Exploration of the Factors Driving Decisions of the Transparency Commission Based on an Analysis of HAS Meeting Transcripts Using Natural Language Processing
Author(s)
Aballea S1, Maszke S2, Shchepetnov A3, Toumi M4
1InovIntell, Rotterdam, South Holland, Netherlands, 2InovIntell, Krakow, NA, Poland, 3InovIntell, Tbilisi, NA, Georgia, 4Aix Marseille University, Marseille, PACA, France
OBJECTIVES: Transcripts of the HAS decision committee meetings are publicly available. Our goal was to investigate the topics brought up during the meetings and how they relate to the ASMR ratings.
METHODS: 428 meeting transcripts corresponding to 309 decisions made by the Transparency Commission (TC) between 2018 and 2023 were downloaded, parsed and tokenized. First, a selection of terms associated with decision outcomes was made using the Naïve Bayes (NB) method and domain knowledge. Second, committee member interventions containing selected terms were then classified as positive, neutral or negative using the GPT 3.5 language model. Third, NB was trained to differentiate outcomes based on (term, sentiment) pairs. Finally, terms were manually grouped by topic to examine the cumulative influence of each topic.
RESULTS: Survival, toxicity, relapse, mortality, and infant were among the terms most likely to be linked to favorable ASMR. Injection, meta-analysis, placebo, and non-inferiority were among the keywords most likely to be linked to unfavorable ASMR. After categorizing keywords by topic, it was found that mentions of terms pertaining to efficacy had the greatest impact on ASMR, whether it was positive or negative, followed by terms pertaining to target population. Discussions of ease of use and mode of administration were linked to unfavorable ASMR. Indirect treatment comparison (ITC)-related interventions also had a primarily negative impact. Patient contributions and patient-reported outcomes discussions had very little impact on decisions.
CONCLUSIONS: This preliminary study corroborates expected trends: ASMR is largely influenced by discussions around efficacy; discussions around indirect treatment comparisons and convenience of use tend to lead to negative decisions; and the patient voice is hardly taken into account. This work is largely exploratory but illustrates the potential of natural language processing to help understanding how HTA agencies make decisions. Further research is suggested to better account for interactions between terms.
Conference/Value in Health Info
Value in Health, Volume 26, Issue 11, S2 (December 2023)
Code
HTA347
Topic
Health Technology Assessment, Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics, Decision & Deliberative Processes
Disease
No Additional Disease & Conditions/Specialized Treatment Areas