Hypertensive intracerebral hemorrhage is a typical cerebrovascular emergency and critical disease in neurosurgery [
However, multiple causes may induce readmission in the postoperative recovery period. The disability and mortality rates are significantly higher than those of the first bleeding, and the prognosis is lacking [
In recent years, medical research has shown that cerebrovascular events’ timing is closely related to the circadian rhythm of human blood pressure [
Machine learning has made great progress in the establishment of various prediction models [
A total of 120 patients with hypertensive intracerebral hemorrhage who were hospitalized in our hospital were selected. There were 74 males and 46 females, aged
The vital sign signals collected from the medical information system in 80 patients were used as the training sets. The remaining 40 patients were used as the testing sets and were randomly divided into two groups. The established model is verified in 20 patients, and they were treated before the high blood pressure appeared (model group), and another 20 patients with a daily fixed treatment scheme were selected as the control group. There was no significant difference in gender, education level, and course of disease between the two groups.
The algorithm flowchart of this study is shown in Figure
Algorithm flowchart of blood pressure modeling.
The signals used in this study include the electrocardiogram (ECG), photoplethysmograph (PPG), and arterial blood pressure (ABP). PPG is the finger volume pulse wave. The pretreatment includes the following stages:
The pulse wave contains rich physiological information. The main characteristic points are pulse wave starting point A, main wave peak B, tide wave starting point C, tide wave ending point D, descending gorge E, and repetition wave peak F (Figure
Single cardiac cycle pulse wave feature points.
The main wave peak B is caused by ventricular contraction, blood from left ventricular to aorta, reflecting the ability of ventricular ejection and compliance of blood vessels, etc. The peak F of the heavy pulse wave is caused by the diastolic period of the heart; the closure of the aortic valve prevents blood from returning to the ventricle, reflecting the elasticity of the artery and the closing function of the active pulse valve. B and F are the two most important characteristic points in a cardiac cycle. Therefore, B and F are selected as the termination points to calculate PTT and are recorded as PTT-p and PTT-d, respectively, as shown in Figure
PTT feature extraction method.
B is obtained by calculating the maximum position of PPG in a cardiac cycle, and F is obtained by continuous wavelet decomposition of PPG with gaus1 wavelet basis. The first zero crossing point after B is E, and the second zero crossing point is F.
This study proposes a new blood pressure prediction model ABP-net, which uses 1D-CNN to automatically extract the waveform features of PPG and predict blood pressure, and solves the problem of feature points being difficult to extract, as shown in Figure
ABP-net network structure.
ABP-net is a convolutional neural network with mixed features, and the grey identification part is a model constructed with PTT features. In this study, PPG segments of the same cardiac cycle are also added to the network as input. One-dimensional convolution is used to extract the waveform features of the pulse wave, and the full connection layer is used to synthesize and select the extracted waveform features, and then, PTT features are used for blood pressure prediction. The model input contains two types of features: PTT is the traditional numerical feature and PPG is the formal feature (such as signal and image). ABP-net effectively integrates numerical features and formal features, which provides a new idea for using different types of features to model together, and improves the effectiveness and accuracy of the model by using the association between the two types of features and the output. ABP-net contains a variety of processing modules, and the dotted part is the residual structure.
The input-output feature maps in Conv-1D are all 1-dimensional; for a single sample, let
BN solves the problem that the distribution of inputs at a certain layer in a deep network changes due to previous changes in network parameters. Let a batch be
The activation function enables the model to obtain nonlinear modeling capability, and the ReLU function
does not cause a gradient disappearance problem and is computationally simple, making the model forward computing faster.
Pooling can reduce feature dimensionality while acquiring the main information in a feature map and is divided into two types, maximum pooling and average pooling.
Pooling of maxima:
Pooling of means:
The other modules are the traditional neural network modules, FC is the fully connected layer, and AVG is the input format that is needed to convert the output of the convolutional layer into a fully connected layer with the method of averaging.
Model training is to update the network parameters iteratively to make the loss function converge to the global minimum in the training set. ABP-net is a regression model with blood pressure as the output. The loss function is defined as the mean square error (MSE). Let
Patients in the control group received the routine blood pressure control measures: antihypertensive drugs three times a day and the rehabilitation did not control the time. The observation group used the model-predicted blood pressure increase pad to give targeted treatment. This seems to be a routine treatment after this.
According to the circadian rhythm fluctuation of blood pressure in patients with hypertensive intracerebral hemorrhage, the model group was given nursing intervention measures such as adjusting medication time and guiding patients’ early limb function training according to the circadian rhythm fluctuation of blood pressure. The detailed biological clock control methods were as follows:
Via follow-up monitoring for 1 year, the antihypertensive effect and recurrence of the cerebrovascular accident in the two groups were monitored and compared. The Barthel score was used to evaluate the ability of daily living. The Fugl-Meyer score was used to evaluate limb motor function.
The blood pressure changes at 6:00, 4:00, 18:00, and 22:00 every day were compared to determine the measurement position, time, and sphygmomanometer, and the therapeutic effect was determined by a specially assigned person as follows: (1)
The age, Barthel score, and Fugl-Meyer score of the two groups were compared by
For each model, the prediction accuracy of DBP is higher, which indicates that DBP has higher correlation with PTT and PPG. For the traditional regression model, the model constructed by PPG is better than the model constructed by PTT, but the accuracy is lower than that of the ABP-net model, which shows that the ABP-net model is effective for the integration of PTT features and PPG features and has greater advantages than other models (Table
Blood pressure model predicts SBP and DBP accuracy comparison.
Methods | SBP (mmHg) | DBP (mmHg) | ||
---|---|---|---|---|
PTT | PPG | PPT | PPG | |
ABP-net | ||||
Linear regression | ||||
Random forest |
The antihypertensive effect of the two groups is shown in Figure
Antihypertensive effect in the model group (a) and control group (b).
Compared with the data before treatment, the daily life self-care ability and limb motor function after treatment in all the two groups were improved. After treatment, the score of the model group was higher than that of the control group (
Comparison of daily life self-care ability and limb motor function in two groups: (a) before treatment; (b) after treatment.
Eight (13.3%) patients in the model group and 17 (28.3%) patients in the control group were readmitted to the hospital because of recurrence of cerebrovascular accident. The recurrent rate between the two groups was statistically significant (
Hypertensive intracerebral hemorrhage is a common primary intracerebral hemorrhage, which is caused by the rupture of blood vessels when blood pressure rises abruptly on the basis of cerebral artery disease caused by hypertension. In recent years, the incidence rate has increased year by year, about 81/105, and the mortality rate of cerebral hemorrhage patients is 38%~43% [
With the exploration of the physiological and pathological rhythm of hypertensive cerebral hemorrhage and the time rhythm of drug action, it has been recognized that the drug treatment effect of hypertensive cerebral hemorrhage is related not only to the pharmacological effect of antihypertensive drugs but also to the blood pressure fluctuation time rule, the medication time rule of patients itself [
The circadian rhythm curve of blood pressure in patients with hypertensive intracerebral hemorrhage is similar to that in normal people, but the overall blood pressure level is higher and the fluctuation range is larger. Even if the blood pressure has decreased after treatment, the rhythm can still exist. At 2:00-3:00 in the morning, it was at the lowest point and then showed an upward trend. After getting up in the morning, it rose rapidly, reaching the first peak at about 8:00-9:00 am and slightly higher at 5:00-6:00 p.m., which was the second peak, and then began to decline slowly. Therefore, the 24 h ambulatory blood pressure monitoring curve showed a double peak and a valley. In the traditional method of administration, the drug was administered by the method of average distribution, three time per day [
In this study, we applied time nursing, according to the time law of circadian blood pressure changes; the patients with hypertensive intracerebral hemorrhage were given medicine three times per day at 6:00, 15:00, and 22:00; two times per day at 6:00 and 22:00; and one time per day at 6:00. The blood pressure of hypertensive cerebral hemorrhage patients increased at 6:00, and the concentration of drugs in the blood was very low after overnight metabolism. At 6:00, the effective pharmacological actions in the body can be supplemented as soon as possible, so that the peak effect of antihypertensive drugs corresponds to the morning blood pressure peak, which is conducive to the control of morning peak blood pressure. At 15:00, the antihypertensive drug is given before the second peak of the blood pressure fluctuation, which can reduce blood pressure in time and reduce the damage to target organs after the blood pressure rises. The effective blood concentration can be maintained at night by administration at 22:00 and can reduce the incidence of complications.
And then, the drug treatment under the guidance of the hybrid feature convolution neural network was used, and the blood pressure prediction model of patients with hypertensive cerebral hemorrhage was established. Using the model to predict the blood pressure increase point of patients, targeted treatment for patients can significantly and smoothly reduce blood pressure, promote health recovery, and reduce the occurrence of cerebrovascular accidents.
Early limb function training can promote the reorganization of the central nervous function, significantly improve the body motor function, reduce the degree of damage and disability, promote the recovery of nerve function, improve the quality of life, and reduce the burden of society [
In summary, mastering the drug treatment under the guidance of a hybrid feature convolution neural network and establishing the blood pressure prediction model of patients with hypertensive intracerebral hemorrhage can reduce the blood pressure of patients with hypertensive intracerebral hemorrhage, promote health recovery, and reduce the occurrence of cerebrovascular accidents. It has very important practical significance to improve the survival quality of patients with hypertensive intracerebral hemorrhage, which is worthy of further clinical investigation.
All data analyzed during this study are available from the corresponding author upon request.
The authors declare that there is no conflict of interest regarding the publication of this article.
Z.Z. and S.L. performed the study and drafted the article. Z.Z. and L.B. conducted data acquisition and data analysis and interpretation. All authors discussed the results and agreed to be accountable for all aspects of the work. All authors read and approved the final manuscript.