Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Language Models for Longitudinal Clinical Prediction

Tananun Songdechakraiwut, Michael Lutz

TL;DR

The paper tackles forecasting longitudinal neuropsychological trajectories in Alzheimer's disease from sparsely observed data. It proposes a data-efficient framework that keeps a large language model frozen while introducing prompt-based conditioning, temporal patch embeddings, and cross-modal semantic mapping to produce multi-target forecasts over horizons from $12$ to $48$ months, trained with a lightweight decoder. Key contributions include demonstrating strong predictive accuracy across four clinically relevant targets on the ADNI dataset, robustness under limited supervision (down to $1\%$ of training data), and an ablation-enabled understanding of how RevIN normalization, prompts, and semantic prototypes drive performance. The approach offers a scalable, deployment-friendly pathway for early-stage Alzheimer's monitoring and personalized planning with minimal annotated data, potentially informing clinical decisions and trial design.

Abstract

We explore a lightweight framework that adapts frozen large language models to analyze longitudinal clinical data. The approach integrates patient history and context within the language model space to generate accurate forecasts without model fine-tuning. Applied to neuropsychological assessments, it achieves accurate and reliable performance even with minimal training data, showing promise for early-stage Alzheimer's monitoring.

Language Models for Longitudinal Clinical Prediction

TL;DR

The paper tackles forecasting longitudinal neuropsychological trajectories in Alzheimer's disease from sparsely observed data. It proposes a data-efficient framework that keeps a large language model frozen while introducing prompt-based conditioning, temporal patch embeddings, and cross-modal semantic mapping to produce multi-target forecasts over horizons from to months, trained with a lightweight decoder. Key contributions include demonstrating strong predictive accuracy across four clinically relevant targets on the ADNI dataset, robustness under limited supervision (down to of training data), and an ablation-enabled understanding of how RevIN normalization, prompts, and semantic prototypes drive performance. The approach offers a scalable, deployment-friendly pathway for early-stage Alzheimer's monitoring and personalized planning with minimal annotated data, potentially informing clinical decisions and trial design.

Abstract

We explore a lightweight framework that adapts frozen large language models to analyze longitudinal clinical data. The approach integrates patient history and context within the language model space to generate accurate forecasts without model fine-tuning. Applied to neuropsychological assessments, it achieves accurate and reliable performance even with minimal training data, showing promise for early-stage Alzheimer's monitoring.

Paper Structure

This paper contains 19 sections, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Longitudinal LLM Framework
  3. Prompt-based clinical conditioning
  4. Temporal embedding of clinical visits
  5. Semantic mapping of temporal patterns
  6. LLM integration and forecasting
  7. Efficient training and clinical adaptability
  8. Forecasting Alzheimer's Progression
  9. Results and clinical interpretation
  10. Ablation study.
  11. Algorithmic Implementation of Longitudinal LLM
  12. Model Configurations
  13. Experimental environment.
  14. ADNI Dataset and Preprocessing
  15. Longitudinal features.
  16. ...and 4 more sections

Figures (2)

  • Figure 1: Schematic of a longitudinal LLM framework.
  • Figure 2: Clinical prompt example.