Development of machine learning-based mpox surveillance models in a learning health system

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Epidemiological Surveillance

Background:

• Mpox (monkeypox) surveillance is crucial for public health.
• Developing automated methods for early detection can enhance response efforts.
• Clinical notes contain valuable information for identifying infectious disease cases.

Purpose of the Study:

• To develop and evaluate machine learning (ML) and deep learning (DL) models for detecting mpox cases using clinical notes.
• To compare the performance of different ML/DL models in identifying mpox for surveillance.

Main Methods:

• Retrospective study of clinical encounters at a major medical center.
• Trained three models: LASSO regression, ClinicalBERT, and ClinicalLongformer.
• Evaluated models using precision, recall, F1 score, AUROC, AUPRC, and RP80.

Main Results:

• LASSO regression demonstrated superior performance over DL models.
• LASSO achieved a precision, recall, and F1 score of 0.93, with an AUROC of 0.97.
• Symptom-related phrases were key predictors in the LASSO model.

Conclusions:

• ML and DL models show promise for mpox case detection and surveillance.
• LASSO regression proved effective in minimizing false positives, outperforming DL models.
• These computational methods can support infectious disease surveillance and quality improvement.