Data complexity signature predicts quantum projected learning benefit for antibiotic resistance

TL;DR

This study evaluates Quantum Projective Learning (QPL) for predicting antibiotic resistance in clinical urine cultures using pre-fault-tolerant quantum hardware. It benchmarks QPL against classical models across simulated and real quantum devices, discovering no global advantage but data-dependent parity or gains in specific splits, notably for Nitrofurantoin. A multivariate data complexity signature—comprising Shannon entropy, Fisher Discriminant Ratio, standard deviation of kurtosis, number of low-variance features, and total correlations—predicts when QPL on quantum hardware will outperform classical models with $AUC = 0.88$, $p$-value $= 0.03$. The results imply complexity-driven adaptive model selection could guide hybrid quantum-classical workflows in healthcare, marking a first practical application of QML in this domain and outlining paths for scaling and error mitigation.

Abstract

This study presents the first large-scale empirical evaluation of quantum machine learning for predicting antibiotic resistance in clinical urine cultures. Antibiotic resistance is amongst the top threats to humanity, and inappropriate antibiotic use is a main driver of resistance. We developed a Quantum Projective Learning (QPL) approach and executed 60 qubit experiments on IBM Eagle and Heron quantum processing units. While QPL did not consistently outperform classical baselines, potentially reflecting current quantum hardware limitations, it did achieve parity or superiority in specific scenarios, notably for the antibiotic nitrofurantoin and selected data splits, revealing that quantum advantage may be data-dependent. Analysis of data complexity measures uncovered a multivariate signature, which comprised Shannon entropy, Fisher Discriminant Ratio, standard deviation of kurtosis, number of low-variance features, and total correlations. The multivariate model accurately (AUC = 0.88, $p$-value = 0.03) distinguished cases wherein QPL executed on quantum hardware would outperform classical models. This signature suggests that quantum kernels excel in feature spaces with high entropy and structural complexity. These findings point to complexity-driven adaptive model selection as a promising strategy for optimizing hybrid quantum-classical workflows in healthcare. Overall, this investigation marks the first application of quantum machine learning in urology, and in antibiotic resistance prediction. Further, this work highlights conditional quantum utility and introduces a principled approach for leveraging data complexity signatures to guide quantum machine learning deployment in biomedical applications.

Figures (9)

• Figure 1: Overview of quantum projection learning workflow. Input data are preprocessed classically and may include normalization, dimensionality reduction, and feature selection. After data are processed, they are quantum projected with a quantum feature map. Measurements are then made in the X, Y, and Z basis. This measured projected data are then analyzed by a suite of classical machine learning classifiers: support vector classifier (SVC), random forest (RF), multilayer perceptron (MLP), logistic regression (LR), and extreme gradient boosting (XGB). These classifiers perform the final antibiotic resistance prediction on the measured data.
• Figure 2: Number of organism–antibiotic susceptibility classifications for each antibiotic.
• Figure 3: Antibiotic resistance prediction performance per antibiotic. A) Predictive performance across four metrics for each of the five QPL types and two classical baselines. B) Performance comparison between the best performing model configuration for QPL, RF, and XGB executed on simulator or QPU for each data split. The best performing model configuration is the model with the maximum median weighted F1 over 10 data splits.
• Figure 4: Predictive performance of QPL versus XGBoost in simulation and hardware. A) Datapoints represent individual data splits across all QPL types for each antibiotics. Weighted F1 of the QPL and XGB are on the y- and x-axis, respectively. Experiments are delineated by QPLs executed on simulator (top) and on QPU (bottom). B) Comparison of the data complexity metric distribution between data splits where QPL outperforms (blue) or under performs (orange) XGB in simulation.
• Figure 5: Relationship of data complexity measures and QPL performance on QPU. A) Boxplot of distribution of metrics values for data splits where QPL outperformed (blue) and underperformed (orange) XGB on IBM Heron R2 QPUs. B) AUC for correctly predicting QPL performance using a logistic regression with ElasticNet penalty in a five fold cross validation. C) Heatmap of metric values found significant by logistic regression. Rows are significant metrics and rows antibiotic data splits. Column color bar indicates whether the QPL outperformed (blue) or underperformed (orange) XGB