TIMER: temporal instruction modeling and evaluation for longitudinal clinical records

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics

Background:

• Electronic health records (EHRs) offer valuable longitudinal patient data.
• Current large language models (LLMs) face challenges in temporal reasoning across multi-visit EHRs.

Purpose of the Study:

• To introduce TIMER (Temporal Instruction Modeling and Evaluation for Longitudinal Clinical Records), a novel method for enhancing LLMs' temporal reasoning capabilities.
• To improve the ability of LLMs to process and interpret patient timelines within EHRs.

Main Methods:

• TIMER employs time-aware instruction tuning to ground LLMs in patient-specific temporal contexts.
• Instruction-response pairs are linked to specific timestamps to ensure temporal fidelity during training.

Main Results:

• TIMER-tuned models demonstrated a 6.6% improvement in completeness compared to conventional medical instruction-tuned models on clinician-curated benchmarks.
• Distribution-matched training with TIMER showed up to a 6.5% advantage in temporal reasoning.
• Qualitative analysis confirmed enhanced temporal boundary adherence, trend detection, and chronological precision.

Conclusions:

• TIMER provides a methodological foundation for developing LLMs adept at handling longitudinal EHR data.
• This approach is crucial for advancing applications like disease trajectory modeling and treatment response monitoring.