Artificial intelligence–enhanced electrocardiography and sudden death risk factors in cardiac arrest survivors diagnosed with hypertrophic cardiomyopathy

Abstract

Background

The characteristics of patients with hypertrophic cardiomyopathy (HCM) who present with sudden cardiac arrest as the initial manifestation of the disease are not well known.

Objective

The purpose of this study was to evaluate artificial intelligence–enhanced electrocardiography (AI-ECG) and sudden cardiac death (SCD) risk factors in patients with HCM presenting with sudden cardiac arrest (SCA).

Methods

We identified patients within the Mayo Clinic enterprise (2001–2022) who were newly diagnosed with HCM after surviving SCA. Index clinical, electrocardiographic, and imaging characteristics were documented. Risk models for SCD in HCM were retrospectively applied. An AI-ECG algorithm designed for the detection of HCM was also applied to the patients’ first ECG obtained at our institution.

Results

Twenty-seven patients met the inclusion criteria. Eight patients (30%) had documented cardiac symptoms preceding SCA. Twenty-four patients (89%) had an abnormal electrocardiogram at index evaluation, and 16 (59%) had ≥1 SCD risk factor that could have qualified them for an implantable cardioverter-defibrillator if HCM had been diagnosed before SCA. Retrospective application of the European Society of Cardiology 5-year HCM Risk-SCD tool yielded a median score of 4.6% (interquartile range 3.2%–7.2%), with 12 patients (44%) having an estimated SCD risk of <4%, implying a low risk. AI-ECG indicated a high prediction score for HCM in 26 patients (96%), suggesting its potential utility as an early detector of the disease if applied before SCA.

Conclusion

Among patients with SCA who were then newly diagnosed with HCM, conventional SCD risk factors were common but not universal in this post hoc assessment. AI-ECG may facilitate the early detection of HCM.

Key Findings.

• ▪Sudden cardiac arrest is an infrequent initial presentation of hypertrophic cardiomyopathy (HCM).
• ▪Abnormal electrocardiograms (ECGs) and preceding cardiac symptoms are common among the survivors of HCM-related sudden cardiac arrest.
• ▪Artificial intelligence–enhanced ECG analysis demonstrated excellent performance in detecting HCM and may represent a paradigm for earlier detection of HCM in broader screening settings.
• ▪In post hoc analysis, more than half of the patients had traditional risk factors for sudden cardiac death that would have warranted consideration of a prophylactic implantable cardioverter-defibrillator if HCM had been diagnosed earlier.

Introduction

Hypertrophic cardiomyopathy (HCM) is one of the most common genetic heart diseases, characterized by pathologic and otherwise unexplained left ventricular hypertrophy (LVH).¹ Disease prevalence has been previously estimated at ∼1 in 500 in the general population, with more recent estimates as high as 1 in 200.² The recent development of artificial intelligence–enhanced electrocardiography (AI-ECG) tools may allow earlier detection of HCM in patients undergoing standard 12-lead ECG. Furthermore, with advances in sudden cardiac death (SCD) risk stratification, HCM therapies, and targeted use of implantable cardioverter-defibrillators (ICDs), HCM-related mortality in adults is comparable to that in the general US population, although younger individuals with HCM continue to experience higher mortality.³ Although most patients with HCM have a good quality of life and overall prognosis, arrhythmic SCD remains one of the most common causes of HCM-related mortality in contemporary cohorts with HCM, and HCM is one of the most commonly identifiable causes of SCD in young people in North America.^4,5

As symptoms of HCM may be absent early in the disease course, many diagnoses are incidental after an abnormal physical examination or screening ECG. Among patients with SCD possibly caused by HCM, as determined by postmortem examination, the majority were not known to have HCM before SCD.⁶ Thus, the true number of patients with HCM who experience SCD is unknown. Some patients may survive an out-of-hospital sudden cardiac arrest (SCA) and then undergo evaluation leading to an HCM diagnosis, allowing a retrospective assessment of SCD risk.

In this study, we sought to assess (1) how often an AI-ECG algorithm could have flagged a high prediction score for HCM as a measure of early disease detection in patients who survived SCA and were subsequently diagnosed with HCM and (2) the prevalence of SCD risk factors and the post hoc performance of existing SCD risk estimators in survivors of HCM-related SCA.

Methods

Study population

This study was approved by the Mayo Clinic Institutional Review Board, and the research conducted adhered to the Declaration of Helsinki. Only patients with available consent on file for research use of their data were included. We identified patients with HCM within the Mayo Clinic enterprise from 2001 to 2022 using a validated approach of natural language processing of electronic medical records, cardiac imaging reports, and diagnosis codes. Among all patients with HCM, this study included those in whom HCM was newly diagnosed during evaluation after resuscitated SCA. We included both patients who presented with SCA and received their initial HCM diagnosis at our institution as well as those referred for subsequent HCM evaluation at our institution after initial SCA evaluation elsewhere. The long-term outcomes of this cohort were recently reported separately.⁷

Data collection

Demographic, genetic, imaging, and clinical information were collected through review of the electronic medical record. Comorbidities present at the time of SCA were documented. Imaging information was obtained from the first transthoracic echocardiogram (TTE) at our institution after SCA. If the first TTE was recorded while the patient was still hemodynamically unstable in the early postarrest period, we selected the next available TTE.

ECG and AI algorithm

As no patients had pre-SCA ECGs for analysis, all analyses were performed using the first ECG recorded at our institution and not necessarily in the immediate postarrest period. A set of predetermined characteristics, including normal ECG, presence of LVH (per Sokolow-Lyon criteria), pathologic Q waves, inferior or lateral T-wave inversions, or nonspecific repolarization abnormality, were analyzed. We did not include ECGs performed sooner than 30 days after SCA to avoid potential confounding conditions related to critical illness.

We documented the prediction score for HCM as determined by a deep-learning AI-ECG algorithm for the detection of HCM that was previously developed at our institution.⁸ As previously reported, this AI-ECG algorithm was developed on the basis of a data set of patients with HCM and controls without HCM who were age- and sex-matched and split into 70% training, 10% validation, and 20% testing data-set groups. Standard 10-second, 12-lead ECGs were converted to a 12 × 5000 matrix, and convolutional neural network methodology was applied, with convolutions within each lead and across different leads of the 12-lead recording. After initial training, further model fine-tuning was performed using an internal validation data set. This algorithm was subsequently validated in a pediatric cohort with HCM from Mayo Clinic and in external, international adult cohorts with HCM.^9,10

The output of this AI-ECG algorithm is expressed as a model prediction score (0%–100%) derived from its final sigmoid activation function that an ECG was recorded from a patient with HCM. Calibration of model outputs based on the observed probability of HCM, as previously described, was not performed.¹¹ Therefore, while this uncalibrated score reflects the model’s relative confidence, it should not be interpreted as an exact likelihood of HCM in an individual patient. In previous work, we determined that the AI-ECG prediction threshold with the optimal trade-off between sensitivity and specificity for a dichotomous prediction of HCM status was 11% (ie, the AI-ECG prediction is considered positive for HCM if the prediction score is >11%).⁸

SCD risk assessment

Each patient was assessed for the presence of SCD risk factors by using the American Heart Association/American College of Cardiology SCD risk stratification scheme to determine whether they would have qualified for a primary prevention ICD had HCM been diagnosed before their sentinel SCA. The following variables were considered: family history of SCD in ≥1 first-degree relative ≤50 years of age, personal history of unexplained syncope within 6 months before SCA, apical aneurysm, ejection fraction (EF) <50%, maximum left ventricular wall thickness (MLVWT) ≥30 mm, late gadolinium enhancement (LGE) ≥15% of left ventricular mass on cardiac magnetic resonance imaging (MRI), and the presence of nonsustained ventricular tachycardia (NSVT).¹ Family history of SCD and history of syncope were determined from clinical notes. EF and MLVWT were determined from the first TTE at our institution because of variable availability of echocardiographic data from outside institutions shortly after SCA. Apical aneurysm was determined from TTE and cardiac MRI. The presence of NSVT was documented from ICD and Holter monitor reports within 1 year of the evaluation at our institution. The extent of LGE was determined from MRI studies available at any time point after SCA.

For each patient, the European Society of Cardiology HCM Risk-SCD model was created using the above clinical and imaging information.¹² A score of <4% suggests that an ICD is generally not indicated, while an ICD may be considered for a score of ≥4–<6%, and an ICD is generally indicated for a score of ≥6%, in conjunction with shared decision making and consideration of other clinical risk markers not included in the HCM Risk-SCD score.¹³

Statistical analysis

For the description of cohort characteristics, categorical data are summarized as means and standard deviations or medians and interquartile ranges (IQRs), as appropriate. The number of SCD risk factors carried by each patient was summarized at both the per-patient and group levels. Data analysis was performed using R version 4.3.0 (R Foundation for Statistical Computing).

Results

Patient characteristics

As previously reported,⁷ of the 6267 patients with HCM who were evaluated at our institution, 27 (0.43%) presented with resuscitated SCA, which subsequently led to the diagnosis of HCM. Baseline characteristics are reported in Table 1. Nineteen patients (70%) were male, and the median age at the time of SCA was 40 years (IQR 25–48 years), with 8 patients younger than 30 years at the time of SCA. Six patients (22%) initially presented with SCA and were diagnosed with HCM at our institution, while the remaining presented for SCA and were diagnosed with HCM elsewhere. The formal diagnosis of HCM was made in the immediate post-SCA period in 25 patients (93%) and ∼0.5 and 1 year later in the remaining 2 patients. Both these patients with delayed diagnosis had apical variant HCM, and the formal diagnosis was made upon evaluation at our institution. The diagnosis of HCM was based on TTE in 23 patients (85%) and on TTE and cardiac MRI in 4 patients (15%). The median time from SCA to the first evaluation at our institution was 2.3 years (IQR 0.2–5.3 years), with 15 patients (56%) first evaluated >1 year after the SCA event.

Table 1.

Patient characteristics (N = 27)

Characteristic	Value
Age (y)	40 (25–48)
Men	19 (70)
Race
Black	3 (11)
White	23 (85)
Asian	1 (4)
Family history of SCD	3 (11)
Genetic testing	14 (52)
Pathogenic or likely pathogenic variant	10 (71)
ACTN2	1
MYBPC3	6
TNNI3	2
TNNT2	1
Comorbidities
Asthma	2 (7)
Coronary artery disease	1 (4)
Hyperlipidemia	4 (15)
Hypertension	6 (22)
Obesity	5 (19)
Obstructive sleep apnea	5 (19)
Smoking	4 (15)
Circumstance when SCA occurred
Awake, at rest	5 (19)
Exercising	8 (30)
Sleeping	1 (4)
Walking	7 (26)
Unknown	6 (22)
Cardiac symptoms before SCA
Chest discomfort	4 (15)
Dyspnea	4 (15)
Palpitations	3 (11)
Presyncope	6 (22)
Syncope	4 (15)
SCA rhythm
Ventricular fibrillation	16 (59)
Ventricular tachycardia	2 (7)
Unknown	9 (33)
Time from SCA to Mayo evaluation (y)	2.3 (0.2–5.3)

Values are presented as median (interquartile range) or n (%).

ACTN2 = actinin alpha 2; MYBPC = myosin-binding protein; SCA = sudden cardiac arrest; SCD = sudden cardiac death; TNNI3 = troponin I3, cardiac type; TNNT2 = troponin T2, cardiac type.

Eight patients (30%) were exercising when SCA occurred, while 7 (26%) experienced SCA during light activity (walking), 5 (19%) were awake at rest, and 1 patient (4%) was asleep. Information about physical activity at the time of SCA was not available in the electronic health record for the remaining patients. Ventricular tachycardia (VT) or ventricular fibrillation was the SCA rhythm in 18 patients (67%), while rhythm documentation was not retrospectively available in the remaining patients. On the basis of electronic health record review, 8 patients (30%) had cardiac symptoms preceding the SCA. The most common symptoms were presyncope in 6 patients (22%) and dyspnea, chest tightness, and syncope in 4 patients (15%) each. All episodes of syncope occurred 3–6 months before SCA, and all other symptoms had been present for at least 3 months before SCA.

Genetic testing was completed in 14 patients (52%), and pathogenic or likely pathogenic variants were identified in 10 (71%), most commonly in the MYBPC3-encoded cardiac myosin-binding protein C (n = 6).

ECG and imaging findings

Table 2 presents ECG, TTE, and cardiac MRI characteristics of the study cohort. Nearly half of the patients (n = 12 [44%]) met ECG criteria for LVH, and more than half (n = 15 [56%]) had T-wave inversions. Twenty-three patients (85%) had an abnormal repolarization pattern, and 3 (11%) patients had a normal ECG. No patients were on cardiac myosin inhibitors at the time of the first ECG at our institution.

Table 2.

ECG and imaging characteristics (N = 27)

Characteristic	Value
ECG
Time from SCA to ECG (y)	2.3 (0.4–6.3)
Normal ECG	3 (11)
Atrial fibrillation/flutter	4 (15)
Ventricular paced ECG	5 (19)
Left bundle branch block	1 (4)
Right bundle branch block	0 (0)
T-wave inversion (inferior or lateral)	15 (56)
LVH (per Sokolow-Lyon criteria)	12 (44)
Pathologic Q waves	3 (11)
Abnormal repolarization pattern	23 (85)
Artificial intelligence–based prediction of HCM (%)	95.6 (68.5–99.8)
Transthoracic echocardiography
Time from SCA to TTE	2.3 (0.2–6.3)
HCM subtype
Apical	6 (22)
Neutral	7 (26)
Reverse curve	11 (41)
Sigmoid	3 (11)
Obstructive HCM	13 (48)
Resting LVOT gradient ≥30 mm Hg	8 (30)
Provoked LVOT gradient ≥30 mm Hg	5 (19)
Septal thickness (mm)	20 (17–26)
Posterior wall thickness (mm)	12 (10–14)
Maximum LV wall thickness ≥30 mm	6 (22)
Mitral regurgitation
None-trivial	9 (33)
Mild	14 (52)
Moderate-severe	4 (15%)
Systolic anterior motion of the mitral valve	16 (59)
Apical aneurysm	5 (19)
Ejection fraction (%)	71 (67–75)
Ejection fraction <50%	1 (4)
LV mass index (g/m2)	150 (132–168)
Left atrial volume index (mL/m2)	45 (38–50)
Left atrial diameter (mm)	53 (46–62)
RVSP (mm Hg)	32 (26–41)
Cardiac MRI
MRI performed	11 (41)
Time from SCA to MRI (y)	0.3 (0.03–7.6)
LGE present	10 (91)
LGE ≥15% of the myocardium	4 (40)

Values are presented as median (interquartile range) or n (%).

Normal ECG was defined as an ECG demonstrating sinus rhythm, sinus arrhythmia, or a heart rate of 50–110 beats/min, with none of the following features: premature atrial or ventricular contractions, complete bundle branch block, ventricular paced rhythm, LVH (per Sokolow-Lyon criteria), T-wave inversions, pathologic Q waves, or artifact.

ECG = electrocardiogram; HCM = hypertrophic cardiomyopathy; LGE = late gadolinium enhancement; LV = left ventricular; LVH = left ventricular hypertrophy; LVOT = left ventricular outflow tract; MRI = magnetic resonance imaging; SCA = sudden cardiac arrest; RVSP = right ventricular systolic pressure; TTE = transthoracic echocardiography.

The model prediction score for HCM determined by the AI-ECG algorithm applied post hoc is shown in Figure 1. The median prediction score was 95.6% (IQR 68.5%–99.8%), with 18 patients (67%) having a prediction score of >90%. When considering a prediction cutoff of 11%, which has been shown to provide the optimal balance of sensitivity and specificity in a previous study,⁸ the AI-ECG algorithm identified HCM in 26 of 27 patients (sensitivity 96%). When limiting this analysis to 9 patients with a Mayo ECG available within 6 months of the SCA event, the median AI-ECG HCM prediction score was 96.7% (IQR 93.7%–98.8%), with 7 of 9 ECGs showing an HCM prediction score of >90%, and 8 of 9 ECGs showing an HCM prediction score of >11%. All 3 patients with normal ECGs by conventional ECG interpretation had a positive HCM prediction score (50.6%, 99.8%, and 93.7%, respectively). It should be noted that none of the patients in the present cohort had been included in the original development of the AI-ECG algorithm.

Figure 1.

Post hoc prediction score for HCM derived by AI analysis of the 12-lead ECG. Box plot demonstrating the distribution of HCM prediction scores based on an AI model for the detection of HCM from the 12-lead ECG. The 11% HCM prediction threshold provides the optimal trade-off between sensitivity and specificity. Open circles represent ECGs that were obtained within 6 months of cardiac arrest. AI = artificial intelligence; ECG = electrocardiogram; HCM = hypertrophic cardiomyopathy; SCA = sudden cardiac arrest.

The most common subtype of HCM in our cohort was the reverse-curve subtype, which was found in 11 patients (41%). Thirteen patients (48%) had obstructive physiology, which was elicited only after provocative maneuvers in 5 patients. The median septal thickness was 20 mm (IQR 17–26 mm). Systolic anterior motion of the mitral valve was present in 16 patients (59%), and 4 (15%) had moderate or severe mitral regurgitation. The median left atrial diameter was 53 mm (IQR 46–62 mm), with a median left atrial volume index of 45 mL/m² (IQR 38–50 mL/m²). Among the 11 patients with available cardiac MRI, LGE was reported in 10 (91%), which was quantified as extensive (>15% of the myocardium) in 4 patients.

SCD risk factors

The number of patients with traditional SCD risk factors assessed post hoc is shown in Figure 2. Sixteen patients (59%) had at least 1 SCD risk factor that might have represented an indication for primary prevention ICD if HCM had been diagnosed before SCA. The most common risk factors present in this cohort were MLVWT >30 mm and NSVT in 6 patients (22%) each, apical aneurysm in 5 patients (19%), and prior unexplained syncope and extensive LGE in 4 patients (14%) each. Three patients (11%) had a family history of SCD, and 1 (4%) had EF < 50%. Excluding NSVT and LGE, 13 patients (48%) had at least 1 risk factor.

Figure 2.

Post hoc patient-level prevalence of SCD risk factors. Distribution of SCD risk factors that would have qualified patients for ICD implantation had an HCM diagnosis been before SCA. Left: Including NSVT and extensive LGE on MRI as SCD risk factors. Right: Excluding NSVT and extensive LGE on MRI as SCD risk factors. HCM = hypertrophic cardiomyopathy; ICD = implantable cardioverter-defibrillator; LGE = late gadolinium enhancement; MRI = magnetic resonance imaging; NSVT = nonsustained ventricular tachycardia; SCA = sudden cardiac arrest; SCD = sudden cardiac death.

Each patient’s 5-year risk based on the HCM Risk-SCD calculator is presented in Figure 3. Twelve patients (44%) had a risk of <4%, while 7 (26%) and 8 (30%) patients had an estimated risk of 4%–<6% and ≥6%, respectively. The median estimated 5-year SCD risk in our cohort was 4.6% (IQR 3.2%–7.2%).

Figure 3.

Post hoc distribution of the European Society of Cardiology HCM Risk-SCD 5-year risk score. Horizontal lines denote 4% and 6% probability thresholds. A score of <4% indicates that ICD is generally not indicated. ICD may be considered for a score of 4–<6%, and ICD is generally indicated for a score of ≥6%. HCM = hypertrophic cardiomyopathy; ICD = implantable cardioverter-defibrillator; SCD = sudden cardiac death.

Discussion

This study provides an in-depth analysis of SCD risk factors in patients who survived SCA and were then newly diagnosed with HCM. While many patients presented to our institution months to years after SCA, key findings of this analysis include the following: (1) approximately one-third of patients had cardiac symptoms preceding the SCA and 9 of 10 patients had an abnormal ECG; (2) an AI-ECG algorithm had a sensitivity of 96% in detecting HCM in this cohort, highlighting the potential use of AI-ECG in early HCM detection; (3) in a post hoc assessment of SCD risk, conventional SCD risk factors were highly prevalent but not universal in this cohort; and (4) based on these risk factors, each of the American Heart Association/American College of Cardiology and European Society of Cardiology risk prediction schemes each would have accurately indicated a high SCD risk in the majority of HCM-related SCA survivors.

SCD risk stratification strategies in HCM have evolved over the past few decades.^12,14, 15, 16, 17 Despite these advances, it should be emphasized that SCD risk is a spectrum and “low risk” is not equivalent to “zero risk.”¹⁸ Ultimately, the decision to implement an ICD strategy for primary prevention should not be based solely on a binary threshold of SCD risk. Rather, it should be a result of shared decision making that considers patient factors (including age, comorbidities, and values) and awareness of the relevant nuances and uncertainties in the estimation of risk. The patients in the present study may represent a distinct subset of HCM with a primary arrhythmic phenotype whose SCD risk is not fully encapsulated in current risk assessments, highlighting the need for ongoing refinement of this approach and identification of new markers of risk. For example, the yield of genetic testing for pathogenic or likely pathogenic sarcomeric variants in this cohort was significantly higher than the ∼30% yield in all-comers with HCM,¹⁹ pointing toward the potential of sarcomeric status as a risk stratifier. It is also striking that the median MLVWT was only 20 mm and fewer than 1 in 4 patients had massive hypertrophy. However, because some of these variables for a subset of patients were collected months to years after SCA, this analysis should be considered exploratory and hypothesis-generating only.

The optimal approach for the detection of yet undiagnosed SCD-predisposing conditions such as HCM is unknown. ECG-based screening has been met with variable adoption because of imperfect accuracy, practical implementation challenges, and the economic and psychosocial impact of false-positive results.²⁰ Traditional computer-based ECG interpretation has limitations in the assessment of HCM because ECG abnormalities can be nonspecific or absent.^21,22 Conversely, deep-learning AI methods can identify ECG signals and patterns that might be unrecognizable by expert human interpreters.²³ The algorithm developed by our group at Mayo Clinic provides an example of such a tool. The high negative predictive value (99%) of this algorithm at a prediction score cutoff of 11% makes it particularly attractive as a rule-out test at the cost of a lower positive predictive value (31%).⁸ Viz HCM (Viz.ai), an Food and Drug Administration–approved software for the detection of HCM, has also demonstrated favorable performance characteristics.²⁴ The Viz HCM software was shown to achieve the highest accuracy at a prediction cutoff score of 85%, and in real-world implementation on consecutive patients, it was found to discover a new diagnosis of HCM in 5%–8% of patients with a positive AI-ECG result.^25,26 In the present study of SCA survivors who were later diagnosed with HCM, the AI-ECG algorithm suggested HCM in almost all patients on the basis of analysis of their subsequent ECGs. This illustrates how an AI-ECG–based screening program could lead to formal diagnostic testing and evaluation by a cardiologist, who could apply SCD risk stratification to newly diagnosed patients with HCM. Based on our post hoc analysis, it is plausible that AI-ECG screening could reduce the risk of SCA/SCD in patients with undiagnosed HCM, although a prospective study is required. These AI-ECG algorithms may also increase diagnostic suspicion for conditions such as HCM in patients who undergo clinically indicated, symptom-driven ECG (∼1 in 3 patients in our cohort were documented to have cardiac symptoms preceding the SCA).

The AI-ECG model applied in this study uses a series of convolutional and pooling layers to detect features of HCM and outputs a prediction of the presence of HCM. Other convolutional-based model architectures have demonstrated similar results in detecting HCM with variations in model design.27, 28, 29, 30 For example, the model by Tison et al²⁷ used hidden Markov models to divide ECGs into segments and demonstrated an area under the curve of 0.91 for the diagnosis of HCM. Other approaches include using ECG images,³¹ convolutional neural networks with tandem logistic regression,³² and supervised machine learning methods to adjust regression coefficients.³³ Although head-to-head comparisons of these models are lacking, it is plausible that combining different model weights across institutions—so-called federated learning—may yield models that are more broadly generalizable.³⁴

Other AI-ECG models have demonstrated additional utility in HCM beyond disease detection,³⁵ for example, in monitoring and quantifying the effect of cardiac myosin inhibitors on obstructive HCM pathophysiology.36, 37, 38 As another example, an algorithm developed by Carrick et al³⁹ can identify high-risk features for SCD, including systolic dysfunction, massive hypertrophy, apical aneurysms, and extensive LGE, in patients with known HCM. This has utility in identifying patients who could benefit from further imaging with regard to SCD risk stratification.

Although the aforementioned examples have a relatively narrow focus, AI-ECG has broader uses and can improve accuracy in global ECG interpretation across all domains (ie, determination of rate, rhythm, axis, and ischemia). For example, transformer-based architectures can leverage beat-to-beat sequence data of ECGs to detect arrhythmias⁴⁰ and broader foundational models synthesize natural language processing data in addition to ECG signal data to output ECG interpretations with high accuracy (area under the curve 0.945 across all diagnostic domains).⁴¹ The effectiveness, risks, and cost implications of such applications of AI-ECG tools are still unknown and require prospective evaluation.

Limitations

This study is retrospective, and several of the included patients were evaluated at our institution several months to years after their SCA and HCM diagnosis. Consequently, assessment of SCD risk factors present at the time of SCA may be limited by incomplete documentation, and we relied on echocardiographic data collected at the first evaluation at our institution. Similarly, documentation of NSVT relied on rhythm monitoring data available after the index evaluation at our institution. This provides only a surrogate for the presence of NSVT closer to the SCA event. However, even excluding NSVT, the prevalence of other SCD risk factors around the time of SCA was still high. Furthermore, many of the ECGs included in our AI analysis of HCM were performed months or years after the index SCA and none of the patients in this cohort had ECGs before the SCA. Therefore, it is impossible to determine what the model would predict for HCM diagnosis on a pre-SCA ECG. It is plausible that our results would differ in pre-SCA ECGs of our study cohort, especially in a progressive disease such as HCM. Finally, this cohort reflects the experience of a large tertiary HCM center, and the results may not be broadly generalizable.

Conclusion

SCA is a rare but important initial presentation of previously undiagnosed HCM. Our study illustrates that more than half of these patients carry SCD risk factors constituting indications for primary prevention ICD therapy if HCM had been diagnosed before SCA. The use of AI-ECG algorithms to diagnose HCM may allow earlier disease detection, risk stratification, and reduced risk of SCD.