Motion monitoring system based on heart rate variability features and fatigue state evaluation method
By extracting multidimensional features from ECG signals and using a random forest classifier, a motion monitoring system based on heart rate variability features was constructed. This solves the problem of the single method for motion state assessment in existing technologies and achieves a more scientific and accurate assessment of fatigue state.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTH CHINA UNIV OF TECH
- Filing Date
- 2024-04-19
- Publication Date
- 2026-05-12
AI Technical Summary
Existing methods for assessing exercise status are based solely on heart rate data, which cannot adapt to the different physical conditions of individuals. Furthermore, existing methods based on heart rate variability still rely on a single indicator and cannot provide in-depth analysis of an exerciser's physical condition.
By deeply extracting the time-domain, frequency-domain, and nonlinear-domain features of HRV from the ECG signals of athletes, a random forest classifier is used for pre-training and classification. Combined with the RPE scale report dataset, an exercise monitoring system and fatigue assessment method based on heart rate variability features are constructed.
It enables a more scientific and accurate assessment of exercise status, improves the accuracy of fatigue assessment, and provides diverse evaluation indicators to suit different people's physical conditions.
Smart Images
Figure CN118383744B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to physiological signal processing and state recognition technology, and in particular to a motion monitoring system and fatigue state assessment method based on heart rate variability characteristics. Background Technology
[0002] During exercise, athletes need to assess their physical condition to decide whether to continue. Incorrect assessments can lead to poor exercise decisions, posing a significant threat to their health, and in severe cases, even their lives.
[0003] Accurate assessment of physical condition requires objective data. Current exercise status assessment technologies primarily rely on heart rate data, judging an exerciser's physical condition based on whether their heart rate falls into different heart rate zones. This assessment method uses a single indicator and has relatively fixed standards, failing to adequately adapt to the diverse physical conditions of individuals. Specifically, we have noticed that some existing heart rate-based exercise status assessment methods acquire the user's heart rate under preset exercise conditions and assess the exerciser's status based on the fluctuation range between the newly acquired heart rate and the preset heart rate. While this method offers greater flexibility than heart rate zone-based assessments, the indicator remains singular and cannot provide a deep analysis of the exerciser's physical condition from multiple dimensions. Summary of the Invention
[0004] To address the problems existing in the prior art, this invention provides a motion monitoring system and fatigue assessment method based on heart rate variability characteristics. By deeply extracting the time-domain, frequency-domain, and nonlinear-domain feature information of HRV in the ECG electrocardiogram of the exerciser, body state data based on heart rate variability characteristics is obtained, making the monitoring and assessment more scientific and accurate.
[0005] The system in this embodiment of the invention is implemented using the following technical solution: a motion monitoring system based on heart rate variability characteristics, comprising:
[0006] ECG data acquisition module, used to acquire ECG signals;
[0007] The ECG data processing module is used to process ECG signals, including sliding window slicing.
[0008] The heart rate variability feature extraction module is used to extract heart rate variability features and create a prediction dataset.
[0009] The body state recognition module is used to pre-train a random forest classifier and classify the prediction dataset based on the pre-trained random forest classifier.
[0010] The ECG data processing module performs sliding window slicing on the ECG signal to obtain ECG data slices; the ECG data slices include complete slices and minimum slices, the complete slices are used to extract frequency domain features and nonlinear domain features, and the minimum slices are used to extract time domain features.
[0011] The heart rate variability feature extraction module first extracts the R-peak from the ECG data slices, then performs time-domain analysis and time-domain feature extraction on the R-peak information of the smallest slice, and also performs frequency-domain analysis and frequency-domain feature extraction, nonlinear domain analysis and nonlinear feature extraction on the R-peak information of the complete slice.
[0012] The heart rate variability feature extraction module also merges the heart rate variability features of multiple ECG data slices into a prediction dataset in chronological order.
[0013] Preferably, the body state recognition module pre-trains a random forest classifier, including:
[0014] Obtain the heart rate variability features of the samples and the corresponding RPE scale report datasets, and merge the heart rate variability features of the samples and the corresponding RPE scale report datasets in chronological order to form the sample dataset;
[0015] The sample dataset is divided into a sample training dataset and a sample test dataset;
[0016] Training samples are randomly sampled from the training set and partitioned based on feature data; at the same time, the Gini coefficient of the decision tree nodes is calculated;
[0017] By iterating through all possible split points of the feature subset of the decision tree, the split point with the smallest Gini coefficient is found as the decision tree node, and the training set is divided into two subsets.
[0018] A decision tree is constructed based on the multiple decision tree nodes found.
[0019] Generate multiple decision trees, construct a random forest classifier model based on decision trees, perform sample partitioning, and memorize the partitioning results of each decision tree;
[0020] Save the random forest classifier model based on decision trees.
[0021] Preferably, the window length of the complete slice is greater than the window length of the minimum slice, and the window length of the minimum slice is greater than or equal to the slice sliding step size of the sliding window slice.
[0022] Preferably, the minimum slice window length ensures that each ECG data slice contains at least four complete heartbeats, thus allowing at least three RR intervals for the calculation of time-domain features.
[0023] Preferably, the window length of a complete slice is 20 to 30 seconds, the window length of the smallest slice is 5 seconds, and the slice sliding step of a sliding window slice is 5 seconds or 4 seconds.
[0024] The method in this embodiment of the invention is implemented using the following technical solution: a fatigue state assessment method based on heart rate variability characteristics, comprising the following steps:
[0025] Obtain real-time ECG signals from athletes;
[0026] The obtained ECG signal is sliced by a sliding window to obtain ECG data slices. The ECG data slices include complete slices and minimum slices. The complete slices are used to extract frequency domain features and nonlinear domain features, and the minimum slices are used to extract time domain features.
[0027] Heart rate variability feature extraction is performed on ECG data slices. First, R-peaks are extracted from the ECG data slices. Then, time-domain analysis and time-domain feature extraction of the minimum slices are performed on the R-peak information. In addition, frequency-domain analysis and frequency-domain feature extraction, nonlinear domain analysis and nonlinear feature extraction of the complete slices are performed on the R-peak information.
[0028] Heart rate variability features from multiple ECG data slices were merged into a prediction dataset in chronological order.
[0029] The predicted dataset is input into a pre-trained random forest classifier for classification;
[0030] Obtain the results of physical exercise fatigue status.
[0031] The present invention has the following advantages and effects compared with the prior art:
[0032] 1. The exercise monitoring system and fatigue assessment method based on heart rate variability (HRV) characteristics proposed in this invention obtains body state data based on HRV characteristics by deeply extracting HRV time-domain, frequency-domain, and nonlinear-domain feature information from the exerciser's ECG electrocardiogram signal. Compared with the single indicator (heart rate) of existing physical exercise fatigue assessment methods, the evaluation indicators of this invention are more diverse, and the monitoring and assessment are more scientific and accurate.
[0033] 2. The pre-trained random forest algorithm classification model used in this invention has higher threshold judgment accuracy than existing evaluation techniques, and the standard training set algorithm model combined with the RPE scale and the Bruce protocol is more reasonable and practical. Attached Figure Description
[0034] Figure 1 This is a framework diagram of a motion monitoring system based on heart rate variability characteristics in an embodiment of the present invention;
[0035] Figure 2This is a flowchart of a fatigue state assessment method based on heart rate variability characteristics in an embodiment of the present invention;
[0036] Figure 3 This is one of the schematic diagrams of sliding window slicing of electrocardiogram signals in an embodiment of the present invention;
[0037] Figure 4 This is a second schematic diagram of sliding window slicing of electrocardiogram signals in an embodiment of the present invention;
[0038] Figure 5 This is a flowchart of heart rate variability feature extraction in an embodiment of the present invention;
[0039] Figure 6 This is a flowchart of the training process for the random forest algorithm model in an embodiment of the present invention;
[0040] Figure 7 This is a schematic diagram of the first classification decision tree of the random forest algorithm model in an embodiment of the present invention. Detailed Implementation
[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments and accompanying drawings. However, the embodiments of this invention are not limited thereto. Furthermore, the step numbers in the embodiments of this invention are only for ease of explanation, and the order of the steps is not strictly limited. The execution order of each step in the embodiments can be adjusted for applicability based on the understanding of those skilled in the art.
[0042] Example
[0043] This embodiment provides a motion monitoring system based on heart rate variability characteristics, such as Figure 1 As shown, the system includes:
[0044] The ECG data acquisition module 101 is used to acquire ECG signals; the ECG data acquisition module also has a database for storing the acquired ECG signals. See step S1 below for detailed procedures.
[0045] The ECG data processing module 102 is used to process ECG signals, including sliding window slicing, cleaning, and filtering preprocessing. See steps S2 and S3 below for detailed procedures.
[0046] The heart rate variability feature extraction module 103 is used to extract heart rate variability features, create a prediction dataset, and preprocess the dataset. See steps S4 and S5 below for detailed procedures.
[0047] The body state recognition module 104 is used to pre-train a random forest classifier and classify the prediction dataset based on the pre-trained random forest classifier. See step S6 below for details.
[0048] The result status display module 105 is used to display the obtained physical fatigue status assessment results. Specifically, it may include a display interface and a display device, which may be a watch, a wristband, and / or a mobile phone.
[0049] This embodiment also provides a fatigue state assessment method based on heart rate variability characteristics, such as... Figure 2 As shown, the method includes the following steps:
[0050] S1. Obtain the real-time ECG signal of the athlete.
[0051] The obtained ECG signal should include a clear and accurate R wave; the length of the ECG signal should be no less than 20 seconds. The device used to acquire the ECG signal can be any single-lead or multi-lead ECG monitoring device, such as a heart rate monitor or ECG patch. This step only requires high signal quality; there are no specific requirements for the signal acquisition equipment.
[0052] S2. The obtained ECG signal is sliced by sliding window to obtain ECG data slices, which include complete slices and minimum slices.
[0053] like Figure 3 , Figure 4 As shown, the sliding window slicing method requires dividing the ECG signal into complete slices and minimum slices, which are used for subsequent frequency domain nonlinear domain feature extraction and time domain feature extraction, respectively. A complete slice refers to an ECG slice with a longer time period, from which relatively stable frequency domain and nonlinear domain features can be extracted, while a minimum slice refers to an ECG slice with a shorter time period, from which relatively stable time domain features can be extracted.
[0054] A signal length of 20 to 30 seconds is sufficient for effective extraction of frequency and nonlinear domain features in step S4. Signals that are too short are difficult to extract these features. However, 20 seconds is too long for time-domain feature extraction, resulting in wasted data. Therefore, this embodiment employs different slicing methods to extract time-domain features: the window length of the complete slice is greater than the window length of the minimum slice, and the window length of the minimum slice is greater than or equal to the sliding window slice's sliding step size. The slicing time for time-domain features (i.e., the minimum slice window length) needs to ensure that each ECG data slice contains at least four complete heartbeats, thus guaranteeing at least three RR intervals (the RR interval being the interval between two heartbeat R peaks) for calculation to complete the time-domain feature calculation. Therefore, this embodiment preferably uses a complete slice window length of 20 seconds and a minimum slice window length of 5 seconds. The sliding window slicing also requires determining the slice sliding step size; this embodiment selects a 5-second sliding step size.
[0055] It should be noted that the data slices in this embodiment are allowed to overlap, meaning that the new data sequences formed after slicing contain reused parts. When the slice sliding step size is exactly the same as the minimum slice window length, the complete slice data overlaps, while the minimum slice data does not overlap, such as... Figure 3 As shown; when the slice sliding step size is less than the minimum slice window length, for example, when the minimum slice window length is selected as 5 seconds and the slice sliding step size is selected as 4 seconds, overlap also occurs between the minimum slice data, such as... Figure 4 As shown.
[0056] S3. Perform data preprocessing such as filtering on each obtained ECG data slice.
[0057] First, power frequency interference needs to be filtered out using 50Hz and 60Hz Butterworth digital notch filters; then, bandpass filtering is performed. Preferably, in this embodiment, a 3-70Hz bandpass filter is selected, which can retain the frequency components of 3-70Hz in the ECG signal and remove the low-frequency part below 3Hz and the high-frequency part above 70Hz. After filtering and other preprocessing, electromechanical interference and baseline drift in the ECG signal can be removed to a certain extent.
[0058] S4. Extract heart rate variability (HRV) features from the pre-filtered ECG data slices.
[0059] like Figure 5 As shown, the time-domain features, frequency-domain features, and nonlinear domain features of the filtered ECG data slices are extracted, specifically including the following steps:
[0060] S401. Extract the R peak from the ECG data slice (i.e., the ECG electrocardiogram data of the slice).
[0061] S402. Perform time-domain analysis and time-domain feature extraction on the extracted R-peak information using the minimum slice (preferably 5 seconds in this embodiment).
[0062] S403. Perform frequency domain analysis and frequency domain feature extraction on the extracted R peak information by performing a complete slice (preferably 20 seconds in this embodiment).
[0063] S404. Perform nonlinear domain analysis and nonlinear feature extraction on the extracted R peak information by performing a complete slice (preferably 20 seconds in this embodiment).
[0064] S405. Save the acquired time-domain features, frequency-domain features, and nonlinear domain features.
[0065] The time-domain, frequency-domain, and nonlinear-domain characteristics of the extracted ECG data slices in this embodiment are shown in Table 1 below.
[0066] Table 1
[0067]
[0068]
[0069] The descriptions of each feature variable in Table 1 are as follows:
[0070] In the frequency domain characteristics, HRV_HF represents the high-frequency band (0.15-0.4Hz) of the HRV power spectrum, HRV_VHF represents the very high-frequency band (0.4-0.5Hz) of the HRV power spectrum, HRV_TP represents the total power of the HRV power spectrum, HRV_HFn represents the normalized value of the high-frequency band (0.15-0.4Hz) of the HRV power spectrum, and HRV_LnHF represents the natural logarithm of the high-frequency band (0.15-0.4Hz) of the HRV power spectrum. These frequency domain characteristics were obtained through statistical analysis following a Fast Fourier Transform.
[0071] In the time-domain features, HRV_MeanNN is the mean of the RR intervals, HRV_SDNN is the standard deviation of the RR intervals, HRV_RMSSD is the square root of the mean of the squares of the consecutive differences between adjacent RR intervals, HRV_SDSD is the standard deviation of the consecutive differences between RR intervals, HRV_CVNN is the standard deviation of the RR intervals (SDNN) divided by the mean of the RR intervals (MeanNN), and HRV_CVSD is the root mean square of the consecutive differences (RMSSD) divided by the mean of the RR intervals (MeanNN). HRV_MedianNN is the median of the RR interval; HRV_SDRMSSD is the ratio of the standard deviation of the RR interval to the root mean square of the continuous differences (SDNN / RMSSD); HRV_Prc20NN is the 20th percentile of the RR interval, HRV_Prc80NN is the 80th percentile of the RR interval, HRV_MinNN is the minimum RR interval; HRV_MaxNN is the maximum RR interval; HRV_HTI is the HRV trigonometric index, which measures the total number of RR intervals divided by the height of the RR interval histogram.
[0072] Nonlinear domain features include other exponential dimensions corresponding to Poincaré plots, entropy, or fractals. Due to the large number, only a portion are explained here. HRV_SD1 is the standard deviation perpendicular to the same line, HRV_SD2 is the standard deviation along the same line, and HRV_SD1SD2 is the ratio of SD1 to SD2; HRV_CSI is the cardiac sympathetic index, a measure of cardiac sympathetic function independent of vagal activity, calculated by dividing the longitudinal variability (4×SD2) of the Poincaré plot by its transverse variability (4×SD1); HRV_CVI is the cardiac vagus index, an index of cardiac parasympathetic function, and is equal to the logarithm of the product of longitudinal variability (4×SD2) and transverse variability (4×SD1).
[0073] S5. For the heart rate variability features of the obtained multiple ECG data slices, perform data merging and data cleaning in chronological order to obtain an effective and usable prediction dataset.
[0074] Short-term data may be unstable, leading to significant differences in model predictions over consecutive time periods. To ensure data stability, heart rate variability data from multiple ECG data slices are merged into a single dataset in chronological order to obtain a usable prediction dataset.
[0075] Preferably, in this embodiment, the HRV feature data obtained every 5 seconds (i.e., one data slice) is considered as one valid data point, and every 12 valid data points are merged into a prediction dataset in chronological order. The merged prediction dataset is then cleaned, and missing values are filled in using the mean of the nearest data points before and after the sequence.
[0076] S6. Input the available prediction dataset into the pre-trained random forest classifier for classification.
[0077] The random forest classifier used in this step is pre-trained and directly loaded into the model for use during the classification task. The training data source and training method for the random forest classifier in this embodiment are as follows:
[0078] Training data source: Treadmill experiment using the modified Bruce protocol. Participants were healthy adults. Participants were instructed to wear a single-lead ECG monitor to record ECG data during exercise. Simultaneously, participants reported their RPE (Recovery Point Expiration) value every minute as a reference for their fatigue level. The Bruce protocol is an extreme exercise test where participants achieve fatigue by increasing the speed and incline of the treadmill every three minutes. The modified Bruce protocol is an improvement on the standard Bruce protocol, featuring a gentler treadmill incline and lower speed, making it more suitable for average-ability individuals. The modified Bruce protocol is shown in Table 2 below.
[0079] Table 2
[0080]
[0081] The RPE (Responsive Exertion) value comes from the Borg scale. The Borg scale, created by physiological psychologist Gunnar Borg in the 1970s, divides exercise intensity into 1 to 20 different levels. 1 represents no effort, and 20 represents extreme effort; the commonly used range starts from 6. During exercise, users need to rate their performance based on their own feelings. The Borg scale used in the experiment is shown in Table 3 below.
[0082] Table 3
[0083]
[0084]
[0085] The training and test datasets used in this embodiment are derived from this experiment.
[0086] The training of a random forest classifier consists of eight steps, such as... Figure 6 As shown, it specifically includes:
[0087] S601. Obtain the heart rate variability features of the sample and the corresponding RPE scale report dataset, and merge the heart rate variability features of the sample and the corresponding RPE scale report dataset in chronological order to form the sample dataset.
[0088] Real-time ECG data were acquired from the ECG acquisition device worn by the subjects. Heart rate variability (HRV) characteristics were calculated using steps S2-S4, specifically including time-domain, frequency-domain, and nonlinear-domain characteristics. The calculated HRV characteristics were then correlated with the subjects' real-time RPE (Recovery Point Expiration) reports recorded during the experiment, chronologically. Simultaneously, fatigue levels were categorized: RPE 6-10 was categorized as Label 1: No fatigue at all; RPE 11-15 as Label 2: Some fatigue; and RPE 16-20 as Label 3: Very fatigued. Finally, the aforementioned HRV characteristics and corresponding fatigue level labels were merged chronologically to form the sample dataset.
[0089] S602. Divide the sample dataset into a sample training dataset and a sample test dataset.
[0090] To train and evaluate the random forest classification model, it is necessary to divide the dataset into training and testing sets. Preferably, in this embodiment, the ratio of the sample dataset to the training set is 7:3, that is, 70% of the sample dataset is divided into the training set and 30% into the testing set. Furthermore, this embodiment uses stratified sampling based on label values for the partitioning.
[0091] S603. Randomly sample training samples from the training set and divide them based on feature data; at the same time, calculate the Gini coefficient (also called Gini index) of the decision tree nodes.
[0092] The Gini coefficient is used for selecting decision trees. The selection criterion for a decision tree is that each child node achieves the highest purity, meaning that all observations falling into a child node belong to the same category. At this point, the Gini coefficient is minimized, purity is maximized, and uncertainty is minimized.
[0093] For a typical decision tree, if there are K classes in total, the probability that a sample belongs to the k-th class is P. k Then the Gini coefficient of this probability distribution is:
[0094]
[0095] S604. Iterate through all possible split points of the feature subset of the decision tree, find the split point with the smallest Gini coefficient as the decision tree node, and divide the training set into two subsets.
[0096] When traversing each node (i.e., split point) of each feature subset, if the training set D is divided into two parts using feature A = a, namely training subset D1 (the set of samples that satisfy A = a) and training subset D2 (the set of samples that do not satisfy A = a), then the Gini coefficient of training set D under the condition of feature A = a is:
[0097]
[0098] Wherein, Gini(D) represents the uncertainty of the training set D; Gini(A,D) represents the uncertainty of the training set D after A=a split.
[0099] S605. Based on the multiple decision tree nodes found, form a decision tree.
[0100] S606. Generate multiple decision trees, construct a random forest classifier model based on decision trees, perform sample partitioning, and memorize the partitioning results of each decision tree.
[0101] S607. Based on the different results of the generated decision trees, a vote is taken, and the category with the most votes is output as the classification result of the random forest classifier.
[0102] S608. Save the random forest classifier model based on decision trees, and use the test set for testing and model evaluation.
[0103] In this embodiment, the constructed random forest classifier model based on decision trees contains 120 classification decision trees, with 147 test samples, 146 of which were correctly classified, resulting in a classifier accuracy of 0.993197. The model of the first classification decision tree is as follows: Figure 7 As shown.
[0104] In step S6, it is only necessary to process the ECG data according to the aforementioned steps to obtain the prediction dataset from step S5 and load it into the trained random forest classifier to obtain the results of physical exercise fatigue status. Users are not required to collect ECG data according to the Bruce treadmill experiment, nor is it necessary to report the RPE value.
[0105] S7. Obtain the results of physical exercise fatigue status, including: not fatigued, fatigued, and very fatigued. After obtaining the results of physical exercise fatigue status, display them through terminal devices such as mobile phones and sports watches, and feed back the results of exercise fatigue status to the operator and exerciser to achieve the purpose of real-time assessment and monitoring.
[0106] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.
Claims
1. A motion monitoring system based on heart rate variability characteristics, characterized in that, include: ECG data acquisition module, used to acquire ECG signals; The ECG data processing module is used to process ECG signals, including sliding window slicing. The heart rate variability feature extraction module is used to extract heart rate variability features and create a prediction dataset. The body state recognition module is used to pre-train a random forest classifier and classify the prediction dataset based on the pre-trained random forest classifier. The ECG data processing module performs sliding window slicing on the ECG signal to obtain ECG data slices; the ECG data slices include complete slices and minimum slices, the complete slices are used to extract frequency domain features and nonlinear domain features, and the minimum slices are used to extract time domain features. The heart rate variability feature extraction module first extracts the R-peak from the ECG data slices, then performs time-domain analysis and time-domain feature extraction on the R-peak information of the smallest slice, and also performs frequency-domain analysis and frequency-domain feature extraction, nonlinear domain analysis and nonlinear feature extraction on the R-peak information of the complete slice. The heart rate variability feature extraction module also merges the heart rate variability features of multiple ECG data slices into a prediction dataset in chronological order; A complete slice refers to an ECG slice with a relatively long time interval, which can extract stable frequency domain and nonlinear domain features; a minimum slice refers to an ECG slice with a relatively short time interval, which can extract stable time domain features. The window length of a complete slice is greater than the window length of the smallest slice, and the window length of the smallest slice is greater than or equal to the slice sliding step size of a sliding window slice.
2. The motion monitoring system according to claim 1, characterized in that, The body state recognition module pre-trains a random forest classifier, including: Obtain the heart rate variability features of the samples and the corresponding RPE scale report datasets, and merge the heart rate variability features of the samples and the corresponding RPE scale report datasets in chronological order to form the sample dataset; The sample dataset is divided into a sample training dataset and a sample test dataset; Training samples are randomly sampled from the sample training dataset and partitioned based on feature data; at the same time, the Gini coefficient of the decision tree nodes is calculated; By iterating through all possible split points of the feature subset of the decision tree, the split point with the smallest Gini coefficient is found as the decision tree node, and the sample training dataset is divided into two subsets. A decision tree is constructed based on the multiple decision tree nodes found. Generate multiple decision trees, construct a random forest classifier model based on decision trees, perform sample partitioning, and memorize the partitioning results of each decision tree; Save the random forest classifier model based on decision trees.
3. The motion monitoring system according to claim 1, characterized in that, The minimum slice window length ensures that each ECG data slice contains at least four complete heartbeats, thus requiring at least three RR intervals for the calculation of time-domain features.
4. The motion monitoring system according to claim 1, characterized in that, The window length for a complete slice is 20 to 30 seconds, the window length for the smallest slice is 5 seconds, and the slide step for a sliding window slice is 5 seconds or 4 seconds.
5. A fatigue state assessment method based on heart rate variability characteristics, characterized in that, Includes the following steps: Obtain real-time ECG signals from athletes; The obtained ECG signal is sliced by a sliding window to obtain ECG data slices. The ECG data slices include complete slices and minimum slices. The complete slices are used to extract frequency domain features and nonlinear domain features, and the minimum slices are used to extract time domain features. Heart rate variability feature extraction is performed on ECG data slices. First, R-peaks are extracted from the ECG data slices. Then, time-domain analysis and time-domain feature extraction of the minimum slices are performed on the R-peak information. In addition, frequency-domain analysis and frequency-domain feature extraction, nonlinear domain analysis and nonlinear feature extraction of the complete slices are performed on the R-peak information. Heart rate variability features from multiple ECG data slices were merged into a prediction dataset in chronological order. The predicted dataset is input into a pre-trained random forest classifier for classification; To obtain results regarding physical fatigue status; A complete slice refers to an ECG slice with a relatively long time interval, which can extract stable frequency domain and nonlinear domain features; a minimum slice refers to an ECG slice with a relatively short time interval, which can extract stable time domain features. The window length of a complete slice is greater than the window length of the smallest slice, and the window length of the smallest slice is greater than or equal to the slice sliding step size of a sliding window slice.
6. The fatigue state assessment method according to claim 5, characterized in that, The minimum slice window length ensures that each ECG data slice contains at least four complete heartbeats, thus requiring at least three RR intervals for the calculation of time-domain features.
7. The fatigue state assessment method according to claim 6, characterized in that, The window length for a complete slice is 20 to 30 seconds, the window length for the smallest slice is 5 seconds, and the slide step for a sliding window slice is 5 seconds or 4 seconds.
8. The fatigue state assessment method according to claim 5, characterized in that, The pre-training process of a random forest classifier includes: Obtain the heart rate variability features of the samples and the corresponding RPE scale report datasets, and merge the heart rate variability features of the samples and the corresponding RPE scale report datasets in chronological order to form the sample dataset; The sample dataset is divided into a sample training dataset and a sample test dataset; Training samples are randomly sampled from the sample training dataset and partitioned based on feature data; at the same time, the Gini coefficient of the decision tree nodes is calculated; By iterating through all possible split points of the feature subset of the decision tree, the split point with the smallest Gini coefficient is found as the decision tree node, and the sample training dataset is divided into two subsets. A decision tree is constructed based on the multiple decision tree nodes found. Generate multiple decision trees, construct a random forest classifier model based on decision trees, perform sample partitioning, and memorize the partitioning results of each decision tree; Save the random forest classifier model based on decision trees.
9. The fatigue state assessment method according to claim 5, characterized in that, Before extracting heart rate variability features from the ECG data slices, each ECG data slice is preprocessed with filtering.