JMIR Publications

ABSTRACT

Background:

Reusing long-term data from electronic health records (EHRs) is essential for training reliable and effective health AI. However, intrinsic changes in health data distributions over time, leading to dataset shifts, can compromise model performance, leading to model obsolescence and inaccurate decisions.

Objective:

In this study, we investigate whether unsupervised, model-agnostic characterization of temporal dataset shifts using the Information-Geometric Temporal (IGT) projection is an early indicator of potential AI performance variations before model development.

Methods:

Using the real-world MIMIC-IV EHR database, encompassing data from over 40,000 patients from 2008 to 2019, we characterized its inherent dataset shift patterns through an unsupervised approach using Information Geometric Temporal (IGT) projections and Data Temporal Heatmaps. We trained and evaluated annually a set of Random Forests and Gradient Boosting models to predict in-hospital mortality. To assess the impact of shifts on model performance, we checked the association between the temporal clusters found in both IGT projections and the inter-time embedding of model performances using Chi-squared test.

Results:

Our results demonstrate a significant relationship between the unsupervised temporal shift patterns identified using the IGT projection method and the performance of the Random Forest and Gradient Boosting models (p < .05). The transition from ICD-9 to ICD-10 was a significant source of dataset shifts, which impacted models’ performance.

Conclusions:

Unsupervised, model-agnostic characterization of temporal shifts via IGT projections can serve as a proactive monitoring tool to anticipate performance drift in clinical AI models. By incorporating early shift detection into the development pipeline, we can enhance decision-making during the training and maintenance of these models. This approach paves the way for more robust, trustworthy, and self-adapting AI systems in healthcare.