A STOCHASTIC SUBSET-SELECTION METHOD FOR ANCHORED POPULATION-ADJUSTED INDIRECT TREATMENT COMPARISONS: SWARM-ITC...
Author(s)
Amit Kumar, BSc1, Mahendra Kumar Rai, PhD2, Deepthi Chinchapattanam, M.Pharm1, Prachi Singh, MS (Pharm)3, Manpreet Singh Kalsey, MBA1.
1Trinity Life Sciences, Mumbai, India, 2Trinity Lifesciences, Singapore, Singapore, 3Trinity Life Sciences, Singapore, Singapore.
1Trinity Life Sciences, Mumbai, India, 2Trinity Lifesciences, Singapore, Singapore, 3Trinity Life Sciences, Singapore, Singapore.
OBJECTIVES: Matching-adjusted indirect comparison (MAIC) reweighs patient-level data (IPD) to match aggregate comparator statistics; however, extreme weights inflate variance and MAIC does not balance covariates across both treatment arms. We developed SWARM-ITC, a stochastic binary particle swarm optimization (BPSO)-based subset-selection method and evaluated it against anchored MAIC in proof-of-concept simulated networks.
METHODS: Two RCTs were simulated: RCT1 (A vs. B; n=300/arm) and RCT2 (B vs. C; n=250/arm) with five covariates: age >60 years, sex, prior stem cell transplant, >3 prior treatment lines, and International Staging System stage; Weibull models generated time-to-event outcomes. SWARM-ITC applied BPSO (V-shaped transfer function; 1,000 stochastic iterations) to select IPD subsets satisfying three constraints: equal arm allocation, aggregate covariate balance across both arms, and a prespecified minimum retained sample size. The objective function minimized squared differences between selected IPD and aggregate statistics. Cox proportional hazards regression was applied per iteration; 95% CIs were derived from empirical HR distributions. Anchored A-vs-C effects used the Bucher method. Scenario 1 used RCT1 IPD with RCT2 aggregate data; Scenario 2 reversed availability. MAIC used Signorovitch method-of-moments approach.
RESULTS: In Scenario 1, SWARM-ITC yielded a hazard ratio (HR) for C vs. A of 0.36 (95% CI: 0.30-0.43; n=304; max covariate difference <0.5%) versus MAIC HR 0.37 (95% CI: 0.16-0.84; ESS=466). In Scenario 2, SWARM-ITC produced HR 0.42 (95% CI: 0.32-0.54; n=254; max covariate difference <0.3%) versus MAIC HR 0.43 (95% CI: 0.19-0.96; ESS=415). Point estimates were consistent; SWARM-ITC empirical CIs were narrower, reflecting avoidance of extreme weight-driven variance inflation; HR and CI bounds stabilized by iteration 900.
CONCLUSIONS: In this proof-of-concept simulation, SWARM-ITC produced estimates consistent with MAIC while enforcing equal covariate balance across both arms without reweighting. The BPSO framework with empirical uncertainty quantification may offer a viable alternative when MAIC weight distributions are unstable. Real-world validation is warranted before HTA adoption.
METHODS: Two RCTs were simulated: RCT1 (A vs. B; n=300/arm) and RCT2 (B vs. C; n=250/arm) with five covariates: age >60 years, sex, prior stem cell transplant, >3 prior treatment lines, and International Staging System stage; Weibull models generated time-to-event outcomes. SWARM-ITC applied BPSO (V-shaped transfer function; 1,000 stochastic iterations) to select IPD subsets satisfying three constraints: equal arm allocation, aggregate covariate balance across both arms, and a prespecified minimum retained sample size. The objective function minimized squared differences between selected IPD and aggregate statistics. Cox proportional hazards regression was applied per iteration; 95% CIs were derived from empirical HR distributions. Anchored A-vs-C effects used the Bucher method. Scenario 1 used RCT1 IPD with RCT2 aggregate data; Scenario 2 reversed availability. MAIC used Signorovitch method-of-moments approach.
RESULTS: In Scenario 1, SWARM-ITC yielded a hazard ratio (HR) for C vs. A of 0.36 (95% CI: 0.30-0.43; n=304; max covariate difference <0.5%) versus MAIC HR 0.37 (95% CI: 0.16-0.84; ESS=466). In Scenario 2, SWARM-ITC produced HR 0.42 (95% CI: 0.32-0.54; n=254; max covariate difference <0.3%) versus MAIC HR 0.43 (95% CI: 0.19-0.96; ESS=415). Point estimates were consistent; SWARM-ITC empirical CIs were narrower, reflecting avoidance of extreme weight-driven variance inflation; HR and CI bounds stabilized by iteration 900.
CONCLUSIONS: In this proof-of-concept simulation, SWARM-ITC produced estimates consistent with MAIC while enforcing equal covariate balance across both arms without reweighting. The BPSO framework with empirical uncertainty quantification may offer a viable alternative when MAIC weight distributions are unstable. Real-world validation is warranted before HTA adoption.
Conference/Value in Health Info
2026-09, ISPOR Asia Pacific 2026, Bangkok, Thailand
Value in Health, Volume 55, Issue S1
Code
MSR23
Topic
Methodological & Statistical Research
Disease
SDC: Oncology