RESULTS
Patient population : We studied 831 patients with a clinical diagnosis of HCM. Atrial fibrillation was diagnosed in 22% of the HCM cohort: 139 patients were diagnosed with AF prior to/at the first clinic visit, and 52 patients were diagnosed with AF during follow up (Supplementary Figure 1). Prevalence of AF varied from 9-30% and increased with age; AF prevalence was highest in the age group of 61-80 years (Figure 1).
Demographic, clinical and imaging features of the HCM cohort at the time of their first clinic visit are presented in Table 4 Patients from the AF group were older, more likely to have obstructive HCM, higher NYHA class, and lower exercise capacity than the No-AF group. The AF group also had larger LA size (Figure 2), greater diastolic dysfunction, worse global longitudinal strain and greater amount of LV replacement fibrosis (reflected by LV-LGE), when compared to the No-AF group, suggesting greater degree of LV myopathy. No difference was observed in LV mass, maximum LV thickness or LVOT gradients between the AF and No-AF groups of HCM patients (Table 4).
Mean follow up was 3.1 years (median = 2.1, 25th-75th percentile = 1.0 - 4.8 years). HCM patients from the AF group had a higher incidence of heart failure and all-cause death, when compared to the No-AF group (Table 4).
Machine learning-based identification of atrial fibrillation cases: Our feature selection process identified 18 clinical variables whose values distinguish AF from No-AF cases within the HCM population. Table 1 provides a list of these predictive variables, along with the corresponding polychoric correlation and P values, indicating their degree of association (or lack thereof) with AF. We identified 7 variables that are negatively correlated with AF, and 11 variables that are positively associated with AF. Left atrial diameter is highly correlated with AF. Several exercise-related parameters, including, lower exercise capacity (reflected by lower METs, exercise time, peak stress heart rate), abnormal BP response to exercise, lower diastolic BP at peak exercise and lower heart rate recovery post-exercise are associated with higher risk for AF. Other predictors of AF include replacement fibrosis in the LV (reflected by LV-LGE), greater diastolic dysfunction (reflected by higher E/A and E/e′, lower LV-SR_E) and worse (more positive) global longitudinal systolic strain rate (LV-SR_S).
Notably, combining the ensemble classifier comprising logistic regression and naïve Bayes with over- and under-sampling led to higher sensitivity and area under the receiver operating curve (AUC), compared to the four simple classifiers (naïve Bayes, logistic regression, decision tree and random forest) alone (Table 2). Figure 3 illustrates the C-index (0.80 ) for our method (HCM-AF-Risk Model ), which assigns an individualized probability to each patient to present with AF.
Comparison of HCM-AF-Risk Model performance with previous AF models : We compared the performance of the HCM-AF-Risk Modelwith that obtained by the Framingham Heart Study,[8] ARIC[10] and CHARGE-AF[9] risk models (Tables 3, 5). The HCM-AF-Risk Model demonstrates significantly higher performance (p < 0.001) across all evaluation metrics including specificity, sensitivity and area under ROC curve (C-index) for HCM patients, when compared with published models [8-10] focused on AF prediction in the general population. The datasets used in these studies are not publicly available, which preclude their use for training/testing on our dataset and comparing their performance according to all the measures we have used. As such, we compared the performance level attained by our model with that reported by the other studies in terms of the C-index.
We also repeated our experiments, using only LA diameter to represent our dataset, based on results from several previous studies [2, 35, 36] that identified LA enlargement as most predictive of AF in HCM. We observed reduction in our model’s performance when LA diameter alone is included in the feature set: specifically, area under ROC curve decreased by 20% (C-index: 0.66 from 0.80), sensitivity decreased by 37% (0.54 from 0.74) and specificity decreased by 14% (0.63 from 0.72).