A103-17 Artificial Intelligence Enabled Electrocardiogram Algorithm for Detection of Arterial Blood PH

Abstract Rationale Arterial blood pH (pHa) is an important indicator of pulmonary ventilation and tissue perfusion. Respiratory acid-base changes and lactic acidosis are frequently encountered in the critical care setting. pHa gudies patient-specific critical care management in acute and chronic acid-base compensation. Arterial blood gas (ABG) analysis is the current standard technique for measuring pHa. Repeated arterial puncture is invasive, painful, and associated with complications such as hematomas. Other methods to estimate pHa include approximation from venous blood gas, which has significant bias in unstable patients, and ventilator simulation models like the physiological dead space model, which require measurement of parameters like dead space using complex techniques. Changes in blood pH affect myocardial ion channels, conductivity, and repolarization, which are reflected in the electrocardiogram (ECG). Neural networks can detect subtle multifocal changes in ECG and stand as promising tools for pHa estimation. Methods Using the MIMIC-IV v1.0 dataset, we identified patients who underwent ABG and had 10-second 12-lead ECG sampled at 5000 Hz, recorded within 3 hours of ABG in the inpatient setting. If more than one ECG existed, we included the one closest to the ABG time. After linking pHa values and ECGs, the dataset was split at the patient level into training (70%), validation (15%), and testing (15%) sets. We trained a convolutional neural network (CNN) with ECG as input and pHa as a continuous numeric output, using the training dataset. The validation dataset was used for hyper-parameter tuning between training steps. Finally, the model was tested on the testing dataset, and the performance was reported as Mean absolute error (MAE). Results 53,096 unique patients with age (65.81 ± 16.29) years, 61.21% non-Hispanic whites with 61,614 ECG and ABG measurements split into training set (n = 37,080; ECGs=43,131), validation set (n = 8,040; ECGs=9,242), and testing set (n = 7,976; ECGs=9,241) were included (Fig. A). A custom CNN with spatial and temporal encoders was trained. The model converged its performance after 20 epochs, with a mean absolute error for pHa of 0.086 on the test dataset (Fig. B). The performance of the model was compared across pHa intervals (Fig. C) Conclusion Our study demonstrates that AI-ECG has the potential to estimate pHa with moderate accuracy and could have large-scale implications for critically ill patients requiring real-time acid-base monitoring for treatment optimization, while also reducing invasiveness and costs associated with traditional ABGs. This abstract is funded by: None