Quantum Reservoir Computing with Neutral Atoms on a Small, Complex, Medical Dataset

TL;DR

This work probes quantum reservoir computing (QRC) for biomarker-based binary classification on a small clinical dataset. It compares six classical models against quantum features generated by both noiseless emulation and hardware execution on Aquila, using SHAP-driven feature ranking to select informative biomarkers. Emulated QRC achieves mean accuracies similar to classical features but shows signs of overfitting; hardware QRC, however, delivers more robust performance and often higher mean accuracy, indicating a hardware-induced regularisation. Mechanistic analysis reveals hardware features undergo a time-dependent, contractive transformation with reduced mutual information relative to emulation, suggesting structured reshaping of the feature space rather than pure noise. Overall, hardware QRC provides robustness and potential gains in this domain, though full benchmarking against classical approaches remains contingent on shot resources and further methodological refinements.

Abstract

Biomarker-based prediction of clinical outcomes is challenging due to nonlinear relationships, correlated features, and the limited size of many medical datasets. Classical machine-learning methods can struggle under these conditions, motivating the search for alternatives. In this work, we investigate quantum reservoir computing (QRC), using both noiseless emulation and hardware execution on the neutral-atom Rydberg processor \textit{Aquila}. We evaluate performance with six classical machine-learning models and use SHAP to generate feature subsets. We find that models trained on emulated quantum features achieve mean test accuracies comparable to those trained on classical features, but have higher training accuracies and greater variability over data splits, consistent with overfitting. When comparing hardware execution of QRC to noiseless emulation, the models are more robust over different data splits and often exhibit statistically significant improvements in mean test accuracy. This combination of improved accuracy and increased stability is suggestive of a regularising effect induced by hardware execution. To investigate the origin of this behaviour, we examine the statistical differences between hardware and emulated quantum feature distributions. We find that hardware execution applies a structured, time-dependent transformation characterised by compression toward the mean and a progressive reduction in mutual information relative to emulation.

Figures (7)

• Figure 1: Performance of six machine-learning models trained on SHAP-ranked classical feature subsets. We use five unique train/test splits of the data. For each model, data split, and feature count, we evaluate performance with optimal machine-learning hyperparameters. Violin plots show variability in test accuracy across the five data splits. The highest mean test accuracy is marked by the black star. The dotted line marks the mean test accuracy when using all 56 data features. In all cases the models exhibit improved performance when using a subset of all available data features.
• Figure 2: Performance of six machine-learning models trained on quantum features generated by an emulated quantum reservoir. The same SHAP-ranked feature subsets as in the classical results are used to encode data into the atoms of the reservoir. We use five unique train/test splits of the data. For each model, data split, and atom count, we evaluate performance with optimal machine-learning hyperparameters. Violin plots show variability in test accuracy across the five data splits. The highest mean test accuracy is marked by the black star. The optimal performance achieved using emulated quantum features is comparable to that obtained with classical features. Models using emulated quantum features generally have higher training accuracies and larger variability across data splits, indicative of overfitting.
• Figure 3: Changes in mean test accuracy (vertical axis) and split-to-split variability (horizontal axis) when moving from emulated to hardware quantum features. We observe that models trained on hardware quantum features often exhibit less variability across data splits than emulated quantum features. Importantly, this is often also accompanied by an increase in mean test accuracy. This indicates that executing the quantum reservoir on hardware helps to mitigate the overfitting tendencies observed with the emulated reservoir.
• Figure 4: Standardised difference (see Eq. \ref{['Eq:standardiseddiff']}) in mean test accuracy between models trained on hardware and emulated quantum features. Points with a black outline correspond to models which also exhibited reduced variability on hardware compared to emulation. Across most model and feature-count combinations, hardware performance is comparable to or exceeds that of emulation, with several configurations exhibiting statistically significant advantages for hardware execution (standardised difference $>2$).
• Figure 5: Comparison of hardware and emulated features across evolution times. Top left: Heatmap of residuals ($\mathrm{HW}-\mathrm{Emu}$), showing structured sample- and feature-dependent deviations. Top right: Mutual information between hardware and emulated features, decreasing at later times. Bottom row: Residuals plotted against emulated values, with $\langle Z \rangle$ and $\langle ZZ \rangle$ features shown separately. Overlaid points indicate equal-population bins (0.01–0.99 quantile range), showing bin means and variances. Residuals follow an approximately linear trend with Gaussian-distributed fluctuations within each bin.