Machine Learning-Based Identification of Patients with Elevated Central Venous Pressure Using Features Extracted from Photoplethysmography Waveforms

Central venous pressure (CVP), a key component of hemodynamic monitoring, is widely used to guide fluid resuscitation in critically ill patients. It is typically measured using central venous line catheterization, which is the gold standard, but this method is invasive, time-consuming, and associated with complications. This study aims to investigate whether machine learning (ML)-based analysis of features extracted from a non-invasive, standard-of-care waveform-the photoplethysmography (PPG) signal-can identify patients with elevated CVP. We trained Light Gradient-Boosting Machine (LightGBM) model using a large perioperative dataset (MLORD), containing 17,327 surgical patients from 2019 to 2022 at UCLA. For this study, we selected 1665 patients with both PPG and CVP waveforms available. A total of 843 PPG features per cardiac cycle (CC) were extracted from the PPG waveforms using a signal processing-based feature extraction tool, along with the simultaneous maximum value calculated from the corresponding CCs in the CVP waveform. Additionally, for each patient, the average and standard deviation of each PPG feature, as well as the mean of the maximum CVP values, were calculated across all cardiac cycles, resulting in 843 averaged PPG features, 843 PPG feature standard deviations, and one averaged maximum CVP value per patient. The average maximum CVP value was used as the ground truth to classify patients as either normal (5 ≤ CVP ≤ 15 mmHg) or elevated (CVP > 15 mmHg). Of the 1,665 patients, 1,182 were normal and 483 were elevated. The dataset was split into 90% for training (1,063 normal and 435 elevated) and 10% for testing (119 normal and 48 elevated). From the 1686 PPG features (843 averaged and 843 standard deviation), 246 were selected for model development using the Recursive Feature Elimination with Cross-Validation (RFECV) approach. To further enhance performance, hyperparameters were tuned through 5-fold cross-validation on the training set. Finally, the best-performing configuration was retrained on the full training data, and its performance was evaluated on the held-out test set. To provide a robust estimate and confidence interval, a bootstrapping procedure with 100 iterations was performed on the test set. The LightGBM classifier achieved a mean area under the receiver operating characteristic curve (AUC) of 0.79 (95% CI: 0.71-0.84) and mean accuracy of 0.71 (95% CI: 0.65-0.77), demonstrating good discriminatory power in distinguishing between patients with normal and elevated CVP. This study highlights the ability of PPG-derived features to discriminate between patients with normal and elevated CVP using ML. These early findings lay the groundwork for future research aimed at developing non-invasive approaches to CVP assessment.