HQML-ICU: hybrid quantum-classical machine learning for ICU mortality prediction using MIMIC-IV electronic health records

Accurate prediction of mortality in the intensive care unit (ICU) is crucial for informing clinical decision-making and optimizing resource allocation. Although machine learning methodologies have enhanced traditional severity-of-illness scoring systems, they encounter intrinsic limitations in modeling complex feature interactions within electronic health record (EHR) data. In this study, we introduced a hybrid quantum-classical machine learning (HQML) framework for predicting ICU mortality, evaluated using the Medical Information Mart for Intensive Care IV (MIMIC-IV, version 2.2) database. Our methodology incorporates a classical autoencoder for dimensionality reduction (from 87 features to 16 latent dimensions), a 4-qubit parameterized quantum circuit (PQC) utilizing ZZFeatureMap encoding and RealAmplitudes variational ansatz (L = 3 layers, 48 trainable parameters), and a classical post-processing layer. In a retrospective cohort of 38,624 ICU admissions (12.8% mortality rate), the HQML model achieved an area under the receiver operating characteristic curve (AUROC) of 0.871 (95% CI: 0.859-0.883), outperforming XGBoost (0.858; 95% CI: 0.844-0.872), deep neural network (0.847), random forest (0.838), logistic regression (0.813), and SAPS-II (0.771). The improvement over XGBoost was statistically significant (DeLong p = 0.012) and corroborated by 5-fold cross-validation. The model demonstrated superior calibration (Brier score 0.089 vs. 0.095) and enhanced data efficiency. Comprehensive ablation experiments have elucidated the effects of circuit depth, quantum feature map selection, autoencoder latent dimensionality, class imbalance management, and quantum noise. Noise analysis suggests that gate error rates below 10-3 are necessary to maintain this advantage. These findings substantiate that hybrid quantum-classical architectures can yield incremental but statistically significant improvements in clinical prediction tasks under idealized simulation conditions. As the study is limited to a single-center dataset (MIMIC-IV), external validation on independent cohorts is required before clinical generalizability can be established.