Using Machine Learning to Identify Heterogeneous Benzodiazepine Use Patterns Following Long-Term Benzodiazepine Therapy
Author(s)
Kitiyaporn Takham, MS1, Weihsuan Jenny Lo-Ciganic, MS, PhD2, Timothy Anderson, MD,MAS2, Tae Woo Park, MD, MSc2.
1Carnegie Mellon University, Pittsburgh, PA, USA, 2University of Pittsburgh School of Medicine, Pittsburgh, PA, USA.
1Carnegie Mellon University, Pittsburgh, PA, USA, 2University of Pittsburgh School of Medicine, Pittsburgh, PA, USA.
OBJECTIVES: Clinical guidelines discourage long-term benzodiazepine therapy (LTBT) and recommend gradual tapering post-LTBT. However, real-world use patterns following LTBT remain poorly characterized. We applied a machine-learning approach to identify post-LTBT benzodiazepine use trajectories and assessed their associations with drug overdose and suicide-related outcomes.
METHODS: We conducted a retrospective cohort study using 2014-2024 data from the All of Us Research Program, a large, diverse U.S. cohort designed to support precision medicine. We included all non-cancer adults (≥18 years) who initiated benzodiazepines. LTBT was defined as ≥3 benzodiazepine orders/fills at least 21 days apart, with ≥84 days supplied. Greedy k-means++ clustering was used to identify 12-month benzodiazepine use trajectories starting from the first day of each LTBT episode. We compared patient characteristics and occurrence of drug overdose and suicide-related emergency department (ED) visits or hospitalizations across trajectories using descriptive statistical methods.
RESULTS: Among 9,449 patients meeting the LTBT definition (age≥65=40.4%, female=72.6%, White=73.3% and Black=9.9%), we identified 4 benzodiazepine use trajectories: Group (1) persistent high-adherence use (53.6% of the cohort); (2): discontinuation within 180 days (18.3%); (3): discontinuation between 180-270 days (10.3%); and (4): intermittent use (14.8%). Compared with Group 1, Group 2 was more likely to be younger and have anxiety disorders; Group 3 was more likely to have anxiety and alcohol use disorders; Group 4 was less likely to have bipolar and opioid use disorders (p<0.05). There was no statistically significant difference in the occurrence of drug overdose (Group 1:0.18%; Group 2:0.35%; Group 3:0.32%; Group 4:0.00%) and suicide-related ED visits and hospitalizations (Group 1:0.14%; Group 2:0.29%; Group 3:0.48%; Group 4:0.14%) during the 12-month trajectory measurement period.
CONCLUSIONS: Machine learning revealed four distinct benzodiazepine use patterns post-LTBT, each associated with different characteristics. More research is needed to inform benzodiazepine deprescribing strategies that might enhance benefits and minimize harms of deprescription.
METHODS: We conducted a retrospective cohort study using 2014-2024 data from the All of Us Research Program, a large, diverse U.S. cohort designed to support precision medicine. We included all non-cancer adults (≥18 years) who initiated benzodiazepines. LTBT was defined as ≥3 benzodiazepine orders/fills at least 21 days apart, with ≥84 days supplied. Greedy k-means++ clustering was used to identify 12-month benzodiazepine use trajectories starting from the first day of each LTBT episode. We compared patient characteristics and occurrence of drug overdose and suicide-related emergency department (ED) visits or hospitalizations across trajectories using descriptive statistical methods.
RESULTS: Among 9,449 patients meeting the LTBT definition (age≥65=40.4%, female=72.6%, White=73.3% and Black=9.9%), we identified 4 benzodiazepine use trajectories: Group (1) persistent high-adherence use (53.6% of the cohort); (2): discontinuation within 180 days (18.3%); (3): discontinuation between 180-270 days (10.3%); and (4): intermittent use (14.8%). Compared with Group 1, Group 2 was more likely to be younger and have anxiety disorders; Group 3 was more likely to have anxiety and alcohol use disorders; Group 4 was less likely to have bipolar and opioid use disorders (p<0.05). There was no statistically significant difference in the occurrence of drug overdose (Group 1:0.18%; Group 2:0.35%; Group 3:0.32%; Group 4:0.00%) and suicide-related ED visits and hospitalizations (Group 1:0.14%; Group 2:0.29%; Group 3:0.48%; Group 4:0.14%) during the 12-month trajectory measurement period.
CONCLUSIONS: Machine learning revealed four distinct benzodiazepine use patterns post-LTBT, each associated with different characteristics. More research is needed to inform benzodiazepine deprescribing strategies that might enhance benefits and minimize harms of deprescription.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
MSR216
Topic
Methodological & Statistical Research, Study Approaches
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
No Additional Disease & Conditions/Specialized Treatment Areas