FedKMeans: uma abordagem baseada em clusterização para agregação em aprendizado federado aplicado à geração de ECG sintético

Abstract

The electrocardiogram (ECG) is an essential tool for diagnosing cardiac conditions, providing detailed information about the heart’s rhythm and electrical activity. However, traditional ECG acquisition is invasive, requiring specialized equipment, electrodes attached to the patient’s body, and active collaboration, which limits its accessibility and continuous monitoring. With the advent of wearable devices, such as smartwatches, it has become possible to passively collect biosignals, such as photoplethysmography (PPG). However, PPG alone does not offer the diagnostic richness of ECG, and direct ECG measurement on these devices still faces technical barriers. This work proposes a machine learning-based approach to estimate ECG signals from PPG data, using Bidirectional Long Short-Term Memory (BiLSTM) neural networks and federated learning techniques to enhance data privacy and model robustness in distributed environments. The methodology includes preprocessing PPG and ECG signals from the MIMIC-II database, modeling the problem as a sequential regression task, and implementing two federated aggregation strategies: FedAvg (standard) and FedKMeans, which introduces a clustering step (K-Means) to group local models with similar characteristics before aggregation. Experiments compared four scenarios: centralized training, individual local training, federated training with FedAvg, and federated training with FedKMeans. Evaluation metrics included RMSE, MSE, Pearson correlation coefficient, execution time, and memory usage. The results demonstrated that FedKMeans offers advantages in heterogeneous scenarios, providing higher accuracy in synthetic ECG generation and better computational efficiency compared to FedAvg. Additionally, the proposed approach overcomes the diagnostic limitations of PPG and addresses ethical concerns associated with centralized medical data collection, ensuring patient privacy. This combination of techniques transforms wearable devices into advanced tools for early detection and management of cardiovascular diseases, with potential for practical and scalable clinical applications. The work contributes to the federated learning field by exploring an efficient and secure solution for generating synthetic medical data, preserving privacy and clinical utility.