Development and validation of a deep learning model for the detection of cardiac sarcoidosis using AI-ECG analysis

Introduction Cardiac sarcoidosis (CS) is a life-threatening manifestation of sarcoidosis. Current screening strategies for patients with extracardiac disease and a suspicion of CS often lead to high rates of advanced cardiac imaging with low diagnostic yield. Artificial intelligence applied to the electrocardiogram (AI-ECG) offers a scalable, low-cost triage tool. While AI-ECG has successfully detected subclinical disease in other cardiomyopathies, its application in sarcoidosis could enable more targeted referrals and earlier identification of clinically relevant cardiac involvement. Purpose To develop and validate a deep learning AI-ECG model specifically designed to detect the presence of cardiac sarcoidosis. Methods A deep convolutional neural network was developed by fine-tuning the ECGFounder foundational model. The training dataset comprised ECGs from patients with probable or definite CS from two Dutch centers, supplemented by patients from the public MIMIC database with ICD-10 codes for CS. Model performance was evaluated in an independent internal test cohort using the Area Under the Receiver Operating Characteristics curve (AUC), sensitivity, specificity, and predictive values. Finally, a clinical utility analysis was performed on a cohort of patients with extra-cardiac sarcoidosis referred for CS diagnostics to estimate the potential reduction in unnecessary imaging. Results The model was trained on 44,700 ECGs from 9,269 patients (median age 63 [IQR 51 - 73], 58% male), including 389 patients with a CS diagnosis. In an internal test cohort of 746 ECGs (33 patients (4.4%) with confirmed CS), the model achieved an AUC of 0.81 (95% CI: 0.73–0.89). At the optimal threshold, the model demonstrated a sensitivity of 0.73 (95% CI: 0.56–0.88), specificity of 0.79 (95% CI: 0.77–0.82), PPV of 0.12 (95% CI: 0.08 – 0.17) and a high NPV of 0.99 (95% CI: 0.98–0.99). In a pilot clinical utility analysis (n = 70, 39% prevalence), the model, calibrated to 90% sensitivity, demonstrated a 21.4% test avoidance rate. Extrapolating these results to the full internal outpatient population (N=741), the model could potentially prevent 137 unnecessary advanced diagnostic scans, misclassifying 2.9% as false negatives. Conclusion Our AI-ECG model accurately identifies patients with CS and provides a scalable tool for automated risk stratification. Given its high NPV, this model may serve as an effective gatekeeper to significantly reduce unnecessary advanced imaging in clinical practice, especially in combination with other diagnostics. Future efforts should focus on external validation and prospective implementation.