Presentation (CTI)

OBJECTIVES: Landmarking is a time-to-event modeling framework where risk predictions are updated at predefined landmark times over fixed windows. Traditional landmarking strategies include landmark supermodeling, which pools data from multiple landmark times into a single extended dataset, and strict landmarking, which fits separate models at each landmark time. The current study additionally evaluates cumulative landmarking, a hybrid approach that progressively expands the training dataset while fitting separate models at successive landmarks. This study compares predictive performance across these landmarking strategies.
METHODS: Arkansas statewide administrative health claims data from November 2018-December 2023 were used to construct discrete-time datasets for new cannabis use disorder diagnosis prediction among medical marijuana (MMJ) cardholders. Five classifiers were evaluated with each landmarking strategy: Random Survival Forest, Support Vector Machine Survival (SVMS), Cox proportional hazards, Random Forest, and Logistic Regression. Models were trained using a 50:50 train-test split and 1:25 random undersampling. Performance was compared using 90-day, 180-day, and 360-day prediction horizons using mean cumulative sensitivity/dynamic specificity area under the receiver operating characteristic (AUC-ROC) and inverse probability of censoring weighting Brier scores (IPCWBS) with 95% confidence intervals (CIs). Both discrete-time and discrete-time-daily frameworks were evaluated for landmark supermodels, with the latter recording follow-up time on a daily scale while maintaining interval-based prediction updates, thereby more precisely representing partial follow-up within the final prediction interval.
RESULTS: Among 54,422 MMJ cardholders, SVMS landmark supermodel trained with 360-day prediction windows achieved the highest discriminative performance [AUC-ROC (95% CI)=0.7993(0.7751,0.8206)] outperforming strict [AUC-ROC (95% CI)=0.6684(0.6087,0.7238)] and cumulative landmarking [AUC-ROC (95% CI)=0.7390(0.7072,0.7680)]. Discriminative performance was comparable between discrete-time and discrete-time-daily frameworks, and calibration was modestly better for discrete-time-daily landmark supermodels (IPCWBS point-estimate range: 0.00385-0.00735) compared to discrete-time (IPCWBS point-estimate range: 0.00465-0.02139).
CONCLUSIONS: Landmark supermodeling was the most discriminating among landmarking strategies. Incorporating a discrete-time-daily representation of follow-up time yielded minor calibration improvements, particularly for longer horizons.