A Personalized and Evidence-Based Clinical Decision Support System Using Ensemble Learning

In clinical practice, medical guidelines and decision support tools help the clinician to ensure good and consistent patient care based on evidence-based knowledge and data science. As the next step in patient care, Digital Twins will provide a holistic patient representation and predict the best decision options available. Towards a patient-centered Digital Twin, we developed a modular, informed, interpretable, personalised, and evidence-adaptive Clinical Decision Support System (CDSS) based upon Ensemble Learning. We compare our approach for different datasets and show the ability to handle medically-provided feature separations, missing data, and the inclusion of evidence-based guidelines. Using the Cleve-land Heart Disease dataset, we reached AUCs of 0.94 (95%CI 0.91-0.97), which is similar to other publications, without extensive preprocessing. Comparing information-based subsets with medical procedure-based subsets, we only find marginal differences with AUCs of 0.92 (0.90-0.94) and 0.94 (0.91-0.96), respectively, emphasizing the variability of our approach. On the TCGA Glioma dataset, we exemplify the benefit of Informed Machine Learning by including guidelines. This results in an AUC of 0.93 (95%CI 0.92-0.95), which is an improvement over the guidelines and other published work. There is also improvement in accuracy over a purely data-driven ensemble, resulting in better accuracy and more balanced precision and recall. We simulated the ensembles behavior under missing feature subspaces and showed consistent improvement over its best performing Base Model. Hence, our CDSS architecture qualifies as a building block for a holistic, patient-journey-representative Digital Twin architecture.Clinical relevance- We provide a modular, informed, interpretable, personalized and evidence-adaptive Clinical Decision Support System, achieving AUCs over 90% on two independent tasks.