Heart rate variability enhances the accuracy of non-invasive continuous blood pressure estimation under blood loss

Background: To propose a new method for real-time monitoring of human blood pressure under blood loss (BPBL), this article combines pulse transit time (PTT) and heart rate variability (HRV) as input parameters in order to establish a model for the estimation of BPBL. Methods: Effective parameters such as PTT, R-R internal (RRI), and HRV were extracted and used to establish the blood pressure (BP) estimation. Three BP estimation models were established: the PTT model, the RRI model, and the HRV model, and they were divided into experimental group and control group. Finally, the effects of different estimation models on the accuracy of BPBL estimation were evaluated based on the experimental results. Results: The Pearson correlation coefficients R were 0.7731, 0.8943 and 0.9169 for the PTT model, the RRI model, and the HRV model, respectively. The root means square error of the estimation set (RMSEP) were 16.83 mmHg, 11.87 mmHg and 10.59 mmHg, respectively. Conclusion: The results suggest that the accuracy of the BPBL estimated by the RRI and HRV models is better than that of the PTT model, which means that both RRI and HRV can enhance the accuracy of BPBL estimation, and HRV seems to be more effective in improving the accuracy of BP prediction compared to RRI. These results provide a new idea for other scholars in the field of BPBL estimation research.

Background 1 continuous monitoring of BP. Others used the PTT method in personal health care to continuously measure human BP [20], 23 they combined the PTT method with a chest sensor and performed a Mean arterial pressure (MAP) estimation on the chest 24 through a corresponding calibration strategy [12], or used the PTT method to measure BP during and after dynamic exercise 25 Although much research has been conducted which has yielded in a series of significant results in this field, there are still 27 some certain limitations in current research outputs. Firstly, there is a large gap in the estimation and analysis of BPBL in 28 terms of accuracy. Given that the scenario of blood loss is very common in the clinic, continuous monitoring and early 29 warning of decreasing BP in the case of blood loss is important, which is what currently is lacking in the relevant BP research. 30 In addition, there are still differences in the PTT-based BP estimation methods which are tied to individual patients, leading 31 to a decrease in the accuracy of the results. Unfortunately, because the physiological states are not exactly the same and its 32 effects cannot be completely eliminated in different individuals, their estimated BP will be different, hence, a standard 33 continuous BP measurement method that is suitable for most people may be difficult to achieve. 34 In this article, R-wave to R-wave internal (RRI) and HRV will be used as input feature parameters to establish the BP 35 estimation model. RRI and HRV are common indicators to measure the physiological status of different individuals, and, 36 when used as input parameters to establish the BP estimation model, the errors caused by individual differences can 37 effectively be reduced and the accuracy of BP estimation can be improved. This study uses RRI and HRV to establish 38 different BP estimation models and evaluates each model based on experimental results, exploring the impact of different 39 input parameters on the accuracy of BP estimation under blood loss. The results of this study may be helpful for other people 40 who are conducting BP estimation research, and it is hoped to provide an additional contribution to the development in this 41 field. 42 4 changes of BPBL, the physiological changes of the human body during blood loss were analyzed, and parameters closely 48 related to blood loss were determined in order to estimate BPBL. the physiological changes of the human body. As shown in Fig. 1 [22], the PPG waveform includes alternate components (AC) and 51 direct components (DC) [23]. AC is depending on the heart rate, and is the pulsating component of the PPG waveform, which 52 superimposes on the DC component [23]. The DC component is related to the blood volume in the blood vessel, and will slowly change 53 due to respiration and changes in the diameter of the blood vessel.

56
Since there is a certain functional relationship between the total blood volume V of the venous blood ( ), the arterial blood ( ) 57 and the blood vessel diameter L, it is assumed that the following formula exists between them: 58 When blood loss occurs, the blood volume in the blood vessel decreases, and according to formula (3), the blood vessel diameter L 62 BPBL, the AC and DC parameters that are closely related to blood loss are of significance. Therefore, in this study the two parameters 66 AC and DC were extracted for the construction of the BPBL model. In addition, since the human body performs autonomous 67 neuromodulation to suppress BP reduction during blood loss, the quantification of the neuromodulation process will help improve the 68 accuracy of the BPBL estimation. One method of assessing autonomic modulation is through HRV, which is usually calculated by RRI 69 and is a common method for monitoring autonomic nerve activity [10][11]. In this study, the time domain parameters (SDNN and 70 RMSSD) and the frequency domain parameters (HF) are used to measure the HRV. Therefore, in addition to AC and DC, RRI and 71 HRV signals were also extracted and utilized to estimate the BPBL.

Experimental process 75
In this study, Landrace pigs were used as experimental objects. They were fixed on the operating table and anesthetized 76 during the experiment. As shown in Fig. 2, three signals were detected during the experiment: the first signal was the ECGs 77 signal detected around the heart, which was obtained via the unipolar lead method. The second signal was the pulse signal 78 located at the nose, which was obtained via the PPG method. The third signal was the invasive continuous blood pressure 79 for modeling calculations and verification. All three signals were collected from the pigs simultaneously and were displayed 81 by the Chengdu Instrument RM6240C multi-channel physiological signal acquisition system, with experimental data being 82 recorded throughout the process, and all signals being sampled at 1000 Hz. 83 The experiments were conducted on a total of five pigs with a weight range of 31 ± 8.5 kg. The specific experimental 84 process is shown in Fig. 3. At the beginning of the experiment, it was crucial to record the baseline physiological data of the 85 pigs in a stable state. It was made sure the pigs were under anesthesia, and 200ml of blood was withdrawn via the pigs' 86 carotid artery. The corresponding time points were recorded and the physiological changes of the pigs were observed. When 87 the BP and HR of the pigs reached a relatively stable state, the pigs were bled again (200ml-400ml) and the corresponding 88 data were recorded. As soon as the pigs lost blood at a certain level, they were given another blood transfusion. The above 89 process was repeated several times, with the number of repetitions determined according to the specific experimental 90 situations. Because the blood volume of the pigs is in linear relationship with their body weight, the number of the 91 bloodletting instances was determined according to the weight of the pigs and the specific experimental conditions, in order 92 to achieve the aim of this study. For this study, a total of 10 sections of valid blood loss data were extracted during the 93 experiment.

Feature parameter extraction 97
After the experiment, the main characteristic parameters such as PTT, RRI, and HRV were extracted from the ECG and PPG 98 signals collected from the experiments. As shown in Fig. 4 (a), The PTT refers to the time that the pulse wave travels from the the pulse wave traveling from the heart to the pig's nose. 101 The acquisition process of PTT was as follows: First, the peak position of the ECG signal and the PPG signal with the 102 corresponding algorithm was located. Due to the ECG and PPG signals being collected simultaneously, and the pulse wave 103 signal being transmitted from the heart to the pig's nose, the ECG signal preceded the PPG signal for a period of time, and 104 there was only one PPG peak between two adjacent R peaks. In the same way, all the data segments that met these criteria 105 were extracted, and the remaining data points were regarded as outliers and discarded. After all the valid data segments were 106 obtained, the R peak point of the ECG signal was used as the starting point, and the peak point of the adjacent PPG signal 107 was used as the endpoint. The time between the starting points and the endpoints is the pulse wave transit time, and all PTT 108 data were extracted to establish the BPBL estimation model.

111
In addition to the PTT parameters, other characteristic parameters were also extracted to establish the BPBL estimation 112 model. As shown in Fig. 5, these parameters include two parts: the first component is the pulse amplitude (AC) and peak 113 (DC) extracted from the PPG signal. AC/DC as the input parameter was also calculated to reduce the interference of 114 individual differences. The second component is the RRI and HRV (SDNN, RMSSD, and HF) extracted from the PPG signal, 115 which is the key parameter of this experiment. Combined with PTT and PTT/ RRI, a total of nine characteristic parameters 116 were extracted in the study which were mainly used as input parameter to establish the BPBL estimation model. respectively, based on the input parameters of the PTT model. In the described experimental procedure of this study, the 126 PTT model was set as the control group, while the RRI model and the HRV model were set as the experimental group. The 127 estimation accuracy of different BP estimation models was compared to evaluate the advantages and disadvantages of the 128 different models and the influence of the two characteristic parameters RRI and HRV on the BPBL estimation accuracy. The 129 input characteristic parameters of the different models are shown in Table 1 The correlation between the estimated value and the actual value was established by evaluating the predictive ability of 138 the model through the following indicators: the root means square error (RMSEC) and the correlation coefficient of the 139 training set (Rt), and the root mean square error (RMSEP) and the correlation coefficient of the estimation set (Re). 140 When the correlation coefficient is larger and the root mean square error is smaller, it indicates that the model has a better 141 predictive ability. 142

Results 143
In the experiment, three different BPBL estimation models were constructed with the input feature parameters mentioned 144 previously based on the partial least squares (PLS). Three different BPBL estimation models were tested with the same 145 experimental data, and the experimental results of the training set and the estimation set of SBP under blood loss were 146 obtained. There were 1229 sample points of all models. The P-values in table 2 shows that all experimental results were 147 statistically significant. The detailed experimental results of each model were shown in Table 2 and Fig. 6. 148 The correlation coefficient (Rp) and the root mean square error (RMSEP) were calculated in our study to assess the 149 accuracy of the estimated BPBL values. According to the Table 2  respectively. RMSEC and RMSEP represent the root mean square error of the training set and prediction set, respectively. 157 158 Fig. 6. The results of the training set and prediction set of three blood loss models 159 Therefore, it is obvious that the correlation estimated by the RRI model and the HRV model is better than that of the PPT 160 model. Moreover, the estimated effect of the BP value of the HRV model is greater than that of the RRI model. 161 help in improving the estimation accuracy of BPBL values. And the results also provide further guidance for the study of 164 BPBL based on PTT and HRV. 165

Discussion 166
The purpose of this study was to estimate the BPBL non-invasively and continuously based on the HRV, PTT, and other 167 parameters. In order to achieve this purpose, an experimental group and a control group were established together with three 168 estimation models. The effect of the estimation models was evaluated through experimental results and the impact of 169 different feature parameters on the accuracy of the BP estimation was explored. 170 171 Fig. 7. Bland-Altman plot of the difference between actual SBP and estimated SBP in three models.

172
The experimental group has a better effect on BP estimation 12 when HRV and R-R intervals are added as input parameters. It could be speculated that this is due to HRV and R-R intervals 180 being more relevant to the characteristics of the human body, and when used as input parameters to establish the estimation 181 model they may reduce the impact of individual differences and increase the accuracy of BP estimation. 182

HRV Enhances the Accuracy of BPBL Estimation 183
According to the experimental results, the HRV model has the highest correlation and the smallest root mean square error, which 184 suggests that HRV greatly promotes the prediction accuracy of BPBL. As already mentioned in the second part of this article, 185 when the BP of the human body decreases, a series of adaptations occurs in the cardiac neuromodulation center to raise the 186 BP, including accelerated blood pumping and blood vessel contraction until the BP rises and eventually becomes stable. 187 Neuromodulation plays a very important role in the process of blood loss, and HRV is one of the important indicators for 188 clinical monitoring of cardiac autonomic nerve activity. From the results of this study it can be concluded that HRV has a 189 strong correlation with BP changes during blood loss, which when used as an input feature parameter to model BPBL, the 190 results will be the most precise among the three models and much higher than the PTT model. Furthermore, HRV can 191 significantly enhance the estimation accuracy of BPBL. 192

HRV is more meaningful than RRI 193
By comparing the estimation results of the RRI model and the HRV model of the experimental group, the Pearson correlation 194 coefficient of SBP and IBP estimated by the HRV model is higher than that of the RRI model, and the mean square error of 195 the estimated value of the HRV model is lower than that of the RRI model. This shows that HRV has a better correlation 196 with BPBL, and can greatly enhance the accuracy of BPBL estimation. 197 The HRV indicators used in this study are calculated by RRI, since RRI may contain more information than HRV, such as 198 heart rate. However, the comparison of the two models in the experimental group indicated that HRV is better than RRI in 199 estimating the BPBL. Therefore, it can be argued that when blood loss in pigs occur, they mainly regulate it through the 200 nervous system, and HRV is more closely related to the nervous system than the RRI. In addition, compared with RRI, HRV

Limitations 204
There are several limitations in this study. The amounts of data used is not sufficient, more blood loss data are needed to 205 increase the reliability of the BPBL estimation model. Future research should obtain more data to generate a more accurate 206 BPBL model. Furthermore, a lot of parameters were extracted for this experiment, but a few of these parameters may only 207 give a small contribution to the overall results, and more relevant parameters have not been identified. The relevance of other 208 parameters to the BPBL needs to be verified, and more effective input parameters are to be used to improve the BPBL model. 209

Conclusions 210
In this article, we established three BPBL estimation models and compared their estimation results. The results of this study 211 and the BPBL estimation model provide evidence that the RRI and HRV are good correlators with the BPBL and enhance 212 the accuracy of the BPBL estimation. The comparison of the three models suggest that the HRV model, based on the 213 characteristic parameters of HRV and PTT, has a better estimation of results during blood loss, thereby improving the 214 accuracy of the BPBL estimation and therefore, being more suitable as a blood loss estimation model. This research on the 215 process of blood loss and BP can be expanded to other fields, which will become one of its future directions. And it is hoped 216 that this article will serve as a guidance for related research attempts in the future. 217 and static components. The near-infrared light emitted by PPG passes through the bloodless tissue layer and blood of the human body 296 and reflects back and is received by the sensor. Since the light intensity change curve of the reflected light is closely related to the 297 pulse, PPG technology can be used to obtain the pulse wave waveform, and the AC and DC can be extracted from the pulse wave. 298 In addition, AC is depending on the heart rate, so we calculated RRI to reflect blood changes together. 299 300 Fig. 2. Experimental scene diagram. The picture shows the side view of the pig on the operating table. During the experiment, we 301 used the Chengdu Instrument RM6240C multi-physiological channel acquisition system to collect signals. We collected the PPG,

328
RMSEC and RMSEP represent the root mean square error of the training set and prediction set, respectively.