Artificial intelligence for phenotyping tetralogy of Fallot patients from electrocardiographic recordings

Background Tetralogy of Fallot (ToF) remains associated with significant morbidity and mortality. Artificial intelligence (AI) is a viable tool for identifying markers: deep learning (DL) can be used to automate measurements on the ECG trace and dimensionality reduction (DR) algorithms can be used for grouping patients based on ECG, imaging and genetic markers. Such an AI-based pipeline can aid in phenotyping by identifying ECG characteristics that correlate with outcome. Methods A cohort of patients with ToF were recruited for the study (n=388). All patients underwent echo- and electrocardiographic exams, and had recorded information on outcome (death, heart failure [HF], history of ventricular tachycardia [VT] or atrial fibrillation [AF]), lifestyle and imaging markers (left ventricular ejection fraction [LVEF] or NYHA score, among others). The analysis of the population consisted in three steps (Figure 1A). Firstly, the ECG was delineated using a DL model to obtain the P, QRS and T onsets/offsets for all cardiac cycles [1]. Secondly, the most stable heartbeat was selected, based on their morphology, for their usage into a DR algorithm [2]. This algorithm allowed combining the information of the different ECG leads, to automatically assess inter-patient similarities. Thirdly, patients were clustered with respect to ECG morphology, as extracted in the previous step, and said clusters were correlated with outcome. Results AI allowed to identify a subset of patients with higher outcome risk (Figure 1A). The blue, green and red clusters contained patients with right bundle branch block (RBBB) but with different morphologies, whereas the orange cluster had a relatively normal morphology (no RBBB) and the lowest QRS width (Figure 1B). With respect to events (Table 1), the red cluster was correlated with a significantly increased risk of all-type negative outcome (36.94%, p-value <0.0001). The orange and red clusters had a higher occurrence of death and AF. The green, blue and red clusters had a higher density of patients with VT. Despite the red, blue and green clusters presenting RBBB and a wide QRS complex, these show very different event rates, hinting at morphology as key to stratify risk: the higher-risk red cluster showed more QRS fractionation and reduced QRS amplitude. Finally, it is important to note the good correspondence with events despite the low correlation with imaging markers (TAPSE, RV diameter), hinting at the complementariness of both modalities. Conclusion AI can help identify clinical markers of interest in patients with ToF, given its ability to agglomerate all morphological information in the different ECG leads. In this work, the ECG markers correlated with clinical data, allowing the identification of a subgroup with increased risk of outcome (death, VT, AF, HF). The analysis shows that a RBBB pattern and increased QRS width are not the sole factors that might affect outcome. Funding Acknowledgement Type of funding sources: Public grant(s) – National budget only. Main funding source(s): Fundaciό La Maratό de TV3