Introduction

Cardiopulmonary resuscitation (CPR) is an emergency procedure to maintain brain function by supplying blood flow to the brain in patients under cardiac arrest [1]. It is important to correctly perform CPR to increase the survival rate of such patients [1]. However, even well-trained CPR professionals fail to consistently act in accordance with the standards of the American Heart Association (AHA) guidelines because each patient has different physical characteristics. [2]. Therefore, a CPR feedback system is necessary to monitor the condition of cardiac-arrest patients during CPR for the sake of its efficient realization.

The gold standard for circulation detection is invasive blood pressure (BP) during CPR. However, it is only feasible at the time in-hospital [3]. The AHA guidelines recommend using a capnogram as a support tool to decide about the return of spontaneous circulation (ROSC) [4]. An increase in end-tidal carbon dioxide (ETCO2) level may serve as an indicator of ROSC [3]. However, a capnogram is only available with intubation and has low accuracy in ROSC detection [3]. Researchers have proposed CPR feedback methods using a transthoracic-thoracic impedance (TTI) [5] or accelerometers [6] to efficiently perform CPR. However, these methods have some limitations such as uncomfortable, time-consuming, or inaccurate procedures. Especially, it is not exactly known whether blood is delivered to the brain during CPR.

Therefore, we propose a new method for CPR feedback based on artificial intelligence using earlobe photoplethysmography (Ear PPG) in this study. Ear PPG is noninvasive and has the advantage of reflecting the blood flow to brain in real time. Besides, BP can be estimated according to a long short-term memory (LSTM) model, which is an artificial intelligence structure, using Ear PPG.

Materials & Methods

Data were collected from two swines (70–75 kg) with cardiac-arrest model after approval by the Institutional Animal Care and Use Committee of Wonju College of Medicine, Yonsei University, for this study. Subjects were placed in a supine position and ventilated using a volume-controlled ventilator (MDS Matrix 3000, Matrix Environmental Technologies Inc. USA) after anesthesia. Electrocardiography (Lead II ECG), ETCO2 (CO2SMO plus, Novametric Medical Systems Inc. USA) and atrial blood pressure (5-Fr micromanometer-tipped catheter, Millar Inc. USA) were continuously monitored using a digital recording system (PowerLab, ADInstruments Ltd. USA). A pacing catheter was positioned in the right ventricle and ventricular fibrillation (VF) and asystole (Asys) was induced by delivering an alternating electrical current at 60 Hz with 60 V and 30 mA to the endocardium. CPR was performed according to the sequence rest (10 s) – CPR (40 s) – rest (10 s). An active compression-decompression (ACD) CPR device (LUCAS2, Physio-Control CO, USA) was used to achieve an uniform chest compression depth and compression rate, which were set at 5 cm and 100/m, respectively. To develop and evaluate the proposed method, a total of 255 datasets were randomly divided as follows: 135 datasets formed a training set (VF/Asys: 87/48) and 80 datasets formed a test set (VF/Asys: 58/32).

Figure 1 shows the method of BP estimation using Ear PPG based on LSTM model for CPR feedback. The analog circuit of the proposed system consists of an earlobe clip-type PPG sensor (EP520, LAXTHA Inc., Republic of Korea), an analog filter and an amplifier. The earlobe PPG sensor is composed of a visible red emitter with a center wavelength of 630 nm, a phototransistor, and an I-V converter. It enables the measurement of changes of blood volume caused by CPR. The PPG signals were amplified with a gain of 100 and filtered with a 0.15–20 Hz band-pass filter to remove power noise and baseline variation. Then, the data were digitized and saved at 10 bits/sample and 360 samples/s using PIC18F4523 (Microchip Technology Inc., USA). The acquired data were first filtered using a digital band-pass filter of bandwidth 0.5–10 Hz and then resampled to 20 Hz. The resulting signal is used as the input for the LSTM model. The LSTM architecture to estimate the BP waveform consists of three layers, namely an input layer, a hidden layer, and an output layer. To determine the optimal number of hidden layers, they were added from one to seven, step by step, while interactively evaluating the performance of each model. The best performance was obtained when using one hidden layer. Finally, two BP parameters (systolic and diastolic blood pressure) were calculated.

Fig. 1
figure 1

Block diagram of BP estimation using Ear PPG based on LSTM model for CPR feedback

Full size image

Results

Table 1 shows the results of statistical analyses for comparison between BP measured by the micromanometer-based gold standard method (BPMEAS) and the proposed Ear PPG-based method (BPEST) during CPR. Pearson’s correlation analysis showed high positive correlations (r = 0.92, p < 0.01) between BPMEAS and BPEST. The paired-samples t-test on the two BP parameters of the two methods indicated no significant differences (p > 0.05). The root mean-squared error (RMSE) of the systolic and diastolic BP was 2.24 ± 1.37 mmHg and 1.90 ± 1.20 mmHg, respectively.

Table 1 Results of statistical analyses for comparison between BP measured by the micromanometer-based gold standard method and the Ear PPG-based proposed method during CPR
Full size table

Discussion & Conclusion

This study proposes single Earlobe PPG-based BP estimation during CPR. We designed the analog circuit to acquire PPG signals during CPR and the LSTM model to estimate the BP waveform. Then, statistical analyses were conducted for comparison between BP measured by the gold standard and the proposed method. The estimated BP showed high positive correlations with the measured BP. The RMSE of the calculated systolic and diastolic BP was 2.24 ± 1.37 mmHg and 1.90 ± 1.20 mmHg, respectively.

Zadi et al. proposed an auto-regressive moving average (ARMA) model for a BP estimation algorithm using a PPG, resulting in an RMSE of 3–8 mmHg in young subjects with no known cardiovascular disorder [7]. Tanveer et al. estimated BP from an ECG and a PPG using a waveform-based ANN-LSTM network. This method achieved a high performance, namely 0.58–1.56 mmHg in multi parameter intelligent monitoring in intensive care (MIMIC) I database [8]. The performance of our study is similar or slightly lower compared with these studies. However, the strength in our study is that BP estimation is possible even during CPR. It enables accurate CPR, thereby increasing ROSC in patients under cardiac arrest. Therefore, we are convinced that the proposed method can be used as an important and useful tool to provide accurate CPR. In a further study, we will test for more subjects and patients under cardiac arrest to ensure the reliability of this method and optimize the LSTM model to improve performance.