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

Area of Science:

• Medical informatics
• Epidemiology
• Machine learning in healthcare

Background:

• Invasive mold infections (IMI) cause significant morbidity and mortality.
• Current public health surveillance for IMI is insufficient.
• Predictive models can aid surveillance and inform case definitions.

Purpose of the Study:

• To develop and validate machine learning models for predicting Invasive Mold Infections (IMI) case status.
• To identify key factors predictive of IMI using medical record data.
• To inform the development of robust public health surveillance for IMI.

Main Methods:

• Utilized least absolute shrinkage and selection operator (LASSO) and random forest machine learning algorithms.
• Modeled medical record data from four Houston medical centers (September 2016 - August 2018).
• Assessed feature importance using leave-one-covariate-out and permutation methods.

Main Results:

• Systemic antifungal medication, IMI-related billing codes, and positive pulmonary histopathology were top predictors.
• Removing these key features reduced model prediction accuracy by 3.6% to 7.6%.
• Certain traditional IMI risk factors, like cancer and corticosteroid use, paradoxically worsened prediction accuracy.

Conclusions:

• Machine learning identified key predictors of IMI that differ from traditional risk factors.
• These findings support the development of feasible prediction models for IMI public health surveillance.
• The study highlights the utility of machine learning in enhancing epidemiological surveillance for infectious diseases.