Example implementations described herein are directed to the problem of dynamically extracting valid sets of clinical treatments and predictive features from electronic health records to construct models that can predict the effect of treatments and hence compare the effect of multiple, putative treatments. By utilizing a data pipelining technique for constructing machine learning models that only utilizes sets of valid treatment options instead of all possible options, the hardware and computational resources required for constructing the machine learning models can thereby be reduced, and the predicted treatment transition outcomes can be traced to valid treatments, thereby allowing the clinician to understand the effects from a clinical perspective.