(A-193) Exploring methods for identifying heterogenous treatment effects (HTEs) for second-line type 2 diabetes medications and cardiovascular outcomes

Background: Understanding who benefits most from different classes of medication is key to personalised therapy. Traditional interaction analyses examine suspected effect-modifiers of the comparative effectiveness of treatments, such as demographic characteristics or comorbidities. Machine learning methods offer a data-driven approach to identifying effect-modifiers by modelling heterogenous treatment effects (HTEs). Many approaches use meta-learners, which are frameworks where users input “learners” (e.g. regression, random forest) for specific predictive tasks. Alternatively, causal survival forests (CSF) provide a single machine learning approach for HTE estimation.

Objectives: • To review methods for identifying effect-modifiers in comparative treatment-effectiveness studies;
• To apply selected methods to a case study assessing comparative effectiveness of sulfonylureas (SUs), dipeptidyl peptidase 4 inhibitors (DPP4is) and sodium-glucose transporter inhibitors (SGLT2is) in preventing cardiovascular outcomes.

Methods: We identified meta-learners for estimating HTEs and selected machine learning components to model outcome, propensity score, censoring mechanism and conditional average treatment effect (CATE, used to estimate HTE). We reviewed meta-learners and CSF based on computational complexity and ability to handle censoring, imbalanced treatment groups, and missing data. Methods were then applied to a case study to identify effect-modifiers of the comparative effectiveness of SUs, DPP4is and SGLT2is on preventing cardiovascular outcomes using CPRD Aurum data. Interaction analyses were first conducted using Cox models with interaction terms. Selected meta-learners and CSF were then used to model HTE at 2 years under each treatment arm. The effect of specific covariates on HTEs were visualised in partial dependency plots.

Results: The S-learner is simpler and more computationally efficient, but cannot directly model CATE or constrain HTE complexity. X- and R- learners and CSF are more computationally intensive, but can directly model the CATE and handle unbalanced treatment groups by downweighing estimations from underrepresented groups Integrating missing data approaches, such as multiple imputation, is complex. Interaction analyses provide more interpretable results and a simpler way to handle missing data, but may miss complexities in the data. Our case study showed discrepancies in method results, prompting us to draw conclusions from multiple models.

Conclusions: Currently available machine learning approaches for assessing HTE provide somewhat inconsistent results, making clinical interpretability unclear.