Temporal Learning with Dynamic Range (TLDR) for modeling recurrent exposure and treatment outcomes.

Area of Science:

• Computational biology and bioinformatics
• Machine learning in healthcare
• Outcomes research and clinical informatics

Background:

• Standard machine learning (ML) models often neglect the temporal nature of clinical events in electronic health records (EHRs).
• This oversight limits the predictive accuracy of models used in outcomes research.
• Accurate prediction of post-acute sequelae of SARS-CoV-2 infection (PASC) requires consideration of temporal event sequencing.

Purpose of the Study:

• To introduce Temporal Learning with Dynamic Range (TLDR), a novel time-sensitive ML framework.
• To identify risk factors for PASC by leveraging longitudinal EHR data.
• To compare the predictive performance and generalizability 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 applied the TLDR framework, a time-sensitive ML approach.
• Compared TLDR's performance against a standard atemporal ML model using metrics like AUROC and PRAUC.

Main Results:

• TLDR achieved a mean AUROC of 0.791, an 18.4% improvement over the benchmark (0.668).
• TLDR's mean PRAUC of 0.590 significantly outperformed the benchmark's 0.421 (40.14% increase).
• The framework demonstrated improved generalizability with a lower mean overfitting index (-0.028) and enhanced interpretability through time-stamped features.

Conclusions:

• TLDR offers a robust and interpretable method for integrating temporal dynamics into predictive modeling for PASC risk.
• The framework effectively captures exposure-outcome associations and provides flexibility in time-stamping strategies.
• TLDR, available in the MLHO R package, supports the exploration of recurrent patterns in clinical settings.