Quantum Machine Learning Edge Emerges in Predicting ICU Mortality From Electronic Health Records

Quantum computing has taken a tentative but measurable step into the intensive care unit. In a study published in Frontiers in Computational Neuroscience, researchers introduced HQML-ICU, a hybrid quantum-classical machine learning framework designed to predict whether a patient will die during an ICU stay. Evaluated on 38,624 adult ICU admissions drawn from the widely used MIMIC-IV electronic health record database, the hybrid model achieved an area under the receiver operating characteristic curve (AUROC) of 0.871, edging out the gradient-boosted powerhouse XGBoost at 0.858, along with deep neural networks, random forests, logistic regression and the classical SAPS-II severity score. The margin is small, but the team reports that it is statistically significant, with a DeLong test p-value of 0.012 and confirmation across five-fold cross-validation.

The motivation behind the work lies in the stubborn limits of conventional models. Traditional severity-of-illness scoring systems such as APACHE, SAPS and SOFA, built on logistic regression with expert-selected variables, have plateaued at AUROCs of roughly 0.75 to 0.85. Modern machine learning methods do better, but they must explicitly enumerate the interactions between features, and those interactions grow combinatorially as data dimensionality rises. Tree ensembles rely on axis-aligned splits that can miss smooth physiological relationships, while deep networks demand enormous datasets. Quantum computing offers a different mathematical arena: n qubits inhabit a Hilbert space of 2^n dimensions, and parameterized quantum circuits can, in principle, represent complex functions of data more compactly than any classical network.

Because today’s noisy intermediate-scale quantum (NISQ) devices cannot accommodate thousands of input features, the researchers built a three-stage pipeline. First, a classical autoencoder compresses 87 electronic health record features into just 16 latent dimensions; the encoder passes data through fully connected layers of 87, 64, 32 and 16 neurons, and a latent size of 16 was chosen as the smallest achieving a reconstruction R-squared above 0.93 on validation data. Next comes the quantum heart of the system: a 4-qubit parameterized quantum circuit that uses a ZZFeatureMap to encode classical values into quantum states via Hadamard gates, parameterized Rz rotations and controlled-Rz gates that entangle every pair of qubits, explicitly writing pairwise feature interactions into a 16-dimensional Hilbert space. A RealAmplitudes variational ansatz with three layers, using Ry rotations and a chain of CNOT entangling gates, then processes the state with 48 trainable rotation angles. Finally, a small classical neural head maps the four measured Pauli-Z expectation values to a calibrated mortality probability. The entire trainable model contains just 145 parameters, compared with 22,145 for the deep neural network baseline.

The cohort behind these experiments is among the largest ever assembled for healthcare quantum machine learning. The team extracted 89 clinical features spanning demographics, 24-hour vital sign summaries, laboratory extremes, medication and intervention indicators such as vasopressors and mechanical ventilation, admission diagnoses, and composite severity scores including SOFA and the Glasgow Coma Scale. Rigorous exclusion criteria removed patients who died or were discharged within the first 24 hours, and data splitting was performed at the patient level with strict prevention of information leakage: imputation via multiple imputation by chained equations, z-score normalization, winsorization and autoencoder training were all fitted exclusively on training data. The final population had a 12.8 percent mortality rate, preserved across stratified training, validation and test splits of 23,254, 7,623 and 7,747 patients respectively.

Beyond raw discrimination, the hybrid model showed the property clinicians arguably value most: honest probabilities. Its Brier score of 0.089 beat XGBoost’s 0.095 and the deep network’s 0.101, and a Hosmer-Lemeshow test yielded a reassuring p-value of 0.493, with a calibration slope of 1.02. When a physician tells a family there is a 30 percent mortality risk, that number should mean something; overconfident neural networks routinely distort such estimates in the highest-risk decile. At an operating threshold of 0.192, the hybrid model reached 71.0 percent sensitivity and 88.6 percent specificity, with a negative predictive value of 95.4 percent, meaning a prediction of survival was correct nearly 96 percent of the time. The authors estimate that in a hospital with 5,000 ICU admissions a year, the sensitivity gain over XGBoost would translate to roughly 20 additional correctly identified high-risk patients.

Interpretability analysis offered a hint that the quantum circuit does something genuinely different rather than merely shuffling the same signals. SHAP-based interaction analysis revealed that the hybrid model captured a stronger interaction between lactate levels and vasopressor use, scoring 0.024 versus 0.015 for XGBoost. That pairing is precisely the clinical signature of refractory septic shock, in which persistent tissue hypoperfusion despite aggressive hemodynamic support signals extremely high mortality risk. The entanglement structure of the circuit, where CNOT gates weave qubits into correlated states, appears naturally suited to encoding such higher-order physiological correlations within a compact parameter space, something classical trees can only approximate by growing ever deeper and less interpretable.

An extensive set of ablation experiments dissects which design choices matter. Circuit depth peaked at three layers, with performance collapsing toward barren-plateau territory at greater depth. Among five encoding strategies, the ZZFeatureMap’s explicit pairwise interaction gates beat Pauli, IQP, amplitude and angle encodings. Latent dimensionality of 16 balanced reconstruction fidelity against the 4-qubit circuit’s processing capacity, and pushing to 32 dimensions paradoxically degraded accuracy, a phenomenon the authors call quantum overfitting. Inverse-frequency class weighting proved the best remedy for the severe class imbalance inherent in mortality prediction. Most strikingly, replacing structured encoding with random rotations slashed the AUROC by 0.073, confirming the quantum feature map’s geometry is doing real work.

The most consequential caveat is that every quantum computation ran on a noiseless statevector simulator on classical hardware, not on a physical quantum processor. The reported 0.871 is therefore a theoretical upper bound. The team’s noise sensitivity analysis shows the advantage over XGBoost survives only if gate error rates fall below about 10^-3 and disappears at 2 x 10^-3, a threshold that current processors from IBM and Google approach but do not consistently meet. Training also took 48 minutes versus 3.2 for XGBoost, because simulating quantum circuits on classical hardware carries exponential cost. On the encouraging side, the hybrid model needed about 28 percent fewer training epochs and performed markedly better in low-data regimes, reaching an AUROC of 0.817 with just 10 percent of training data compared with 0.775 for the deep network.

Fairness checks add credibility. Across 12 patient subgroups defined by sex, age, ICU type and illness severity, the hybrid model improved AUROC by 0.010 to 0.016 over XGBoost in every stratum, and equalized odds and demographic parity differences showed no amplification of demographic bias. The authors are candid about remaining limitations, however: the study relies on a single-center dataset from Beth Israel Deaconess Medical Center, so external validation on multi-center databases such as eICU, HiRID or AmsterdamUMCdb is essential before any clinical deployment, and the internal representations of the quantum circuit remain a black box even when SHAP explains individual features.

Whether a modest AUROC gain justifies quantum machinery is context-dependent, and the researchers themselves frame the work as demonstrating potential under idealized conditions rather than proven hardware superiority. Yet as a blueprint, the study matters. It shows a complete, leakage-controlled, statistically disciplined path from raw electronic health records through autoencoder compression into an entangled quantum feature space and back out to calibrated clinical probabilities, executed at a scale of nearly 39,000 patients. As hardware error rates march toward 10^-4 and error mitigation techniques mature, the same architecture could be rerun on real quantum processors, and future fault-tolerant machines may allow deeper circuits operating on full-dimensional clinical data. For now, the quantum advantage in the ICU remains incremental and simulated, but it is no longer hypothetical in structure, and that alone marks a milestone for quantum machine learning in medicine.

Subject of Research: Hybrid quantum-classical machine learning for ICU mortality prediction using electronic health records

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

Article References: Abdel-Aty, A.-H., Kadry, H., Omer, N., Shaaban, S. M., Eltayeb, H., Ghorashi, S. A., Hafez, M., Voon, B. W., & Farsi, M. (2026). HQML-ICU: hybrid quantum-classical machine learning for ICU mortality prediction using MIMIC-IV electronic health records. Frontiers in Computational Neuroscience, 20, Article 1900553. https://doi.org/10.3389/fncom.2026.1900553

Image Credits: AI Generated

DOI: 10.3389/fncom.2026.1900553

Keywords: quantum machine learning, ICU mortality prediction, MIMIC-IV, electronic health records, parameterized quantum circuits, hybrid quantum-classical computing, XGBoost, autoencoder, clinical prediction, variational quantum circuits, critical care, Brier score