Using Machine Learning Algorithms to Identify Key Predictors of Invasive Mold Infection Surveillance

Background Invasive mold infections (IMI) can lead to severe morbidity and mortality, but routine public health surveillance is lacking. Although extensive evaluation is needed for clinical diagnosis, case classification prediction models may inform surveillance efforts, which are essential to better characterize epidemiologic trends and assess the value of a more inclusive IMI case definition. Methods We modeled medical record data of potential IMI cases from 4 medical centers in Houston, Texas, during September 2016 to August 2018. We used least absolute shrinkage and selection operator and random forest machine learning methods to identify key host and clinical factors, mycological evidence, diagnostics, and health care exposures predictive of IMI case versus noncase status using both conventional and novel definitions. We assessed feature importance by measuring each variable's impact on prediction error using leave-one-covariate-out and permutation feature importance approaches. Results Receipt of systemic antifungal medication, hospital billing codes related to IMI, and positive pulmonary histopathology results were identified as the most important predictors of IMI case status across all measures. Removal of these features from the models resulted in reductions to prediction accuracy ranging from 3.6% (95% confidence interval [CI], 3.2%–3.2%) to 7.6% (95% CI, 7.2%–8.0%). Some IMI risk factors, including cancer diagnosis and prolonged receipt of corticosteroid medications, worsened prediction in several assessments of feature importance. Conclusions Features identified as important predictors of IMI case status using machine learning methods deviated from classic IMI risk factors. Our results will inform robust and feasible IMI case prediction models for public health surveillance.