AI-ASSISTED SELECTION OF MODEL INPUTS IN PHARMACOECONOMIC EVALUATION: A CASE STUDY OF CARDIOVASCULAR DISEASES...
Author(s)
Li Ma, BSc1, Yifan Wu, BSc1, Jin Peng, PhD2, Siqi Zhou, MSc2, Najwa Ali Yasin, MSc3, Minghuan Jiang, PhD2.
1School of Pharmacy, Xi'an Jiaotong University, Xi’an, China, 2School of Pharmacy, Xi'an Jiaotong University, Xi'an, China, 3School of Pharmacy, Xi'an JiaoTong University, Xi'an, China.
1School of Pharmacy, Xi'an Jiaotong University, Xi’an, China, 2School of Pharmacy, Xi'an Jiaotong University, Xi'an, China, 3School of Pharmacy, Xi'an JiaoTong University, Xi'an, China.
OBJECTIVES: Core model parameters in pharmacoeconomic evaluation exhibit considerable variability across studies. This study assessed heterogeneity in key model parameters and developed a human-AI collaborative framework to support parameter identification and selection for pharmacoeconomic evaluation.
METHODS: A systematic search of CNKI and Medline was conducted to identify cardiovascular pharmacoeconomic evaluation studies published between 2021 and 2025. Eligible studies were reviewed to extract core model parameters: transition probabilities, costs, and utility values. For each parameter, data on value, health state, and original source were recorded. Candidate sources were appraised against four predefined criteria: methodological quality, timeliness, population relevance, and reporting completeness. Three AI tools, ChatGPT, DeepSeek, and Doubao, were applied to full-text source documents using standardized prompts based on the RISEN framework. Their performance was compared with manual review.
RESULTS: Seventy cardiovascular pharmacoeconomic studies were included, yielding 424 core parameters. Substantial heterogeneity was observed across studies.For AI extraction performance, ChatGPT achieved completion rates of 76.0%, 85.2%, and 98.3% for cost, utility, and transition probability parameters respectively;DeepSeek achieved 70.0%, 59.3%, and 36.7%. Regarding accuracy, both ChatGPT and DeepSeek reached 100% consistency with human extraction results, whereas Doubao performed markedly worse. For extraction speed, DeepSeek had the shortest average processing time per task (12-19s), followed by ChatGPT (27-79s), with Doubao being the slowest (97-257s). For the concordance of optimal source recommendations compared with human judgment, ChatGPT achieved 9/13, DeepSeek 8/13, and Doubao only 4/13. Based on the above findings, this study established a four-step human-AI collaborative pathway.
CONCLUSIONS: AI tools can support parameter selection in pharmacoeconomic modeling; however, expert oversight remains indispensable for adjudicating parameter applicability. The proposed human-AI framework holds potential to enhance the efficiency, transparency, and reproducibility of parameter selection in health economics and outcomes research.
METHODS: A systematic search of CNKI and Medline was conducted to identify cardiovascular pharmacoeconomic evaluation studies published between 2021 and 2025. Eligible studies were reviewed to extract core model parameters: transition probabilities, costs, and utility values. For each parameter, data on value, health state, and original source were recorded. Candidate sources were appraised against four predefined criteria: methodological quality, timeliness, population relevance, and reporting completeness. Three AI tools, ChatGPT, DeepSeek, and Doubao, were applied to full-text source documents using standardized prompts based on the RISEN framework. Their performance was compared with manual review.
RESULTS: Seventy cardiovascular pharmacoeconomic studies were included, yielding 424 core parameters. Substantial heterogeneity was observed across studies.For AI extraction performance, ChatGPT achieved completion rates of 76.0%, 85.2%, and 98.3% for cost, utility, and transition probability parameters respectively;DeepSeek achieved 70.0%, 59.3%, and 36.7%. Regarding accuracy, both ChatGPT and DeepSeek reached 100% consistency with human extraction results, whereas Doubao performed markedly worse. For extraction speed, DeepSeek had the shortest average processing time per task (12-19s), followed by ChatGPT (27-79s), with Doubao being the slowest (97-257s). For the concordance of optimal source recommendations compared with human judgment, ChatGPT achieved 9/13, DeepSeek 8/13, and Doubao only 4/13. Based on the above findings, this study established a four-step human-AI collaborative pathway.
CONCLUSIONS: AI tools can support parameter selection in pharmacoeconomic modeling; however, expert oversight remains indispensable for adjudicating parameter applicability. The proposed human-AI framework holds potential to enhance the efficiency, transparency, and reproducibility of parameter selection in health economics and outcomes research.
Conference/Value in Health Info
2026-09, ISPOR Asia Pacific 2026, Bangkok, Thailand
Value in Health, Volume 55, Issue S1
Code
MSR21
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
No Additional Disease & Conditions/Specialized Treatment Areas, SDC: Cardiovascular Disorders (including MI, Stroke, Circulatory)