Explainable machine learning model to predict refeeding hypophosphatemia

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Prediction Models

Background:

• Refeeding syndrome (RFS) occurs when feeding resumes, altering metabolism from catabolic to anabolic states.
• RFS can range from asymptomatic to fatal, making early detection challenging.
• Explainable machine learning offers potential for improved RFS diagnosis and treatment.

Purpose of the Study:

• To develop and evaluate a machine learning model for predicting refeeding hypophosphatemia, a key indicator of RFS.
• To compare the performance of the machine learning model against conventional regression techniques.
• To interpret the machine learning model's predictions using Shapley additive explanations (SHAP) values.

Main Methods:

• A retrospective study of 806 patients with prolonged nothing-by-mouth orders was conducted.
• Patients were classified into hypophosphatemia (n=367) and non-hypophosphatemia (n=439) groups based on phosphate levels.
• An extreme gradient boosting (XGBoost) model was developed and compared with logistic, Lasso, and ridge regression models using Area Under the Receiver Operating Characteristic Curve (AUC).

Main Results:

• The XGBoost model achieved an AUC of 0.950, significantly outperforming logistic (0.760), Lasso (0.751), and ridge (0.758) regression models.
• Key predictors for refeeding hypophosphatemia identified by SHAP values included low initial phosphate, recent weight loss, high creatinine, and diabetes mellitus with insulin use.
• Other significant predictors included low hemoglobin A1c, furosemide use, ICU admission, specific blood urea nitrogen levels, parenteral nutrition, abnormal magnesium levels, low potassium, and older age.

Conclusions:

• Machine learning, specifically XGBoost, offers superior effectiveness in predicting RFS compared to conventional regression methods.
• An accurate, explainable machine learning tool can facilitate early identification of patients at risk for RFS.
• Implementing such a tool can guide nutrition management and monitoring in ICUs, potentially reducing RFS-related morbidity and mortality.