Configuration of CP ECMO
As shown in Fig. 5, the p-ECMO system for CP included a patient monitor that collects invasive blood pressure (IBP) and ECG data, a bio-signal analysis AI program, controller, dual pulsatile pump, and an oxygenator. The p-ECMO for CP was used as the VA-ECMO, wherein blood was drawn from the vein through an inlet catheter, passed through a blood pump and an oxygenator, and then ejected into the artery through an outlet catheter. IBP/ECG measurements were taken using a patient monitor, and the bio-signal analysis AI program was used to analyze the BP waveforms to determine the timing of the pulses of heart and of the p-ECMO. In this study, a custom-made valveless pulsatile blood pump, powered by air pressure was used, and the p-ECMO controller was used to regulate the CP of the dual pulsatile pump, based on signals received through the bio-signal analysis AI program.
p-ECMO
The p-ECMO, shown in Fig. 6, can provide blood flow of up to 8 L/min instantaneously and allows for blood flow adjustment according to HRs of 40 to 100 bpm. Preliminary research confirmed its ability to maintain an average blood flow of 3.5 to 5 L/min under an average BP of 100 mmHg, and ongoing animal experiments are evaluating its biocompatibility [18, 19]. The f-NN for the BP waveform analysis was executed by a computer, while a custom-built controller, incorporating an embedded STM32F103VET6 ARM microcontroller (STMicroelectronics, Switzerland) with serial communication capabilities, was used to regulate the CP using the PLL method by receiving signals for pulse rate adjustment and ECMO pulsation timing control to adjust the overall pulse rate and timing.
CP Algorithm
As shown in Fig. 7, the program installed on the computer included two f-NNs, whose role was to distinguish between the pulses generated by the heart model and p-ECMO. For this purpose, information from the BP waveform and ECMO pulsation timing was used as input; after removing the DC offset from the dual BP waveforms, they were differentiated and standardized to emphasize changes in the BP waveform due to the heart model's pulse. After the pulse determination by the f-NNs, the HR was calculated using the index value of that pulse, and the algorithm would decide whether to advance or delay the ECMO pulsation timing to reduce the time to reach CP. For example, when the interval between two consecutive pulses generated by the heart model was taken as 1, the system would maintain the "stay mode" without adjusting the timing of the ECMO pulsation, if the ECMO pulses fell within the range of 0.3 to 0.4 of the heart pulse intervals; this was identified as CP. However, if the ECMO pulses occurred within an interval shorter than 0.3 of the heart pulse intervals, it was determined to emit a 'lag' signal; and if they occurred beyond 0.4 of the heart pulse intervals, it was determined to emit a 'lead' signal. When a 'lead' signal was passed to the p-ECMO, its bpm would be increased by 2 bpm for one pulse interval. Conversely, when a 'lag' signal was passed, the p-ECMO’s bpm would be decreased by 2 bpm for one pulse interval. This adjustment would allow for the synchronization between the heart model and p-ECMO.
The two f-NNs, which were designed to distinguish between the heart model’s pulse and p-ECMO’s pulse, consisted of an input layer, two hidden layers, and an output layer. ReLU functions were employed as activation functions, and SoftMax functions were utilized in the output layer. To train the f-NNs responsible for discerning the pulses of the heart model and p-ECMO, the heart model's bpm was varied in 10 bpm increments within a range of 40 to 100 bpm. Additionally, the p-ECMO was set to operate with a bpm differing by 1 bpm from that of the heart model. Both devices were then activated to record the BP data and the operational signal data of the p- ECMO. Subsequently, the stored data were processed to create input data, totaling 80,002-point training data set by combining 125 BP data points with 180 activation signal data points from the two pumps of the p-ECMO into a set of 305 data points at a time. To prevent overfitting during learning, the dropout rate was set to 0.7. Cross-entropy loss function and Adam Optimizer were utilized for optimization during machine learning, conducted with 1000 epochs. The accuracy of the neural network model for the heart pulse detection was 87.54%, while that for the ECMO pulse detection was 88.75%.
In-vitro Experimental Setup
The in-vitro experimental apparatus for evaluating the proposed system under conditions similar to those in an in-vivo application is shown in Fig. 8. Typically, an ECMO device is connected to the femoral vein using a catheter to draw blood, and then it injects the blood into the aorta through a catheter connected to the femoral artery, while the patient is lying down. Therefore, the in-vitro experimental apparatus was composed of a cylindrical open chamber, mimicking the venous pressure of a patient, and a closed chamber replicating the arterial pressure in the aorta. In the closed chamber, a space was provided to inject or expel air, thereby allowing for the adjustment of chamber compliance. Between the venous and arterial chambers, a tube and clamp system were provided that simulated the vascular resistance in the human body. By partially closing the clamp on the tube, the vascular resistance could be adjusted, thereby allowing for the regulation of the average BP in the arterial chamber to 100 mmHg. To mimic the ejection of blood from the patient's heart, an artificial heart model (LibraHeart I; Cervika, Korea), capable of providing pulsatile blood flow, was attached between the venous and arterial chambers.
The heart model's HR was set at 70 bpm, and the ejection volume was adjusted to maintain an average blood flow of 4 L/min. The clamp was adjusted to achieve an average pressure of 100 mmHg in the arterial chamber, and the air volume in the arterial chamber was regulated to maintain a BP change from 80 to 120 mmHg. These conditions simulated the same blood flow and pressure as in adult patients. The p-ECMO proposed in this study was connected to the heart and circulation model. The p-ECMO was connected to the venous chamber using a 22 Fr catheter and to the arterial chamber using a 15 Fr catheter. The connection between the device and catheters was established using Tygon tubes. As the p-ECMO drew blood from the venous chamber, the blood flow entering the heart model decreased. To maintain an overall flow of approximately 4 L/min, the ejection volume of the heart model was adjusted. The heart model's ejection volume was set to 30% of the total blood flow, with the p-ECMO's ejection volume adjusted to 70%. Subsequently, during the in-vitro experiment, the heart model's HR was adjusted from 55 to 75 bpm. While the p-ECMO operated in synchronous or asynchronous modes, variations in the mean BP, peak-to-peak amplitude, and flow were observed. However, to analyze the effect of the p-ECMO and its CP control, variations in blood flow and pressure waveforms were observed without altering the compliance and clamp settings.
In-vitro Experimental Procedure
All experiments were initiated under co-pulsation conditions, and how the p-ECMO maintained the CP in response to changes in the HR of the heart model was evaluated. For real-time CP control evaluation, both the heart model and p-ECMO were operated by initially setting them to the same bpm of 55. Next, the bpm of heart model was gradually increased by 5 bpm per min until it reached 75 bpm, followed by a gradual decrease of 5 bpm back to 55 bpm. During this process, the maintenance of CP and the occurrence rate of CP were measured. Additionally, the heart model and p-ECMO were separately operated at bpm values of 40, 50, 60, 70, 80, and 90. Subsequently, while maintaining the heart model at its initial bpm, it was incrementally adjusted by 5 or 10 bpm to observe how CP control was executed by the p-ECMO. To compare the CP occurrence rates, the bpm of the heart model was fixed at 65, while the pulse rate of the p-ECMO was set 5 or 10 bpm higher or lower than the HR of the heart model, and both were operated without CP control to measure the CP occurrence rate.