Interpretable machine-learning prognosis of mycetoma from routine clinical data

Area of Science:

• Medical Informatics
• Infectious Diseases
• Machine Learning

Background:

• Mycetoma is a neglected tropical infection characterized by chronicity.
• Current tools for predicting mycetoma treatment outcomes are limited, hindering effective patient management.

Purpose of the Study:

• To evaluate supervised machine learning models for predicting treatment outcomes in mycetoma patients.
• To identify key predictors of treatment success, recurrence, and disability.

Main Methods:

• Utilized routinely collected data from 1,084 mycetoma patients.
• Compared logistic regression, support vector machine, and random forest models for binary and three-class outcome prediction.
• Employed stratified five-fold cross-validation with data preprocessing, imputation, scaling, and class balancing.

Main Results:

• Key predictors identified include treatment mode, duration, disease duration, adherence, lesion size, and imaging findings.
• Random forest model demonstrated superior performance in predicting binary outcomes for eumycetoma and actinomycetoma.
• For the three-class task, random forest excelled in eumycetoma, while logistic regression was optimal for actinomycetoma.

Conclusions:

• Clinical and imaging variables are valuable for machine learning-based risk stratification in mycetoma.
• These models show potential for improving early management and follow-up strategies.
• Further external validation and clinical utility assessments are necessary.