Quantum ML for ICU Mortality Prediction¶
A reproducible benchmark of Quantum SVMs, a VQC, and a QNN against classical baselines for ICU mortality prediction. The pipeline is built for the WiDS Datathon 2020 ICU dataset, but every result on this site comes from its schema-matched synthetic fallback, not the real WiDS data. Built on Qiskit, fully seeded, and honest about what it found.
Headline: on the synthetic data, Logistic Regression reaches ROC-AUC 0.817 [0.787, 0.845]. With the pipeline's unscaled angle encoding every quantum model's 95% confidence interval includes 0.5. A bandwidth ablation traces the QSVM result to that encoding: the kernels sit at the random-state value, and with the input scale picked on validation data the QSVMs reach test ROC-AUC 0.70 to 0.80, the best of them level with logistic regression on the same 200 rows. An RBF SVM with its bandwidth tuned the same way (a post-hoc control) reaches 0.810 there, so the rescaling removes an encoding artifact rather than showing a quantum benefit. On those rows no model, classical or quantum, scores above the APACHE risk column alone. See the findings.
At a glance¶
Main results table as a chart¶
Each bar spans the 95% bootstrap confidence interval for ROC-AUC. The dashed line marks 0.5, the score of a coin flip. Classical bars sit clearly to the right of chance; with the pipeline's encoding every quantum bar crosses it. For the rescaled QSVMs and the tuned RBF control, see the bandwidth ablation.
Explore¶
Not a clinical tool
This is a research and portfolio project on synthetic or de-identified competition data. Nothing here is validated for clinical use.