Beyond Predictions: A Study of Clinical Model Development, Missing Data Impact Analysis, and Healthcare Provider Perspectives

Machine learning (ML) models are increasingly used in healthcare to support clinical decision making, yet their implementation in the real world remains limited by various challenges such as data quality issues, lack of interpretability, integration difficulties with existing clinical workflows, regulatory barriers and generalizability. This research aims to explore how ML models behave under conditions common in clinical practice – specifically missing data and the need for transparent, patient specific explanations – and how clinicians interpret and respond to these model outputs. To explore these issues, we applied our framework (sensitivity analysis framework, imputation and explainability framework) to the prediction of venous thromboembolism (VTE), a preventable but serious complication in surgical patients.

We developed ML models to predict VTE during post-surgical hospitalization and after discharge using structured electronic health records (EHR) data from diverse surgical populations. In addition to optimizing the predictive performance, we also investigate how the missing clinical features at the time of prediction can affect the model’s output. We simulate a range of plausible values for missing features and observe the variation in risk scores. This allows us to quantify how sensitive or uncertain a prediction is. By identifying features that drive this uncertainty, we offer a new layer of transparency into the model’s reliability at the patient level.

To understand how the clinicians interpret these model outputs, we conducted a study in which clinicians review patient cases with and without access to model explanations and uncertainty estimates. Using statistical assessments, we measure how their risk perception, confidence, and decision-making change in response to additional information.

This work tests the hypothesis that ML models that combine accurate predictions with tailored explanations and visibility into the effects of missing data can improve clinician understanding and change clinician behavior. While VTE serves as the motivating application, the broader contribution is in establishing a general framework for evaluating ML model behavior under conditions common in real-world healthcare and for studying how such models are perceived by clinical end-users.