Utilizing Probabilistic Topic Model To Detect And Analyze The Evolving Trends Of Medical Behaviors

Clinical pathways (CPs), as multidisciplinary care plans that details the care steps in appropriate time stamps to help patients with a specific condition or a set of symptoms, are playing significant roles in the delivery of high quality and cost-effective health services. Medical behaviors keep improving with the continuous development of medical technologies. As well, medical strategies keep changing over time, which will end up to changes of medical behaviors in the pathway. Timely discovery of the evolvement patterns of treatment behaviors in CPs is beneficial for the improvement of CPs. It should be noted that most of existing approaches to generate "simple" reports on medical behaviors change trends are based on experiences and knowledge of clinical experts, or are oriented to clinical data statistical analysis. In such techniques, analysts interpret large amounts of collected medical behaviors, and elaborate patient cases, piece after piece. Obviously, it can be a very time-consuming and tedious process. In this regard, this study proposes an automatic approach, which based on probabilistic topic model and statistical analysis, to discover and analyze the evolving trends of significant medical behaviors.In detail, the proposed approach is consisted of 4 steps. Firstly, based on the established clinical term dictionary, all the medical behaviors in patient care journeys, extracted from the electronic medical records (EMRs), are mapped into unified abstraction levels. Then, we apply a probabilistic topic model, i.e., Latent Dirichlet allocation (LDA), to disclose yearly treatment patterns with regard to the risk stratification of patients from a large volume of EMRs. Subsequently, the evolving trends of yearly occurring probabilities of medical behaviors in corresponding treatment patterns can be clustered into six content evolvement patterns (i.e., growing, declining, up-down, down-up, steady, and jumping) and three occurring time evolvement patterns (i.e., early-implemented, steady-implemented and delay-implemented). Finally, a significance analysis method, i.e., run test is adopted to discover the significantly changed medical behaviors of respective treatment patterns. Once significantly changed medical behaviors have been derived, the relevancy of these medical behaviors, computed from correlation analysis method, can be categorized into six correlation patterns (i.e., linearly correlated, significantly correlated, highly correlated, weakly correlated, slightly correlated).The proposed method was evaluated via a real clinical dataset pertaining to the unstable angina pathway with time arranges of 10 years. The experimental results indicate that the proposed approach is efficient to mine the significant changes of medical behaviors and execution time of treatment interventions in CPs, as well as the correlations between these medical behaviors. Thus these detected changes and correlation analysis results can be potentially exploited to serve for CP redesign and improvement.