DISCUSSION
The HCM-AF-Risk Model is the first machine learning-based method
for the identification of AF cases and clinical features associated with
higher/lower risk of AF in HCM, using electronic health record data. In
our model, individual patient data is represented as an N-dimensional
vector, and the model output is a probability score for AF (AF risk) in
HCM. We identified 18 clinical variables that are highly associated
(positively/negatively) with AF in HCM patients. In addition to age,
NYHA class, LA size and LV fibrosis that have been previously associated
with AF in HCM, we found additional clinical features such as LV
diastolic dysfunction/lower LV-systolic strain that are positively
associated with AF, and greater exercise capacity that is negatively
associated with AF in HCM.
HCM-AF-Risk Model : Employing a statistical machine learning
method is advantageous as it allows automatic quantification of the
likelihood of an event (AF, in this case) based on the combination of
eature values - as obtained from the patient’s electronic health records
- and their level of association with the event. Moreover, unlike
traditional rule based models, machine learning methods are robust in
the face of new data, as these methods can tune and update their
parameters, which govern the classification algorithm based on the added
data. Thus, machine learning methods are well-suited for use in the
clinical setting we additional patients’ data is frequently accumulated.
In contrast to the majority of current ‘black-box’ machine learning
methods that are based on artificial neural networks [29, 37, 38]
and whose output decision typically cannot be explained, our method is
based on modeling a clear probabilisitic decision process, that can be
tracked back and used to justifiy the decision – we believe that this
is a critical aspect when using machine learning for supporting clinical
decisions. Our HCM-AF-Risk Model addresses data imbalance, and
utilizes a set of 18 clinical variables to identify AF cases, and
clinical features associated with higher/lower risk for AF in HCM
patients.
We note that heart failure, along with VT/VF and stroke, were not
included in the list of clinical variables considered by our method.
This is because our goal is to identify demographic, clinical, and
imaging features that
predict adverse
outcomes (AF in this case) in HCM patients, and using such adverse
outcomes as predictors defeats this purpose. However, as several
previous studies including the Framingham Heart Study,[8] ARIC
[10] and CHARGE-AF [9] have shown heart failure to be a
predictor of risk for AF, we assessed the performance of our model while
including heart failure. Inclusion of heart failure did not increase AUC
or sensitivity, but led to a slight increase in specificity of our
model, from 0.72 to 0.73.
Clinical predictors of AF in HCM using the HCM-AF-Risk Model:Left atrial diameter is the strongest predictor of AF in our study. The
association between LA size and AF has been extensively documented in
the general population[39-43] and HCM patients.[2, 44-48] The
association between LA enlargement and AF has been attributed to
stretch-induced LA structural and electrophysiologic remodeling.[49]
In the case of HCM, since most causal HCM mutations are expressed in
both atrial and ventricular myocytes, atrial myopathy and LV diastolic
dysfunction could underlie the high prevalence of AF in HCM.
Our HCM-AF-Risk Model indicates an association between diastolic
dysfunction and AF in HCM. We found that higher values for E/A,
E/e′[50] and lower (worse) global diastolic strain rate reflecting
greater degree of diastolic dysfunction are associated with higher risk
for AF in HCM. Similar results have been reported in studies conducted
in non-HCM patients.[39, 51, 52] The mechanism whereby diastolic
dysfunction has been proposed to predispose to AF is by increasing LA
preload (stretch), afterload and wall stress (dilation), which lead to
ion channel remodeling, fibrosis and increase susceptibility for
reentrant arrhythmias such as atrial fibrillation/flutter. [51]
Left ventricular fibrosis (LV-LGE) and worse LV global longitudinal peak
systolic strain rate, which reflect greater degree of LV myopathy are
associated with AF in our model. Several previous studies have detected
an association between LV fibrosis and AF in HCM.[53-55] A recent
CMR study in HCM patients reported greater amounts of LA fibrosis in HCM
patients with PAF, as well as a positive association between atrial and
ventricular fibrosis (LGE).[56] Since fibrosis slows conduction and
predisposes to reentry, LA fibrosis would be expected to increase risk
for AF.
Lower exercise capacity, lower chronotropic response/heart rate
recovery, abnormal BP response to exercise and lower diastolic BP at
peak exercise are associated with higher risk for AF in our study.
Similar results of exercise intolerance in HCM patients with PAF have
been reported in a previous study of 265 HCM patients during sinus
rhythm[57] – here, the authors did not observe an association
between lower exercise capacity and diastolic dysfunction or LA volume.
Additionally, ECHO[58] and CMR[5, 56] studies in HCM patients
have revealed greater impairment of LA function and greater degree of LA
fibrosis in HCM patients with PAF, suggesting that PAF is a marker of LA
myopathy.
One mechanism underlying reduced exercise capacity in HCM patients (with
PAF), even during sinus rhythm[57] could be impairment of LV
hemodynamics in the setting of LA myopathy, since the LA modulates LV
performance by its reservoir function during ventricular systole,
conduit function during early ventricular diastole and booster pump
function durimg late ventricular diastole. A second possibility is
higher pulmonary capillary wedge pressure (PCWP) in HCM patients with
AF, based on results of a study in 123 patients who underwent
simultaneous left and right heart catheterization, where PCWP was higher
than LV end-diastolic pressure (LVEDP) among AF patients and lower than
LVEDP among patients in sinus rhythm.[59] Other contributors to
lower exercise capacity in HCM patients with AF include sympathovagal
imbalance[60] leading to systemic vasodilation, chronotropic
incompetence induced by atrial remodeling/medications, and greater
degree of LV myopathy.
Comparison of predictors for atrial fibrillation and ventricular
arrhythmias identified by the HCM-AF-Risk and HCM-VAr-Risk Models: In
an earlier study[14] we developed a machine-learning based model for
predicting lethal ventricular arrhythmias (VT/VF) in HCM patients. We
identified 5 predictors (exercise time, METs, E/e′ ratio, LV
global longitudinal peak systolic strain rate and LV global longitudinal
early diastolic strain rate) that are common in the two models and13 variables associated with AF, but not VT/VF (Supplementary
Table 3).
Higher age is associated with increased risk for AF, but lower risk for
VT/VF, which has been confirmed by other studies.[2, 61] HCM type
(non-obstructive), family history of HCM or sudden cardiac death and
non-sustained VT are associated with VT/VF but not AF, which may reflect
differences in arrhythmic substrate in the LV and LA in HCM. Notably, LV
hypertrophy (max IVS thickness, IVS/PW ratio) is associated with VT/VF
but not AF – higher risk for VT/VF but not AF could be attributed to
greater degree of myocardial ischemia,[62] interstitial
fibrosis[63] and myocyte disarray [64] in the hypertrophied
LV.[65] The association of replacement fibrosis (LV-LGE) with AF but
not VT/VF could reflect the impact of greater degree of diastolic
dysfunction induced by LV fibrosis resulting in LA dilatation/remodeling
and AF. Taken together, our results suggest distinct pathophysiologic
mechanisms underlying atrial and ventricular arrhythmias in HCM
(Supplementary Table 3).