OUP accepted manuscript

The application of explainable artificial intelligence (XAI) in electronic health record research: A scoping review

Machine Learning (ML) and Deep Learning (DL) models show potential in surpassing traditional methods including generalised linear models for healthcare predictions, particularly with large, complex datasets. However, low interpretability hinders practical implementation. To address this, Explainable Artificial Intelligence (XAI) methods are proposed, but a comprehensive evaluation of their effectiveness is currently limited. The aim of this scoping review is to critically appraise the application of XAI methods in ML/DL models using Electronic Health Record (EHR) data. In accordance with PRISMA scoping review guidelines, the study searched PUBMED and OVID/MEDLINE (including EMBASE) for publications related to tabular EHR data that employed ML/DL models with XAI. Out of 3220 identified publications, 76 were included. The selected publications published between February 2017 and June 2023, demonstrated an exponential increase over time. Extreme Gradient Boosting and Random Forest models were the most frequently used ML/DL methods, with 51 and 50 publications, respectively. Among XAI methods, Shapley Additive Explanations (SHAP) was predominant in 63 out of 76 publications, followed by partial dependence plots (PDPs) in 11 publications, and Locally Interpretable Model-Agnostic Explanations (LIME) in 8 publications. Despite the growing adoption of XAI methods, their applications varied widely and lacked critical evaluation. This review identifies the increasing use of XAI in tabular EHR research and highlights a deficiency in the reporting of methods and a lack of critical appraisal of validity and robustness. The study emphasises the need for further evaluation of XAI methods and underscores the importance of cautious implementation and interpretation in healthcare settings.

Multistate Modeling of Clinical Trajectories and Outcomes in the ICU: A Proof-of-Concept Evaluation of Acute Kidney Injury Among Critically Ill Patients With COVID-19

Multistate models yield high-fidelity analyses of the dynamic state transition and temporal dimensions of a clinical condition's natural history, offering superiority over aggregate modeling techniques for addressing these types of problems.

OBJECTIVES

To demonstrate the utility of these models in critical care, we examined acute kidney injury (AKI) development, progression, and outcomes in COVID-19 critical illness through multistate analyses.

DESIGN SETTING AND PARTICIPANTS

Retrospective cohort study at an urban tertiary-care academic hospital in the United States. All patients greater than or equal to 18 years in an ICU with COVID-19 in 2020, excluding patients with preexisting end-stage renal disease.

MAIN OUTCOMES AND MEASURES

Using electronic health record data, we determined AKI presence/stage in discrete 12-hour time windows and fit multistate models to determine longitudinal transitions and outcomes.

RESULTS

Of 367 encounters, 241 (66%) experienced AKI (maximal stages: 88 stage-1, 49 stage-2, 104 stage-3 AKI [51 received renal replacement therapy (RRT), 53 did not]). Patients receiving RRT overwhelmingly received invasive mechanical ventilation (IMV) (n = 60, 95%) compared with the AKI-without-RRT (n = 98, 53%) and no-AKI groups (n = 39, 32%; p < 0.001), with similar mortality patterns (RRT: n = 36, 57%; AKI: n = 74, 40%; non-AKI: n = 23, 19%; p < 0.001). After 24 hours in the ICU, almost half the cohort had AKI (44.9%; 95% CI, 41.6-48.2%). At 7 days after stage-1 AKI, 74.0% (63.6-84.4) were AKI-free or discharged. By contrast, fewer patients experiencing stage-3 AKI were recovered (30.0% [24.1-35.8%]) or discharged (7.9% [5.2-10.7%]) after 7 days. Early AKI occurred with similar frequency in patients receiving and not receiving IMV: after 24 hours in the ICU, 20.9% of patients (18.3-23.6%) had AKI and IMV, while 23.4% (20.6-26.2%) had AKI without IMV.

CONCLUSIONS AND RELEVANCE

In a multistate analysis of critically ill patients with COVID-19, AKI occurred early and heterogeneously in the course of critical illness. Multistate methods are useful and underused in ICU care delivery science as tools for understanding trajectories, prognoses, and resource needs.