JMIR Publications

ABSTRACT

Background:

Skin Neglected Tropical Diseases (skin NTDs) are the most prevalent diseases worldwide, affecting people living in resource-limited areas with low healthcare services and trained professionals. While machine learning (ML)-based diagnostic tools can be used for initial clinical assessment and patient screening, especially in resource-limited areas (including in Ethiopia), little efforts have been done in the area.

Objective:

This study, thus, proposes a skin NTDs classification model using clinical patient metadata.

Methods:

NTDs diagnostic data, collected from the affected areas in the Southwest of Ethiopia, is acquired and used in this study. The initial dataset (IDS), and 3 additional versions created by applying manual preprocessing, random oversampling, and SMOTE methods on IDS are used to train 8 selected models, comparatively analyze performances, and finally identify the best-performing model. The model trainings on the newly created datasets have led the models to overfit, given the smaller size of the dataset.

Results:

While the hold-out method on all forms of the dataset leads to overfitting, the use of K-fold method on IDS allowed the models to achieve reliable results, with SVM (linear) outperforming the other models with overall accuracy of 76.92% and balanced accuracy 58.87%, followed by random forest (RF) with 74.25% and 54.73%. Comparatively, the RF model performs consistently higher across all trainings, hence, identified as the best-performing model, as it achieves 76.25%, 59.75%, 0.64, 0.70 in overall accuracy, balanced accuracy, F1-score, and G-mean, respectively after hyperparameter tuning.

Conclusions:

Overall, this study has been highly challenged by data scarcity and class imbalances, hence, suggests future studies to confirm the results on larger datasets having all the classes of the skin NTDs.