Temporal Learning with Dynamic Range (TLDR) for Modeling Recurrent Exposure and Treatment Outcomes

Area of Science:

• Machine Learning
• Medical Informatics
• Computational Biology

Background:

• Standard machine learning (ML) models often neglect the temporal sequence of clinical events in electronic health records (EHRs).
• This oversight limits the predictive accuracy of ML models in outcomes research.
• Accurate prediction requires capturing the dynamic nature of patient health trajectories.

Purpose of the Study:

• To introduce Temporal Learning with Dynamic Range (TLDR), a novel time-sensitive ML framework.
• To identify risk factors for post-acute sequelae of SARS-CoV-2 infection (PASC) using longitudinal EHR data.
• To compare the performance of TLDR against conventional atemporal ML models.

Main Methods:

• Utilized longitudinal EHR data from over 85,000 patients in the Precision PASC Research Cohort (P2RC).
• Developed and implemented the Temporal Learning with Dynamic Range (TLDR) framework.
• Compared TLDR's predictive performance against a benchmark atemporal ML model.

Main Results:

• TLDR demonstrated superior predictive performance with a 18.4% improvement in AUROC (0.791 vs. 0.668) and a 40.14% increase in PRAUC (0.590 vs. 0.421).
• The framework showed improved generalizability with a lower mean overfitting index (-0.028), indicating robustness.
• TLDR's time-stamped features enhanced interpretability, providing more precise patient record characterization.

Conclusions:

• TLDR effectively captures exposure-outcome associations and offers flexible time-stamping strategies for clinical research.
• The framework provides a simple yet effective approach for integrating dynamic temporal windows into predictive modeling.
• TLDR is available in the MLHO R package to support exploration of recurrent patterns in clinical settings.