A method for assessing sports risk based on machine learning

Abstract

The present application belongs to the field of medical care information technology, and particularly relates to a movement risk assessment method based on machine learning, wherein historical data and injury records are arranged to obtain a data set, high correlation indexes related to various movement injuries are obtained through feature screening, and independent regression models are established between the high correlation indexes and the risk of various injuries; the physical quality and physiological index system data of a new training personnel group are obtained through testing, the regression models are used to predict the injury risk indexes of each part, the training personnel are grouped and trained through K-means clustering algorithm according to the injury risk indexes; and the historical training load data, functional action screening scores and the next day training plan content are fused to construct a model to evaluate the injury risk index of the training personnel on the next day, so that the present method is suitable for large-scale training groups, the index dimensions are comprehensive and targeted, and the accuracy and efficiency of risk assessment are improved.

Description

Technical Field

[0001] This invention belongs to the field of healthcare information technology, specifically relating to a machine learning-based method for assessing exercise risk. Background Technology

[0002] Sports risk assessment systematically identifies potential injury triggers during exercise, which helps prevent sports injuries and provides core assurance for training safety and effectiveness. Its importance is particularly evident in military and sports training scenarios. The incidence of training injuries in the armed forces directly affects combat effectiveness, and accurate assessment can reduce non-combat casualties through early intervention. In the field of public sports, regular assessments can promptly locate weaknesses such as abnormal joint mobility and muscle imbalances. Early screening for problems such as scoliosis in adolescents and knee strain in runners can prevent injury from worsening and optimize training efficiency.

[0003] Traditional assessment methods primarily rely on personal feelings and human experience, including Functional Movement Screening (FMS), expert assessment, and biomechanical analysis. FMS requires professionally trained personnel, its scoring is influenced by subjective judgment, and results from different assessors on the same subject show low consistency, making it impossible to generate scientifically targeted training or recovery recommendations. Expert assessment is also mainly qualitative, lacking biological quantitative indicators and making it difficult to accurately predict individual risk thresholds. Biomechanical analysis measures parameters such as joint angles through 3D modeling, but the equipment is expensive and complex, limiting its application to laboratory environments and failing to meet the real-time assessment needs of military field training and mass sports activities.

[0004] Artificial intelligence (AI) is a system that can utilize perceived information and take actions based on the highest probability to achieve goals. It can overcome the technical bottlenecks of traditional methods through data-driven approaches and has already seen initial applications in the field of sports risk assessment. Chinese invention patent application CN119943385A discloses a method and system for military physical training risk assessment based on deep learning. The system includes: collecting fixed influencing factors, calculating altitude attenuation coefficient functions and terrain attenuation coefficient functions, constructing a deep learning model, collecting vital sign data, risk assessment, training feedback, and updating the subjective attenuation coefficient function. The system comprises modules for influencing data collection, altitude, terrain, model construction, vital sign data collection, risk assessment, training feedback, and updating. By introducing the influencing factors of terrain flatness and altitude, and calculating terrain attenuation coefficients and altitude attenuation coefficients, it improves the comprehensiveness and accuracy of risk assessment, enhancing the safety and efficiency of training. Chinese invention patent application CN118053579A discloses a comprehensive assessment method for exercise risk in community fitness centers based on machine learning. The method includes screening influencing factors of exercise risk in community fitness centers, collecting relevant data on exercise risk in community fitness centers, establishing a comprehensive assessment index system for exercise risk in community fitness centers, obtaining exercise risk assessment index data for community fitness centers, establishing a general comprehensive assessment model for exercise risk in community fitness centers based on machine learning methods, conducting a comprehensive assessment of exercise risk in community fitness centers, selecting appropriate risk prevention and control strategies, and establishing an exercise risk assessment index system for community fitness centers from three aspects: individual conditions of exercisers, exercise behavior, and exercise environment.

[0005] However, there are still some problems with the application of artificial intelligence in sports risk assessment. For example, for large-scale training personnel, such as military units or schools, due to the requirements of venues and equipment, it is usually necessary to implement a consistent training plan. However, since the physical condition of each individual trainee is different, their tendency to experience sports risks is also different. As a result, the current sports risk assessment lacks comprehensiveness and specificity, which limits the generalization ability of the model and results in insufficient accuracy of group assessment. Summary of the Invention

[0006] This invention provides a motion risk assessment method based on machine learning, and the complete technical solution includes:

[0007] A machine learning-based method for sports risk assessment involves organizing historical data from a physical fitness and physiological indicator system and sports injury records to obtain a first dataset. The physical fitness and physiological indicator system includes body morphology indicators, body composition indicators, and athletic performance indicators. The sports injury records include the number of injuries to seven areas: neck, shoulder, elbow, lower back, abdomen, knee, and ankle.

[0008] Feature screening was performed to obtain high-correlation indicators for sports injuries in different parts of the body, and independent risk regression models were established for each high-correlation indicator and each sports injury in different parts of the body.

[0009] The physical fitness and physiological indicators of the new group of trainees were tested to obtain systematic data. The injury risk index of each part of their body was analyzed by XGBoost risk regression model. Based on the injury risk index, the trainees were grouped by K-means clustering algorithm, and the key sports risk parts of each group were determined according to the average risk index within the group.

[0010] During the training process, an XGBoost prediction model is constructed to assess the injury risk index of key sports risk areas for each group of trainees by integrating historical training load data, functional movement screening scores, and the training plan for the next day.

[0011] The body morphology indicators include: age, height, weight, chest circumference, waist circumference, hip circumference, upper arm circumference, thigh circumference, calf circumference, and body mass index.

[0012] The body composition indicators include: protein content, inorganic content, water content, muscle mass, and lean body mass.

[0013] The athletic performance indicators include:

[0014] (1) Strength category:

[0015] 1) Maximum strength: Seated chest press, seated leg extension, seated back pull;

[0016] 2) Swift power: Reverse jump, standing long jump, medicine ball throw;

[0017] 3) Strength and endurance: Pull-ups, dips, plank;

[0018] (2) Endurance: 3000-meter run;

[0019] (3) Speed: 30m sprint;

[0020] (4) Flexibility: Seated forward bend and back-grabbing test;

[0021] (5) Sensitivity: Illinois test, 30×2 shuttle run;

[0022] (6) Balance ability: YBT test.

[0023] The scores or results of the above tests.

[0024] The highly correlated indicators include: 3000-meter run, plank, YBT test results, body mass index (BMI), body fat percentage, seated leg extension maximum strength, standing long jump, thigh circumference, waist circumference, back grab test, reverse jump, muscle mass, and Illinois test results.

[0025] Furthermore, for trainees with multiple group preferences, first calculate the Euclidean distance from each trainee to all cluster centers. If the difference between the largest and second largest distances is less than... This is considered a tendency towards multiple groups. At this point, the key movement risk areas of the corresponding clusters are matched according to the area with the highest risk coefficient, and finally a unique group is determined.

[0026] By integrating historical training load data, functional movement screening scores and the training plan content for the next day, an XGBoost prediction model was constructed to assess the injury risk index of seven parts of the trainees the next day, including: (1) quantification of training load, (2) model input features, (3) model training, and (4) sports risk assessment.

[0027] Training load quantification includes:

[0028] (1) Determine the exercise load index: Label all training exercises and define the load index of each exercise on 7 body parts, with a value range of [0,5].

[0029] (2) Determine the total load of a single movement: For a single movement, calculate the total load of key movement risk areas within the group based on training parameters. :

[0030]

[0031] in: This represents the load index on the target area caused by the action. For the number of groups, For each set of numbers, Duration of a single action, in seconds. This represents the action order, with a value ranging from 1.0 to 1.5.

[0032] (3) Determine the total daily load: The total daily load for a certain part is the sum of the total load of all movements on the key movement risk parts on that day. :

[0033]

[0034] in, The number of training exercises for the day. For the first The total load on a certain part of the body by a single action.

[0035] Historical training load data includes cumulative load, average load, maximum load, load growth rate, and number of consecutive days with high load.

[0036] The FMS test includes squats, hurdles, straight lunges, shoulder flexibility, active straight leg raises, trunk stability push-ups, and rotational stability.

[0037] After grouping, training was conducted using standard movements from the standard movement library, and training plans were designed for each group.

[0038] All training movements are labeled, and the load index of each movement on 7 body parts is defined.

[0039] Furthermore, the input features of the XGBoost prediction model are the key motion risk areas for each group of trainees:

[0040] Cumulative load : , For the first Total daily workload for each action;

[0041] Average load ;

[0042] Maximum load ;

[0043] Load growth rate ;

[0044] Number of consecutive high-load days In the past 7 days The number of days;

[0045] Historical load and FMS interaction items: , , For the first i The score value of each action;

[0046] Next-day load to historical load ratio: , The predicted total load for the key movement risk areas of the trainees in the next day's training;

[0047] The output feature is the injury risk index of the key sports risk areas of each group of trainees on the following day.

[0048] The advantages of this invention over the prior art are:

[0049] 1. This invention is based on machine learning thinking, using supervised learning (regression prediction) + unsupervised learning (clustering and grouping) as its core technical approach and framework. At the supervised learning level, it achieves feature selection, filtering highly correlated features to improve model efficiency and avoid the curse of dimensionality. Compared with traditional correlation analysis-based selection (which usually only captures linear correlations), it can capture the non-linear correlation between features and labels through decision trees, and the selected features are more consistent with the complex patterns of actual injury occurrence. At the unsupervised learning level, soldiers with similar risk patterns are clustered and grouped, realizing the rapid division of large groups into small groups. Targeted assessments are performed on key movement risk areas within the group, avoiding the arbitrariness of manual grouping and completing the transformation from individual prediction to group management. For samples with minimal distance differences to multiple cluster centers (fuzzy samples in unsupervised learning), combined with the highest risk area predicted by supervised learning, they are matched to the cluster where the key movement risk area is consistent with the soldier's highest risk area, solving the fuzziness problem of unsupervised clustering and ensuring that each sample is uniquely grouped. The XGBoost risk regression model is combined with the K-means clustering model to form a closed loop of static risk assessment, group training and dynamic risk monitoring. The model design is in line with the time continuity and risk dynamics of troop training, improves training effectiveness and is in line with the development direction of scientific and standardized training.

[0050] 2. For large-scale training groups, such as military units that require unified training and testing, basic indicators such as body shape, body composition, and athletic performance, as well as the load index of training movements, are labeled and quantified. This breaks through the limitations of traditional single-dimensional indicators that only consider athletic performance, and constructs a multi-dimensional indicator system. The data examination is more comprehensive and systematic, and it fully captures the main risk sources that may induce injuries in military training, such as static basic defects, dynamic overload, and insufficient functional status. The data dimensions are more comprehensive.

[0051] 3. By filtering and extracting basic indicators, the prediction accuracy is ensured, the data structure is optimized, redundancy in data dimensions of large-scale troops is avoided, and the data collection method is reasonable, which can support the simultaneous collection of data by large-scale training groups.

[0052] 4. Traditional methods often use multi-output regression models, employing a single model to predict risks in multiple body parts simultaneously. However, the causes of injury vary greatly across different body parts, making it difficult for a single model to simultaneously optimize the prediction loss for multiple labels. This invention constructs independent regression prediction models for each of the seven body parts, overcoming the problems of general assessment and coarse intervention in existing technologies. It achieves precise risk management at the body part level, resulting in higher prediction accuracy: avoiding risk interference between body parts, improving the accuracy of single-body prediction, and more accurate feature matching: conforming to the biomechanical characteristics of the body parts, avoiding the problem of one-size-fits-all general features, and providing more refined and focused differentiation in the prediction of motion risks.

[0053] 5. By integrating historical training load with FMS for risk assessment, a full-chain correlation of load accumulation, functional status, and injury risk is achieved. This is particularly suitable for high-intensity training scenarios in the armed forces, making up for the limitations of single indicators, capturing non-linear risk correlations, supporting dynamic and precise intervention, and adapting to real-time adjustments in training scenarios.

[0054] 6. Addressing the problem that existing technologies focus only on single-intensity or cumulative load, neglecting dynamic load changes, leading to incomplete and delayed predictions of injury risk, this invention employs a multi-indicator synergistic assessment in selecting historical training load data. This aligns with the injury mechanism of "long-term accumulation + short-term triggering." For example, for acute injuries (such as muscle strains), risk can be identified through maximum load (single excessive load) and load growth rate (sudden increase), with intervention prioritizing reduction of peak load and control of the growth rate. For chronic injuries (such as stress fractures), risk can be identified through cumulative load (total excessive load) and the number of consecutive days of high load (insufficient recovery). This multi-dimensional indicator selection, compared to the single-indicator selection of existing technologies, more accurately captures the complex mechanisms of load-induced injuries in troop training, providing more effective input features for subsequent risk prediction models, ultimately improving the accuracy of injury risk assessment and the targeting of interventions. Attached Figure Description

[0055] To more clearly illustrate the content and specific embodiments of this invention, the accompanying drawings required for the description process are introduced. The following drawings are merely for the purpose of illustrating the subject matter of this invention. Those skilled in the art can readily provide other similar drawings based on the drawings in this invention without any inventive effort.

[0056] Figure 1 This is a schematic diagram of the motion risk assessment method based on machine learning of the present invention. Detailed Implementation

[0057] The present invention will now be described in detail with reference to embodiments and accompanying drawings. However, it should be understood that the embodiments and drawings are for illustrative purposes only and do not constitute any limitation on the scope of protection of the present invention. All reasonable modifications and combinations included within the inventive spirit of the present invention fall within the scope of protection of the present invention.

[0058] like Figure 1 As shown, a machine learning-based sports risk assessment method is designed for groups such as the military and schools that require large-scale physical fitness or physical training. Based on historical data, a correlation model is established between physical fitness indicators and physiological indicators and sports injury records of different parts of the body. Key indicators are selected and scientifically grouped. The training process is monitored and the risk assessment of sports injuries is carried out.

[0059] First, historical data is analyzed. The current physical fitness and physiological indicators in this implementation include:

[0060] 1. Body morphology indicators (10 items): age, height (cm), weight (kg), chest circumference (cm), waist circumference (cm), hip circumference (cm), upper arm circumference (cm, relaxed state), thigh circumference (cm, mid-thigh), calf circumference (cm, calf at its thickest point), body mass index (BMI, kg / m²).

[0061] 2. Body composition indicators (6 items): protein content (kg), inorganic content (bone, minerals, kg), body fat percentage (%), water content (kg), muscle mass (kg), lean body mass (kg).

[0062] 3. Performance indicators (16 items), including: (1) Strength: 1) Maximum strength: seated chest press (kg), seated leg extension (kg), seated pull-back (kg); 2) Speed ​​strength: reverse jump (height, cm), standing long jump (m), medicine ball throw (m); 3) Strength endurance: pull-ups (times), dips (times), plank (minutes). (2) Endurance: 3000-meter run (minutes). (3) Speed: 30-meter sprint (seconds). (4) Flexibility: seated forward bend (static, cm), back grab test (hands clasped behind the back, middle finger spacing, cm). (5) Agility: Illinois test (seconds), 30×2 shuttle run (seconds). (6) Balance ability: YBT test (dominant leg, forward, cm).

[0063] Injury records include the number of injuries to different parts of the body, including: upper limbs: neck, shoulder, elbow; core: lower back, abdomen; lower limbs: knee, ankle.

[0064] Numerical examples are shown in Table 1:

[0065] Table 1 Examples of Historical Data

[0066]

[0067] First, the above data is preprocessed. The values ​​are normalized to the [0,1] interval, and missing values ​​are filled with the mean. The injury records are then labeled, with each of the following areas—neck, shoulder, elbow, waist, abdomen, knee, and ankle—as a single label (the number represents the number of injuries, 0 = no injury), resulting in a 7-dimensional vector of labels. For example, [0,1,0,2,0,0,0] indicates one injury to the shoulder, two injuries to the waist, and no injuries to other areas.

[0068] Subsequently, feature filtering was performed, calculating the importance score of each indicator or parameter, and excluding irrelevant indicators to reduce model complexity. Specifically:

[0069] Input features X: preprocessed physical fitness and physiological indicators from 800 soldier data points. Output label y: a 7-dimensional vector of injury frequency.

[0070] XGBoost models were trained for each of the seven body parts, with the objective function of the model being... The sum of the loss function and the regularization term:

[0071]

[0072] For loss function, For regularization terms, This is the set of all parameters of the model.

[0073] Where the loss function Using squared loss, , m For the total number of soldiers, For the first The actual number of injuries and illnesses suffered by each soldier. For the first The predicted number of injuries or illnesses for each soldier.

[0074] Feature importance is quantified using split gain: the importance score of each feature is equal to its total gain when it is a split node in all decision trees. The importance scores of each feature in the seven models are summed, and the average is taken as the final importance. Ranked by average importance, the 13 most relevant indicators to sports injuries are obtained: 3000-meter run time (minutes), plank time (minutes), YBT test result (cm), body mass index (BMI), body fat percentage (%), seated leg extension maximum force (kg), standing long jump (m), thigh circumference (cm), waist circumference (cm), back grab test (cm), reverse jump height (cm), muscle mass (kg), and Illinois test (seconds).

[0075] Based on the above results, further analysis revealed highly correlated indicators for the occurrence of injuries and illnesses in various locations:

[0076] Table 2 High-correlation indicators for injuries and illnesses at various sites

[0077]

[0078] Based on the above results, the XGBoost risk regression model was adopted to establish mapping models between the 13 highly correlated indicators and the risk of injuries at various locations. During training, the input features were a 13-dimensional vector composed of the normalized values ​​of the highly correlated indicators for each soldier. For the output parameters, the number of historical injuries was normalized and used as the risk index. Used for model training. For the first Risk index of each sample This represents the maximum number of injuries or illnesses to that area. The minimum number of injuries to this site is defined. Mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) are used as model evaluation metrics. Risk levels are defined as: low risk: [0, 0.3), medium risk: [0.3, 0.7), and high risk: [0.7, 1].

[0079] After measuring the indicators of new trainees, a trained XGBoost risk regression model is used to predict the injury risk index at various locations. The input features are a 13-dimensional vector composed of highly correlated indicator values ​​for each new recruit, and the output is a multi-dimensional vector composed of the risk index of each body part, as shown in the example below:

[0080] Table 3 Examples of Risk Indices at Various Locations

[0081]

[0082] Subsequently, based on the predicted injury risk index of new recruits, the K-means clustering algorithm was used to group trainees with similar risk patterns into one group (10-25 people per group), and the problem of multiple grouping was solved by prioritizing the highest risk body parts.

[0083] The similarity of risk patterns refers to the consistent distribution trend of risk indices across the seven body parts of trainees (e.g., all showing high risk in the knee and ankle or high risk in the waist and neck). The tendency towards multiple groups refers to situations where, in the clustering results, the distance differences to multiple cluster centers are extremely small (distance differences are all less than 0.1), making it impossible to directly determine a unique group based on distance. In such cases, the key movement risk areas of the corresponding cluster are matched with the body parts with the highest risk coefficients, ultimately determining a unique group.

[0084] First, determine the optimal number of clusters. Based on the total number of trainees and the range of people in each group, determine the range of the number of clusters. In this implementation, K∈[20,50]. The process of handling trainees who are biased towards multiple groups is as follows: First, calculate the Euclidean distance from each trainee to all cluster centers. If the difference between the largest distance and the second largest distance is less than... If so, it is considered a tendency towards multiple groups, as described in this embodiment. If the value is 0.1, then the key movement risk areas of the corresponding clusters are matched according to the area with the highest risk coefficient, and finally a unique group is determined. Some grouping results are shown in the table below.

[0085] Table 4 Grouping Examples

[0086]

[0087] To simplify the subsequent analysis and prediction process, the key sports risk areas here refer to the areas with an average risk index ≥ 0.65 in each group. For groups without an average risk index ≥ 0.65, the area with the highest average risk index in that group is taken as the key sports risk area.

[0088] After grouping, design training plans based on the standard movements in the standard movement library and conduct training. An example of the standard movement library is shown in the table below:

[0089] Table 5 Examples of Standard Action Library Classification

[0090]

[0091] By collecting and integrating historical training load data, Functional Movement Screening (FMS) scores, and the next day's training plan, a dataset (as shown in Table 6) was constructed. An XGBoost prediction model was then built to assess the injury risk index of seven body parts for trainees the following day. In this embodiment, the historical training load data includes multiple quantitative loads on key high-risk areas over the past seven days, as well as indicators fused with FMS test scores. This achieves a dynamic correlation between training load, functional status, and injury risk, providing data support for daily training adjustments and ultimately reducing the probability of injury to high-risk areas.

[0092] Table 6 Dataset Contents

[0093]

[0094] Specifically, it includes:

[0095] 1. Quantification of training exercise load

[0096] (1) Determine the motion load index

[0097] All training movements are labeled, and the load index (LI) of each movement on 7 body parts is defined, with a value range of [0,5]: 0: no load, 1-2: low load, 3-4: medium load, 5: high load.

[0098] The load index is determined by professional coaches in conjunction with biomechanical analysis of movement (refer to the section on movement load grading standards in "Functional Training: Movement Exercises and Program Design to Improve Athletic Performance"). Some examples are shown in Table 7.

[0099] Table 7 Example of Load Index

[0100]

[0101] (2) Determine the total load of a single action

[0102] For a single movement, the total load on key movement risk areas within the group is calculated based on training parameters. :

[0103]

[0104] in: This represents the load index on the target area caused by the action. For the number of groups, For each set of numbers, Duration of a single action (in seconds). The exercise level represents the difficulty; the higher the level, the greater the difficulty. The value is 1.0-1.5, which can be adjusted according to weight / speed. You can refer to the intensity rating in the "NSCA-CPT National Strength and Conditioning Association Personal Trainer Certification Guide" and have it determined by an expert.

[0105] (3) Determine the total daily load

[0106] The total daily load on a certain body part is the sum of the total load on the key movement risk areas from all movements on that day. :

[0107]

[0108] in, The number of training exercises for the day. For the first The total load on a certain part of the body by a single action.

[0109] (4) FMS test score

[0110] The FMS test was used to assess and score the quality of human movement patterns through seven movements: squat, hurdle step, straight lunge, shoulder flexibility, active straight leg raise, trunk stability push-up, and rotational stability. Each movement was scored from 1 to 3 points, with a score of 0 for each instance of pain during the movement. The performance of the left and right limbs was compared during the testing and scoring process; significant differences between the left and right sides indicated a defect in the movement pattern. Specifically:

[0111] During the test, observe and compare the performance of the left and right limbs in the following aspects:

[0112] 1) Range of motion: such as the distance the left and right shoulders touch when they extend backward, and the angle at which the left and right legs are raised straight.

[0113] 2) Movement stability: For example, when squatting to the left or right, can the body remain upright without swaying?

[0114] 3) Compensation situation: For example, whether only one knee is buckled inward or only one heel cannot touch the ground.

[0115] The following factors will be considered when assigning scores:

[0116] 1) If the difference between the left and right sides is slight (e.g., amplitude difference < 2.5 cm and no obvious compensation), and one side can reach the 3-point standard, it is judged as 3 points.

[0117] 2) If there is a significant difference between the left and right sides (e.g., amplitude difference ≥ 2.5 cm or one side shows compensation), and the action can be completed on one side, the action is scored as 1 point.

[0118] 3) If there is a significant difference between the left and right sides (e.g., one side can complete the task while the other side cannot), or if pain occurs on one side, the task will be scored as 0 points.

[0119] Calculate using the above method:

[0120] 1) Individual scores for each of the 7 actions;

[0121] 2) Total Score , For the first i The score for each action.

[0122] 2. Constructing model input features

[0123] (1) Constructing historical load characteristics of the past 7 days

[0124] For each group of trainees, the key risk areas of movement were identified, and the following information was obtained:

[0125] Cumulative load : , For the first Total daily workload for each action.

[0126] Average load ;

[0127] Maximum load ;

[0128] Load growth rate ;

[0129] Number of consecutive high-load days In the past 7 days The number of days.

[0130] Historical load and FMS interaction items: This feature reflects the degree of matching between load and functional status;

[0131] Next-day load to historical load ratio: This feature reflects the risk of sudden changes in the load. The predicted total load for the key movement risk areas of the trainees in the next day's training is calculated using the same method as the daily total load.

[0132] Data preprocessing was performed. For missing values, FMS scores were imputed using the most recent test value, and missing training loads were imputed using the mean of the same group for the trainee. For outliers, such as samples with total loads exceeding the 95th percentile within the group, the 95th percentile value was used for truncation. After normalizing all numerical features, a one-model-per-location strategy was adopted to build and train XGBoost prediction models for each of the seven locations, with the cumulative loads mentioned above as the input features. Average load Maximum load Load growth rate Number of consecutive high-load days Historical load and FMS interaction items Next day load to historical load ratio The output feature is the injury risk index of that part of the body the next day.

[0133] In another alternative implementation, during actual verification, it was found that using the XGBoost model for risk index prediction has issues such as a long data collection period, insufficient data volume in the early stages of model training, and difficulty in meeting the accuracy requirements for risk prediction and assessment. Therefore, a method for direct risk assessment in the early stages of model training was also designed. For the key motion risk areas of each group, the motion risk index for the next day is calculated in the following way:

[0134]

[0135] In the formula, This is a risk index for injury to a single body part the following day, with a value range of [0,1]. , , , , All are weighting coefficients, among which the preferred ones are... , , , , ; This is the single-site load tolerance threshold, which is the same as the cumulative load of the 95th percentile of the group. It can be determined through a small-sample pre-survey. For individual FMS scores related to a single body part, expert evaluation can be used to determine the score. For example, the knee corresponds to the scores of hurdles and straight lunges, while the waist corresponds to the scores of squats and trunk stability push-ups. The threshold for the safe load growth rate is 0.1, with the preferred value being 0.1.

[0136] The above supplements address the issues of long data collection cycles and insufficient initial model accuracy during model training. The calculations are intuitive, the implementation cost is extremely low, the interpretability is strong, and targeted interventions are easy to implement. The results show that for a sports group of 500-800 people, using the above method for sports risk assessment in the first four weeks of training, and then using the trained XGBoost prediction model for sports risk assessment after four weeks, and implementing targeted interventions based on the assessment results, resulted in a 45.2% reduction in the average number of injuries in the training group compared to the same period in previous years during the total three-month training period. This indicates that the above sports risk assessment method has a good effect, significantly improves training effectiveness, and avoids unnecessary casualties.

[0137] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A motion risk assessment method based on machine learning, characterized in that, The first dataset was obtained by organizing historical data and sports injury records of physical fitness and physiological indicators. The physical fitness and physiological indicators system includes body morphology indicators, body composition indicators, and athletic performance indicators. The sports injury records include the number of injuries to seven areas: neck, shoulder, elbow, lower back, abdomen, knee, and ankle. The body morphology indicators include: age, height, weight, chest circumference, waist circumference, hip circumference, upper arm circumference, thigh circumference, calf circumference, and body mass index; The body composition indicators include: protein content, inorganic content, water content, muscle mass, and lean body mass. The athletic performance indicators include: (1) Strength category: 1) Maximum strength: Seated chest press, seated leg extension, seated back pull; 2) Swift power: Reverse jump, standing long jump, medicine ball throw; 3) Strength and endurance: Pull-ups, dips, plank; (2) Endurance: 3000-meter run; (3) Speed: 30m sprint; (4) Flexibility: Seated forward bend and back-grabbing test; (5) Sensitivity: Illinois test, 30×2 shuttle run; (6) Balance ability: YBT test; Feature screening was performed to obtain high-correlation indicators for sports injuries in various parts of the body, and independent regression models were established between the high-correlation indicators and the risk of sports injuries in each part of the body. The highly correlated indicators include: 3000-meter run, plank, YBT test results, body mass index, body fat percentage, seated leg extension maximum strength, standing long jump, thigh circumference, waist circumference, back grab test, reverse jump, muscle mass, and Illinois test results. A new group of trainees was tested to obtain systematic data on their physical fitness and physiological indicators. The XGBoost risk regression model was used to analyze the injury risk index of each body part. Based on the injury risk index, the trainees were grouped using the K-means clustering algorithm, and the key sports risk areas within each group were determined according to the average risk index within the group. Specifically, the XGBoost risk regression model was used to establish mapping models between 13 highly correlated indicators and the injury risk of each location. During training, the input features were a 13-dimensional vector composed of the normalized values ​​of the highly correlated indicators for each trainee. For the output parameters, the historical number of injuries was normalized to serve as the injury risk index. Used for model training. For the first Injury risk index for each sample This represents the maximum number of injuries or illnesses to that area. The minimum number of injuries to this part of the body is determined by the mean absolute error, root mean square error, and coefficient of determination as model evaluation indicators. The risk levels are defined as: low risk: [0, 0.3), medium risk: [0.3, 0.7), and high risk: [0.7, 1]. After the indicators are measured for new trainees, the trained XGBoost risk regression model is used to predict the injury risk index at each part of their body. Subsequently, K-means clustering algorithm was used to group trainees with similar risk patterns into one group, and the problem of multiple grouping tendency was solved by prioritizing the highest risk body parts. The similarity of risk patterns means that the distribution trend of injury risk index of trainees across 7 body parts is consistent. The multiple grouping tendency means that the distance difference to multiple cluster centers is less than 0.1 in the clustering results. In this case, the body parts with the highest risk coefficient are matched with the key sports risk body parts of the corresponding cluster, and finally a unique group is determined. The key sports risk body parts are the body parts with an average risk index ≥ 0.65 in each group. For groups without an average risk index ≥ 0.65, the body part with the highest average risk index in that group is taken as the key sports risk body part. During the training process, the XGBoost prediction model is constructed. By integrating historical training load data, functional movement screening scores and the training plan content of the next day, the injury risk index of key sports risk areas of each group of trainees is evaluated. This includes: (1) quantification of training load, (2) model input features, (3) model training, and (4) sports risk assessment. The training load quantization includes: (1) Determine the exercise load index: Label all training exercises and define the load index of each exercise on 7 body parts, with a value range of [0,5]. (2) Determine the total load of a single movement: For a single movement, calculate the total load of key movement risk areas within the group based on training parameters. : in: This represents the load index on the target area caused by the action. For the number of groups, For each set of numbers, Duration of a single action, in seconds. This represents the action order, with a value between 1.0 and 1.

5. (3) Determine the total daily load: The total daily load for a certain part is the sum of the total load of all movements on the key movement risk parts on that day. : in, The number of training exercises for the day. For the first The total load on a certain part of the body by a single action.

2. The motion risk assessment method based on machine learning according to claim 1, characterized in that, Historical training load data includes the cumulative load, average load, maximum load, load growth rate, and number of consecutive high load days over the past seven days.

3. The motion risk assessment method based on machine learning according to claim 2, characterized in that, The FMS test includes squats, hurdles, straight lunges, shoulder flexibility, active straight leg raises, trunk stability push-ups, and rotational stability.

4. The motion risk assessment method based on machine learning according to claim 3, characterized in that, After grouping, the training of trainees will be conducted using standard movements from the standard movement library, and training plans will be designed to train the trainees.

5. The motion risk assessment method based on machine learning according to claim 4, characterized in that, All training movements are labeled, and the load index of each movement on 7 body parts is defined.

6. The motion risk assessment method based on machine learning according to claim 5, characterized in that, The input features for the XGBoost prediction model are the key motion risk areas for each group of trainees: Cumulative load : , For the first Total daily workload for each action; Average load ; Maximum load ; Load growth rate ; Number of consecutive high-load days In the past 7 days The number of days; Historical load and FMS interaction items: , , For the first i The score value of each action; Next-day load to historical load ratio: , The predicted total load for the key movement risk areas of the trainees in the next day's training; The output feature is the injury risk index of the key sports risk areas of each group of trainees on the following day.

Images