Forecasting Clinical Risk from Textual Time Series: Structuring Narratives for Temporal AI in Healthcare.

Area of Science:

• Artificial Intelligence
• Clinical Informatics
• Natural Language Processing

Background:

• Clinical case reports contain valuable temporal patient data.
• Traditional machine learning methods often fail to leverage this temporal information effectively.
• Large language models (LLMs) offer new possibilities for analyzing unstructured clinical text.

Purpose of the Study:

• To introduce and address the forecasting problem using textual time series from clinical notes.
• To evaluate various machine learning models for predicting patient trajectories.
• To compare the performance of different model architectures on distinct temporal prediction tasks.

Main Methods:

• Extracted timestamped clinical findings using an LLM-assisted annotation pipeline.
• Developed a textual time series dataset from clinical notes.
• Systematically evaluated encoder-based transformers and decoder-based LLMs on event prediction, temporal ordering, and survival analysis tasks.

Main Results:

• Encoder-based models outperformed others in event occurrence prediction and temporal ordering tasks (higher F1 scores, better temporal concordance).
• Fine-tuned masking approaches improved ranking performance.
• Instruction-tuned decoder models showed advantages in early prognosis within survival analysis.

Conclusions:

• Time ordering of clinical text is crucial for accurate temporal predictions, outperforming simple text ordering.
• LLMs can be effectively adapted for clinical time series analysis, enhancing patient trajectory forecasting.
• This work underscores the value of time-ordered clinical data for advancing AI in healthcare.