Comparison of an interpretable data-driven approach with state of the art classifiers: Application to cardiovascular risk assessment

Data-driven approaches can be very useful in the extraction of relevant knowledge from clinical datasets. In fact, this knowledge may have an important role in the support of clinical decision in the daily clinical practice. In the context of cardiovascular disease (the lead cause of death worldwide), it is critical to identify the risk of a patient suffering from a cardiovascular event. This allows to stratify patients according to their risk and to decide on the best patients' treatment. This work compares an innovative data driven approach based on a clustering strategy with some well-known classifiers (decision trees, discriminant analysis and neural networks). This assessment is based not only on the performance but also on the interpretability. This study was applied to patients that were admitted to the emergency service with a first event of myocardial infarction. It is supported by a real patients' dataset with Acute Coronary Syndromes, provided by the Portuguese Society of Cardiology considering the risk of death within 30 days. The validation of this work was performed based on three different metrics: sensitivity, specificity and geometric mean. The obtained results showed that the performance of the proposed clustering approach is comparable to the best classifiers while maintaining the interpretability aspects. Namely, it provided a very well balanced classifier obtaining a geometric mean value of 79%).