Spatial Determinants of Pathogenic Leptospira Viability in Paddy Field Surface Waters: Mechanistic Insights from Machine Learning and Geographically Weighted Logistic Regression

Abstract

Pathogenic Leptospira spp. persist in surface waters, forming environmental reservoirs that sustain leptospirosis transmission in agricultural landscapes. However, the physicochemical controls governing environmental viability and their spatial variability remain poorly resolved. This study integrated controlled laboratory experiments with machine learning and geographically weighted logistic regression (GWLR) to quantify determinants of Leptospira viability in paddy field surface waters from northeastern Thailand. Surface water samples (n = 126) were analyzed for pH, calcium, magnesium, and iron, and Leptospira viability was experimentally assessed over 20 days. Median viability across days 5–20 was used to classify samples into high- and low-persistence outcomes. Machine learning classifiers achieved moderate discrimination in independent testing (AUC = 0.64–0.70), with a histogram-based gradient boosting machine (HistGBM) providing the strongest performance (AUC = 0.70, 95% CI: 0.493–0.888). SHAP analysis identified pH as the dominant predictor, followed by magnesium and calcium, with iron contributing more weakly. GWLR revealed significant spatial nonstationarity in physicochemical effects and identified a characteristic interaction scale of approximately 2700–3000 m. Together, these results demonstrate that Leptospira persistence is governed by spatially heterogeneous water chemistry gradients and highlight the value of integrating interpretable machine learning with spatially adaptive regression for environmental pathogen risk assessment.