Interpretable machine-learning prognosis of mycetoma from routine clinical data

Abstract

BACKGROUND: Mycetoma is a chronic tropical infection with limited tools for predicting treatment outcomes. METHODS: Routinely collected demographic, clinical, imaging, etiological, treatment, and follow-up data from 1,084 patients at the Mycetoma Research Centre were used to evaluate supervised machine learning models for outcome prediction. Binary and three-class tasks (cured, recurrence, disability) were assessed using stratified five-fold cross-validation with preprocessing, imputation, scaling, and class balancing. Logistic regression, support vector machine, and random forest models were compared. RESULTS: Treatment mode, treatment duration, disease duration at presentation, adherence, lesion size, and imaging findings were the most important predictors. In eumycetoma, random forest achieved the best binary performance (accuracy 0.760 ± 0.036; ROC-AUC 0.835 ± 0.036). In actinomycetoma, random forest achieved accuracy of 0.719 ± 0.115 and ROC-AUC of 0.779 ± 0.138. For the three-class task, random forest performed best in eumycetoma (accuracy 0.612 ± 0.033; macro-AUC 0.777 ± 0.029), while logistic regression performed best in actinomycetoma (accuracy 0.735 ± 0.097; macro-AUC 0.713 ± 0.092). CONCLUSIONS: Standard clinical and imaging variables can support machine learning-based risk stratification in mycetoma, potentially improving early management and follow-up. External validation and clinical utility assessment are still required.