(P26-005-23) To Quantitatively Visualize Clinically Relevant Non-linear Function of Malnutrition Risk Factors Using AI Model of Explainable Boosting Machine

Objectives: Malnutrition is strongly associated with poorer hospital outcomes. However, no current definition of malnutrition identifies, with adequate sensitivity, patients who will respond to nutrition interventions. This retrospective study aims to reveal non-linear complex functions between clinical characteristics and impending malnutrition risk, using interpretable artificial intelligence technique of Explainable Boosting Machine (EBM). We hypothesize that quantifying the non-linear complex function of clinical signs indicative of malnutrition might shed light on the ultimate goal of better identification of patients who are likely to respond to nutrition support (i.e., those truly malnourished).

Methods: Electronic medical record (EMR) data for adult inpatients discharged from New York-Presbyterian Hospital between January 1, 2021 to March 1, 2022 was extracted. Those diagnosed with malnutrition, based on dietitians’ nutrition diagnosis notes and following published guidelines, were identified. The cohort was randomly split (75:25) into training and test sets. EBM was developed and cross validated on the training data, and the non-linear function of malnutrition risk factors was derived from EBM.

Results: In all, 6.30% of total 183,678 discharges were associated with malnutrition. EBM provided 0.834±0.015 area under the receiver operating curve (AUC) and 0.829±0.014 area under precision-recall curve (AUPRC) to identify malnutrition. The model also created quantitative and clinically meaningful visualization of non-linear function of risk factors such as age, BMI, albumin, hemoglobin, BUN/creatine ratio, platelet, and vital signs.

Conclusions: We have quantified the complex non-linear function of malnutrition risk factors using machine learning AI method of EBM. This analysis is a step further toward our development of a novel machine learning AI algorithm to predict response to nutrition intervention in a large cohort.

Funding Sources: None