Introduction: Current mechanical CPR devices deliver one-size-fits-all CPR, following internationally recognized CPR guidelines. There is evidence that patient tailored CPR is more effective than one-size-fits-all CPR and that CPR efficacy changes with time. There is a need for a mechanical CPR approach that uses patient physiology to guide chest compression delivery.
Hypothesis: We hypothesize that machine learning will predict the carotid blood flow generated by mechanical chest compressions delivered at different rates.
Methods: Carotid blood flow was measured during MCC in nine domestic swine (~30 kg). MCC were delivered at a depth of 2” and at a rate of 50, 75, 100, 125, or 150 cpm. Rates were randomized and changed every 2 min. A regression tree algorithm was developed to model the relationship between the compression waveform and net carotid flow. The regression tree was modeled using the following chest compression parameters: chest compression rate, compression count within the 2 min epoch, and compression count within the experiment. The animal-specific performance of the algorithm was computed by using a 4-fold cross validation procedure.
Results: In the figure below, net carotid blood flow per compression for a single animal is shown in blue, and the prediction of the model is shown in red. The average out-of-sample, normalized mean squared error for this subject is 10 ± 1.2 μl/comp. The model captures the significant features of the carotid blood flow over time, particularly the collapse of blood flow after about 700 compressions.
Conclusions: This model accurately predicted the carotid blood flow generated by mechanical chest compressions. Future development of this model will allow us to answer the inverse problem, namely which chest compression waveform will generate the best carotid blood flow. This approach can be used to develop a closed-loop algorithm that delivers adaptive, automated mechanical chest compression based on patient physiology.