Plant disease epidemiologists often work with datasets smaller than ideal for data-hungry machine learning (ML) algorithms, thereby risking overfitting. We demonstrate how an interpretation-guided modeling approach, leveraging complex ML primarily for insight generation, can overcome this challenge, using white mold (caused by Sclerotinia sclerotiorum) in snap beans (Phaseolus vulgaris) as a case study. An observational dataset of white mold prevalence across 356 commercial snap bean fields in central and western New York State (2006-2008) was augmented by merging georeferenced observations with POLARIS soils data and engineered features from downscaled ERA5-Land environmental data. Functional data analysis identified weather periods associated with white mold risk, and random forests (RF), used interpretatively, identified key predictors. While RF models showed high apparent performance, they exhibited significant overfitting and poor calibration. Insights from RF interpretation (via SHapley Additive exPlanations analysis) guided the development of a simpler, four-predictor logistic regression model using restricted cubic splines. This simpler model was better calibrated and had acceptable discrimination (internally-validated C-statistic = 0.77). For smaller epidemiological datasets, our results advocate using ML primarily as an interpretive tool to guide the development of simpler, less data-intensive, yet robust predictive models better suited for practical disease management decisions.
Alves et al. (Sat,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: