Machine learning based decision support systems (DSS) for heart disease diagnosis: a review

The current review contributes with an extensive overview of decision support systems in diagnosing heart diseases in clinical settings. The investigators independently screened and abstracted studies related to heart diseases-based clinical decision support system (DSS) published until 8-June-2015 in PubMed, CINAHL and Cochrane Library. The data extracted from the twenty full-text articles that met the inclusion criteria was classified under the following fields; heart diseases, methods for data sets formation, machine learning algorithms, machine learning-based DSS, comparator types, outcome evaluation and clinical implications of the reported DSS. Out of total of 331 studies 20 met the inclusion criteria. Most of the studies relate to ischemic heart diseases with neural network being the most common machine learning (ML) technique. Among the ML techniques, ANN classifies myocardial infarction with 97% and myocardial perfusion scintigraphy with 87.5% accuracy, CART classifies heart failure with 87.6%, neural network ensembles classifies heart valve with 97.4%, support vector machine classifies arrhythmia screening with 95.6%, logistic regression classifies acute coronary syndrome with 72%, artificial immune recognition system classifies coronary artery disease with 92.5% and genetic algorithms and multi-criteria decision analysis classifies chest-pain patients with 91% accuracy respectively. There were 55% studies that validated the results in clinical settings while 25% validated the results through experimental setups. Rest of the studies (20%) did not report the applicability and feasibility of their methods in clinical settings. The study categorizes the ML techniques according to their performance in diagnosing various heart diseases. It categorizes, compares and evaluates the comparator based on physician's performance, gold standards, other ML techniques, different models of same ML technique and studies with no comparison. It also investigates the current, future and no clinical implications. In addition, trends of machine learning techniques and algorithms used in the diagnosis of heart diseases along with the identification of research gaps are reported in this study. The reported results suggest reliable interpretations and detailed graphical self-explanatory representations by DSS. The study reveals the need for establishment of non-ambiguous real-time clinical data for proper training of DSS before it can be used in clinical settings. The future research directions of the ML-based DSS is mostly directed towards development of generalized systems that can decide on clinical measurements which are easily accessible and assessable in real-time.