Background : Previous studies have demonstrated the ability of Artificial Intelligence (AI) algorithms to accurately interpret electrocardiogram (ECG) records. The highly demanded accuracy, sensitivity, and specificity can be achieved with more data input and recalibration of the model.
Methods : A machine learning based model (recurrent neural networks [RNN]) was created to perform ECG delineation. RNN was trained to identify six different segments of the ECG: P wave, PQ, QR, RS, ST, and TP segments. Dataset : 2-lead ECG from the Physiobank QT Database, sampled at 250Hz. Each of the 105 records had at least 30 heartbeats manually labeled by cardiologists with the aforementioned segments. Preprocessing : The first 1.3s of each record
were removed as they frequently contain artifacts. Machine learning: A RNN with bidirectional recurrent layers was used as architecture. The model was trained with 1,542 labeled records using Adam Optimizer and tested on 282 records. Experiment Strategy : the algorithm was trained with 1,260 labeled records using Adam Optimizer, and validated and tested with 282 not-labeled records each; all records contained 325 samples each. The records included in the training were exclusively used in this step. Class assignment strategy: For each point of each record, the model computes the probability of belonging to one of each six segments. This prediction was then compared to the actual segment according to the manual labels (Ground
Truth).
Results : The AI algorithm was able to predict the ECG segments with high accuracy when compared to manually labeled ECG segments.
Conclusions : An effective and accurate AI-powered tool can be applied to recognize specific ECG segments, These features can be applied to enhance ECG interpretation and flatten the expertise disparities between physicians.

Application of Machine Learning to ECG Segment Delineation