FACTORS ASSOCIATED WITH TRANSPARENCY COMMITTEE ACCEPTABILITY OF COMPOSITE ENDPOINTS: A MULTI-LLM ARTIFICIAL INTELLIGENCE ANALYSIS OF FRENCH HTA OPINIONS (2018-2025)
Author(s)
Ludovic Lamarsalle, MSc, PharmD1, Caroline Heng, Pharm D, MSc2.
1Founder, HEALSTRA, Lyon, France, 2Johnson & Johnson Innovative Medicine, Issy-les-moulineaux, France.
1Founder, HEALSTRA, Lyon, France, 2Johnson & Johnson Innovative Medicine, Issy-les-moulineaux, France.
OBJECTIVES: Many pivotal trials use a composite endpoint as their primary endpoint, yet how the French Transparency Committee (TC, HAS) appraises them is poorly characterized. We aimed to identify factors associated with the acceptability of composite scores as primary endpoints and their relationship with clinical benefit (SMR) and added clinical benefit (ASMR) ratings, using a systematic artificial-intelligence pipeline.
METHODS: We screened all TC opinions issued 2018-2025 (n=2,281). After excluding generics and biosimilars and manually retaining chronic-disease therapeutic areas likely to use composite endpoints, 864 opinions remained, of which 87 used a composite score as primary endpoint and were analyzed together with their meeting transcripts. Two large language models (GPT-5 mini, Claude Sonnet 4.5) extracted predefined fields in parallel; a third (Mistral Large) arbitrated divergences (14.8%) as quality control, yielding 853 coded statements. Each statement was a positive or negative appraisal of the composite score (CS+/CS-) or of study results (R+/R-). Per opinion, the CS+/CS- and R+/R- ratios were correlated with SMR and ASMR using Spearman's rho.
RESULTS: Result statements were predominantly positive (R+/R-=1.85), whereas composite scores were appraised more negatively (CS+/CS-=0.65). Important SMR was granted in 60% of opinions and ASMR V in 63%; only 16% combined an Important SMR with ASMR II-IV. Methodological concerns appeared in 90% of opinions; about 60% lacked a defined minimal clinically important difference and 83% lacked long-term data on the primary endpoint. A higher R ratio correlated significantly with a more favorable SMR (ρ=+0.345); the CS-SMR, R-ASMR and CS-ASMR associations were non-significant.
CONCLUSIONS: TC decisions were associated with the strength of clinical results rather than with the methodological appraisal of the composite score itself; criticism of the score was not prohibitive when efficacy evidence was convincing. The AI approach enabled high-dimensional analysis of eight years of HTA opinions. Findings are observational and association-based and require confirmation.
METHODS: We screened all TC opinions issued 2018-2025 (n=2,281). After excluding generics and biosimilars and manually retaining chronic-disease therapeutic areas likely to use composite endpoints, 864 opinions remained, of which 87 used a composite score as primary endpoint and were analyzed together with their meeting transcripts. Two large language models (GPT-5 mini, Claude Sonnet 4.5) extracted predefined fields in parallel; a third (Mistral Large) arbitrated divergences (14.8%) as quality control, yielding 853 coded statements. Each statement was a positive or negative appraisal of the composite score (CS+/CS-) or of study results (R+/R-). Per opinion, the CS+/CS- and R+/R- ratios were correlated with SMR and ASMR using Spearman's rho.
RESULTS: Result statements were predominantly positive (R+/R-=1.85), whereas composite scores were appraised more negatively (CS+/CS-=0.65). Important SMR was granted in 60% of opinions and ASMR V in 63%; only 16% combined an Important SMR with ASMR II-IV. Methodological concerns appeared in 90% of opinions; about 60% lacked a defined minimal clinically important difference and 83% lacked long-term data on the primary endpoint. A higher R ratio correlated significantly with a more favorable SMR (ρ=+0.345); the CS-SMR, R-ASMR and CS-ASMR associations were non-significant.
CONCLUSIONS: TC decisions were associated with the strength of clinical results rather than with the methodological appraisal of the composite score itself; criticism of the score was not prohibitive when efficacy evidence was convincing. The AI approach enabled high-dimensional analysis of eight years of HTA opinions. Findings are observational and association-based and require confirmation.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
HTA85
Topic
Health Technology Assessment, Methodological & Statistical Research
Topic Subcategory
Decision & Deliberative Processes