A GPT-based EHR modeling system for unsupervised novel disease detection

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Epidemiology

Background:

• Early detection of novel diseases and emerging outbreaks is crucial for public health.
• Physician-led anomaly detection can be enhanced by advanced computational tools.
• Electronic Health Records (EHR) contain vast data for disease pattern recognition.

Purpose of the Study:

• To develop an Artificial Intelligence (AI)-based anomaly detection model.
• To complement physician expertise in identifying novel disease cases within hospitals.
• To prevent emerging infectious disease outbreaks through early detection.

Main Methods:

• A Generative Pre-trained Transformer (GPT)-based clinical anomaly detection system was developed.
• The system modeled hospitalized patients' Electronic Health Records (EHR) using Empirical Risk Minimization (ERM).
• Methods inspired by Large Language Models (LLMs) were used to compute Out-Of-Distribution (OOD) anomaly scores.

Main Results:

• The GPT model predicted Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) hospitalizations with 92.2% Area Under the ROC Curve (AUC) in an unsupervised setting.
• Individual patient anomaly detection and mortality prediction achieved AUCs of 78.3% and 94.7%, outperforming linear models.
• The model captured diverse SARS-CoV-2 clinical trajectories and offered interpretable detections.

Conclusions:

• A GPT model can accurately detect emerging outbreaks within hospitals by analyzing EHR time sequences.
• The AI system can identify anomalous patient cases where outcomes deviate from model predictions.
• The GPT model's ability to forecast clinical variables aids in personalized treatment planning.