Knowledge-Driven Online Multimodal Automated Phenotyping System.

Area of Science:

• Biomedical Informatics
• Clinical Data Science
• Artificial Intelligence in Healthcare

Background:

• Electronic health record (EHR) systems offer vast clinical data but face challenges in phenotyping due to inaccurate records and feature relevance.
• Current EHR-based phenotyping often requires human-labeled training sets, limiting scalability and efficiency.

Purpose of the Study:

• To introduce the knowledge-driven online multimodal automated phenotyping (KOMAP) system.
• To enable efficient and accurate phenotyping from EHR data using summary information and automated feature selection.

Main Methods:

• Developed the KOMAP system integrating an online narrative and codified feature search engine (ONCE).
• Utilized composite knowledge from EHRs, online articles, and large language models for feature generation.
• Trained multimodal phenotyping algorithms using summary data, bypassing the need for patient-level data and gold-standard labels.

Main Results:

• Features selected by ONCE demonstrated high concordance with state-of-the-art AI models (GPT4, ChatGPT).
• KOMAP generated efficient phenotyping algorithms with robust performance, validated across four healthcare centers.
• The system successfully reduced the feature set size while maintaining high relevance for large-scale phenotyping.

Conclusions:

• KOMAP offers a significant advancement for large-scale, multi-center phenotyping by overcoming limitations of traditional EHR-based methods.
• The fully online nature of KOMAP facilitates collaboration and reduces the dependency on manual data labeling and patient-level data access.