Quantum Machine Learning for Predicting Anastomotic Leak: A Clinical Study

TL;DR

This study evaluates quantum neural networks for predicting anastomotic leaks after colorectal surgery, comparing their performance to classical models and highlighting their potential for capturing complex clinical data patterns.

Contribution

It introduces a quantum neural network framework for clinical prediction, benchmarking it against classical models, and analyzes the impact of different quantum encodings and optimizers on performance.

Findings

Quantum models achieved highest AUC of 0.7966 with EfficientSU2-BFGS.

QNNs with RealAmplitudes and CMA-ES excelled in Average Precision.

Effective optimizer convergence correlates with better classification metrics.

Abstract

Anastomotic leak (AL) is a life-threatening complication following colorectal surgery, and its accurate prediction remains a significant clinical challenge. This study explores the potential of Quantum Neural Networks (QNNs) for AL prediction, presenting a rigorous benchmark against hyperparameter-tuned classical models including logistic regression, multilayer perceptrons, and boosting algorithms. Using a clinical dataset of 200 patients and four key predictors identified through statistical analysis, we evaluated QNNs with ZZFeatureMap encoding and EfficientSU2 and RealAmplitudes ans\"atze simulated under realistic hardware noise models. Our framework emphasizes robustness, with performance metrics averaged over 10 independent optimization runs using multiple algorithms. The EfficientSU2-BFGS combination achieved the highest mean AUC of $0.7966 \pm 0.0237$, while RealAmplitudes with CMA-ES excelled in Average Precision ($0.5041 \pm 0.1214$), critical for imbalanced medical datasets. We establish a direct link between optimizer convergence and model efficacy, where effective variational parameter optimization translates to improved classification metrics. Interpretability analysis suggests QNNs may capture complex, non-linear feature relationships not evident in classical linear models. This work highlights QNNs' potential as screening tools while underscoring the need for model selection based on specific clinical goals, pending validation on larger datasets.