The Internet of Medical Things (IoMT) allows ongoing monitoring and automatic analysis of physiological signals, e.g., electrocardiogram (ECG) or similar ones. Nevertheless, the high level of classification, feature representation, and computational viability in the IoMT resource-constrained environment remains a challenge. Traditional machine learning algorithms have been characterized by poor scalability and poor inter-feature modeling in ECG signals. To overcome these constraints, the present research proposes an ECG classification model based on a Similarity Directed Graph Neural Network (SDGNN) that encodes ECG features as graph-structured data to model their relationships explicitly. To improve classification efficiency and convergence stability, a Mountaineering Team-Based Optimization (MTBO) algorithm is used to optimise parameters and fine-tune models. The experimental assessment of the benchmark ECG datasets shows that the suggested SDGNN-MTBO framework is even more accurate and precise than the regular methods, while consuming less computing resources. The framework achieves 99% classification accuracy, indicating its suitability for conducting a reliable analysis of the ECG signal in a healthcare monitoring system that employs the IoMT.