Artificial intelligence (XGBoost) in predicting outcomes among CAR-T therapy patients: The impact of malnutrition and comorbidities using the National Inpatient Sample (2020-2022).

Abstract

2536 Background: Chimeric Antigen Receptor T-cell (CAR-T) therapy has revolutionized hematologic malignancy treatment but remains costly, with limited access and complications like prolonged hospitalization, sepsis, and mortality. Malnutrition, common in cancer patients, worsens these outcomes. Despite AI’s growing role in oncology, its use in risk stratification for malnourished CAR-T recipients is underexplored. This study leverages the National Inpatient Sample (NIS) 2020-2022 to develop AI-driven models predicting length of stay (LOS), mortality, and sepsis, incorporating the Charlson Comorbidity Index and other factors. Methods: Using the NIS database, adult CAR-T therapy patients were identified with ICD-10 codes. Key variables included demographics (age, gender, race/ethnicity, income), clinical factors (Charlson Comorbidity Index, sepsis, admission type), and hospital characteristics (size, teaching status). AI models (XGBoost, Random Forest, Neural networks) were trained on the 2020 dataset and validated on 2020-2022 data. Hyperparameter tuning via grid search was performed to optimize model performance. LOS was modeled as a continuous outcome, while mortality and sepsis were classified as binary outcomes. Data preprocessing included handling missing values, one-hot encoding of categorical variables, and standardizing continuous variables. SHapley Additive exPlanations (SHAP) were used to interpret feature importance. Results: The study analyzed 1,912 CAR-T hospitalizations over three years, with 11.5% identified as malnourished. AI models demonstrated strong predictive performance, with XGBoost (RMSE: 3.5 days, R² = 0.82) for LOS, Random Forest (AUC: 0.91) for mortality, and Neural Networks (AUC: 0.87) for sepsis. Malnutrition significantly worsened outcomes, increasing LOS by 14.2 days (p < 0.001) and mortality risk by 3.2-fold (p < 0.001). Patients with Charlson Comorbidity Index scores ≥3 had 9.8-day longer LOS and 2.9-fold higher mortality risk (p < 0.001). Racial disparities were evident, with Black patients at 25% higher risk of prolonged LOS and Hispanic patients at increased risk of sepsis (p < 0.05). Malnourished patients in non-teaching hospitals with high comorbidity burdens had the worst outcomes, emphasizing the need for targeted interventions in high-risk populations. Conclusions: AI-driven models incorporating malnutrition and Charlson Comorbidity Index accurately predict LOS, mortality, and sepsis in CAR-T patients. Early identification and management of malnutrition and comorbidities, particularly in racially diverse populations, are critical to improving outcomes. Future research should focus on prospective validation and AI integration into clinical workflows to mitigate disparities.