Skip to content

Back to the catalog

Quantum Feature Selection for Biomedical Data Analysis
arxiv.orgpaper

Quantum Feature Selection for Biomedical Data Analysis

This paper formulates metabolomic feature selection as a Quadratic Unconstrained Binary Optimization (QUBO) problem, balancing feature relevance against redundancy to prepare high-dimensional data for machine learning models like computational biomarkers. The authors test Bias-Field Digitized Counterdiabatic Quantum Optimization (BF-DCQO) and QAOA on gate-based quantum hardware, comparing against classical filter, wrapper, and embedded feature selection methods (including exhaustive search and Iterative Tabu Search) on three Autism Spectrum Disorder metabolomic datasets. Results show reduced runtime versus exhaustive/Tabu search and competitive classifier performance, though the authors explicitly do not claim quantum advantage, framing the work as a NISQ hardware feasibility demonstration.

TL;DR Quantum QUBO-based feature selection for metabolomic ML data, tested on real quantum hardware without claiming quantum advantage.
Open original ↗
Document

What it claims

The paper's actual novel contribution is a specific QUBO coefficient formulation that the existing summary leaves generic: it uses the Fisher score to measure class-discriminative relevance (linear terms) and absolute covariance to penalize pairwise redundancy (quadratic terms), replacing the mutual-information coefficients of the prior MIQUBO framework (Mücke et al. 2023). The benchmark comparison is against MIQUBO specifically, plus classical filter/wrapper/embedded methods, on three ASD metabolomic datasets.

How it works

The objective is Q(x,α) = −α Σ Iᵢxᵢ + (1−α) Σ Rᵢⱼxᵢxⱼ, where Iᵢ is the Fisher score normalized to [0,1] and Rᵢⱼ is normalized absolute covariance with Rᵢᵢ=0. A single parameter α does double duty: it balances relevance against redundancy and controls subset size — α is tuned per target size until the optimal solution contains the desired number of features, avoiding an explicit cardinality penalty. The QUBO maps to an Ising Hamiltonian via xᵢ→(I−σᶻᵢ)/2 and is solved by either BF-DCQO (non-variational, iteratively updated bias fields seeded from prior-iteration spin expectations) or QAOA (here p=2 layers, 4 variational parameters, optimized with COBYLA). For both solvers the lowest-QUBO-energy sampled bitstring is taken as the selected subset.

The MIQUBO baseline discretizes each continuous feature into B=20 quantile bins to estimate the MI terms.

Evidence

The one hard hardware number the page gives is a transpiled-circuit comparison on a 24-feature QUBO instance: BF-DCQO needed 4,675 gates at circuit depth 777, while QAOA needed 8,731 gates at depth 2,011 — roughly 2× the gates and 2.6× the depth. Both mapped to 120 physical qubits, circuit width 144. Selected subsets were then scored across five classifiers: LDA, linear SVM, logistic regression, random forest, and PLS-DA. Classical baselines were Pearson-correlation filtering, exhaustive SVM-wrapper search, SVM-RFE (embedded), and Iterated Tabu Search as a classical QUBO solver.

Limits

The fetched text cuts off in the Performance Evaluation section (mid-sentence on LDA), so the actual accuracy tables, per-dataset results, and dataset sizes are not present in this capture — the runtime and "competitive performance" claims cannot be verified from what is shown here. The circuit-resource figure is for a single 24-feature instance only; scaling behavior is not given. The authors explicitly disclaim quantum advantage, framing this as NISQ feasibility. Affiliation note for adopters: co-author is from Kipu Quantum (BF-DCQO's originators).

Organisations named here

  • Autism Spectrum Disorder

Related in research