Development of a Discrete-Time-Updating Algorithm to Predict Cannabis Use Disorder Among Arkansas Medical Marijuana Cardholders
Author(s)
Allen M. Smith, PharmD, Horacio Gomez-Acevedo, PhD, Corey J. Hayes, MPH, PharmD, PhD, Melody Greer, PhD, Bradley C. Martin, RPh, PharmD, PhD;
University of Arkansas for Medical Sciences, Little Rock, AR, USA
University of Arkansas for Medical Sciences, Little Rock, AR, USA
Presentation Documents
OBJECTIVES: A discrete-time-updating algorithm was developed to predict cannabis use disorder (CUD) risk within the next 90 days among Arkansas medical marijuana (MMJ) cardholders without a history of CUD.
METHODS: Statewide health insurance claims data linked to MMJ cardholder data between November 2018 - December 2023 was utilized. Subjects were followed from the date they could first legally purchase MMJ in Arkansas until they received a new CUD diagnosis or were lost to follow-up. A person-period dataset with 90-day time intervals was then constructed with a binary indicator (1: CUD|0: No CUD) for each interval. Features were engineered from demographics, comorbidities, and prescription and healthcare utilization characteristics. Prediction of CUD in each time interval was informed by the previous 6 months of features for each subject. Data were randomly split 70:30 at the subject-level into training and test sets. A 1:100 random undersampling was performed on the person-period training set followed by hyperparameter tuning with 10-fold cross validation for training random survival forest (RSF), Cox proportional hazards (CPH), random forest (RF), and logistic regression (LR) models. Model performance was evaluated on the complete test set using the cumulative/dynamic area under the receiver-operating characteristic (C/D AUC-ROC) curve.
RESULTS: 54,422 Arkansas MMJ cardholders met eligibility and 1,479 received a new CUD diagnosis during the study period. With a mean C/D AUC-ROC of 0.832, the LR-model achieved the highest discrimination in predicting CUD in the next 90 days, followed by the CPH-model (mean C/D AUC-ROC=0.792), RSF-model (mean C/D AUC-ROC=0.737), and RF-model (mean C/D AUC-ROC=0.708).
CONCLUSIONS: A discrete-time-updating prediction tool achieved high levels of discrimination for predicting a new CUD diagnosis among Arkansas MMJ cardholders. This tool could be used to identify persons authorized to purchase MMJ that are at risk of developing CUD to develop targeted screening and treatment programs.
METHODS: Statewide health insurance claims data linked to MMJ cardholder data between November 2018 - December 2023 was utilized. Subjects were followed from the date they could first legally purchase MMJ in Arkansas until they received a new CUD diagnosis or were lost to follow-up. A person-period dataset with 90-day time intervals was then constructed with a binary indicator (1: CUD|0: No CUD) for each interval. Features were engineered from demographics, comorbidities, and prescription and healthcare utilization characteristics. Prediction of CUD in each time interval was informed by the previous 6 months of features for each subject. Data were randomly split 70:30 at the subject-level into training and test sets. A 1:100 random undersampling was performed on the person-period training set followed by hyperparameter tuning with 10-fold cross validation for training random survival forest (RSF), Cox proportional hazards (CPH), random forest (RF), and logistic regression (LR) models. Model performance was evaluated on the complete test set using the cumulative/dynamic area under the receiver-operating characteristic (C/D AUC-ROC) curve.
RESULTS: 54,422 Arkansas MMJ cardholders met eligibility and 1,479 received a new CUD diagnosis during the study period. With a mean C/D AUC-ROC of 0.832, the LR-model achieved the highest discrimination in predicting CUD in the next 90 days, followed by the CPH-model (mean C/D AUC-ROC=0.792), RSF-model (mean C/D AUC-ROC=0.737), and RF-model (mean C/D AUC-ROC=0.708).
CONCLUSIONS: A discrete-time-updating prediction tool achieved high levels of discrimination for predicting a new CUD diagnosis among Arkansas MMJ cardholders. This tool could be used to identify persons authorized to purchase MMJ that are at risk of developing CUD to develop targeted screening and treatment programs.
Conference/Value in Health Info
2025-05, ISPOR 2025, Montréal, Quebec, CA
Value in Health, Volume 28, Issue S1
Code
MSR68
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
SDC: Systemic Disorders/Conditions (Anesthesia, Auto-Immune Disorders (n.e.c.), Hematological Disorders (non-oncologic), Pain)