Method for quality assessment of continuous physiological signals

Abstract

The present application discloses a method for evaluating the quality of continuous physiological signals, which comprises: segmenting the original physiological signals by windowing; preprocessing the segmented signals; the preprocessing comprises removing the signal baseline, bandpass filtering, and median filtering; extracting eigenvalues ​​of the preprocessed signals; inputting the extracted eigenvalues ​​into a pretrained model to obtain a scoring result score; and then judging the scoring result according to set thresholds L1 and L2 to obtain a signal quality evaluation result; the pretrained model is an isolation forest model.

Description

Technical Field

[0001] The present application relates to the detection of physiological signals, and in particular to the quality assessment of physiological signals. Background Art

[0002] Continuous physiological signals refer to continuously recorded signals such as ECG, respiration, pulse, etc. that are highly related to human health and disease status. Currently, their value in disease diagnosis, adverse event prediction, human status assessment, and rehabilitation prognosis evaluation is increasingly recognized. For example, the deterioration of some clinical diseases is usually manifested in physiological signals 8-24 hours before causing more serious clinical outcomes (such as sudden cardiac death). Clinicians and researchers in this field are increasingly aware of the importance of continuous monitoring and in-depth mining and analysis of physiological signals.

[0003] In recent years, a variety of measurement methods and sensors have developed rapidly and integrated into wearable devices, greatly enriching the means for people to continuously obtain physiological signals. For example, electrocardiography (ECG) is a technology that uses electrodes attached to the surface of human skin to record the overall potential changes of the heart and usually displays it as a waveform. It has been widely recognized in the clinic and is used in the diagnosis and analysis of heart diseases in medicine; Photoplethysmography (PPG) is a method that uses photoelectric means to measure blood volume changes in living tissues. This method uses a fixed wavelength of light to pass through the skin, and each time the heart beats, the blood vessels will contract and expand, which in turn affects the light reflection. The intensity of the light reflection recorded by the sensor reflects the characteristics of blood flow. This technology is currently the core monitoring technology of various health monitoring devices, especially health monitoring bracelets, and has broad application prospects and rich scientific research value; Respiratory Inductive Plethysmography (RIP) is a method of recording the chest and abdominal movements of the subject by wrapping two elastic straps around the chest and abdomen of the subject respectively, and the breathing movement drives the elastic straps to expand and contract, and the coil area changes. This method provides a non-invasive and indirect measurement method for respiratory movement, which has the advantages of convenience, low load and high accuracy.

[0004] However, although there are many ways to measure and record physiological signals, most physiological signals have small amplitude, low frequency, and are susceptible to interference. In addition, the sources of noise in the real world are complex and uncontrollable, so physiological signals measured in daily life are often drowned by noise, which greatly affects the process and accuracy of subsequent signal in-depth mining and analysis. Therefore, signal quality assessment is usually the first step in the subsequent research and analysis of continuous physiological signals. Automatically screening out signal segments with high signal quality from long-range continuous physiological signals is a prerequisite for related subsequent work.

[0005] For a specific physiological signal, researchers usually use the following methods to evaluate signal quality: 1. Threshold method. That is, extract features based on the characteristics of the signal. Once the feature exceeds the predetermined threshold, the signal is determined to be of poor quality; 2. Establish a corresponding signal quality score (Signal Quality Index, SQI). For example, for ECG signals, people use the features of signal calculations and add certain weight coefficients to develop kSQI, sSQI, pSQI and other scoring methods; 3. Use artificial intelligence, machine learning and deep learning methods to build models. The current mainstream model building method is still supervised learning, that is, the model learns a large amount of pre-labeled data, summarizes the rules and uses them for subsequent signal quality evaluation and analysis.

[0006] However, the above methods all have strong limitations: First, the signals in early studies / patents mostly come from bedside monitors, which collect signals that are fundamentally different from real-world signals, and the source and intensity of noise are different from the problems faced by wearable devices; second, despite the rapid development of high-performance machine learning / deep learning models, most of these methods require a large number of labels to complete the model learning steps. Insufficient number of labels or too strong limitations of model training samples will lead to insufficient generalization performance of the model, making it difficult to meet the requirements of use. The problem is that it is relatively easy to obtain massive amounts of data, but it takes a lot of manpower and time to label all of this data, and the rapid iteration of wearable device hardware and software has further increased the task of labeling. Summary of the invention

[0007] In view of the above problems, this application aims to propose a continuous physiological signal quality assessment method.

[0008] The continuous physiological signal quality assessment method of the present application comprises:

[0009] The original physiological signal is segmented by adding windows;

[0010] Preprocessing the segmented signal; the preprocessing includes removing the signal baseline, bandpass filtering, and median filtering;

[0011] Extracting feature values ​​from the preprocessed signal;

[0012] The extracted feature values ​​are input into the pre-trained model to obtain the score result; the score result is then judged according to the set thresholds L1 and L2 to obtain the evaluation result of the signal quality;

[0013] The pre-trained model is the Isolation Forest model.

[0014] Preferably, the pre-trained model is trained on a large sample unlabeled training set, and then verified and the threshold is determined on a small sample manually annotated verification set.

[0015] Preferably, the pre-trained model returns a scoring result of [-1, 1] on the validation set;

[0016] When score ≥ L2, it is judged as category 1, the signal quality is good, and it returns to category 1;

[0017] When L1≤score<L2, it is judged as category 2, the signal quality is medium, and it returns to category 2;

[0018] When score is less than L1, a larger window mean filter is used to remove the baseline for the data segment with score less than L1, and then feature extraction is performed. The extracted feature values ​​are input into the pre-trained model and scored again. If the score is still less than L1, it is judged as category 3 with poor signal quality; if the score is greater than L1, it is judged as category 2 with medium signal quality.

[0019] Preferably, when the original physiological signal is segmented by windowing, the length of the window is fixed;

[0020] If the physiological signal is an electrocardiogram signal, the window length is 10 seconds; if the physiological signal is a respiratory signal, the window length is 30 seconds.

[0021] Preferably, when removing the signal baseline, a 2-s sliding window is used for scanning to obtain the median of the signal within 2 seconds, that is, the baseline of the current signal, and the signal baseline is removed by subtracting the baseline from the original signal;

[0022] In the bandpass filtering, a bandpass filter is used for filtering, and the passband is selected according to the characteristics of the signal; if the physiological signal is an electrocardiogram signal, the passband is set to 0.1-100 Hz; if the physiological signal is a respiratory signal, the passband is set to 0.01-20 Hz;

[0023] In median filtering, a median filter is used to remove outliers, and the filter length of the median filter is set to 40 data points.

[0024] Preferably, the characteristic values ​​include skewness, kurtosis, and the distance between two adjacent waveforms calculated based on a dynamic time warping algorithm.

[0025] Preferably, the skewness is in accordance with Perform calculations;

[0026] Kurtosis according to Perform calculations;

[0027] Among them, μ is the sample mean, σ is the sample standard deviation, N is the number of samples, Skewness is the skewness, and Kurtosis is the kurtosis.

[0028] Preferably, the extracted eigenvalues ​​are normalized using a linear function.

[0029] Preferably, the eigenvalue is a eigenvector;

[0030] use Perform normalization processing;

[0031] Among them, X norm is the normalized eigenvector, X is the original eigenvector, X min and X max are the minimum and maximum values ​​in the eigenvector respectively.

[0032] The method of this application is the first attempt to apply the idea of ​​anomaly detection and the isolation forest model to the quality assessment of continuous physiological signals; the method of this application is based on the unsupervised model isolation forest, which avoids a lot of data labeling work, and the model is flexible and easy to adjust, with strong generalization performance; the application results of the method of this application on ECG and respiratory signals show that the patented algorithm can effectively distinguish the signal quality level, and the test results are good. Even if there are some pathological changes in the signal, the signal quality can still be correctly distinguished. The method of this application can also be extended to the quality evaluation of other physiological signals, with a wide range of application scenarios and great potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A flowchart of a method for evaluating quality of physiological signals;

[0034] Figure 2 It is the original signal diagram of ECG and respiratory wave;

[0035] Figure 3 To verify the quality results of the ECG signal;

[0036] Figure 4 This is the breathing signal quality result of the test set;

[0037] Figure 5 It is the result of ECG signal quality assessment;

[0038] Figure 6 It is the result of respiratory signal quality assessment;

[0039] Figure 7 is the signal quality assessment result;

[0040] Figure 8 Assess the results for signal quality in the presence of pathology; DETAILED DESCRIPTION

[0041] Below, the present application is described in detail with reference to the accompanying drawings.

[0042] This application invents an algorithm for evaluating the quality of continuous physiological signals based on the idea of ​​anomaly detection. The core of the algorithm is to use an unsupervised machine learning model, Isolation Forest, to extract features through window splitting, then read the model to score the degree of abnormality of the signal, and classify the signal according to the model scoring result.

[0043] The continuous physiological signal quality assessment method of the present application (ECG and respiratory signals are used as examples in this example) splits the signal by windowing, removes the signal baseline, filters, removes outliers, extracts features and normalizes the features in the window, and then reads the signal into the isolation forest model for classification evaluation, and evaluates the signal quality based on the scoring results.

[0044] The signal quality assessment process is as follows: Figure 1 shown.

[0045] 1. The original signal of the signal in this example is as follows Figure 2 shown.

[0046] 2. Window splitting

[0047] The physiological signals are segmented with fixed time window lengths, in this case 10 seconds for cardiac electrocardiography and 30 seconds for respiration. The signal quality of the signal in each time window is determined.

[0048] 3. Signal preprocessing

[0049] Scan the signal with a 2S sliding window to obtain the median of the signal within 2S, that is, the baseline of the current signal. Subtract the baseline from the original signal to remove the signal baseline.

[0050] Use a bandpass filter to filter the signal and select the passband according to the signal characteristics. In this example, the central electrical signal passband is set to 0.1-100Hz, and the respiratory signal passband is set to 0.01-20Hz;

[0051] A median filter was used to remove outliers from the signal, and the filter length was set to 40 data points;

[0052] 4. Feature extraction and normalization

[0053] Features used include skewness and kurtosis:

[0054] Skewness:

[0055]

[0056] Kurtosis:

[0057]

[0058] Where μ is the sample mean, δ is the sample standard deviation, and N is the number of samples.

[0059] Extract characteristic points from physiological signals (such as R waves in ECG signals, peaks and troughs in respiratory signals, etc.), cut the signal in the window into several parts, that is, each part represents a complete physiological signal waveform (such as cardiac cycle and respiratory cycle), calculate the DTW distance of the waveforms of two adjacent parts, and form a sequence of waveform similarity in the window. Select Q1, median, and Q3 quantiles in the sequence as features.

[0060] Other features include the entropy of the signal within the window, spectrum distribution, energy proportion of each frequency band, etc., which need to be determined according to the specific signal.

[0061] Linear function normalization (Min-Max scaling) is used to normalize the features, that is, for specific features, the following normalization is performed:

[0062]

[0063] Where X norm is the normalized eigenvector, X is the original eigenvector, X min and X max are the minimum and maximum values ​​in the eigenvector respectively.

[0064] 5. Model construction and evaluation

[0065] 1) Training the model

[0066] The data for training the model mainly comes from the patient and healthy person data collected by the research team in advance. The data collection time for each case is about 24 hours. The specific model training process is as follows:

[0067] · 10 examples of signal data that are considered to be of relatively high quality were selected, and the signals were split according to the predetermined window length, and about 100,000 data sets were selected for training;

[0068] Extract signal features within the window and form a feature matrix;

[0069] ·Training model, the model selected is the isolation forest model, and the parameter settings are: n_estimators = 2000, contamination = 0.05. The model returns a score of -1 to 1, reflecting the quality of the signal.

[0070] 2) Determination of classification threshold

[0071] In order to improve the robustness of the model and map the scoring results to the three-category results, we set the classification thresholds L1 and L2. To determine the values ​​of L1 and L2, we construct a labeled model validation set. In this example, the validation set of ECG data contains 3460 data with a window length of 10s, and the respiratory data contains 2086 data with a window length of 30s. These data are annotated in advance by clinical and biomedical engineering experts on the signal quality of the data, and the annotation results are good signal quality (Good), medium quality (Acceptable), and poor quality (Unacceptable). At the same time, we test the performance of the model on this verification and quantitatively obtain the test results, which are quantified using a confusion matrix. The specific operation process is as follows:

[0072] Preprocess and extract features of the validation set signals to form a feature matrix;

[0073] Input the feature matrix into the trained model to obtain the model signal quality score for each window data;

[0074] L1, L2 starts with the lowest model score and divides the threshold with a step size of 0.001. When score ≥ L2, it is judged as category 1, the signal quality is good (Good), and 1 is returned; L1 ≤ score < L2, it is judged as category 2, the signal quality is medium (Acceptable), and 2 is returned; score < L1, the signal quality is poor (Unacceptable). The performance results of the model are obtained and evaluated by the accuracy of the model classification (Accuracy).

[0075] According to the best classification accuracy, determine L1 and L2. In this example, for ECG signals, L1 = 0.0, L2 = 0.03; for respiratory signals, L1 = -0.002, L2 = 0.042.

[0076] 3) Model fine-tuning

[0077] To reduce the risk of misclassifying pathological changes as poor signal quality, we further fine-tuned the model to ensure that the parts classified as poor signal quality have higher credibility, and further processed the data segments with model score < L1:

[0078] For the data segment with score < L1, the original data is further filtered using mean filtering, and then the feature extraction in step 3 is performed. After the feature value is input into the model, the classification score is obtained. If the score is < L1, it is judged as category 3 with poor signal quality (Unacceptable). If the score is > L1, it is judged as category 2 with medium signal quality (Acceptable).

[0079] 4) Test results

[0080] Embodiment 1

[0081] For ECG signals, the 3460 labeled data were used as a validation set and feature extraction was performed and input into the model. The model scoring results were compared with the manually labeled labels to obtain the confusion matrix, as shown in Table 1. The accuracy rate reached 94.97%. Figure 3 shown.

[0082] Table 1. Test set ECG signal quality results

[0083]

[0084] Accuracy: 94.97%

[0085] Embodiment 2

[0086] For respiratory signals, the 2086 labeled data were used as validation sets and feature extraction was performed and input into the model. The thresholds L1 = -0.002 and L2 = 0.042 were determined. The model scoring results were compared with the manually labeled labels to obtain the confusion matrix, as shown in Table 2. The accuracy rate reached 81.06%. The respiratory signal quality results are shown in Figure 4 shown.

[0087] Table 2. Validation set respiratory signal quality results

[0088]

[0089] Accuracy: 81.06%

[0090] Examples:

[0091] Wang XX, male, 176cm, 53 years old, extracted features from his monitored ECG and respiratory signals and input them into the model to obtain the signal quality evaluation results as follows Figure 5-8 shown.

[0092] Unless otherwise defined, all technical and / or scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which the invention relates. The materials, methods and examples mentioned in this application are illustrative only and not limiting.

[0093] Although the present invention has been described in conjunction with specific embodiments, within the spirit of the invention of this application, those skilled in the art may make appropriate substitutions, modifications and changes, and such substitutions, modifications and changes still fall within the scope of protection of this application.

Claims

1. A method for evaluating the quality of a continuous physiological signal, comprising: The original physiological signal is segmented by adding windows; Preprocessing the segmented signal; The preprocessing includes signal baseline removal, bandpass filtering and median filtering; Extracting characteristic values ​​from the preprocessed signal; the characteristic values ​​include skewness, kurtosis, and the distance between two adjacent waveforms calculated based on a dynamic time warping algorithm; The extracted feature values ​​are input into the pre-trained model to obtain the score result; the score result is then judged according to the set thresholds L1 and L2 to obtain the evaluation result of the signal quality; The pre-trained model is an isolation forest model; When the original physiological signal is segmented by windowing, the length of the window is fixed; When removing the signal baseline, scan with a 2-second sliding window to obtain the median of the signal within 2 seconds, that is, the baseline of the current signal. The original signal minus the baseline is used to remove the signal baseline. Skewness according to Perform calculations; Kurtosis according to Perform calculations; Among them, μ is the sample mean, σ is the sample standard deviation, N is the number of samples, Skewness is the skewness, and Kurtosis is the kurtosis.

2. The continuous physiological signal quality assessment method according to claim 1, characterized in that: The pre-trained model is trained on a large sample unlabeled training set, and then verified and the threshold is determined on a small sample manually annotated verification set.

3. The continuous physiological signal quality assessment method according to claim 1, characterized in that: The pre-trained model returns a scoring result of [-1, 1] on the validation set; When score ≥ L2, it is judged as category 1, the signal quality is good, and it returns to category 1; When L1≤score<L2, it is judged as category 2, the signal quality is medium, and it returns to category 2; When score is less than L1, a larger window mean filter is used to remove the baseline for the data segment with score less than L1, and then feature extraction is performed. The extracted feature values ​​are input into the pre-trained model and scored again. If the score is still less than L1, it is judged as category 3 with poor signal quality; if the score is greater than L1, it is judged as category 2 with medium signal quality.

4. The continuous physiological signal quality assessment method according to claim 1, characterized in that: If the physiological signal is an electrocardiogram signal, the window length is 10 seconds; if the physiological signal is a respiratory signal, the window length is 30 seconds.

5. The continuous physiological signal quality assessment method according to claim 1, characterized in that: In the bandpass filtering, a bandpass filter is used for filtering, and the passband is selected according to the characteristics of the signal; if the physiological signal is an electrocardiogram signal, the passband is set to 0.1-100 Hz; if the physiological signal is a respiratory signal, the passband is set to 0.01-20 Hz; In median filtering, a median filter is used to remove outliers, and the filter length of the median filter is set to 40 data points.

6. The continuous physiological signal quality assessment method according to claim 1, characterized in that: The extracted eigenvalues ​​are normalized using a linear function.

7. The continuous physiological signal quality assessment method according to claim 6, characterized in that: The eigenvalue is an eigenvector; use Perform normalization processing; Among them, X norm is the normalized eigenvector, X is the original eigenvector, X min and X max are the minimum and maximum values ​​in the eigenvector respectively.

Images