Quantum Projective Learning Achieves Parity with 60 Qubit Experiments for Antibiotics

Researchers are tackling the growing crisis of antibiotic resistance by exploring whether quantum machine learning can improve prediction of resistance in urinary tract infections. Kahn Rhrissorrakrai, Filippo Utro, and Alex Milinovich, from IBM Research and the Cleveland Clinic, alongside colleagues including Sandip Vasavada, Daniel Rhoads, and Laxmi Parida et al., have conducted the first large-scale empirical evaluation of this approach using clinical urine culture data. Their work demonstrates that, although current quantum hardware doesn’t consistently outperform classical methods, a specific ‘data complexity signature’ , encompassing measures like Shannon entropy and structural complexity , accurately predicts when quantum algorithms will excel. This discovery is significant because it offers a principled way to select the best machine learning model for the job, potentially optimising hybrid classical-quantum workflows and paving the way for more effective antibiotic stewardship in healthcare.

Quantum machine learning predicts urine antibiotic

Scientists have undertaken the first large-scale empirical evaluation of quantum machine learning applied to predicting antibiotic resistance in clinical urine cultures. Antibiotic resistance represents a critical threat to global health, and inappropriate antibiotic use is a major contributing factor to its proliferation. Researchers developed a Quantum Projective Learning (QPL) approach, executing 60 qubit experiments on both IBM Eagle and Heron quantum processing units to assess its potential. While QPL did not consistently surpass classical machine learning baselines, potentially due to limitations in current quantum hardware, it achieved comparable or superior results in specific instances, particularly with the antibiotic nitrofurantoin and certain data splits, suggesting that quantum advantage is data-dependen.
The study unveiled a multivariate signature, comprising Shannon entropy, Fisher Discriminant Ratio, standard deviation of kurtosis, number of low-variance features, and total correlations, accurately distinguishing scenarios where QPL outperformed classical models with an AUC of 0.88 and a p-value of 0.03. This signature indicates that quantum kernels perform optimally in feature spaces exhibiting high entropy and structural complexity. These findings establish complexity-driven adaptive model selection as a promising strategy for optimising hybrid quantum-classical workflows within healthcare settings. This investigation marks a pioneering application of quantum machine learning not only in urology, but also in the crucial field of antibiotic resistance prediction.

Experiments involved a quantum feature map followed by measurements in the X, Y, and Z bases, with the resulting projected data analysed by classical machine learning classifiers including support vector classifiers, random forests, multilayer perceptrons, logistic regression, and extreme gradient boosting. Unlike variational quantum circuits requiring iterative parameter optimisation, QPL avoids issues like barren plateaus, thereby reducing computational overhead on pre-fault tolerant quantum devices. Recent benchmarking has indicated that quantum kernels can outperform classical radial basis function kernels on complex, high-dimensional datasets, but practical quantum advantage requires scaling to larger quantum devices. Furthermore, the research addresses the urgent need for rapid and accurate prediction of antibiotic resistance in urine cultures, a critical factor in improving antibiotic selection and reducing selective pressure for resistant organisms. With antibiotic resistance linked to over 4 million deaths annually, early prediction offers significant potential to improve patient outcomes and future resistance profiles. The team’s work introduces a principled approach for leveraging data complexity signatures to guide quantum machine learning deployment in biomedical applications, highlighting conditional quantum utility and opening avenues for future research in this rapidly evolving field.

Quantum Projective Learning for Antibiotic Resistance Prediction

Scientists embarked on a large-scale empirical evaluation of machine learning techniques to predict antibiotic resistance in clinical urine cultures. The research team developed a Projective Learning (QPL) approach, executing 60 qubit experiments on both Eagle and Heron processing units to assess its performance against classical algorithms. Data preparation commenced with a clinical dataset comprising 2,723,116 organism, antibiotic susceptibility classifications from six antibiotics, meticulously documented in a referenced study. Categorical variables, totaling 192, underwent binary encoding using the category-encoders package and BinaryEncoder, transforming them into a format suitable for machine learning models.

To generate representative training and testing sets, the study pioneered the application of CORDS2, employing a feedforward neural network with two fully connected layers, the first mapping inputs to a 512-dimensional embedding and the second producing predictions for two classes. This network utilized a cross entropy loss function and a ‘PerClass’ selection type, generating ten coresets collections for each antibiotic, each containing a 300-sample training set and a 75-sample testing set. Dimensionality reduction was then performed using three distinct methods: NMF, PCA, and UMAP, each applied after min-max scaling the data to optimize feature representation. NMF, implemented using the scikit-learn package, was capped at a maximum of 200 components, further refining the dataset for analysis.

The QPL approach leverages quantum projected data, embedding it via a quantum feature map and subsequently analyzing it with both classical kernel methods and four additional classical machine learning models. This innovative design circumvents the need for iterative parameter optimization inherent in variational quantum circuits and quantum neural networks, mitigating issues like barren plateaus that often hinder trainability. Experiments were conducted on both quantum simulators and physical quantum devices, allowing researchers to directly assess the impact of data characteristics on predictive performance on Programmable Feature-based Trapped Ion Quantum Devices (PFTQDs). The team then evaluated the performance of QPL, comparing its accuracy and speed against established classical machine learning algorithms, including gradient boosting (XGB), to identify scenarios where quantum methods yielded improvements.

Quantum machine learning for urine resistance prediction

Scientists achieved a groundbreaking evaluation of machine learning techniques for predicting antibiotic resistance in clinical urine cultures, marking the first large-scale empirical study of its kind in urology. The research team developed a Projective Learning (QPL) approach and executed experiments utilising 60 qubit processing units, specifically, the Eagle and Heron platforms, to assess its efficacy. While QPL did not consistently surpass classical machine learning baselines, it demonstrated parity or superiority in specific instances, particularly when analysing the antibiotic nitrofurantoin and certain data splits, suggesting performance is data-dependent. This finding reveals that quantum machine learning advantage may not be universal, but rather contingent on the characteristics of the data being analysed.

Experiments meticulously measured data complexity using a multivariate signature comprising Shannon entropy, Fisher Discriminant Ratio, standard deviation of kurtosis, the number of low-variance features, and total correlations. This signature accurately distinguished scenarios where QPL, when executed on quantum hardware, would outperform classical models, achieving an Area Under the Curve (AUC) of 0.88 with a p-value of 0.03. Measurements confirm that kernels perform optimally in feature spaces exhibiting high entropy and structural complexity, providing valuable insight into the conditions under which quantum machine learning can excel. The team’s analysis of Ampicillin data revealed a geometric separation between classical and quantum-projected radial basis function kernels of 14.833, approaching the expected value of √N, which was calculated as 17.321.

Further tests demonstrated a classical model complexity of 0.6681 compared to 1.3249 for the quantum-projected kernel, indicating potential for quantum advantage, though not consistently across all antibiotics tested. For another antibiotic, the geometric separation was significantly lower at 0.0079, with classical and quantum model complexities of 1.000 and 0.447 respectively. These detailed measurements highlight the conditional utility of QPL and introduce a principled approach for leveraging data complexity signatures to guide machine learning deployment in biomedical applications. The study utilised Pauli twirling to tailor noise in gate and measurement operations to a level of 0.2, and optimised hyperparameters via RandomizedSearchCV using five-fold cross validation over 40 iterations for various machine learning methods including Support Vector Classifier, Random Forest, Multilayer Perceptron, Logistic Regression, and Extreme Gradient Boosting. This work establishes a foundation for complexity-driven adaptive model selection, promising to optimise hybrid classical-quantum workflows in healthcare and beyond.

QPL Predicts Nitrofurantoin Resistance with Signatures in Urinary

Scientists have undertaken the first large-scale empirical evaluation of machine learning techniques to predict antibiotic resistance in urine cultures. Researchers developed a Projective Learning (QPL) approach and ran 60 qubit experiments using Eagle and Heron processing units to assess its performance. Although QPL did not consistently surpass classical machine learning models overall, it achieved comparable or superior results in specific instances, particularly with the antibiotic nitrofurantoin and certain data subsets, suggesting performance is data-dependent. Analysis revealed a multivariate signature, comprising Shannon entropy, Fisher Discriminant Ratio, standard deviation of kurtosis, number of low-variance features, and total correlations, that accurately distinguished scenarios where QPL outperformed classical models, with an AUC of 0.88 and a p-value of 0.03.

This signature indicates that these kernels perform well in feature spaces exhibiting high entropy and structural complexity. These findings demonstrate a promising strategy for optimizing hybrid quantum-classical workflows in healthcare through complexity-driven adaptive model selection.This investigation represents the initial application of machine learning within urology and antibiotic resistance prediction, highlighting the conditional utility of quantum machine learning and introducing a principled method for leveraging data complexity signatures to guide deployment in biomedical applications. The authors acknowledge limitations related to current quantum hardware capabilities and focused solely on binary classification for antibiotic resistance. Future research should explore optimized trotterization schemes and advanced error mitigation techniques to improve quantum performance. Additionally, extending QPL to multi-class classification, survival analysis, or longitudinal prediction tasks may offer further insights into its applicability in healthcare.