Temporal Modeling in Clinical Artificial Intelligence, Decision-Making, and Cognitive Computing: Empirical Exploration of Practical Challenges

Temporal modeling holds great promise for healthcare, where treatment decisions must be made over time, and where continually re-evaluating ongoing treatment is critical to optimizing clinical care for individual patients. Tremendous advances have been made in data mining and temporal modeling of healthcare data, but practical challenges exist in moving these advances from the laboratory/theoretical setting to applied settings with real patients. In this paper, we address a number of these challenges. First, we provide empirical evidence for calculating the optimal trade-off between costs and outcomes in temporal modeling, suggesting that it may be a dynamical system of relative values of costs and effects between treatment actions (rather than absolute values). Such an approach may allow optimal reward functions to be derived from clinical data. Second, we evaluate the effects of finite horizon levels on both cost effectiveness and outcome change. Finally, we provide a proof-ofconcept application for integrating machine-learningclassifier-based (ML) transition models into temporal models (e.g. Markov Decision Processes). The results showed that even a relatively poor classifier can produce small gains in performance and highlights the potential of such an approach for further exploration. Individualized transition models via such ML integration provide a potential practical avenue for implementation of personalized medicine approaches in EHRs and realworld clinical practice. We also discuss a number of future directions for research, such as inclusion of patient safety and treatment non-adherence, and temporal modeling of the clinical process as a basis for cognitive