Artificial Intelligence-Based Hospital Malnutrition Screening: Validation of a Novel Machine Learning Model.

Abstract

Background: Despite its morbidity, mortality, and financial burden, in-hospital malnutrition remains underdiagnosed and undertreated. Artificial intelligence offers a promising clinical informatics solution for identifying malnutrition risk and one that can be coupled with clinician-delivered patient care.

Objective: The objective of the study was to evaluate an artificial intelligence-based hospital malnutrition screening model in a large and diverse inpatient population and to compare it to the currently used clinician-delivered malnutrition screening tool.

Methods: We studied the performance of a gradient-boosted decision tree model incorporating a large language model (LLM) for feature extraction using the electronic medical record data of 106,449 patients over 3.75 years.

Results: The model's area under the receiver operating curve was 0.92 (95% CI: 0.91-0.92) on the first day of hospitalization and rose to 0.95 (95% CI: 0.95-0.96) using the maximum risk predicted for each patient throughout hospitalization, indexed against discharge-coded malnutrition. Similar results were observed when indexed against dietitian-recorded malnutrition. The model outperformed a nurse-administered, modified version of the Malnutrition Screening Tool (MST) and patients identified by the model had higher likelihoods of readmission and death compared to patients identified by the nurse-administered screener.

Conclusion: Our study findings provide validation for a novel model's use in the prediction of in-hospital malnutrition.