Introduction and Aim. Traditional AI models in cardiology are often criticized for their "black-box" nature which obscures the decision-making process. This study introduces a novel explainable AI (xAI) framework aimed at demystifying the mechanisms behind the differential diagnosis of wide complex tachycardias (WCT) which remains a significant challenge in arrhythmology. We highlight the importance of transparency by annotating diagnostic features in ECG data. Our approach seeks to bridge the gap between electrophysiological findings of WCT and specific ECG patterns by providing transparency and interpretability through annotated diagnostic features within ECG data, an aspect not addressed by conventional NN-based algorithms thereby enhancing physician understanding and trust in the diagnostic process.
Materials and Methods. The xAI model was trained on a dataset comprising various classes of WCT sourced from the PTB-XL database. In total, 11207 cases were used for training, 6232 for validation and 4360 for testing ensuring a comprehensive training process that captures the unique characteristics of each WCT class. We employed a proprietary developed artificial NN with an architecture specifically designed for the use of various xAI techniques.
Results. Our xAI model demonstrated a high mean accuracy of 94% and a mean ROC-AUC of 93% in identifying crucial WCT across 4360 test cases. It significantly improved diagnostic confidence by clarifying the reasoning behind each classification decision as evidenced by enhanced physician reviews. In addition, we provided comparisons between various xAI approaches and highlighted their differences, possible applications, strengths and weaknesses.
Conclusion. The implementation of xAI in differential diagnosis of WCT bridges the gap between advanced AI technologies and clinical usability. By providing clear annotated explanations of AI decisions our model fosters greater trust and reliability in automated ECG interpretation potentially leading to broader acceptance and application in clinical practice.