READMISSION RISK PREDICTION USING EXPLAINABLE AI BASED ON MULTIMODAL EHR DATA

Abstract

Hospital readmission within thirty days of discharge is a major indicator of healthcare quality. Accurate Readmission Risk Prediction (RRP) is therefore important for patient outcomes, mortality, and cost of care. Modern hospitals routinely capture heterogeneous clinical evidence: unstructured clinical narratives, time-series vital signs, and structured International Classification of Diseases (ICD) diagnosis codes. Prior deep-learning approaches to RRP have largely treated these modalities in isolation, ignored their cross-modal dependencies, or omitted explainability, which yields brittle predictions and limited clinical trust. This thesis proposes a multimodal deep-learning framework that integrates all three modalities for interpretable 30-day RRP. The framework has five new components. Non-linear Ridge ClinicalBERT (NR-ClinicalBERT) performs Semantic Contextual Analysis (SCA) on clinical notes and ICD diagnosis text; the non-linear ridge function augments ClinicalBERT to capture non-linear relationships while controlling feature variance. A Co-Attention Variational AutoEncoder (Co AVAE) performs Cross-Model Integration Analysis (CMIA) over the clinical, vital, and ICD modalities; the co-attention mechanism captures local cross-modal correlations that the conventional Variational AutoEncoder cannot. A Logarithmic Dual Sigmoid Fuzzy Logic (LDS-FL) module labels risk using the Charlson Comorbidity Index (CCI) and a National Early Warning Score 2 (NEWS2) based Vital Instability Score; the LDS membership function reduces categorisation bias near decision boundaries. A Wald Conditional Random Field (WCRF) models the irregular temporal patterns in vital signs that conventional CRFs miss. A Hessian Smooth Maximum Bidirectional Gated Recurrent Unit (HSM-BiGRU) produces the final readmission prediction using the Smooth Maximum Unit (SMU) activation and Hessian regularization for stability and non-linear feature learning. Local Interpretable Model-agnostic Explanations (LIME) then provide feature-level explanations to support clinician trust. The framework was implemented in Python and evaluated on the Medical Information Mart for Intensive Care IV (MIMIC-IV) dataset (version 3.1), with 299,712 patient records split 80:20 for training and testing. The proposed HSM BiGRU classifier reached a readmission prediction accuracy of 98.67%, Root Mean Squared Error (RMSE) of 0.154, Mean Absolute Error (MAE) of 0.118, Cohen's Kappa of 0.961, and Receiver Operating Characteristic Area Under the Curve (ROC AUC) of 0.9867, beating the BiGRU, BiLSTM, LSTM, and RNN baselines. Co AVAE obtained a Mean Squared Error (MSE) of 0.011 against 0.092 for the conventional VAE, and NR-ClinicalBERT reached a Mean Reciprocal Rank (MRR) of 0.93 and Mean Average Precision (MAP) of 0.91, ahead of ClinicalBERT, RoBERTa, DistilBERT, and BERT. LDS-FL achieved a defuzzification time of 235 vi ms. WCRF obtained a log-likelihood of 1.845 with a Sequence Error Rate (SER) of 0.08. In practice, the framework supports faster and more accurate decision-making for discharge planning and post-discharge resource allocation. Future work can extend the input set to include laboratory reports, medication histories, radiology findings, wearable-sensor data, and genomic information for more complete patient profiling and earlier intervention.