Abstract Body: Introduction:
Invasive arterial blood pressure can help optimize cardiopulmonary resuscitation due to the relationship of diastolic blood pressure with coronary perfusion pressure. The expulsion of blood from the thoracic cavity during chest compressions can result in retrograde blood flow during decompression causing incorrect extraction of diastolic pressure by monitors. To address this challenge, we evaluated several neural networks for automated extraction of diastolic blood pressure in swine cardiac arrest models.
Methods:
Fifty-two Yorkshire swine were anesthetized, instrumented, and subjected to 5 or 10 minutes of ventricular fibrillation or sham intervention. Cardiac arrest animals were resuscitated with mechanical chest compression and defibrillations, then monitored for 24 hours after return of spontaneous circulation. Systolic and diastolic locations were labeled for twenty seconds of data at each minute of resuscitation and each hour of monitoring. Several neural networks used in similar applications including ECG peak detection were built with Python’s TensorFlow library. A 70% train-test split was performed, and 5-fold cross validation was used to determine optimal hyperparameters. Models were evaluated on the final test set using F1 score.
Results:
The four models had similar performance on systolic peak detection with mean (SD) F1 scores of 0.99 (0.002) and 0.98 (0.002) for compression and non-compression waveforms respectively. The diastolic detection task was more difficult, with non-compression diastolic detection F1 score at 0.96 (0.006) and compression detection F1 score of 0.91 (0.01). Model size had a negligible relationship with F1 score. Compatible models were converted to the smaller TensorFlow Lite format and showed no drop in effectiveness despite a large reduction in memory requirement.
Conclusion:
Small models performed similarly to larger ones while requiring less memory. These models will allow easy analysis of compression effectiveness in cardiac arrest and are small enough to be integrated into a microcomputer onboard a portable monitor, enabling real-time resuscitation feedback. Additional data will be integrated into models, along with evaluation on human-derived waveform data.