Systematic Review of Approaches to Preserve Machine Learning Performance in the Presence of Temporal Dataset Shift in Clinical Medicine

Pathophysiological Features in Electronic Medical Records Sustain Model Performance under Temporal Dataset Shift

Access to real-world data streams like electronic medical records (EMRs) has accelerated the development of supervised machine learning (ML) models for clinical applications. However, few studies investigate the differential impact of particular features in the EMR on model performance under temporal dataset shift. To explain how features in the EMR impact models over time, this study aggregates features into feature groups by their source (e.g. medication orders, diagnosis codes and lab results) and feature categories based on their reflection of patient pathophysiology or healthcare processes. We adapt Shapley values to explain feature groups' and feature categories' marginal contribution to initial and sustained model performance. We investigate three standard clinical prediction tasks and find that while feature contributions to initial performance differ across tasks, pathophysiological features help mitigate temporal discrimination deterioration. These results provide interpretable insights on how specific feature groups contribute to model performance and robustness to temporal dataset shift.

Sustainable deployment of clinical prediction tools-a 360° approach to model maintenance

BACKGROUND

As the enthusiasm for integrating artificial intelligence (AI) into clinical care grows, so has our understanding of the challenges associated with deploying impactful and sustainable clinical AI models. Complex dataset shifts resulting from evolving clinical environments strain the longevity of AI models as predictive accuracy and associated utility deteriorate over time.

OBJECTIVE

Responsible practice thus necessitates the lifecycle of AI models be extended to include ongoing monitoring and maintenance strategies within health system algorithmovigilance programs. We describe a framework encompassing a 360° continuum of preventive, preemptive, responsive, and reactive approaches to address model monitoring and maintenance from critically different angles.

DISCUSSION

We describe the complementary advantages and limitations of these four approaches and highlight the importance of such a coordinated strategy to help ensure the promise of clinical AI is not short-lived.

Performance of risk models to predict mortality risk for patients with heart failure: evaluation in an integrated health system

BACKGROUND

Referral of patients with heart failure (HF) who are at high mortality risk for specialist evaluation is recommended. Yet, most tools for identifying such patients are difficult to implement in electronic health record (EHR) systems.

OBJECTIVE

To assess the performance and ease of implementation of Machine learning Assessment of RisK and EaRly mortality in Heart Failure (MARKER-HF), a machine-learning model that uses structured data that is readily available in the EHR, and compare it with two commonly used risk scores: the Seattle Heart Failure Model (SHFM) and Meta-Analysis Global Group in Chronic (MAGGIC) Heart Failure Risk Score.

DESIGN

Retrospective, cohort study.

PARTICIPANTS

Data from 6764 adults with HF were abstracted from EHRs at a large integrated health system from 1/1/10 to 12/31/19.

MAIN MEASURES

One-year survival from time of first cardiology or primary care visit was estimated using MARKER-HF, SHFM, and MAGGIC. Discrimination was measured by the area under the receiver operating curve (AUC). Calibration was assessed graphically.

KEY RESULTS

Compared to MARKER-HF, both SHFM and MAGGIC required a considerably larger amount of data engineering and imputation to generate risk score estimates. MARKER-HF, SHFM, and MAGGIC exhibited similar discriminations with AUCs of 0.70 (0.69-0.73), 0.71 (0.69-0.72), and 0.71 (95% CI 0.70-0.73), respectively. All three scores showed good calibration across the full risk spectrum.

CONCLUSIONS

These findings suggest that MARKER-HF, which uses readily available clinical and lab measurements in the EHR and required less imputation and data engineering than SHFM and MAGGIC, is an easier tool to identify high-risk patients in ambulatory clinics who could benefit from referral to a HF specialist.

Monitoring performance of clinical artificial intelligence: a scoping review protocol

OBJECTIVE

The objective of this scoping review is to describe the scope and nature of research on the monitoring of clinical artificial intelligence (AI) systems. The review will identify the various methodologies used to monitor clinical AI, while also mapping the factors that influence the selection of monitoring approaches.

INTRODUCTION

AI is being used in clinical decision-making at an increasing rate. While much attention has been directed toward the development and validation of AI for clinical applications, the practical implementation aspects, notably the establishment of rational monitoring/quality assurance systems, has received comparatively limited scientific interest. Given the scarcity of evidence and the heterogeneity of methodologies used in this domain, there is a compelling rationale for conducting a scoping review on this subject.

INCLUSION CRITERIA

This scoping review will include any publications that describe systematic, continuous, or repeated initiatives that evaluate or predict clinical performance of AI models with direct implications for the management of patients in any segment of the health care system.

METHODS

Publications will be identified through searches of the MEDLINE (Ovid), Embase (Ovid), and Scopus databases. Additionally, backward and forward citation searches, as well as a thorough investigation of gray literature, will be conducted. Title and abstract screening, full-text evaluation, and data extraction will be performed by 2 or more independent reviewers. Data will be extracted using a tool developed by the authors. The results will be presented graphically and narratively.

REVIEW REGISTRATION

Open Science Framework https://osf.io/afkrn.

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.

Clinical outcome prediction using observational supervision with electronic health records and audit logs

OBJECTIVE

Audit logs in electronic health record (EHR) systems capture interactions of providers with clinical data. We determine if machine learning (ML) models trained using audit logs in conjunction with clinical data ("observational supervision") outperform ML models trained using clinical data alone in clinical outcome prediction tasks, and whether they are more robust to temporal distribution shifts in the data.

MATERIALS AND METHODS

Using clinical and audit log data from Stanford Healthcare, we trained and evaluated various ML models including logistic regression, support vector machine (SVM) classifiers, neural networks, random forests, and gradient boosted machines (GBMs) on clinical EHR data, with and without audit logs for two clinical outcome prediction tasks: major adverse kidney events within 120 days of ICU admission (MAKE-120) in acute kidney injury (AKI) patients and 30-day readmission in acute stroke patients. We further tested the best performing models using patient data acquired during different time-intervals to evaluate the impact of temporal distribution shifts on model performance.

RESULTS

Performance generally improved for all models when trained with clinical EHR data and audit log data compared with those trained with only clinical EHR data, with GBMs tending to have the overall best performance. GBMs trained with clinical EHR data and audit logs outperformed GBMs trained without audit logs in both clinical outcome prediction tasks: AUROC 0.88 (95% CI: 0.85-0.91) vs. 0.79 (95% CI: 0.77-0.81), respectively, for MAKE-120 prediction in AKI patients, and AUROC 0.74 (95% CI: 0.71-0.77) vs. 0.63 (95% CI: 0.62-0.64), respectively, for 30-day readmission prediction in acute stroke patients. The performance of GBM models trained using audit log and clinical data degraded less in later time-intervals than models trained using only clinical data.

CONCLUSION

Observational supervision with audit logs improved the performance of ML models trained to predict important clinical outcomes in patients with AKI and acute stroke, and improved robustness to temporal distribution shifts.

Utilizing Mixed Training and Multi-Head Attention to Address Data Shift in AI-Based Electromagnetic Solvers for Nano-Structured Metamaterials

When designing nano-structured metamaterials with an iterative optimization method, a fast deep learning solver is desirable to replace a time-consuming numerical solver, and the related issue of data shift is a subtle yet easily overlooked challenge. In this work, we explore the data shift challenge in an AI-based electromagnetic solver and present innovative solutions. Using a one-dimensional grating coupler as a case study, we demonstrate the presence of data shift through the probability density method and principal component analysis, and show the degradation of neural network performance through experiments dealing with data affected by data shift. We propose three effective strategies to mitigate the effects of data shift: mixed training, adding multi-head attention, and a comprehensive approach that combines both. The experimental results validate the efficacy of these approaches in addressing data shift. Specifically, the combination of mixed training and multi-head attention significantly reduces the mean absolute error, by approximately 36%, when applied to data affected by data shift. Our work provides crucial insights and guidance for AI-based electromagnetic solvers in the optimal design of nano-structured metamaterials.

Data drift in medical machine learning: implications and potential remedies

Data drift refers to differences between the data used in training a machine learning (ML) model and that applied to the model in real-world operation. Medical ML systems can be exposed to various forms of data drift, including differences between the data sampled for training and used in clinical operation, differences between medical practices or context of use between training and clinical use, and time-related changes in patient populations, disease patterns, and data acquisition, to name a few. In this article, we first review the terminology used in ML literature related to data drift, define distinct types of drift, and discuss in detail potential causes within the context of medical applications with an emphasis on medical imaging. We then review the recent literature regarding the effects of data drift on medical ML systems, which overwhelmingly show that data drift can be a major cause for performance deterioration. We then discuss methods for monitoring data drift and mitigating its effects with an emphasis on pre- and post-deployment techniques. Some of the potential methods for drift detection and issues around model retraining when drift is detected are included. Based on our review, we find that data drift is a major concern in medical ML deployment and that more research is needed so that ML models can identify drift early, incorporate effective mitigation strategies and resist performance decay.

Machine learning in precision diabetes care and cardiovascular risk prediction

Artificial intelligence and machine learning are driving a paradigm shift in medicine, promising data-driven, personalized solutions for managing diabetes and the excess cardiovascular risk it poses. In this comprehensive review of machine learning applications in the care of patients with diabetes at increased cardiovascular risk, we offer a broad overview of various data-driven methods and how they may be leveraged in developing predictive models for personalized care. We review existing as well as expected artificial intelligence solutions in the context of diagnosis, prognostication, phenotyping, and treatment of diabetes and its cardiovascular complications. In addition to discussing the key properties of such models that enable their successful application in complex risk prediction, we define challenges that arise from their misuse and the role of methodological standards in overcoming these limitations. We also identify key issues in equity and bias mitigation in healthcare and discuss how the current regulatory framework should ensure the efficacy and safety of medical artificial intelligence products in transforming cardiovascular care and outcomes in diabetes.

Comparative Analysis of the Clustering Quality in Self-Organizing Maps for Human Posture Classification

The objective of this article is to develop a methodology for selecting the appropriate number of clusters to group and identify human postures using neural networks with unsupervised self-organizing maps. Although unsupervised clustering algorithms have proven effective in recognizing human postures, many works are limited to testing which data are correctly or incorrectly recognized. They often neglect the task of selecting the appropriate number of groups (where the number of clusters corresponds to the number of output neurons, i.e., the number of postures) using clustering quality assessments. The use of quality scores to determine the number of clusters frees the expert to make subjective decisions about the number of postures, enabling the use of unsupervised learning. Due to high dimensionality and data variability, expert decisions (referred to as data labeling) can be difficult and time-consuming. In our case, there is no manual labeling step. We introduce a new clustering quality score: the discriminant score (DS). We describe the process of selecting the most suitable number of postures using human activity records captured by RGB-D cameras. Comparative studies on the usefulness of popular clustering quality scores-such as the silhouette coefficient, Dunn index, Calinski-Harabasz index, Davies-Bouldin index, and DS-for posture classification tasks are presented, along with graphical illustrations of the results produced by DS. The findings show that DS offers good quality in posture recognition, effectively following postural transitions and similarities.

DEPLOYR: a technical framework for deploying custom real-time machine learning models into the electronic medical record

OBJECTIVE

Heatlhcare institutions are establishing frameworks to govern and promote the implementation of accurate, actionable, and reliable machine learning models that integrate with clinical workflow. Such governance frameworks require an accompanying technical framework to deploy models in a resource efficient, safe and high-quality manner. Here we present DEPLOYR, a technical framework for enabling real-time deployment and monitoring of researcher-created models into a widely used electronic medical record system.

MATERIALS AND METHODS

We discuss core functionality and design decisions, including mechanisms to trigger inference based on actions within electronic medical record software, modules that collect real-time data to make inferences, mechanisms that close-the-loop by displaying inferences back to end-users within their workflow, monitoring modules that track performance of deployed models over time, silent deployment capabilities, and mechanisms to prospectively evaluate a deployed model's impact.

RESULTS

We demonstrate the use of DEPLOYR by silently deploying and prospectively evaluating 12 machine learning models trained using electronic medical record data that predict laboratory diagnostic results, triggered by clinician button-clicks in Stanford Health Care's electronic medical record.

DISCUSSION

Our study highlights the need and feasibility for such silent deployment, because prospectively measured performance varies from retrospective estimates. When possible, we recommend using prospectively estimated performance measures during silent trials to make final go decisions for model deployment.

CONCLUSION

Machine learning applications in healthcare are extensively researched, but successful translations to the bedside are rare. By describing DEPLOYR, we aim to inform machine learning deployment best practices and help bridge the model implementation gap.

Algorithmic fairness in artificial intelligence for medicine and healthcare

In healthcare, the development and deployment of insufficiently fair systems of artificial intelligence (AI) can undermine the delivery of equitable care. Assessments of AI models stratified across subpopulations have revealed inequalities in how patients are diagnosed, treated and billed. In this Perspective, we outline fairness in machine learning through the lens of healthcare, and discuss how algorithmic biases (in data acquisition, genetic variation and intra-observer labelling variability, in particular) arise in clinical workflows and the resulting healthcare disparities. We also review emerging technology for mitigating biases via disentanglement, federated learning and model explainability, and their role in the development of AI-based software as a medical device.

Interpretable clinical phenotypes among patients hospitalized with COVID-19 using cluster analysis

Background

Multiple clinical phenotypes have been proposed for coronavirus disease (COVID-19), but few have used multimodal data. Using clinical and imaging data, we aimed to identify distinct clinical phenotypes in patients admitted with COVID-19 and to assess their clinical outcomes. Our secondary objective was to demonstrate the clinical applicability of this method by developing an interpretable model for phenotype assignment.

Methods

We analyzed data from 547 patients hospitalized with COVID-19 at a Canadian academic hospital. We processed the data by applying a factor analysis of mixed data (FAMD) and compared four clustering algorithms: k-means, partitioning around medoids (PAM), and divisive and agglomerative hierarchical clustering. We used imaging data and 34 clinical variables collected within the first 24 h of admission to train our algorithm. We conducted a survival analysis to compare the clinical outcomes across phenotypes. With the data split into training and validation sets (75/25 ratio), we developed a decision-tree-based model to facilitate the interpretation and assignment of the observed phenotypes.

Results

Agglomerative hierarchical clustering was the most robust algorithm. We identified three clinical phenotypes: 79 patients (14%) in Cluster 1, 275 patients (50%) in Cluster 2, and 203 (37%) in Cluster 3. Cluster 2 and Cluster 3 were both characterized by a low-risk respiratory and inflammatory profile but differed in terms of demographics. Compared with Cluster 3, Cluster 2 comprised older patients with more comorbidities. Cluster 1 represented the group with the most severe clinical presentation, as inferred by the highest rate of hypoxemia and the highest radiological burden. Intensive care unit (ICU) admission and mechanical ventilation risks were the highest in Cluster 1. Using only two to four decision rules, the classification and regression tree (CART) phenotype assignment model achieved an AUC of 84% (81.5-86.5%, 95 CI) on the validation set.

Conclusions

We conducted a multidimensional phenotypic analysis of adult inpatients with COVID-19 and identified three distinct phenotypes associated with different clinical outcomes. We also demonstrated the clinical usability of this approach, as phenotypes can be accurately assigned using a simple decision tree. Further research is still needed to properly incorporate these phenotypes in the management of patients with COVID-19.

Foundation models for generalist medical artificial intelligence

The exceptionally rapid development of highly flexible, reusable artificial intelligence (AI) models is likely to usher in newfound capabilities in medicine. We propose a new paradigm for medical AI, which we refer to as generalist medical AI (GMAI). GMAI models will be capable of carrying out a diverse set of tasks using very little or no task-specific labelled data. Built through self-supervision on large, diverse datasets, GMAI will flexibly interpret different combinations of medical modalities, including data from imaging, electronic health records, laboratory results, genomics, graphs or medical text. Models will in turn produce expressive outputs such as free-text explanations, spoken recommendations or image annotations that demonstrate advanced medical reasoning abilities. Here we identify a set of high-impact potential applications for GMAI and lay out specific technical capabilities and training datasets necessary to enable them. We expect that GMAI-enabled applications will challenge current strategies for regulating and validating AI devices for medicine and will shift practices associated with the collection of large medical datasets.

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.

Evaluation of Feature Selection Methods for Preserving Machine Learning Performance in the Presence of Temporal Dataset Shift in Clinical Medicine

BACKGROUND

 Temporal dataset shift can cause degradation in model performance as discrepancies between training and deployment data grow over time. The primary objective was to determine whether parsimonious models produced by specific feature selection methods are more robust to temporal dataset shift as measured by out-of-distribution (OOD) performance, while maintaining in-distribution (ID) performance.

METHODS

 Our dataset consisted of intensive care unit patients from MIMIC-IV categorized by year groups (2008-2010, 2011-2013, 2014-2016, and 2017-2019). We trained baseline models using L2-regularized logistic regression on 2008-2010 to predict in-hospital mortality, long length of stay (LOS), sepsis, and invasive ventilation in all year groups. We evaluated three feature selection methods: L1-regularized logistic regression (L1), Remove and Retrain (ROAR), and causal feature selection. We assessed whether a feature selection method could maintain ID performance (2008-2010) and improve OOD performance (2017-2019). We also assessed whether parsimonious models retrained on OOD data performed as well as oracle models trained on all features in the OOD year group.

RESULTS

 The baseline model showed significantly worse OOD performance with the long LOS and sepsis tasks when compared with the ID performance. L1 and ROAR retained 3.7 to 12.6% of all features, whereas causal feature selection generally retained fewer features. Models produced by L1 and ROAR exhibited similar ID and OOD performance as the baseline models. The retraining of these models on 2017-2019 data using features selected from training on 2008-2010 data generally reached parity with oracle models trained directly on 2017-2019 data using all available features. Causal feature selection led to heterogeneous results with the superset maintaining ID performance while improving OOD calibration only on the long LOS task.

CONCLUSIONS

 While model retraining can mitigate the impact of temporal dataset shift on parsimonious models produced by L1 and ROAR, new methods are required to proactively improve temporal robustness.

Artificial intelligence in bronchopulmonary dysplasia- current research and unexplored frontiers

Provide an overview of bronchopulmonary dysplasia, its definitions, and their shortcomings. Explore the areas where machine learning may be used to further our understanding of bronchopulmonary dysplasia.

Why did AI get this one wrong? - Tree-based explanations of machine learning model predictions

Increasingly complex learning methods such as boosting, bagging and deep learning have made ML models more accurate, but harder to interpret and explain, culminating in black-box machine learning models. Model developers and users alike are often presented with a trade-off between performance and intelligibility, especially in high-stakes applications like medicine. In the present article we propose a novel methodological approach for generating explanations for the predictions of a generic machine learning model, given a specific instance for which the prediction has been made. The method, named AraucanaXAI, is based on surrogate, locally-fitted classification and regression trees that are used to provide post-hoc explanations of the prediction of a generic machine learning model. Advantages of the proposed XAI approach include superior fidelity to the original model, ability to deal with non-linear decision boundaries, and native support to both classification and regression problems. We provide a packaged, open-source implementation of the AraucanaXAI method and evaluate its behaviour in a number of different settings that are commonly encountered in medical applications of AI. These include potential disagreement between the model prediction and physician's expert opinion and low reliability of the prediction due to data scarcity.

Targeted validation: validating clinical prediction models in their intended population and setting

Clinical prediction models must be appropriately validated before they can be used. While validation studies are sometimes carefully designed to match an intended population/setting of the model, it is common for validation studies to take place with arbitrary datasets, chosen for convenience rather than relevance. We call estimating how well a model performs within the intended population/setting "targeted validation". Use of this term sharpens the focus on the intended use of a model, which may increase the applicability of developed models, avoid misleading conclusions, and reduce research waste. It also exposes that external validation may not be required when the intended population for the model matches the population used to develop the model; here, a robust internal validation may be sufficient, especially if the development dataset was large.

Narrative Review of Machine Learning in Rheumatic and Musculoskeletal Diseases for Clinicians and Researchers: Biases, Goals, and Future Directions

There has been rapid growth in the use of artificial intelligence (AI) analytics in medicine in recent years, including in rheumatic and musculoskeletal diseases (RMDs). Such methods represent a challenge to clinicians, patients, and researchers, given the "black box" nature of most algorithms, the unfamiliarity of the terms, and the lack of awareness of potential issues around these analyses. Therefore, this review aims to introduce this subject area in a way that is relevant and meaningful to clinicians and researchers. We hope to provide some insights into relevant strengths and limitations, reporting guidelines, as well as recent examples of such analyses in key areas, with a focus on lessons learned and future directions in diagnosis, phenotyping, prognosis, and precision medicine in RMDs.

Open questions and research gaps for monitoring and updating AI-enabled tools in clinical settings

As the implementation of artificial intelligence (AI)-enabled tools is realized across diverse clinical environments, there is a growing understanding of the need for ongoing monitoring and updating of prediction models. Dataset shift-temporal changes in clinical practice, patient populations, and information systems-is now well-documented as a source of deteriorating model accuracy and a challenge to the sustainability of AI-enabled tools in clinical care. While best practices are well-established for training and validating new models, there has been limited work developing best practices for prospective validation and model maintenance. In this paper, we highlight the need for updating clinical prediction models and discuss open questions regarding this critical aspect of the AI modeling lifecycle in three focus areas: model maintenance policies, performance monitoring perspectives, and model updating strategies. With the increasing adoption of AI-enabled tools, the need for such best practices must be addressed and incorporated into new and existing implementations. This commentary aims to encourage conversation and motivate additional research across clinical and data science stakeholders.

Advances in Machine Learning Approaches to Heart Failure with Preserved Ejection Fraction

Heart failure with preserved ejection fraction (HFpEF) represents a prototypical cardiovascular condition in which machine learning may improve targeted therapies and mechanistic understanding of pathogenesis. Machine learning, which involves algorithms that learn from data, has the potential to guide precision medicine approaches for complex clinical syndromes such as HFpEF. It is therefore important to understand the potential utility and common pitfalls of machine learning so that it can be applied and interpreted appropriately. Although machine learning holds considerable promise for HFpEF, it is subject to several potential pitfalls, which are important factors to consider when interpreting machine learning studies.

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.