Introduction: Hypertrophic cardiomyopathy (HCM) is a heart muscle disease characterized by left ventricular (LV) hypertrophy without a systemic etiology and is associated with heart failure, stroke and sudden death. Disease prevalence is estimated at 1:500, but ~84% remain undiagnosed. Patients with obstructive HCM (oHCM) have dynamic obstruction of the LV outflow tract and characteristic abnormalities in arterial bloodflow patterns.
Hypothesis: Arterial pulsewaves recorded with a wearable biosensor and analyzed with machine learning algorithms could identify a signature of oHCM when compared to unaffected controls.
Methods: We compared baseline arterial pulse wave morphology, obtained by photoplethysmography using an investigational wristworn biosensor (Wavelet Health, Mtn. View, CA), from oHCM patients enrolled in a digital health substudy of PIONEER HCM (NCT02842242) to unaffected controls from a Wavelet Health database. Five minute recordings were obtained at rest, and data sets were divided into training and validation cohorts. A beat-by-beat machine learning model was developed using a predefined feature set to calculate an HCM probability score, and an optimal threshold score was determined. The model was evaluated using summary statistics and an ROC area-under-curve metric.
Results: Arterial pulsewave recordings were obtained from 14 patients with oHCM at rest and 81 unaffected controls. An oHCM machine learning classifier was developed based on 42 calculated metrics. After training and cross-validation (n=9 oHCM, n=48 control), the model achieved 98% accuracy. Application of this model to a validation cohort (n=5 oHCM, n=33 control) confirmed an increased probability in oHCM patients compared to unaffected controls (0.40 ± 0.13 vs. 0.18 ± 0.10; p=0.006). Analysis of the ROC curve in the pooled cohort shows an area under the curve of 0.88.
Conclusion: This first-of-its-kind study suggests that a signature of arterial bloodflow in oHCM can be identified with the combination of a wristworn biosensor and machine learning algorithms. These data raise the possibility of a novel approach to the non-invasive detection of oHCM.