Dementia risk prediction using decision-focused content selection from medical notes.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• General-purpose language models (LMs) show promise in text analysis but require adaptation for the medical domain.
• Electronic Health Records (EHRs) contain valuable patient history in medical notes, but LM input length limitations necessitate content selection.
• Existing methods like truncation or summarization risk information loss, redundancy, or irrelevance.

Purpose of the Study:

• To propose a novel decision-focused content selection technique for medical notes.
• To develop a dementia risk prediction model utilizing this technique and the Longformer LM architecture.
• To evaluate the model's performance against existing content selection methods.

Main Methods:

• A decision-focused content selection methodology was developed to identify sentences relevant to a target outcome within a specific observation period.
• The selected content was used to train a dementia risk prediction model based on the Longformer LM architecture.
• The model's performance was assessed using Area Under the Curve (AUC) metrics.

Main Results:

• The proposed framework achieved an AUC of 78.43 with a summary restricted to 1024 tokens.
• This performance surpassed previously proposed content selection techniques.
• The model accurately estimates dementia risk with a one-year prediction horizon using only one year of medical notes.

Conclusions:

• The decision-focused content selection technique effectively addresses the challenges of processing long medical notes for LMs.
• The developed dementia risk prediction model demonstrates superior performance and preserves contextual information, unlike tabular methods.
• This approach circumvents LM computational complexity limitations and enables feature engineering from raw text.