From Logistic Regression to Foundation Models: Factors Associated With Improved Forecasts

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Epidemiology

Background:

• Electronic health records (EHR) are crucial for chronic disease risk modeling, impacting screening and resource allocation.
• Clinical utility of these models depends on calibration, transportability, and decision utility (net benefit).

Purpose of the Study:

• To comparatively evaluate classical regression, gradient-boosted decision trees (GBDTs), deep neural networks (DNNs), and foundation backbones for chronic disease risk prediction.
• To assess model performance regarding calibration, transportability, and decision utility.

Main Methods:

• A narrative synthesis of comparative studies published between January 2019 and October 2025.
• Appraisal of logistic regression, GBDT, DNN, and foundation backbone models using metrics like Brier score, calibration error (slope, intercept), and decision curves (net benefit).

Main Results:

• Modern tree-based methods (GBDTs) often outperformed logistic regression in Brier scores and external calibration error.
• Logistic regression demonstrated superior calibration slope stability under temporal drift.
• Deep neural networks (DNNs) tended to underestimate risk in high-risk groups; foundation backbones required recalibration for improved performance.
• Recalibration was essential for models to achieve a net benefit, particularly when expected calibration error (ECE) was maintained below 0.03.

Conclusions:

• While advanced models offer potential, logistic regression shows robustness in calibration slope.
• Effective recalibration strategies are critical for enhancing the clinical utility and decision-making of risk prediction models.
• Operational acceptance criteria should integrate calibration slope (0.90-1.10), performance thresholds, and monitoring schedules.