Use of Machine Learning Models to Predict Microaspiration Measured by Tracheal Pepsin A.

Abstract

Background: Enteral feeding intolerance, a common type of gastrointestinal dysfunction leading to underfeeding, is associated with increased mortality. Tracheal pepsin A, an indicator of microaspiration, was found in 39% of patients within 24 hours of enteral feeding. Tracheal pepsin A is a potential biomarker of enteral feeding intolerance.

Objective: To identify predictors of microaspiration (tracheal or oral pepsin A). It was hypothesized that variables predicting the presence of tracheal pepsin A might be similar to predictors of enteral feeding intolerance.

Methods: In this secondary analysis, machine learning models were fit for 283 adults receiving mechanical ventilation who had tracheal and oral aspirates obtained every 12 hours for up to 14 days. Pepsin A levels were measured using the proteolytic enzyme assay method, and values of 6.25 ng/mL or higher were classified as indicating microaspiration. Demographics, comorbidities, and variables associated with enteral feeding were analyzed with 3 machine learning models-random forest, XGBoost, and support vector machines with recursive feature elimination-using 5-fold cross-validation tuning.

Results: Random forest for tracheal pepsin A was the best-performing model (area under the curve, 0.844 [95% CI, 0.792-0.897]; accuracy, 87.55%). The top 20 predictors of tracheal pepsin A were identified.

Conclusion: Four predictor variables for tracheal pepsin A (microaspiration) are also reported predictors of enteral feeding intolerance, supporting the exploration of tracheal pepsin A as a potential biomarker of enteral feeding intolerance. Identification of predictor variables using machine learning models may facilitate treatment of patients at risk for enteral feeding intolerance.