements OHDSI best practices, and is used in almost all published OHDSI studies, including some that have directly informed regulatory decisions.

Improving machine learning with ensemble learning on observational healthcare data

Ensemble learning is a powerful technique for improving the accuracy and reliability of prediction models, especially in scenarios where individual models may not perform well. However, combining models with varying accuracies may not always improve the final prediction results, as models with lower accuracies may obscure the results of models with higher accuracies. This paper addresses this issue and answers the question of when an ensemble approach outperforms individual models for prediction. As a result, we propose an ensemble model for predicting patients at risk of postoperative prolonged opioid. The model incorporates two machine learning models that are trained using different covariates, resulting in high precision and recall. Our study, which employs five different machine learning algorithms, shows that the proposed approach significantly improves the final prediction results in terms of AUROC and AUPRC.

Attention-based neural networks for clinical prediction modelling on electronic health records

BACKGROUND

Deep learning models have had a lot of success in various fields. However, on structured data they have struggled. Here we apply four state-of-the-art supervised deep learning models using the attention mechanism and compare against logistic regression and XGBoost using discrimination, calibration and clinical utility.

METHODS

We develop the models using a general practitioners database. We implement a recurrent neural network, a transformer with and without reverse distillation and a graph neural network. We measure discrimination using the area under the receiver operating characteristic curve (AUC) and the area under the precision recall curve (AUPRC). We assess smooth calibration using restricted cubic splines and clinical utility with decision curve analysis.

RESULTS

Our results show that deep learning approaches can improve discrimination up to 2.5% points AUC and 7.4% points AUPRC. However, on average the baselines are competitive. Most models are similarly calibrated as the baselines except for the graph neural network. The transformer using reverse distillation shows the best performance in clinical utility on two out of three prediction problems over most of the prediction thresholds.

CONCLUSION

In this study, we evaluated various approaches in supervised learning using neural networks and attention. Here we do a rigorous comparison, not only looking at discrimination but also calibration and clinical utility. There is value in using deep learning models on electronic health record data since it can improve discrimination and clinical utility while providing good calibration. However, good baseline methods are still competitive.

The added value of text from Dutch general practitioner notes in predictive modeling

OBJECTIVE

This work aims to explore the value of Dutch unstructured data, in combination with structured data, for the development of prognostic prediction models in a general practitioner (GP) setting.

MATERIALS AND METHODS

We trained and validated prediction models for 4 common clinical prediction problems using various sparse text representations, common prediction algorithms, and observational GP electronic health record (EHR) data. We trained and validated 84 models internally and externally on data from different EHR systems.

RESULTS

On average, over all the different text representations and prediction algorithms, models only using text data performed better or similar to models using structured data alone in 2 prediction tasks. Additionally, in these 2 tasks, the combination of structured and text data outperformed models using structured or text data alone. No large performance differences were found between the different text representations and prediction algorithms.

DISCUSSION

Our findings indicate that the use of unstructured data alone can result in well-performing prediction models for some clinical prediction problems. Furthermore, the performance improvement achieved by combining structured and text data highlights the added value. Additionally, we demonstrate the significance of clinical natural language processing research in languages other than English and the possibility of validating text-based prediction models across various EHR systems.

CONCLUSION

Our study highlights the potential benefits of incorporating unstructured data in clinical prediction models in a GP setting. Although the added value of unstructured data may vary depending on the specific prediction task, our findings suggest that it has the potential to enhance patient care.

Development and Validation of a Prognostic Classification Model Predicting Postoperative Adverse Outcomes in Older Surgical Patients Using a Machine Learning Algorithm: Retrospective Observational Network Study

BACKGROUND

Older adults are at an increased risk of postoperative morbidity. Numerous risk stratification tools exist, but effort and manpower are required.

OBJECTIVE

This study aimed to develop a predictive model of postoperative adverse outcomes in older patients following general surgery with an open-source, patient-level prediction from the Observational Health Data Sciences and Informatics for internal and external validation.

METHODS

We used the Observational Medical Outcomes Partnership common data model and machine learning algorithms. The primary outcome was a composite of 90-day postoperative all-cause mortality and emergency department visits. Secondary outcomes were postoperative delirium, prolonged postoperative stay (≥75th percentile), and prolonged hospital stay (≥21 days). An 80% versus 20% split of the data from the Seoul National University Bundang Hospital (SNUBH) and Seoul National University Hospital (SNUH) common data model was used for model training and testing versus external validation. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) with a 95% CI.

RESULTS

Data from 27,197 (SNUBH) and 32,857 (SNUH) patients were analyzed. Compared to the random forest, Adaboost, and decision tree models, the least absolute shrinkage and selection operator logistic regression model showed good internal discriminative accuracy (internal AUC 0.723, 95% CI 0.701-0.744) and transportability (external AUC 0.703, 95% CI 0.692-0.714) for the primary outcome. The model also possessed good internal and external AUCs for postoperative delirium (internal AUC 0.754, 95% CI 0.713-0.794; external AUC 0.750, 95% CI 0.727-0.772), prolonged postoperative stay (internal AUC 0.813, 95% CI 0.800-0.825; external AUC 0.747, 95% CI 0.741-0.753), and prolonged hospital stay (internal AUC 0.770, 95% CI 0.749-0.792; external AUC 0.707, 95% CI 0.696-0.718). Compared with age or the Charlson comorbidity index, the model showed better prediction performance.

CONCLUSIONS

The derived model shall assist clinicians and patients in understanding the individualized risks and benefits of surgery.

Prognostic models of in-hospital mortality of intensive care patients using neural representation of unstructured text: A systematic review and critical appraisal

OBJECTIVE

To review and critically appraise published and preprint reports of prognostic models of in-hospital mortality of patients in the intensive-care unit (ICU) based on neural representations (embeddings) of clinical notes.

METHODS

PubMed and arXiv were searched up to August 1, 2022. At least two reviewers independently selected the studies that developed a prognostic model of in-hospital mortality of intensive-care patients using free-text represented as embeddings and extracted data using the CHARMS checklist. Risk of bias was assessed using PROBAST. Reporting on the model was assessed with the TRIPOD guideline. To assess the machine learning components that were used in the models, we present a new descriptive framework based on different techniques to represent text and provide predictions from text. The study protocol was registered in the PROSPERO database (CRD42022354602).

RESULTS

Eighteen studies out of 2,825 were included. All studies used the publicly-available MIMIC dataset. Context-independent word embeddings are widely used. Model discrimination was provided by all studies (AUROC 0.75-0.96), but measures of calibration were scarce. Seven studies used both structural clinical variables and notes. Model discrimination improved when adding clinical notes to variables. None of the models was externally validated and often a simple train/test split was used for internal validation. Our critical appraisal demonstrated a high risk of bias in all studies and concerns regarding their applicability in clinical practice.

CONCLUSION

All studies used a neural architecture for prediction and were based on one publicly available dataset. Clinical notes were reported to improve predictive performance when used in addition to only clinical variables. Most studies had methodological, reporting, and applicability issues. We recommend reporting both model discrimination and calibration, using additional data sources, and using more robust evaluation strategies, including prospective and external validation. Finally, sharing data and code is encouraged to improve study reproducibility.

Assessment of Prediction Tasks and Time Window Selection in Temporal Modeling of Electronic Health Record Data: a Systematic Review

Temporal electronic health record (EHR) data are often preferred for clinical prediction tasks because they offer more complete representations of a patient's pathophysiology than static data. A challenge when working with temporal EHR data is problem formulation, which includes defining the time windows of interest and the prediction task. Our objective was to conduct a systematic review that assessed the definition and reporting of concepts relevant to temporal clinical prediction tasks. We searched PubMed® and IEEE Xplore® databases for studies from January 1, 2010 applying machine learning models to EHR data for patient outcome prediction. Publications applying time-series methods were selected for further review. We identified 92 studies and summarized them by clinical context and definition and reporting of the prediction problem. For the time windows of interest, 12 studies did not discuss window lengths, 57 used a single set of window lengths, and 23 evaluated the relationship between window length and model performance. We also found that 72 studies had appropriate reporting of the prediction task. However, evaluation of prediction problem formulation for temporal EHR data was complicated by heterogeneity in assessing and reporting of these concepts. Even among studies modeling similar clinical outcomes, there were variations in terminology used to describe the prediction problem, rationale for window lengths, and determination of the outcome of interest. As temporal modeling using EHR data expands, minimal reporting standards should include time-series specific concerns to promote rigor and reproducibility in future studies and facilitate model implementation in clinical settings.

Supplementary Information

The online version contains supplementary material available at 10.1007/s41666-023-00143-4.

Ten Topics to Get Started in Medical Informatics Research

The vast and heterogeneous data being constantly generated in clinics can provide great wealth for patients and research alike. The quickly evolving field of medical informatics research has contributed numerous concepts, algorithms, and standards to facilitate this development. However, these difficult relationships, complex terminologies, and multiple implementations can present obstacles for people who want to get active in the field. With a particular focus on medical informatics research conducted in Germany, we present in our Viewpoint a set of 10 important topics to improve the overall interdisciplinary communication between different stakeholders (eg, physicians, computational experts, experimentalists, students, patient representatives). This may lower the barriers to entry and offer a starting point for collaborations at different levels. The suggested topics are briefly introduced, then general best practice guidance is given, and further resources for in-depth reading or hands-on tutorials are recommended. In addition, the topics are set to cover current aspects and open research gaps of the medical informatics domain, including data regulations and concepts; data harmonization and processing; and data evaluation, visualization, and dissemination. In addition, we give an example on how these topics can be integrated in a medical informatics curriculum for higher education. By recognizing these topics, readers will be able to (1) set clinical and research data into the context of medical informatics, understanding what is possible to achieve with data or how data should be handled in terms of data privacy and storage; (2) distinguish current interoperability standards and obtain first insights into the processes leading to effective data transfer and analysis; and (3) value the use of newly developed technical approaches to utilize the full potential of clinical data.

Privacy-Preserving Federated Model Predicting Bipolar Transition in Patients With Depression: Prediction Model Development Study

BACKGROUND

Mood disorder has emerged as a serious concern for public health; in particular, bipolar disorder has a less favorable prognosis than depression. Although prompt recognition of depression conversion to bipolar disorder is needed, early prediction is challenging due to overlapping symptoms. Recently, there have been attempts to develop a prediction model by using federated learning. Federated learning in medical fields is a method for training multi-institutional machine learning models without patient-level data sharing.

OBJECTIVE

This study aims to develop and validate a federated, differentially private multi-institutional bipolar transition prediction model.

METHODS

This retrospective study enrolled patients diagnosed with the first depressive episode at 5 tertiary hospitals in South Korea. We developed models for predicting bipolar transition by using data from 17,631 patients in 4 institutions. Further, we used data from 4541 patients for external validation from 1 institution. We created standardized pipelines to extract large-scale clinical features from the 4 institutions without any code modification. Moreover, we performed feature selection in a federated environment for computational efficiency and applied differential privacy to gradient updates. Finally, we compared the federated and the 4 local models developed with each hospital's data on internal and external validation data sets.

RESULTS

In the internal data set, 279 out of 17,631 patients showed bipolar disorder transition. In the external data set, 39 out of 4541 patients showed bipolar disorder transition. The average performance of the federated model in the internal test (area under the curve [AUC] 0.726) and external validation (AUC 0.719) data sets was higher than that of the other locally developed models (AUC 0.642-0.707 and AUC 0.642-0.699, respectively). In the federated model, classifications were driven by several predictors such as the Charlson index (low scores were associated with bipolar transition, which may be due to younger age), severe depression, anxiolytics, young age, and visiting months (the bipolar transition was associated with seasonality, especially during the spring and summer months).

CONCLUSIONS

We developed and validated a differentially private federated model by using distributed multi-institutional psychiatric data with standardized pipelines in a real-world environment. The federated model performed better than models using local data only.

Evaluation of reporting quality of cohort studies using real-world data based on RECORD: systematic review

OBJECTIVE

Real-world data (RWD) and real-world evidence (RWE) have been paid more and more attention in recent years. We aimed to evaluate the reporting quality of cohort studies using real-world data (RWD) published between 2013 and 2021 and analyze the possible factors.

METHODS

We conducted a comprehensive search in Medline and Embase through the OVID interface for cohort studies published from 2013 to 2021 on April 29, 2022. Studies aimed at comparing the effectiveness or safety of exposure factors in the real-world setting were included. The evaluation was based on the REporting of studies Conducted using Observational Routinely-collected health Data (RECORD) statement. Agreement for inclusion and evaluation was calculated using Cohen's kappa. Pearson chi-square test or Fisher's exact test and Mann-Whitney U test were used to analyze the possible factors, including the release of RECORD, journal IFs, and article citations. Bonferroni's correction was conducted for multiple comparisons. Interrupted time series analysis was performed to display the changes in report quality over time.

RESULTS

187 articles were finally included. The mean ± SD of the percentage of adequately reported items in the 187 articles was 44.7 ± 14.3 with a range of 11.1-87%. Of 23 items, the adequate reporting rate of 10 items reached 50%, and the reporting rate of some vital items was inadequate. After Bonferroni's correction, the reporting of only one item significantly improved after the release of RECORD and there was no significant improvement in the overall report quality. For interrupted time series analysis, there were no significant changes in the slope (p = 0.42) and level (p = 0.12) of adequate reporting rate. The journal IFs and citations were respectively related to 2 areas and the former significantly higher in high-reporting quality articles.

CONCLUSION

The endorsement of the RECORD cheklist was generally inadequate in cohort studies using RWD and has not improved in recent years. We encourage researchers to endorse relevant guidelines when utilizing RWD for research.

Bottom-up and top-down paradigms of artificial intelligence research approaches to healthcare data science using growing real-world big data

OBJECTIVES

As the real-world electronic health record (EHR) data continue to grow exponentially, novel methodologies involving artificial intelligence (AI) are becoming increasingly applied to enable efficient data-driven learning and, ultimately, to advance healthcare. Our objective is to provide readers with an understanding of evolving computational methods and help in deciding on methods to pursue.

TARGET AUDIENCE

The sheer diversity of existing methods presents a challenge for health scientists who are beginning to apply computational methods to their research. Therefore, this tutorial is aimed at scientists working with EHR data who are early entrants into the field of applying AI methodologies.

SCOPE

This manuscript describes the diverse and growing AI research approaches in healthcare data science and categorizes them into 2 distinct paradigms, the bottom-up and top-down paradigms to provide health scientists venturing into artificial intelligent research with an understanding of the evolving computational methods and help in deciding on methods to pursue through the lens of real-world healthcare data.

Predicting patients who are likely to develop Lupus Nephritis of those newly diagnosed with Systemic Lupus Erythematosus

Patients diagnosed with systemic lupus erythematosus (SLE) suffer from a decreased quality of life, an increased risk of medical complications, and an increased risk of death. In particular, approximately 50% of SLE patients progress to develop lupus nephritis, which oftentimes leads to life-threatening end stage renal disease (ESRD) and requires dialysis or kidney transplant1. The challenge is that lupus nephritis is diagnosed via a kidney biopsy, which is typically performed only after noticeable decreased kidney function, leaving little room for proactive or preventative measures. The ability to predict which patients are most likely to develop lupus nephritis has the potential to shift lupus nephritis disease management from reactive to proactive. We present a clinically useful prediction model to predict which patients with newly diagnosed SLE will go on to develop lupus nephritis in the next five years.

A standardized framework for risk-based assessment of treatment effect heterogeneity in observational healthcare databases

Treatment effects are often anticipated to vary across groups of patients with different baseline risk. The Predictive Approaches to Treatment Effect Heterogeneity (PATH) statement focused on baseline risk as a robust predictor of treatment effect and provided guidance on risk-based assessment of treatment effect heterogeneity in a randomized controlled trial. The aim of this study is to extend this approach to the observational setting using a standardized scalable framework. The proposed framework consists of five steps: (1) definition of the research aim, i.e., the population, the treatment, the comparator and the outcome(s) of interest; (2) identification of relevant databases; (3) development of a prediction model for the outcome(s) of interest; (4) estimation of relative and absolute treatment effect within strata of predicted risk, after adjusting for observed confounding; (5) presentation of the results. We demonstrate our framework by evaluating heterogeneity of the effect of thiazide or thiazide-like diuretics versus angiotensin-converting enzyme inhibitors on three efficacy and nine safety outcomes across three observational databases. We provide a publicly available R software package for applying this framework to any database mapped to the Observational Medical Outcomes Partnership Common Data Model. In our demonstration, patients at low risk of acute myocardial infarction receive negligible absolute benefits for all three efficacy outcomes, though they are more pronounced in the highest risk group, especially for acute myocardial infarction. Our framework allows for the evaluation of differential treatment effects across risk strata, which offers the opportunity to consider the benefit-harm trade-off between alternative treatments.

Integrating real-world data from Brazil and Pakistan into the OMOP common data model and standardized health analytics framework to characterize COVID-19 in the Global South

OBJECTIVES

The aim of this work is to demonstrate the use of a standardized health informatics framework to generate reliable and reproducible real-world evidence from Latin America and South Asia towards characterizing coronavirus disease 2019 (COVID-19) in the Global South.

MATERIALS AND METHODS

Patient-level COVID-19 records collected in a patient self-reported notification system, hospital in-patient and out-patient records, and community diagnostic labs were harmonized to the Observational Medical Outcomes Partnership common data model and analyzed using a federated network analytics framework. Clinical characteristics of individuals tested for, diagnosed with or tested positive for, hospitalized with, admitted to intensive care unit with, or dying with COVID-19 were estimated.

RESULTS

Two COVID-19 databases covering 8.3 million people from Pakistan and 2.6 million people from Bahia, Brazil were analyzed. 109 504 (Pakistan) and 921 (Brazil) medical concepts were harmonized to Observational Medical Outcomes Partnership common data model. In total, 341 505 (4.1%) people in the Pakistan dataset and 1 312 832 (49.2%) people in the Brazilian dataset were tested for COVID-19 between January 1, 2020 and April 20, 2022, with a median [IQR] age of 36 [25, 76] and 38 (27, 50); 40.3% and 56.5% were female in Pakistan and Brazil, respectively. 1.2% percent individuals in the Pakistan dataset had Afghan ethnicity. In Brazil, 52.3% had mixed ethnicity. In agreement with international findings, COVID-19 outcomes were more severe in men, elderly, and those with underlying health conditions.

CONCLUSIONS

COVID-19 data from 2 large countries in the Global South were harmonized and analyzed using a standardized health informatics framework developed by an international community of health informaticians. This proof-of-concept study demonstrates a potential open science framework for global knowledge mobilization and clinical translation for timely response to healthcare needs in pandemics and beyond.

EHR foundation models improve robustness in the presence of temporal distribution shift

Temporal distribution shift negatively impacts the performance of clinical prediction models over time. Pretraining foundation models using self-supervised learning on electronic health records (EHR) may be effective in acquiring informative global patterns that can improve the robustness of task-specific models. The objective was to evaluate the utility of EHR foundation models in improving the in-distribution (ID) and out-of-distribution (OOD) performance of clinical prediction models. Transformer- and gated recurrent unit-based foundation models were pretrained on EHR of up to 1.8 M patients (382 M coded events) collected within pre-determined year groups (e.g., 2009-2012) and were subsequently used to construct patient representations for patients admitted to inpatient units. These representations were used to train logistic regression models to predict hospital mortality, long length of stay, 30-day readmission, and ICU admission. We compared our EHR foundation models with baseline logistic regression models learned on count-based representations (count-LR) in ID and OOD year groups. Performance was measured using area-under-the-receiver-operating-characteristic curve (AUROC), area-under-the-precision-recall curve, and absolute calibration error. Both transformer and recurrent-based foundation models generally showed better ID and OOD discrimination relative to count-LR and often exhibited less decay in tasks where there is observable degradation of discrimination performance (average AUROC decay of 3% for transformer-based foundation model vs. 7% for count-LR after 5-9 years). In addition, the performance and robustness of transformer-based foundation models continued to improve as pretraining set size increased. These results suggest that pretraining EHR foundation models at scale is a useful approach for developing clinical prediction models that perform well in the presence of temporal distribution shift.

Machine Learning and Real-World Data to Predict Lung Cancer Risk in Routine Care

BACKGROUND

This study used machine learning to develop a 3-year lung cancer risk prediction model with large real-world data in a mostly younger population.

METHODS

Over 4.7 million individuals, aged 45 to 65 years with no history of any cancer or lung cancer screening, diagnostic, or treatment procedures, with an outpatient visit in 2013 were identified in Optum's de-identified Electronic Health Record (EHR) dataset. A least absolute shrinkage and selection operator model was fit using all available data in the 365 days prior. Temporal validation was assessed with recent data. External validation was assessed with data from Mercy Health Systems EHR and Optum's de-identified Clinformatics Data Mart Database. Racial inequities in model discrimination were assessed with xAUCs.

RESULTS

The model AUC was 0.76. Top predictors included age, smoking, race, ethnicity, and diagnosis of chronic obstructive pulmonary disease. The model identified a high-risk group with lung cancer incidence 9 times the average cohort incidence, representing 10% of patients with lung cancer. Model performed well temporally and externally, while performance was reduced for Asians and Hispanics.

CONCLUSIONS

A high-dimensional model trained using big data identified a subset of patients with high lung cancer risk. The model demonstrated transportability to EHR and claims data, while underscoring the need to assess racial disparities when using machine learning methods.

IMPACT

This internally and externally validated real-world data-based lung cancer prediction model is available on an open-source platform for broad sharing and application. Model integration into an EHR system could minimize physician burden by automating identification of high-risk patients.

Association between impaired glucose metabolism and long-term prognosis at the time of diagnosis of depression: Impaired glucose metabolism as a promising biomarker proposed through a machine-learning approach

BACKGROUND

Predicting the course of depression is necessary for personalized treatment. Impaired glucose metabolism (IGM) was introduced as a promising depression biomarker, but no consensus was made. This study aimed to predict IGM at the time of depression diagnosis and examine the relationship between long-term prognosis and predicted results.

METHODS

Clinical data were extracted from four electronic health records in South Korea. The study population included patients with depression, and the outcome was IGM within 1 year. One database was used to develop the model using three algorithms. External validation was performed using the best algorithm across the three databases. The area under the curve (AUC) was calculated to determine the model's performance. Kaplan-Meier and Cox survival analyses of the risk of hospitalization for depression as the long-term outcome were performed. A meta-analysis of the long-term outcome was performed across the four databases.

RESULTS

A prediction model was developed using the data of 3,668 people, with an AUC of 0.781 with least absolute shrinkage and selection operator (LASSO) logistic regression. In the external validation, the AUCs were 0.643, 0.610, and 0.515. Through the predicted results, survival analysis and meta-analysis were performed; the hazard ratios of risk of hospitalization for depression in patients predicted to have IGM was 1.20 (95% confidence interval [CI] 1.02-1.41, p = 0.027) at a 3-year follow-up.

CONCLUSIONS

We developed prediction models for IGM occurrence within a year. The predicted results were related to the long-term prognosis of depression, presenting as a promising IGM biomarker related to the prognosis of depression.

Exploring the potential of OMOP common data model for process mining in healthcare

BACKGROUND AND OBJECTIVE

Recently, Electronic Health Records (EHR) are increasingly being converted to Common Data Models (CDMs), a database schema designed to provide standardized vocabularies to facilitate collaborative observational research. To date, however, rare attempts exist to leverage CDM data for healthcare process mining, a technique to derive process-related knowledge (e.g., process model) from event logs. This paper presents a method to extract, construct, and analyze event logs from the Observational Medical Outcomes Partnership (OMOP) CDM for process mining and demonstrates CDM-based healthcare process mining with several real-life study cases while answering frequently posed questions in process mining, in the CDM environment.

METHODS

We propose a method to extract, construct, and analyze event logs from the OMOP CDM for process types including inpatient, outpatient, emergency room processes, and patient journey. Using the proposed method, we extract the retrospective data of several surgical procedure cases (i.e., Total Laparoscopic Hysterectomy (TLH), Total Hip Replacement (THR), Coronary Bypass (CB), Transcatheter Aortic Valve Implantation (TAVI), Pancreaticoduodenectomy (PD)) from the CDM of a Korean tertiary hospital. Patient data are extracted for each of the operations and analyzed using several process mining techniques.

RESULTS

Using process mining, the clinical pathways, outpatient process models, emergency room process models, and patient journeys are demonstrated using the extracted logs. The result shows CDM's usability as a novel and valuable data source for healthcare process analysis, yet with a few considerations. We found that CDM should be complemented by different internal and external data sources to address the administrative and operational aspects of healthcare processes, particularly for outpatient and ER process analyses.

CONCLUSION

To the best of our knowledge, we are the first to exploit CDM for healthcare process mining. Specifically, we provide a step-by-step guidance by demonstrating process analysis from locating relevant CDM tables to visualizing results using process mining tools. The proposed method can be widely applicable across different institutions. This work can contribute to bringing a process mining perspective to the existing CDM users in the changing Hospital Information Systems (HIS) environment and also to facilitating CDM-based studies in the process mining research community.

External validation of existing dementia prediction models on observational health data

BACKGROUND

Many dementia prediction models have been developed, but only few have been externally validated, which hinders clinical uptake and may pose a risk if models are applied to actual patients regardless. Externally validating an existing prediction model is a difficult task, where we mostly rely on the completeness of model reporting in a published article. In this study, we aim to externally validate existing dementia prediction models. To that end, we define model reporting criteria, review published studies, and externally validate three well reported models using routinely collected health data from administrative claims and electronic health records.

METHODS

We identified dementia prediction models that were developed between 2011 and 2020 and assessed if they could be externally validated given a set of model criteria. In addition, we externally validated three of these models (Walters' Dementia Risk Score, Mehta's RxDx-Dementia Risk Index, and Nori's ADRD dementia prediction model) on a network of six observational health databases from the United States, United Kingdom, Germany and the Netherlands, including the original development databases of the models.

RESULTS

We reviewed 59 dementia prediction models. All models reported the prediction method, development database, and target and outcome definitions. Less frequently reported by these 59 prediction models were predictor definitions (52 models) including the time window in which a predictor is assessed (21 models), predictor coefficients (20 models), and the time-at-risk (42 models). The validation of the model by Walters (development c-statistic: 0.84) showed moderate transportability (0.67-0.76 c-statistic). The Mehta model (development c-statistic: 0.81) transported well to some of the external databases (0.69-0.79 c-statistic). The Nori model (development AUROC: 0.69) transported well (0.62-0.68 AUROC) but performed modestly overall. Recalibration showed improvements for the Walters and Nori models, while recalibration could not be assessed for the Mehta model due to unreported baseline hazard.

CONCLUSION

We observed that reporting is mostly insufficient to fully externally validate published dementia prediction models, and therefore, it is uncertain how well these models would work in other clinical settings. We emphasize the importance of following established guidelines for reporting clinical prediction models. We recommend that reporting should be more explicit and have external validation in mind if the model is meant to be applied in different settings.

Establishment of an International Evidence Sharing Network Through Common Data Model for Cardiovascular Research

A retrospective observational study is one of the most widely used research methods in medicine. However, evidence postulated from a single data source likely contains biases such as selection bias, information bias, and confounding bias. Acquiring enough data from multiple institutions is one of the most effective methods to overcome the limitations. However, acquiring data from multiple institutions from many countries requires enormous effort because of financial, technical, ethical, and legal issues as well as standardization of data structure and semantics. The Observational Health Data Sciences and Informatics (OHDSI) research network standardized 928 million unique records or 12% of the world's population into a common structure and meaning and established a research network of 453 data partners from 41 countries around the world. OHDSI is a distributed research network wherein researchers do not own or directly share data but only analyzed results. However, sharing evidence without sharing data is difficult to understand. In this review, we will look at the basic principles of OHDSI, common data model, distributed research networks, and some representative studies in the cardiovascular field using the network. This paper also briefly introduces a Korean distributed research network named FeederNet.

Semantic Data Visualisation for Biomedical Database Catalogues

Biomedical databases often have restricted access policies and governance rules. Thus, an adequate description of their content is essential for researchers who wish to use them for medical research. A strategy for publishing information without disclosing patient-level data is through database fingerprinting and aggregate characterisations. However, this information is still presented in a format that makes it challenging to search, analyse, and decide on the best databases for a domain of study. Several strategies allow one to visualise and compare the characteristics of multiple biomedical databases. Our study focused on a European platform for sharing and disseminating biomedical data. We use semantic data visualisation techniques to assist in comparing descriptive metadata from several databases. The great advantage lies in streamlining the database selection process, ensuring that sensitive details are not shared. To address this goal, we have considered two levels of data visualisation, one characterising a single database and the other involving multiple databases in network-level visualisations. This study revealed the impact of the proposed visualisations and some open challenges in representing semantically annotated biomedical datasets. Identifying future directions in this scope was one of the outcomes of this work.

Development and external validation of prediction models for adverse health outcomes in rheumatoid arthritis: A multinational real-world cohort analysis

BACKGROUND

Identification of rheumatoid arthritis (RA) patients at high risk of adverse health outcomes remains a major challenge. We aimed to develop and validate prediction models for a variety of adverse health outcomes in RA patients initiating first-line methotrexate (MTX) monotherapy.

METHODS

Data from 15 claims and electronic health record databases across 9 countries were used. Models were developed and internally validated on Optum® De-identified Clinformatics® Data Mart Database using L1-regularized logistic regression to estimate the risk of adverse health outcomes within 3 months (leukopenia, pancytopenia, infection), 2 years (myocardial infarction (MI) and stroke), and 5 years (cancers [colorectal, breast, uterine] after treatment initiation. Candidate predictors included demographic variables and past medical history. Models were externally validated on all other databases. Performance was assessed using the area under the receiver operator characteristic curve (AUC) and calibration plots.

FINDINGS

Models were developed and internally validated on 21,547 RA patients and externally validated on 131,928 RA patients. Models for serious infection (AUC: internal 0.74, external ranging from 0.62 to 0.83), MI (AUC: internal 0.76, external ranging from 0.56 to 0.82), and stroke (AUC: internal 0.77, external ranging from 0.63 to 0.95), showed good discrimination and adequate calibration. Models for the other outcomes showed modest internal discrimination (AUC < 0.65) and were not externally validated.

INTERPRETATION

We developed and validated prediction models for a variety of adverse health outcomes in RA patients initiating first-line MTX monotherapy. Final models for serious infection, MI, and stroke demonstrated good performance across multiple databases and can be studied for clinical use.

FUNDING

This activity under the European Health Data & Evidence Network (EHDEN) has received funding from the Innovative Medicines Initiative 2 Joint Undertaking under grant agreement No 806968. This Joint Undertaking receives support from the European Union's Horizon 2020 research and innovation programme and EFPIA.

Standard Vocabularies to Improve Machine Learning Model Transferability With Electronic Health Record Data: Retrospective Cohort Study Using Health Care-Associated Infection

BACKGROUND

With the widespread adoption of electronic healthcare records (EHRs) by US hospitals, there is an opportunity to leverage this data for the development of predictive algorithms to improve clinical care. A key barrier in model development and implementation includes the external validation of model discrimination, which is rare and often results in worse performance. One reason why machine learning models are not externally generalizable is data heterogeneity. A potential solution to address the substantial data heterogeneity between health care systems is to use standard vocabularies to map EHR data elements. The advantage of these vocabularies is a hierarchical relationship between elements, which allows the aggregation of specific clinical features to more general grouped concepts.

OBJECTIVE

This study aimed to evaluate grouping EHR data using standard vocabularies to improve the transferability of machine learning models for the detection of postoperative health care-associated infections across institutions with different EHR systems.

METHODS

Patients who underwent surgery from the University of Utah Health and Intermountain Healthcare from July 2014 to August 2017 with complete follow-up data were included. The primary outcome was a health care-associated infection within 30 days of the procedure. EHR data from 0-30 days after the operation were mapped to standard vocabularies and grouped using the hierarchical relationships of the vocabularies. Model performance was measured using the area under the receiver operating characteristic curve (AUC) and F₁-score in internal and external validations. To evaluate model transferability, a difference-in-difference metric was defined as the difference in performance drop between internal and external validations for the baseline and grouped models.

RESULTS

A total of 5775 patients from the University of Utah and 15,434 patients from Intermountain Healthcare were included. The prevalence of selected outcomes was from 4.9% (761/15,434) to 5% (291/5775) for surgical site infections, from 0.8% (44/5775) to 1.1% (171/15,434) for pneumonia, from 2.6% (400/15,434) to 3% (175/5775) for sepsis, and from 0.8% (125/15,434) to 0.9% (50/5775) for urinary tract infections. In all outcomes, the grouping of data using standard vocabularies resulted in a reduced drop in AUC and F₁-score in external validation compared to baseline features (all P<.001, except urinary tract infection AUC: P=.002). The difference-in-difference metrics ranged from 0.005 to 0.248 for AUC and from 0.075 to 0.216 for F₁-score.

CONCLUSIONS

We demonstrated that grouping machine learning model features based on standard vocabularies improved model transferability between data sets across 2 institutions. Improving model transferability using standard vocabularies has the potential to improve the generalization of clinical prediction models across the health care system.

Generic Medical Concept Embedding and Time Decay for Diverse Patient Outcome Prediction Tasks

Many fields, including Natural Language Processing (NLP), have recently witnessed the benefit of pre-training with large generic datasets to improve the accuracy of prediction tasks. However, there exist key differences between the longitudinal healthcare data (e.g., claims) and NLP tasks, which make the direct application of NLP pre-training methods to healthcare data inappropriate. In this article, we developed a pre-training scheme for longitudinal healthcare data that leverages the pairing of medical history and a future event. We then conducted systematic evaluations of various methods on ten patient-level prediction tasks encompassing adverse events, misdiagnosis, disease risks, and readmission. In addition to substantially reducing model size, our results show that a universal medical concept embedding pretrained with generic big data as well as carefully designed time decay modeling improves the accuracy of different downstream prediction tasks.

•

This work develops a pre-training scheme for longitudinal healthcare data

•

The method leverages the pairing of medical history and a future event

•

We created a universal medical concept embedding pretrained with generic data

•

We designed a time-decay method for medical concept data

Development of an Interoperable and Easily Transferable Clinical Decision Support System Deployment Platform: System Design and Development Study

BACKGROUND

A clinical decision support system (CDSS) is recognized as a technology that enhances clinical efficacy and safety. However, its full potential has not been realized, mainly due to clinical data standards and noninteroperable platforms.

OBJECTIVE

In this paper, we introduce the common data model-based intelligent algorithm network environment (CANE) platform that supports the implementation and deployment of a CDSS.

METHODS

CDSS reasoning engines, usually represented as R or Python objects, are deployed into the CANE platform and converted into C# objects. When a clinician requests CANE-based decision support in the electronic health record (EHR) system, patients' information is transformed into Health Level 7 Fast Healthcare Interoperability Resources (FHIR) format and transmitted to the CANE server inside the hospital firewall. Upon receiving the necessary data, the CANE system's modules perform the following tasks: (1) the preprocessing module converts the FHIRs into the input data required by the specific reasoning engine, (2) the reasoning engine module operates the target algorithms, (3) the integration module communicates with the other institutions' CANE systems to request and transmit a summary report to aid in decision support, and (4) creates a user interface by integrating the summary report and the results calculated by the reasoning engine.

RESULTS

We developed a CANE system such that any algorithm implemented in the system can be directly called through the RESTful application programming interface when it is integrated with an EHR system. Eight algorithms were developed and deployed in the CANE system. Using a knowledge-based algorithm, physicians can screen patients who are prone to sepsis and obtain treatment guides for patients with sepsis with the CANE system. Further, using a nonknowledge-based algorithm, the CANE system supports emergency physicians' clinical decisions about optimum resource allocation by predicting a patient's acuity and prognosis during triage.

CONCLUSIONS

We successfully developed a common data model-based platform that adheres to medical informatics standards and could aid artificial intelligence model deployment using R or Python.

Performance of an electronic health record-based predictive model to identify patients with atrial fibrillation across countries

Background

Atrial fibrillation (AF) burden on patients and healthcare systems warrants innovative strategies for screening asymptomatic individuals.

Objective

We sought to externally validate a predictive model originally developed in a German population to detect unidentified incident AF utilising real-world primary healthcare databases from countries in Europe and Australia.

Methods

This retrospective cohort study used anonymized, longitudinal patient data from 5 country-level primary care databases, including Australia, Belgium, France, Germany, and the UK. The study eligibility included adult patients (≥45 years) with either an AF diagnosis (cases) or no diagnosis (controls) who had continuous enrolment in the respective database prior to the study period. Logistic regression was fitted to a binary response (yes/no) for AF diagnosis using pre-determined risk factors.

Results

AF patients were from Germany (n = 63,562), the UK (n = 42,652), France (n = 7,213), Australia (n = 2,753), and Belgium (n = 1,371). Cases were more likely to have hypertension or other cardiac conditions than controls in all validation datasets compared to the model development data. The area under the receiver operating characteristic (ROC) curve in the validation datasets ranged from 0.79 (Belgium) to 0.84 (Germany), comparable to the German study model, which had an area under the curve of 0.83. Most validation sets reported similar specificity at approximately 80% sensitivity, ranging from 67% (France) to 71% (United Kingdom). The positive predictive value (PPV) ranged from 2% (Belgium) to 16% (Germany), and the number needed to be screened was 50 in Belgium and 6 in Germany. The prevalence of AF varied widely between these datasets, which may be related to different coding practices. Low prevalence affected PPV, but not sensitivity, specificity, and ROC curves.

Conclusions

AF risk prediction algorithms offer targeted ways to identify patients using electronic health records, which could improve screening number and the cost-effectiveness of AF screening if implemented in clinical practice.

Developing Clinical Prediction Models Using Primary Care Electronic Health Record Data: The Impact of Data Preparation Choices on Model Performance

Objective

To quantify prediction model performance in relation to data preparation choices when using electronic health records (EHR).

Study Design and Setting

Cox proportional hazards models were developed for predicting the first-ever main adverse cardiovascular events using Dutch primary care EHR data. The reference model was based on a 1-year run-in period, cardiovascular events were defined based on both EHR diagnosis and medication codes, and missing values were multiply imputed. We compared data preparation choices based on (i) length of the run-in period (2- or 3-year run-in); (ii) outcome definition (EHR diagnosis codes or medication codes only); and (iii) methods addressing missing values (mean imputation or complete case analysis) by making variations on the derivation set and testing their impact in a validation set.

Results

We included 89,491 patients in whom 6,736 first-ever main adverse cardiovascular events occurred during a median follow-up of 8 years. Outcome definition based only on diagnosis codes led to a systematic underestimation of risk (calibration curve intercept: 0.84; 95% CI: 0.83-0.84), while complete case analysis led to overestimation (calibration curve intercept: -0.52; 95% CI: -0.53 to -0.51). Differences in the length of the run-in period showed no relevant impact on calibration and discrimination.

Conclusion

Data preparation choices regarding outcome definition or methods to address missing values can have a substantial impact on the calibration of predictions, hampering reliable clinical decision support. This study further illustrates the urgency of transparent reporting of modeling choices in an EHR data setting.

Learning patient-level prediction models across multiple healthcare databases: evaluation of ensembles for increasing model transportability

Background

Prognostic models that are accurate could help aid medical decision making. Large observational databases often contain temporal medical data for large and diverse populations of patients. It may be possible to learn prognostic models using the large observational data. Often the performance of a prognostic model undesirably worsens when transported to a different database (or into a clinical setting). In this study we investigate different ensemble approaches that combine prognostic models independently developed using different databases (a simple federated learning approach) to determine whether ensembles that combine models developed across databases can improve model transportability (perform better in new data than single database models)?

Methods

For a given prediction question we independently trained five single database models each using a different observational healthcare database. We then developed and investigated numerous ensemble models (fusion, stacking and mixture of experts) that combined the different database models. Performance of each model was investigated via discrimination and calibration using a leave one dataset out technique, i.e., hold out one database to use for validation and use the remaining four datasets for model development. The internal validation of a model developed using the hold out database was calculated and presented as the ‘internal benchmark’ for comparison.

Results

In this study the fusion ensembles generally outperformed the single database models when transported to a previously unseen database and the performances were more consistent across unseen databases. Stacking ensembles performed poorly in terms of discrimination when the labels in the unseen database were limited. Calibration was consistently poor when both ensembles and single database models were applied to previously unseen databases.

Conclusion

A simple federated learning approach that implements ensemble techniques to combine models independently developed across different databases for the same prediction question may improve the discriminative performance in new data (new database or clinical setting) but will need to be recalibrated using the new data. This could help medical decision making by improving prognostic model performance.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12911-022-01879-6.

Training prediction models for individual risk assessment of postoperative complications after surgery for colorectal cancer

BACKGROUND

The occurrence of postoperative complications and anastomotic leakage are major drivers of mortality in the immediate phase after colorectal cancer surgery. We trained prediction models for calculating patients' individual risk of complications based only on preoperatively available data in a multidisciplinary team setting. Knowing prior to surgery the probability of developing a complication could aid in improving informed decision-making by surgeon and patient and individualize surgical treatment trajectories.

METHODS

All patients over 18 years of age undergoing any resection for colorectal cancer between January 1, 2014 and December 31, 2019 from the nationwide Danish Colorectal Cancer Group database were included. Data from the database were converted into Observational Medical Outcomes Partnership Common Data Model maintained by the Observation Health Data Science and Informatics initiative. Multiple machine learning models were trained to predict postoperative complications of Clavien-Dindo grade ≥ 3B and anastomotic leakage within 30 days after surgery.

RESULTS

Between 2014 and 2019, 23,907 patients underwent resection for colorectal cancer in Denmark. A Clavien-Dindo complication grade ≥ 3B occurred in 2,958 patients (12.4%). Of 17,190 patients that received an anastomosis, 929 experienced anastomotic leakage (5.4%). Among the compared machine learning models, Lasso Logistic Regression performed best. The predictive model for complications had an area under the receiver operating characteristic curve (AUROC) of 0.704 (95%CI 0.683-0.724) and an AUROC of 0.690 (95%CI 0.655-0.724) for anastomotic leakage.

CONCLUSIONS

The prediction of postoperative complications based only on preoperative variables using a national quality assurance colorectal cancer database shows promise for calculating patient's individual risk. Future work will focus on assessing the value of adding laboratory parameters and drug exposure as candidate predictors. Furthermore, we plan to assess the external validity of our proposed model.

Mapping Cancer Registry Data to the Episode Domain of the Observational Medical Outcomes Partnership Model (OMOP)

A great challenge in the use of standardized cancer registry data is deriving reliable, evidence-based results from large amounts of data. A solution could be its mapping to a common data model such as OMOP, which represents knowledge in a unified semantic base, enabling decentralized analysis. The recently released Episode Domain of the OMOP CDM allows episodic modelling of a patient’ disease and treatment phases. In this study, we mapped oncology registry data to the Episode Domain. A total of 184,718 Episodes could be implemented to concepts of cancer drug treatment. Additionally, source data were mapped to new terminologies as part of the release. It was possible to map ≈ 73.8% of the source data to the respective OMOP standard. Best mapping was achieved in the Procedure Domain with 98.7%. To evaluate the implementation, the survival probabilities of the CDM and source system were calculated (n = 2756/2902, median OAS = 82.2/91.1 months, 95% Cl = 77.4–89.5/84.4–100.9). In conclusion, the new release of the CDM increased its applicability, especially in observational cancer research. Regarding the mapping, a higher score could be achieved if terminologies which are frequently used in Europe are included in the Standardized Vocabulary Metadata Repository.

Trajectories: a framework for detecting temporal clinical event sequences from health data standardized to the Observational Medical Outcomes Partnership (OMOP) Common Data Model

Objective

To develop a framework for identifying temporal clinical event trajectories from Observational Medical Outcomes Partnership-formatted observational healthcare data.

Materials and Methods

A 4-step framework based on significant temporal event pair detection is described and implemented as an open-source R package. It is used on a population-based Estonian dataset to first replicate a large Danish population-based study and second, to conduct a disease trajectory detection study for type 2 diabetes patients in the Estonian and Dutch databases as an example.

Results

As a proof of concept, we apply the methods in the Estonian database and provide a detailed breakdown of our findings. All Estonian population-based event pairs are shown. We compare the event pairs identified from Estonia to Danish and Dutch data and discuss the causes of the differences. The overlap in the results was only 2.4%, which highlights the need for running similar studies in different populations.

Conclusions

For the first time, there is a complete software package for detecting disease trajectories in health data.

Patient-Level Fall Risk Prediction Using the Observational Medical Outcomes Partnership's Common Data Model: Pilot Feasibility Study

BACKGROUND

Falls in acute care settings threaten patients' safety. Researchers have been developing fall risk prediction models and exploring risk factors to provide evidence-based fall prevention practices; however, such efforts are hindered by insufficient samples, limited covariates, and a lack of standardized methodologies that aid study replication.

OBJECTIVE

The objectives of this study were to (1) convert fall-related electronic health record data into the standardized Observational Medical Outcome Partnership's (OMOP) common data model format and (2) develop models that predict fall risk during 2 time periods.

METHODS

As a pilot feasibility test, we converted fall-related electronic health record data (nursing notes, fall risk assessment sheet, patient acuity assessment sheet, and clinical observation sheet) into standardized OMOP common data model format using an extraction, transformation, and load process. We developed fall risk prediction models for 2 time periods (within 7 days of admission and during the entire hospital stay) using 2 algorithms (least absolute shrinkage and selection operator logistic regression and random forest).

RESULTS

In total, 6277 nursing statements, 747,049,486 clinical observation sheet records, 1,554,775 fall risk scores, and 5,685,011 patient acuity scores were converted into OMOP common data model format. All our models (area under the receiver operating characteristic curve 0.692-0.726) performed better than the Hendrich II Fall Risk Model. Patient acuity score, fall history, age ≥60 years, movement disorder, and central nervous system agents were the most important predictors in the logistic regression models.

CONCLUSIONS

To enhance model performance further, we are currently converting all nursing records into the OMOP common data model data format, which will then be included in the models. Thus, in the near future, the performance of fall risk prediction models could be improved through the application of abundant nursing records and external validation.

Evaluation of domain generalization and adaptation on improving model robustness to temporal dataset shift in clinical medicine

Temporal dataset shift associated with changes in healthcare over time is a barrier to deploying machine learning-based clinical decision support systems. Algorithms that learn robust models by estimating invariant properties across time periods for domain generalization (DG) and unsupervised domain adaptation (UDA) might be suitable to proactively mitigate dataset shift. The objective was to characterize the impact of temporal dataset shift on clinical prediction models and benchmark DG and UDA algorithms on improving model robustness. In this cohort study, intensive care unit patients from the MIMIC-IV database were categorized by year groups (2008-2010, 2011-2013, 2014-2016 and 2017-2019). Tasks were predicting mortality, long length of stay, sepsis and invasive ventilation. Feedforward neural networks were used as prediction models. The baseline experiment trained models using empirical risk minimization (ERM) on 2008-2010 (ERM[08-10]) and evaluated them on subsequent year groups. DG experiment trained models using algorithms that estimated invariant properties using 2008-2016 and evaluated them on 2017-2019. UDA experiment leveraged unlabelled samples from 2017 to 2019 for unsupervised distribution matching. DG and UDA models were compared to ERM[08-16] models trained using 2008-2016. Main performance measures were area-under-the-receiver-operating-characteristic curve (AUROC), area-under-the-precision-recall curve and absolute calibration error. Threshold-based metrics including false-positives and false-negatives were used to assess the clinical impact of temporal dataset shift and its mitigation strategies. In the baseline experiments, dataset shift was most evident for sepsis prediction (maximum AUROC drop, 0.090; 95% confidence interval (CI), 0.080-0.101). Considering a scenario of 100 consecutively admitted patients showed that ERM[08-10] applied to 2017-2019 was associated with one additional false-negative among 11 patients with sepsis, when compared to the model applied to 2008-2010. When compared with ERM[08-16], DG and UDA experiments failed to produce more robust models (range of AUROC difference, - 0.003 to 0.050). In conclusion, DG and UDA failed to produce more robust models compared to ERM in the setting of temporal dataset shift. Alternate approaches are required to preserve model performance over time in clinical medicine.

Seek COVER: using a disease proxy to rapidly develop and validate a personalized risk calculator for COVID-19 outcomes in an international network

Background

We investigated whether we could use influenza data to develop prediction models for COVID-19 to increase the speed at which prediction models can reliably be developed and validated early in a pandemic. We developed COVID-19 Estimated Risk (COVER) scores that quantify a patient’s risk of hospital admission with pneumonia (COVER-H), hospitalization with pneumonia requiring intensive services or death (COVER-I), or fatality (COVER-F) in the 30-days following COVID-19 diagnosis using historical data from patients with influenza or flu-like symptoms and tested this in COVID-19 patients.

Methods

We analyzed a federated network of electronic medical records and administrative claims data from 14 data sources and 6 countries containing data collected on or before 4/27/2020. We used a 2-step process to develop 3 scores using historical data from patients with influenza or flu-like symptoms any time prior to 2020. The first step was to create a data-driven model using LASSO regularized logistic regression, the covariates of which were used to develop aggregate covariates for the second step where the COVER scores were developed using a smaller set of features. These 3 COVER scores were then externally validated on patients with 1) influenza or flu-like symptoms and 2) confirmed or suspected COVID-19 diagnosis across 5 databases from South Korea, Spain, and the United States. Outcomes included i) hospitalization with pneumonia, ii) hospitalization with pneumonia requiring intensive services or death, and iii) death in the 30 days after index date.

Results

Overall, 44,507 COVID-19 patients were included for model validation. We identified 7 predictors (history of cancer, chronic obstructive pulmonary disease, diabetes, heart disease, hypertension, hyperlipidemia, kidney disease) which combined with age and sex discriminated which patients would experience any of our three outcomes. The models achieved good performance in influenza and COVID-19 cohorts. For COVID-19 the AUC ranges were, COVER-H: 0.69–0.81, COVER-I: 0.73–0.91, and COVER-F: 0.72–0.90. Calibration varied across the validations with some of the COVID-19 validations being less well calibrated than the influenza validations.

Conclusions

This research demonstrated the utility of using a proxy disease to develop a prediction model. The 3 COVER models with 9-predictors that were developed using influenza data perform well for COVID-19 patients for predicting hospitalization, intensive services, and fatality. The scores showed good discriminatory performance which transferred well to the COVID-19 population. There was some miscalibration in the COVID-19 validations, which is potentially due to the difference in symptom severity between the two diseases. A possible solution for this is to recalibrate the models in each location before use.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12874-022-01505-z.

Guidelines for artificial intelligence in medicine: A literature review and content analysis of frameworks (Preprint)

Background

Artificial intelligence (AI) is rapidly expanding in medicine despite a lack of consensus on its application and evaluation.

Objective

We sought to identify current frameworks guiding the application and evaluation of AI for predictive analytics in medicine and to describe the content of these frameworks. We also assessed what stages along the AI translational spectrum (ie, AI development, reporting, evaluation, implementation, and surveillance) the content of each framework has been discussed.

Methods

We performed a literature review of frameworks regarding the oversight of AI in medicine. The search included key topics such as “artificial intelligence,” “machine learning,” “guidance as topic,” and “translational science,” and spanned the time period 2014-2022. Documents were included if they provided generalizable guidance regarding the use or evaluation of AI in medicine. Included frameworks are summarized descriptively and were subjected to content analysis. A novel evaluation matrix was developed and applied to appraise the frameworks’ coverage of content areas across translational stages.

Results

Fourteen frameworks are featured in the review, including six frameworks that provide descriptive guidance and eight that provide reporting checklists for medical applications of AI. Content analysis revealed five considerations related to the oversight of AI in medicine across frameworks: transparency, reproducibility, ethics, effectiveness, and engagement. All frameworks include discussions regarding transparency, reproducibility, ethics, and effectiveness, while only half of the frameworks discuss engagement. The evaluation matrix revealed that frameworks were most likely to report AI considerations for the translational stage of development and were least likely to report considerations for the translational stage of surveillance.

Conclusions

Existing frameworks for the application and evaluation of AI in medicine notably offer less input on the role of engagement in oversight and regarding the translational stage of surveillance. Identifying and optimizing strategies for engagement are essential to ensure that AI can meaningfully benefit patients and other end users.

Trends in the conduct and reporting of clinical prediction model development and validation: a systematic review

Objectives

This systematic review aims to provide further insights into the conduct and reporting of clinical prediction model development and validation over time. We focus on assessing the reporting of information necessary to enable external validation by other investigators.

Materials and Methods

We searched Embase, Medline, Web-of-Science, Cochrane Library, and Google Scholar to identify studies that developed 1 or more multivariable prognostic prediction models using electronic health record (EHR) data published in the period 2009–2019.

Results

We identified 422 studies that developed a total of 579 clinical prediction models using EHR data. We observed a steep increase over the years in the number of developed models. The percentage of models externally validated in the same paper remained at around 10%. Throughout 2009–2019, for both the target population and the outcome definitions, code lists were provided for less than 20% of the models. For about half of the models that were developed using regression analysis, the final model was not completely presented.

Discussion

Overall, we observed limited improvement over time in the conduct and reporting of clinical prediction model development and validation. In particular, the prediction problem definition was often not clearly reported, and the final model was often not completely presented.

Conclusion

Improvement in the reporting of information necessary to enable external validation by other investigators is still urgently needed to increase clinical adoption of developed models.

Psychosis Relapse Prediction Leveraging Electronic Health Records Data and Natural Language Processing Enrichment Methods

BACKGROUND

Identifying patients at a high risk of psychosis relapse is crucial for early interventions. A relevant psychiatric clinical context is often recorded in clinical notes; however, the utilization of unstructured data remains limited. This study aimed to develop psychosis-relapse prediction models using various types of clinical notes and structured data.

METHODS

Clinical data were extracted from the electronic health records of the Ajou University Medical Center in South Korea. The study population included patients with psychotic disorders, and outcome was psychosis relapse within 1 year. Using only structured data, we developed an initial prediction model, then three natural language processing (NLP)-enriched models using three types of clinical notes (psychological tests, admission notes, and initial nursing assessment) and one complete model. Latent Dirichlet Allocation was used to cluster the clinical context into similar topics. All models applied the least absolute shrinkage and selection operator logistic regression algorithm. We also performed an external validation using another hospital database.

RESULTS

A total of 330 patients were included, and 62 (18.8%) experienced psychosis relapse. Six predictors were used in the initial model and 10 additional topics from Latent Dirichlet Allocation processing were added in the enriched models. The model derived from all notes showed the highest value of the area under the receiver operating characteristic (AUROC = 0.946) in the internal validation, followed by models based on the psychological test notes, admission notes, initial nursing assessments, and structured data only (0.902, 0.855, 0.798, and 0.784, respectively). The external validation was performed using only the initial nursing assessment note, and the AUROC was 0.616.

CONCLUSIONS

We developed prediction models for psychosis relapse using the NLP-enrichment method. Models using clinical notes were more effective than models using only structured data, suggesting the importance of unstructured data in psychosis prediction.

Development and Validation of the Radiology Common Data Model (R-CDM) for the International Standardization of Medical Imaging Data

PURPOSE

Digital Imaging and Communications in Medicine (DICOM), a standard file format for medical imaging data, contains metadata describing each file. However, metadata are often incomplete, and there is no standardized format for recording metadata, leading to inefficiency during the metadata-based data retrieval process. Here, we propose a novel standardization method for DICOM metadata termed the Radiology Common Data Model (R-CDM).

MATERIALS AND METHODS

R-CDM was designed to be compatible with Health Level Seven International (HL7)/Fast Healthcare Interoperability Resources (FHIR) and linked with the Observational Medical Outcomes Partnership (OMOP)-CDM to achieve a seamless link between clinical data and medical imaging data. The terminology system was standardized using the RadLex playbook, a comprehensive lexicon of radiology. As a proof of concept, the R-CDM conversion process was conducted with 41.7 TB of data from the Ajou University Hospital. The R-CDM database visualizer was developed to visualize the main characteristics of the R-CDM database.

RESULTS

Information from 2801360 cases and 87203226 DICOM files was organized into two tables constituting the R-CDM. Information on imaging device and image resolution was recorded with more than 99.9% accuracy. Furthermore, OMOP-CDM and R-CDM were linked to efficiently extract specific types of images from specific patient cohorts.

CONCLUSION

R-CDM standardizes the structure and terminology for recording medical imaging data to eliminate incomplete and unstandardized information. Successful standardization was achieved by the extract, transform, and load process and image classifier. We hope that the R-CDM will contribute to deep learning research in the medical imaging field by enabling the securement of large-scale medical imaging data from multinational institutions.

OUP accepted manuscript

Objective

This systematic review aims to assess how information from unstructured text is used to develop and validate clinical prognostic prediction models. We summarize the prediction problems and methodological landscape and determine whether using text data in addition to more commonly used structured data improves the prediction performance.

Materials and Methods

We searched Embase, MEDLINE, Web of Science, and Google Scholar to identify studies that developed prognostic prediction models using information extracted from unstructured text in a data-driven manner, published in the period from January 2005 to March 2021. Data items were extracted, analyzed, and a meta-analysis of the model performance was carried out to assess the added value of text to structured-data models.

Results

We identified 126 studies that described 145 clinical prediction problems. Combining text and structured data improved model performance, compared with using only text or only structured data. In these studies, a wide variety of dense and sparse numeric text representations were combined with both deep learning and more traditional machine learning methods. External validation, public availability, and attention for the explainability of the developed models were limited.

Conclusion

The use of unstructured text in the development of prognostic prediction models has been found beneficial in addition to structured data in most studies. The text data are source of valuable information for prediction model development and should not be neglected. We suggest a future focus on explainability and external validation of the developed models, promoting robust and trustworthy prediction models in clinical practice.