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

Data: MLORD perioperative waveform dataset (17,327 surgical patients); 918 patients with concurrent PPG, ABP and CVP. Features: landmark-based extraction of 843 PPG and 846 ABP features per cardiac cycle (time, frequency, morphology); summarised per patient as mean and SD → 3,378 candidates; 91 selected by permutation importance with an events-per-variable cap. Labels: patient-level mean of per-cycle maximum CVP; normal 5–15 mmHg vs elevated >15 mmHg. Split: 90% train (balanced 413/413) / 10% held-out test (46/46); 3-fold grid search; 100-iteration bootstrap on the test set.

Background Central venous pressure (CVP) is widely used to guide fluid resuscitation and assess right ventricular preload in critically ill and perioperative patients, but requires invasive catheterization associated with procedural risk and resource use. Despite the ubiquity of photoplethysmography (PPG) and arterial blood pressure (ABP) monitoring, their combined use for CVP assessment remains underexplored. Methods We investigated whether machine learning (ML) models using features extracted from PPG and ABP waveforms can discriminate between patients with normal and elevated CVP. Using a large perioperative waveform dataset (MLORD, 17,327 surgical patients), we identified 918 patients with concurrent PPG, ABP, and CVP recordings. For each cardiac cycle, a landmark-based feature extraction tool derived 843 PPG and 846 ABP features, spanning time-, frequency-, and morphology-domain characteristics. Cardiac cycle–level features were summarized at the patient level by computing the mean and standard deviation of each PPG and ABP feature. The maximum CVP in each cardiac cycle was also determined and averaged across all cycles to yield a patient-level mean of cycle-level maximum CVP. Patients were classified as having normal CVP (5–15 mmHg) or elevated CVP (> 15 mmHg) based on this patient-level mean. Model development: We evaluated nine ML classifiers-Light Gradient Boosting Machine (LightGBM), eXtreme Gradient Boosting (XGBoost), Random Forest, histogram-based gradient boosting, multilayer perceptron (MLP), k-nearest neighbors (kNN), Gaussian naive Bayes, support vector classifier (SVC), and logistic regression-for binary classification of normal versus elevated CVP. From 3,378 candidate waveform features per patient (1,686 PPG and 1,692 ABP), we selected 91 features using permutation importance constrained by an events-per-variable rule (maximum number of features set to 10% of the total number of patients). Model hyperparameters were tuned using 3-fold grid search on the training set. The optimal configuration for each classifier was then retrained on the full training set (90% of patients; 413 normal, 413 elevated), and performance was evaluated on a held-out test set (10% of patients; 46 normal, 46 elevated). To obtain robust estimates and confidence intervals, we applied a 100‑iteration bootstrapping procedure to the test set. Results Seven of nine classifiers achieved mean area under the receiver operating characteristic (ROC) curve (AUC) ≥ 0.80 on the test set. Although SVC and XGBoost produced the highest mean AUC and area under the precision–recall curve (AUPRC), LightGBM demonstrated the most balanced overall performance across threshold-dependent metrics at a multi-metric optimized classification threshold. LightGBM achieved a mean AUC of 0.835 (standard deviation 0.042; 95% confidence interval [CI] 0.749–0.916), mean precision 0.827 (0.060; 95% CI 0.727–0.933), mean recall 0.806 (0.054; 95% CI 0.689–0.896), mean F1‑score 0.815 (0.043; 95% CI 0.725–0.893), mean AUPRC 0.841 (0.051; 95% CI 0.739–0.938), mean accuracy 0.817 (0.038; 95% CI 0.739–0.891), and mean specificity 0.830 (0.056; 95% CI 0.722–0.931) which was the best overall performance. Interpretability We applied SHapley Additive exPlanations (SHAP) to the final LightGBM model to quantify the contribution of individual PPG and ABP features to its predictions. Both PPG and ABP features, including both mean and standard deviation measures, had substantial impact on classification, with top features reflecting venous-influenced diastolic PPG morphology and ABP frequency/morphology characteristics consistent with physiologic ventricular interdependence. Conclusions In a large perioperative cohort, combined PPG and ABP waveform features enabled ML-based discrimination between normal and elevated CVP with good performance. These findings suggest that routinely acquired PPG and ABP signals may support minimally invasive screening for elevated CVP and motivate future work on multicenter validation and continuous, noninvasive CVP estimation.

Verbatim from the publisher's DOI record (opens in a new tab) · Publisher licence: https://creativecommons.org/licenses/by/4.0/

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.