Background/Objectives: Hemodynamic monitoring is crucial for managing critically ill patients and those undergoing major surgeries. Cardiac output (CO) is an essential marker for diagnosing hemodynamic deterioration and guiding interventions. The gold standard thermodilution method for measuring CO is invasive, prompting a search for non-invasive alternatives. This pilot study aimed to develop a non-invasive algorithm for classifying the cardiac index (CI) into low and non-low categories using finger photoplethysmography (PPG) and a machine learning model. Methods: PPG and continuous thermodilution CO data were collected from patients undergoing off-pump coronary artery bypass graft surgery. The dataset underwent preprocessing, and features were extracted and selected using the Relief algorithm. A CatBoost machine learning model was trained and evaluated using a validation and testing phase approach. Results: The developed model achieved an accuracy of 89.42% in the validation phase and 87.57% in the testing phase. Performance was balanced across low and non-low CO categories, demonstrating robust classification capabilities. Conclusions: This study demonstrates the potential of machine learning and non-invasive PPG for accurate CO classification. The proposed method could enhance patient safety and comfort in critical care and surgical settings by providing a non-invasive alternative to traditional invasive CO monitoring techniques. Further research is needed to validate these findings in larger, diverse patient populations and clinical scenarios.
Read full abstract