Expert opinion elicitation for assisting deep learning based Lyme disease classifier with patient data

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Dermatology

Background:

• Erythema migrans (EM) is the primary early symptom of Lyme disease.
• Deep learning for EM diagnosis is limited by a lack of comprehensive patient data.
• Clinical diagnosis relies on both lesion images and patient history.

Purpose of the Study:

• To develop a method for calculating Lyme disease probability using patient data.
• To enhance deep learning models for EM diagnosis by incorporating clinical insights.
• To address data scarcity challenges in medical AI.

Main Methods:

• Expert elicitation from fifteen doctors using a structured questionnaire.
• Conversion of doctor evaluations to probability scores via Gaussian mixture density estimation.
• Validation and explanation using formal concept analysis and decision trees.
• Development of an algorithm to merge multi-modal probability estimates.

Main Results:

• Successfully elicited expert opinions to build a patient data-based EM probability scoring model.
• Demonstrated the feasibility of quantifying clinical expertise for AI integration.

Conclusions:

• The elicited probability scores and algorithm can improve the robustness of image-based deep learning Lyme disease pre-scanners.
• The proposed method offers a practical approach for medical diagnosis problems with limited patient data.
• This expert elicitation framework is user-friendly for clinicians.