Quantum-enhanced optimization for patient stratification in clinical trials
Clinical trials are a vital step in drug development, yet nearly 90 \% fail to reach approval [1], despite costing between {\}29M and {\}135M per therapy [2]. A major contributor to these failures is poor patient stratification-the suboptimal assignment of patients to treatment and control groups-re...
Uloženo v:
| Vydáno v: | Proceedings / IEEE Computer Society Annual Symposium on VLSI Ročník 1; s. 1 |
|---|---|
| Hlavní autoři: | , |
| Médium: | Konferenční příspěvek |
| Jazyk: | angličtina |
| Vydáno: |
IEEE
06.07.2025
|
| Témata: | |
| ISSN: | 2159-3477 |
| On-line přístup: | Získat plný text |
| Tagy: |
Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
|
| Shrnutí: | Clinical trials are a vital step in drug development, yet nearly 90 \% fail to reach approval [1], despite costing between {\}29M and {\}135M per therapy [2]. A major contributor to these failures is poor patient stratification-the suboptimal assignment of patients to treatment and control groups-resulting in imbalances across critical attributes such as age, biomarkers, or genetic markers. These discrepancies can obscure therapeutic effects and lead to inconclusive outcomes. Traditional randomization, while unbiased in expectation, often fails to balance covariates in practice. We address this issue by formulating patient stratification as a constrained combinatorial optimization problem. Each patient is described by a set of covariates, and the objective is to divide the cohort into equal-sized groups that minimize discrepancies in means, variances, and covariances. The resulting mixed integer linear optimization problem, subject to group-size and assignment constraints, becomes computationally intractable at large scales classically. To investigate practical solutions, we conduct experiments on a real-world dataset from a Mayo Clinic trial on primary biliary cholangitis [3], involving 312 patients. Focusing on three key covariates (age, alkaline phosphatase, and prothrombin time), we benchmark six optimization methods across patient sizes from 10 to 200. These include Gurobi, a classical solver that guarantees optimality but struggles with scalability; Graver Augmented Multi-start Algorithm (GAMA), which generates feasible seeds using D-Wave's quantum annealer and refines them with Graver basis augmentations; HybridCQM, a D-Wave hybrid solver that partitions and solves subproblems using a mix of classical and quantum resources; Kerberos, which tackles a QUBO formulation using Tabu search, simulated annealing, and quantum subproblem sampling; Tabu Search, a classical metaheuristic; and Quantum Approximate Optimization Algorithm (QAOA), a variational quantum algorithm simulated classically. |
|---|---|
| ISSN: | 2159-3477 |
| DOI: | 10.1109/ISVLSI65124.2025.11130253 |