Real-time heart rhythm analysis based on PPG signals with artificial ıntelligence

This study aims to collect raw photoplethysmography (PPG) signals using a sensor, process them on a Raspberry Pi 3 platform, and perform real-time rhythm classification with artificial-intelligence-based methods. The raw PPG data streamed from the sensor to the Pi are cleaned in Python; band-pass filtering, baseline-wander removal, and motion-artefact suppression are applied. The thesis consists of two major parts: theory and application. In the theoretical part, a literature survey is presented, covering PPG technology, data mining, time-series mining, and signal-based emotion-like rhythm analysis. Within the scope of the application, more than 60,000 PPG samples were recorded in three‑minute sessions. From these, 1,250 windows (250 samples per window) were randomly selected and manually labelled into three classes—Normal Sinus, Tachycardia, and Artefact. The labelled dataset was fed to a LightGBM classifier and, for comparison, to an LSTM‑based deep‑learning model. Both models achieved accuracies close to 90 %, yet the LSTM showed a markedly higher F1‑score in the Artefact class. These results demonstrate that even lightweight, edge‑compatible models can deliver clinically acceptable performance for real‑time PPG monitoring systems. Consequently, the proposed Pi infrastructure proves that PPG data can be transformed through AI techniques into an on‑the‑fly rhythm‑monitoring and alert system. The thesis serves as a valuable reference for future research in wearable health technology and cardiology informatics, guiding data acquisition, preprocessing, and model deployment.