OBJECTIVES: We present a novel causal tool for the characterization of sub-populations demonstrating higher affinity to one treatment over another.

METHODS: The tool consists of machine learning, causal inference, and visualization modules developed by our team and tested on observational data. Using Truven MarketScan® database, we defined a cohort of patients undergoing total hip/knee arthroplasty (THA/TKA) and treated with anticoagulants – either enoxaparin or direct Xa inhibitors (“xabans”) - for prophylaxis against venous thromboembolism (VTE). The outcome was defined as either VTE or major bleeding during the three months post-surgery.

RESULTS: Our tool was demonstrated on the task of identifying and characterizing a sub-population that better responds to “xabans” versus enoxaparin. The cohort comprised of 90,000 patients, randomly divided into train and test sets (63,000 and 27,000 patients, respectively). Utilizing our tool, 34 candidate variables were extracted and used in model training. Of all variables, four effect modifiers were identified by our tool (THA/TKA, previous major bleeding, number of surgical visits, previous use of "xabans") and their weights were estimated. Using those weights, the tool calculates a score for each patient, representing their affinity to “xabans” over enoxaparin. The user may set different score thresholds, using interactive visualization, determining what is considered high affinity. That threshold impacts the benefitting population’s size and odds ratio. The high/low affinity populations are displayed in a parallel coordinates chart, where their characterizations may be compared. An additional visualization shows patients conversion potential – which patients currently on one treatment would benefit from switching to the alternative.

CONCLUSIONS: Our tool enables the introduction of many candidate variables into the analysis, resulting in a small set of variables which are the causal effect modifiers. This, along with the interactive visualization, makes the model and its results easier to interpret, and the sub-populations easier to characterize.