P066 - RESEARCH ON THE INTERPRETABILITY OF AN ENSEMBLE LEARNING-BASED MODEL FOR PREDICTING ALBUMIN LEVELS IN ELDERLY INPATIENTS

P066

RESEARCH ON THE INTERPRETABILITY OF AN ENSEMBLE LEARNING-BASED MODEL FOR PREDICTING ALBUMIN LEVELS IN ELDERLY INPATIENTS

J. Du¹, J. HU ², F. Teng¹, N. Lin^2,*

¹School of Computer and Artificial Intelligence, Southwest Jiaotong University, ²Department of Clinical Nutrition, The General Hospital of Western Theater Command, Chengdu, China

Rationale: Albumin level is an important indicator of blood protein status, However, less is known about how to choose artificial nutrition (AN) intervention based on clinical indexes for improving albumin in ederly inpatients.

Methods: Due to lack of albumin level prediction dataset, a complete AN intervention dataset was constructed including the elderly inpatients data from January 1st, 2016 to December 31st, 2021 in a third tertiary referral Hospital. Demographic information, blood test indicators, nutritional medication, and other medication were collected. Enteral nutrition included all types of formula used both oral and tube feeding, and parenteral nutrition included fat emulsions, amino acid and glucose preparations. Based on the timestamp of the intervention, the average blood test data within 72 hours before the timestamp was taken as the feature, and the average albumin level within 72 hours after the timestamp was taken as the intervention results. Encoding variables as binary dependent variable, the TPE-XGBoost model was finally used after comparing with TabTransformer, CatBoost, LightGBM, XgBoost and the model interpretation algorithm SHAP was combined to select features and enhance the interpretability of the model from the global and local levels.

Results: Among all 22 features, the model selected hemoglobin concentration (MCHC), aspartate aminotransferase (AST), total bilirubin (TBIL), HsCPR according to ranking of feature importance. The prediction accuracy was 87.8% and was AUC 87.47%. In addition, the interpretable analysis of SHAP obtained above 4 key factors and the threshold effects of TBIL and HsCPR could be mutually verified with clinical studies.

Conclusion: The TPE-XGBoost model revealed that TBIL and HsCPR could be reference indexes for clinicians to chose different AN intervention according to their suggested thresholds.

Disclosure of Interest: None declared