Background: Various cardiomyopathies can present with a hypertrophic phenotype, complicating the diagnosis of the underlying diseases and delaying disease-specific therapy. Widely available diagnostic tools such as an ECG paired with artificial intelligence have the potential to shorten diagnostic pathways. The present study aims to establish a diagnostic approach (i) to distinguish between healthy controls and cardiomyopathies and (ii) to differentiate cardiomyopathies with left ventricular hypertrophy (LVH) using a machine learning (ML) - based ECG analysis.
Methods: The retrospective DETECT study enrolled patients with LVH defined as a left ventricular wall thickness (IVSD) ≥12 mm due to aortic stenosis (AS), cardiac amyloidosis (CA), hypertrophic cardiomyopathy (HCM), or Fabry disease (FD) as well as healthy controls (HC) without LVH. Patients with combined cardiomyopathies were excluded. All ECGs were analyzed using the ML-based Glasgow algorithm. In addition to ECG data, 21 selected baseline characteristics including comorbidities, laboratory and echocardiographic measurements were collected. Parameters with >75% missing values, without variance and a high correlation (> 0.85) were removed. All 376 remaining variables were integrated into the final ML algorithm. A 2-step model approach was implemented using random forest (RF) as a classifier and recursive feature elimination (RFE) for further variable selection. The first model (HC model) separates the healthy cohort from patients with cardiomyopathies, and the second model (CMP model) differentiates between cardiomyopathies. The model was trained with 80% and tested with 20% of the data set.
Results: From 2000 to 2022, 548 patients including 200 (37%) with AS, 100 (18%) with CA, 97 (18%) with HCM, and 82 (15%) with FD as well as 69 (13%) HC were enrolled in the study. AS patients had a median age of 82 (78-85) years and 45% were female. The CA cohort had a median age of 80 (72-83) years and included 18% women. Patients with HCM and FD were younger with a median age of 61 (50-70) years and 56 (48-64) years, respectively. 42% of HCM and 49% of FD patients were female. The HC group presented the youngest cohort (median age of 48 [44-54] years) with 38% women. The final ML algorithm comprised a two-step approach: First, healthy controls were separated from patients with cardiomyopathy with an accuracy of 1.00 using only ECG parameters (HC model). In a second step, AS, CA, HCM, and FD were differentiated with an overall accuracy of 0.76 using ECGs, comorbidities, laboratory and echocardiographic measurements (CMP model). The CMP model achieved an accuracy of 0.82 to discriminate AS from other cardiomyopathies, 0.84 for CA, 0.89 for HCM, and 0.91 for FD.
Conclusion: To conclude, our two-step ML algorithm based on automated ECG parameters augmented with comorbidities, laboratory and echocardiographic measurements enables (i) the differentiation of healthy controls and cardiomyopathies and (ii) the discrimination of cardiomyopathies with a hypertrophic phenotype. The implementation of the approach into clinical routine may lead to a faster diagnosis followed by earlier initiation of disease-specific therapy.
