An Interpretable AI Framework to Predict Post-ICU Mental Health Risks employing EHR Data and Vitamin D Biomarkers

Abstract

Mental complications (delirium, depression, anxiety) are common to patients in the intensive care unit (ICU) and a cause of post-intensive care syndrome (PICS) but are difficult to predict at the earliest stages because of sparse coding and minimal emphasis on modifiable risks. To construct and test classical machine learning models to predict ICU mental health outcomes based on MIMIC-III database, including a clear consideration of vitamin D deficiency proxies as one of the modifiable predictors. We examined 29, 214 unique first adult ICU admissions (2001-2012). The demographics, vital/laboratory aggregates, medications, comorbidities, and clinical notes were features of the first 48 hours (NLP-augmented). The deficiency of vitamin D was measured through multimodal proxies (labs, ICD-9, prescriptions, NLP). There were 4 models (Logistic Regression, Random Forest, XGBoost, LightGBM), which were trained on 70% data, tuned by nested cross-validation and Optuna, and tested on a 15% hold-out test data. Measurement of performance was done using AUROC, AUPRC, calibration, SHAP interpretability, ablation experiments, and fairness measures. XGBoost using vitamin D proxies had the best performance: delirium AUROC 0.859 [0.851-0.867], depression 0.791 [0.777-0.805], anxiety 0.806 [0.793-0.819] and so forth. Vitamin D proxies were found to be of incremental use (DAUROC +0.026 in the case of delirium, p<0.001) and had a SHAP importance of 4th-7th. Close relations were found (deficiency x age [?]65, deficiency x inflammation). Fairness analysis indicated that there were few differences after mitigation (equalized odds -0.04 between genders/ethnicities). The finding: Multimodal models of classical ML with vitamin D proxies provide powerful and interpretable early prediction of ICU mental health outcomes. Deficiency of vitamin D becomes a clinically significant, practically useful risk factor, especially in older and inflamed patients. These findings support development of decision support tools for targeted screening and supplementation in critical care.