CHiLL: Zero-shot Custom Interpretable Feature Extraction from Clinical Notes with Large Language Models

Area of Science:

• Artificial Intelligence
• Clinical Informatics
• Machine Learning

Background:

• Electronic Health Records (EHR) contain vast amounts of data.
• Extracting clinically meaningful features from EHR is challenging.
• Physician expertise is crucial for effective risk prediction.

Purpose of the Study:

• To introduce CHiLL (Crafting High-Level Latents), a novel approach for natural-language specification of features for linear models.
• To enable physicians to leverage their domain knowledge for feature engineering from EHR data.
• To improve interpretability and performance in clinical predictive modeling.

Main Methods:

• CHiLL prompts large language models (LLMs) with expert-crafted queries to generate features from health records.
• The generated features are used to train simple linear classifiers.
• The approach was evaluated using MIMIC-III and MIMIC-CXR datasets for tasks like 30-day readmission prediction.

Main Results:

• Linear models trained with CHiLL-generated features demonstrated performance comparable to models using reference features.
• CHiLL provided greater interpretability compared to linear models using "Bag-of-Words" features.
• Learned feature weights showed strong alignment with clinical expectations.

Conclusions:

• CHiLL offers an effective method for generating interpretable features from EHR data using LLMs.
• The approach empowers clinicians by integrating their expertise into the feature engineering process.
• CHiLL shows promise for enhancing clinical risk prediction models with improved interpretability and performance.