arXiv:2609.30256
Quantum Feature Selection for Biomedical Data Analysis
Hongbin Liu, Robert Lahmann, Benjamin Campbell, Zhemin Zhang, Zhiding Liang, Siona Bapat, Juergen Hahn
Abstract
Feature selection is an essential step for reducing complexity of high dimensional data, usually in preparation for developing machine learning models such as computational biomarkers. However, there are limitations associated with feature selection such as for metabolomic data where there are hundreds or even thousands of features per study participant, while the available number of participants in a clinical trials is limited. Classical methods such as exhaustive search require evaluation of all possible feature combinations, making them costly in terms of computation and runtime, or even infeasible, as feature dimensionality increases. In this study, we propose a novel Quadratic Unconstrained Binary Optimization (QUBO) coefficient formulation and pose metabolomic feature selection as a QUBO problem that selects a specified number of features by balancing their relevance against redundancy among the selected variables. To evaluate our proposed QUBO objective function, we conducted a series of experiments using Bias-Field Digitized Counterdiabatic Quantum Optimization (BF-DCQO) and Quantum Approximate Optimization Algorithm (QAOA) on a quantum gate based computer. We also compared the method to several classical methods on three metabolomic datasets associated with Autism Spectrum Disorder (ASD) on a classical computer. Our method reduces runtime compared with exhaustive search and Iterative Tabu Search (ITS) and achieves competitive performance across classifiers compared to classical filter, wrapper, and embedded methods. These results do not claim quantum advantage; rather, they establish hardware feasibility and demonstrate the current capabilities of Noisy Intermediate-Scale Quantum (NISQ) devices.
AI Summary
The paper studies QUBO-based feature selection for small-sample biomedical classification. It replaces mutual-information relevance and redundancy coefficients with normalized Fisher scores and absolute covariances, minimizing $-\alpha\sum_i I_i x_i+(1-\alpha)\sum_{ij}R_{ij}x_ix_j$. The parameter $\alpha$ is adjusted to obtain five-feature subsets without an explicit cardinality penalty. Experiments use three autism-related datasets containing 24, 56, and 54 features and 159, 38, and 38 subjects, respectively; the latter two datasets undergo supervised prescreening. QUBOs are solved on IBM gate-based processors using bias-field digitized counterdiabatic quantum optimization (BF-DCQO), with five iterations and 1,000 shots per sampling step. Depth-two QAOA is evaluated only on the 24-feature dataset, using six optimizer evaluations and 10,000 shots per evaluation. Selected subsets are assessed with five classifiers under leave-one-out cross-validation and compared with mutual-information QUBO, Pearson ranking, exhaustive wrapper search, and SVM recursive feature elimination. BF-DCQO returns the same subsets as classical iterated tabu search (ITS) in all tested cases. The proposed coefficients outperform mutual-information QUBO in 13 of 15 classifier–dataset comparisons; ten QAOA runs also yield higher mean accuracies and lower variability across all five classifiers. No selection method consistently performs best. Total BF-DCQO runtimes are 83, 490, and 383.5 seconds, versus 20.14, 789.35, and 1,031.50 seconds for ITS. Stated limitations include small or reduced-dimensional datasets, hardware noise and circuit-depth constraints, and omission of higher-order feature interactions. Globally optimal solutions are not independently established, preventing time-to-solution benchmarking; the runtime comparisons do not establish quantum advantage.
Central claim may not hold: The abstract's runtime-reduction statement needs qualification because BF-DCQO total runtime exceeds both classical baselines on Dataset 1, although it is lower on Datasets 2 and 3. The reported LOOCV protocol does not establish that supervised feature selection and label-based prescreening exclude held-out subjects, leaving the claimed predictive-performance evidence vulnerable to information leakage. The proposed QUBO does not enforce a specified feature count, and tuning its relevance–redundancy parameter is not a general guarantee of attaining every requested cardinality, despite the reported five-feature outputs.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.