Detection of arterial pressure waveform error using machine learning trained algorithms

Abstract

In critically ill and high-risk surgical room patients, an invasive arterial catheter is often inserted to continuously measure arterial pressure (AP). The arterial waveform pressure measurement, however, may be compromised by damping or inappropriate reference placement of the pressure transducer. Clinicians, decision support systems, or closed-loop applications that rely on such information would benefit from the ability to detect error from the waveform alone. In the present study we hypothesized that machine-learning trained algorithms could discriminate three types of transducer error from accurate monitoring with receiver operator characteristic (ROC) curve areas greater than 0.9. After obtaining written consent, patient arterial line waveform data was collected in the operating room in real-time during routine surgery requiring arterial pressure monitoring. Three deliberate error conditions were introduced during monitoring: Damping, Transducer High, and Transducer Low. The waveforms were split up into 10 s clips that were featurized. The data was also either calibrated against the patient’s own baseline or left uncalibrated. The data was then split into training and validation sets, and machine-learning algorithms were run in a Monte-Carlo fashion on the training data with variable sized training sets and hyperparameters. The algorithms with the highest balanced accuracy were pruned, then the highest performing algorithm in the training set for each error state (High, Low, Damped) for both calibrated and uncalibrated data was finally tested against the validation set and the ROC and precision-recall curve area-under the curve (AUC) calculated. 38 patients were enrolled in the study with a mean age of 52 ± 15 years. A total of 40 h of monitoring time was recorded with approximately 120,000 heart beats featurized. For all error states, ROC AUCs for algorithm performance on classification of the state were greater than 0.9; when using patient-specific calibrated data AUCs were 0.94, 0.95, and 0.99 for the transducer low, transducer high, and damped conditions respectively. Machine-learning trained algorithms were able to discriminate arterial line transducer error states from the waveform alone with a high degree of accuracy.