| Physicians are constantly faced with making decisions under uncertainty, and despite the extraordinary advancements in the field of clinical informatics, there is a significant void about how to build simple and trustworthy clinical decision support systems. This dissertation focusses on investigating whether a hybrid recommender framework approach can exceed conventional data analysis techniques in order to provide physicians with accurate insights. The research questions explored a novel hybrid recommender framework that improves upon common clinical recommendation practices such as data driven, case base reasoning and machine learning techniques by integrating them into a unified data model. Conceptually this study was framed within theories of probability, numerical analysis, case base reasoning, machine learning, clinical decision support and recommendation systems. Experiments demonstrate that the proposed hybrid recommender framework is more accurate and effective than common baseline techniques. We evaluate the framework by implementing a prototype and experimenting with an outstanding clinical problem: how to reduce the number of unnecessary pre-operative blood tests for pediatric neurosurgical patients. We analyze heterogeneous databases containing 359,475 patient encounters at The Children's Hospital of Philadelphia from 2001 to 2014. Experimental analysis shows that our hybrid approach has a sensitivity of 0.80, a specificity of 0.85 and a mean absolute error of 0.875.;Finally, we demonstrate preliminary results of a real-world implementation by embedding the recommendations into the physician's workflow in the production environment of the hospital's electronic health record. The application shows a reduction of ordering unnecessary tests by ∼ 25% in the first quarter of 2016 and a 100% adoption rate by the user base. This result suggests that our approach helps in improving the quality of physician's decisions with a positive impact on outcomes. |
