Can machine learning help to identify heart failure with preserved ejection fraction?

Abstract : Purpose: No consensus exists for the diagnosis of heart failure with preserved ejection fraction (HFPEF). Current studies recommend stress echocardiography protocols, but single standard peak or timing measurements are limited towards the complexity of the disease. We show the contribution of a combined analysis of multiple myocardial velocity patterns and investigate their relative relevance for the characterization of HFPEF. Methods: Velocity traces from 55 subjects were examined (69±6 years, 22 healthy, 19 HFPEF and 14 breathless subjects). Data came from tissue Doppler acquisitions at rest and exercise at the basal septum and lateral wall, and were temporally aligned to a common reference for comparison. Each phase of the cycle at each stage of the protocol was identified for analysis (figure). Unsupervised machine learning (multiple kernel learning) was used to characterize this population and automatically determine the relevance of each velocity pattern. Results: The learning found a discrimination algorithm that performed well in agreement with diagnosis based on current guidelines (sens=78.9%, spec=86.3%, K=0.65). The importance given to early diastole at exercise was substantially higher, while the isovolumic contraction was the lowest contributor. The breathless subjects were associated to their closest subgroup. Conclusion: The characterization of HFPEF is improved by a combined analysis of multiple velocity traces from stress echo studies using machine learning. The method additionally suggests features of interest to be used in clinical diagnosis.