Feature importance: Opening a soil-transmitted helminth machine learning model via SHAP.

Area of Science:

• Epidemiology
• Geospatial Health
• Machine Learning

Background:

• Hookworm infection is a significant public health issue in Ethiopia, ranking third highest in Sub-Saharan Africa.
• Landscape epidemiology utilizes machine learning for modeling disease risk scenarios.

Purpose of the Study:

• To demystify machine learning (ML) "black box" models in geospatial health.
• To determine variable contributions in ML models for hookworm prevalence in Ethiopia using SHAP.
• To compare ML models with classical statistical models.

Main Methods:

• Employed XGBoost, a popular machine learning model, for data analysis.
• Utilized the Python SHAP library to interpret variable importance and interactions within the ML model.
• Applied various plotting methods to visualize variable contributions to predictions.

Main Results:

• Machine learning models demonstrated superior performance compared to classical statistical models.
• SHAP analysis successfully explained the influence and interactions of variables in predicting hookworm prevalence.
• Identified key variables contributing to hookworm distribution in the study area.

Conclusions:

• Machine learning, enhanced by SHAP, offers a powerful and interpretable approach to geospatial health modeling.
• The findings provide valuable insights into Ethiopia's hookworm epidemiology.
• This methodology serves as a valuable tool for future epidemiological studies.