This research introduces a novel approach to medical forecasting by extracting and analyzing time-series data from clinical text reports rather than structured data alone.

• Establishes a framework for predicting clinical outcomes from timestamped textual findings
• Evaluates performance of fine-tuned decoder-based LLMs versus encoder-based models
• Demonstrates an LLM-assisted annotation pipeline for extracting clinical findings
• Applies the approach to critical forecasting tasks including event prediction and survival analysis

This advancement matters because it unlocks the rich temporal information in narrative clinical documentation that traditional structured data approaches miss, potentially improving clinical decision support and patient outcome prediction.

Forecasting from Clinical Textual Time Series: Adaptations of the Encoder and Decoder Language Model Families