Determining the Level of Reaction to a Moving Object in Combat Sports Athletes Using an Composite Index and Machine Learning

1. Introduction

In modern combat sports, the effectiveness of competitive performance is largely determined not only by the athlete’s level of physical, technical-tactical, and functional preparedness, but also by the ability to quickly perceive the changing situation of a bout, anticipate the opponent’s actions, and implement a motor response in a timely manner [1,2].

Combat sports are characterized by high variability of motor actions, limited time for decision-making, and constant changes in distance, direction of attack, tempo, and rhythm of interaction with the opponent [3,4,5]. Under such conditions, sensorimotor and perceptual-cognitive skills become particularly important, including reaction speed, accuracy in selecting the moment of action, anticipatory ability, and stability of the motor response [6,7]. One important component of sensorimotor preparedness in combat sports athletes is the reaction to a moving object (RMO) [8]. In combat sports practice, this ability is manifested when responding to an opponent’s attacking or defensive actions, changes in distance, the initiation of an attack, movement, and the selection of the optimal moment for a counterattack. Therefore, the reaction to a moving object may be considered an informative psychophysiological indicator that reflects the quality of spatiotemporal anticipation and sensorimotor control in athletes. Recent studies confirm that elite combat sports athletes have advantages in perceptual anticipation compared with lower-level athletes [9,10]. They respond more quickly and accurately to an opponent’s attacking actions. In particular, one study showed that elite combat athletes demonstrate a better ability to anticipate an opponent’s actions, make decisions faster, and focus their attention on a smaller number of the most informative visual cues [11]. This indicates that response effectiveness in combat sports is associated not only with the mechanical speed of the motor response, but also with the quality of the prior analysis of the emerging competitive situation.

The specific feature of the reaction to a moving object is that it combines several interrelated components, such as the perception of movement speed and trajectory, estimation of the time required for the object to reach a certain position, anticipation of the response moment, and execution of the motor action [12]. For combat sports, this is of fundamental importance, since successful performance often depends not on the fastest possible response, but on timely reaction. A premature response may lead to an incorrect technical-tactical decision, whereas a delayed reaction reduces the effectiveness of defense or counterattack.

Studies involving combat sports athletes confirm the feasibility of using computerized methods to assess sensorimotor reactions. In particular, one study substantiated a method for assessing the reaction to a moving object in combat sports athletes using computer technologies. The authors emphasize that this form of testing makes it possible to obtain quantitative indicators that characterize athletes’ response patterns under dynamic conditions close to the specific demands of combat sports [8]. The importance of reaction speed and accuracy is also confirmed by studies in various combat sports. Thus, recent studies involving boxers and athletes from Olympic combat sports have shown that visuomotor reaction may vary depending on functional state, chronotype, mental fatigue, and task conditions. This indicates that sensorimotor indicators are sensitive to the athlete’s current state and may be used not only for one-time diagnostics but also for ongoing monitoring of preparedness [13,14]. An athlete’s reaction under competitive conditions has a multicomponent nature and requires comprehensive analysis [15,16]. In combat sports, not only the ability to react quickly is important, but also the ability to identify the most informative cues in the opponent’s movement. Studies of the perception of attacking actions in combat sports demonstrate that athletes must anticipate the target and direction of the opponent’s attack using information about the position of the body, head, limbs, and other biomechanical cues. Under such conditions, reaction effectiveness depends on the integration of visual perception, attention, anticipation, and motor response [17,18].

The traditional presentation of results for the assessment of reaction to a moving object in the form of separate indicators does not always allow for a quick and unambiguous interpretation of the level of its manifestation. Mean reaction time, standard deviation, and the percentages of premature, delayed, and accurate responses reflect different aspects of sensorimotor activity; however, they require generalization for practical use by a coach or researcher. Therefore, the development of an integral Reaction to a Moving Object index (RMO_Index), which combines the key characteristics of the reaction, namely speed and stability, is a promising approach. This approach makes it possible to move from an isolated analysis of individual indicators to a comprehensive assessment of the level of reaction to a moving object.

At the same time, the modern development of digital technologies and machine learning methods creates new opportunities for the automated interpretation of psychophysiological testing results. In contemporary sport, intelligent data analysis methods are considered one of the promising directions of smart sport training, as they enable the processing of large sets of indicators, the identification of hidden patterns, and support for training decision-making [19,20]. The use of ML models in sports monitoring makes it possible not only to process multidimensional data but also to generate classification decisions that can be applied to the assessment of sports performance, functional state, and risks associated with the training process [21]. In addition, recent reviews emphasize that the combination of sensor technologies, digital platforms, and artificial intelligence algorithms creates conditions for more personalized athlete monitoring and rapid interpretation of the obtained data [22]. For training practice, this is of considerable importance, as it enables rapid interpretation of results, determination of the level of manifestation of the studied function, and formation of an individual athlete profile.

Thus, the study of reaction to a moving object in combat sports athletes is relevant from both theoretical and practical perspectives. Its theoretical significance lies in deepening the understanding of the sensorimotor and perceptual-cognitive mechanisms that ensure the effectiveness of technical-tactical actions under the dynamic conditions of a bout. Its practical significance is associated with the possibility of creating an integral assessment of the manifestation of the studied function and the subsequent use of a machine learning model for the automated determination of the level of reaction to a moving object in combat sports athletes.

The aim of the study was to determine the level of reaction to a moving object in combat sports athletes based on an composite index and to develop a machine learning model for the automated classification of testing results.

2. Materials and Methods

Participants. The study involved 312 combat sports athletes (taekwondo, karate, Greco-Roman and freestyle wrestling, and judo) aged 10-25 years and of different sports qualifications (Mean = 17.5, SD = 4.33 years).

Procedure. The indicators of reaction to a moving object in combat sports athletes (n = 312) were determined using the specialized “Reaction RMO Pro” test for mobile devices running iPadOS [8]. An iPad (9th generation) with a 10.2-inch screen was used as the testing device. The test consisted of three stages. The specific feature of the test exercise was the gradual increase in task complexity. In the first stage, the participants had to respond to a slowly moving object; in the second stage, to a fast-moving object; and in the third stage, to a fast-moving object in the presence of interfering visual stimuli. The number of attempts at each stage was 20, with a total of 60 attempts per test. Psychophysiological testing was conducted in the second half of the day. To increase interest and conscious participation, the athletes were informed in advance about the aim of the study, the content of the test tasks, and the practical value of the obtained results for assessing reaction to a moving object. Before the main examination, a practice test was conducted, allowing the participants to become familiar with the task procedure, adapt to the testing conditions, and verify that they understood the instructions correctly. Only those athletes who confirmed that they understood the task procedure and were ready to perform it according to the instructions were admitted to the main testing.

The study was conducted in accordance with the Declaration of Helsinki and approved by the Bioethics Commission of the National University of Ukraine on Physical Education and Sport (Protocol No. 5 dated July 29, 2026). All participants were informed about the aim of the study, the specifics of the testing procedure, the nature of the data collected, and their right to withdraw from participation at any stage without giving reasons. For underage athletes, written informed consent was obtained from parents or legal representatives, who were present during testing. At the time of the study, all participants were medically cleared to perform the test tasks.

Sample of Variables. At each stage of the test exercise, indicators of speed (reaction time, ms) and performance variability (standard deviation, ms) were recorded. The indicators of mean reaction time by stages were as follows: RT1_mean, RT2_mean, and RT3_mean. The indicators of reaction variability by stages were as follows: RT1_SD, RT2_SD, and RT3_SD. The indicators of reaction direction were: Premature_% - premature reactions; Delayed_% - delayed reactions; and Accurate_% - accurate reactions. It should be noted that Premature_%, Delayed_%, and Accurate_% together amounted to 100%. For model development, the variable names were adapted to the CSV format: Premature_pct, Delayed_pct, and Accurate_pct. The integral RMO index was formed based on two main characteristics: RT_avg, the mean reaction time for the entire test, and RT_SD, the standard deviation of reaction time for the entire test, according to the following formula:

In the present study, the term reaction time in the RMO test does not refer to conventional simple reaction time measured from stimulus onset to the initiation of a motor response. Instead, it represents the temporal deviation between the participant’s response and the predicted moment at which the moving object reaches the target position. Negative values indicate premature responses, positive values indicate delayed responses, and values falling within the predefined accuracy interval are classified as accurate responses. For the calculation of the mean reaction-time magnitude, absolute temporal deviations were used; therefore, smaller values indicate more precise temporal matching of the response to the target moment.

RMO_Index = -[z(RT_avg) + z(RT_SD)],

Standardization was performed using the following formulas:

z(RT_avg) = (RT_avg - mean(RT_avg)) / SD(RT_avg);

z(RT_SD) = (RT_SD - mean(RT_SD)) / SD(RT_SD)

where z(RT_avg) is the standardized value of the mean reaction time for the entire test; z(RT_SD) is the standardized value of reaction variability.

In these formulas, the difference between the individual value of the indicator and its mean value in the sample reflects the deviation of a particular athlete’s result from the group level, while division by the standard deviation converts the indicators into a single standardized scale. The integral RMO index reflects the combination of reaction speed and stability. The direction of the index was inverted because lower values of reaction time and its variability characterize a better result. The higher the RMO_Index, the better the result; a lower RMO_Index indicates a slower and less stable reaction. The machine learning model was developed in the Xcode Create ML environment using the Tabular Classification task type. The target variable, RMO_Level_Target, was formed based on the integral RMO index. The levels were determined according to tertiles: High, Medium, and Low. A Boosted Trees algorithm was used as the machine learning method. This algorithm is effective for tabular data and is capable of modeling nonlinear relationships between features. The model parameters (Table 1) were aimed at ensuring a balance between classification accuracy and reducing the risk of overfitting. Limiting the tree depth to 3 levels helped preserve a relatively simple model structure, whereas the use of partial selection of rows and predictors at each iteration increased its generalization ability.

The task of the model was to automatically assign a class according to the existing scaling algorithm. The class distribution in the full sample (n = 312) was balanced: High - n = 104; Medium - n = 104; Low - n = 104. This approach ensured class balance, which is important for training an ML model. Nine predictors were included in the model: mean reaction time (RT1_mean, RT2_mean, RT3_mean); reaction variability (RT1_SD, RT2_SD, RT3_SD); and reaction direction (Premature_pct, Delayed_pct, Accurate_pct). The indicators RT_avg, RT_SD, z_RT_avg, z_RT_SD, and RMO_Index were not included as predictors, as they were used to form the target variable.

Statistical Analysis and Machine Learning Procedure. To evaluate the quality of the classification model, a stratified split of the sample into a training-validation subset and an independent test subset was applied. The training-validation subset included 264 observations, while the independent test subset included 48 observations. Stratification ensured the same ratio of the High, Medium, and Low classes in each part of the sample, which reduced the risk of biased model evaluation associated with class imbalance. In the training-validation subset, each class was represented by 88 observations, whereas in the independent test subset, each class was represented by 16 observations. The quality of the classification model was assessed using accuracy, precision, recall, and F1-score. Accuracy was used for the overall assessment of the proportion of correctly classified results. Precision characterized the proportion of correct predictions among all cases assigned by the model to a particular class. Recall reflected the model’s ability to identify all observations that actually belonged to the corresponding class. F1-score was used as a generalized metric combining precision and recall and is informative for comparing classification quality across individual classes. Additional statistical calculations, such as descriptive statistics and determination of the composite index of reaction to a moving object, were performed in RStudio (version 2026.01.0+392). Descriptive statistics are presented as mean (Mean), standard deviation (SD), first and third quartiles (Q1 and Q3), median (Median), and minimum and maximum values (Min and Max).

The 264 observations constituted a single training-validation dataset used for model development in Create ML. The validation reported by Create ML was performed internally within this same training-validation dataset and therefore does not represent an additional independent sample of 264 participants. The independent test subset (n = 48) was kept separate from model development and was used only for the final evaluation of classification performance.

3. Results

To characterize the study sample and ensure the reproducibility of the obtained results, descriptive statistics of the study participants (n = 312) were calculated and are presented in Table 2.

The analysis of descriptive statistics for the results of the reaction to a moving object test provided a general characterization of the study sample. Across the sample, premature responses accounted for 48.18 ± 10.18% of the trials, delayed responses for 44.66 ± 10.45%, and accurate responses for 7.16 ± 4.40%. Thus, premature and delayed responses were substantially more frequent than accurate responses.

For a generalized assessment of the testing results, an integral Reaction to a Moving Object index (RMO_Index) was developed. Its calculation was based on two main indicators characterizing the quality of performance in the RMO test: the mean reaction time for the entire test (RT_avg) and the standard deviation of reaction time (RT_SD). The RT_avg indicator reflected the athlete’s reaction speed, whereas RT_SD characterized the stability of task performance. Lower values of both indicators indicated a better result, as they corresponded to a faster and less variable reaction to a moving object. Before being combined into a single index, the RT_avg and RT_SD indicators were standardized using z-transformation. This made it possible to convert them into a common dimensionless scale and ensure their correct combination in one integral indicator. Based on the RMO_Index values, three levels of reaction to a moving object were formed: High, Medium, and Low. To ensure class balance, tertile-based division was applied. As a result, the target variable RMO_Level_Target was formed, in which each class was represented by an equal number of observations (Table 3).

After the formation of the integral RMO_Index and the target variable RMO_Level_Target, a machine learning model was developed for the automated classification of the level of reaction to a moving object. The model assigned an athlete’s result to one of three levels: High, Medium, or Low. To build the model, indicators characterizing RMO test performance at individual stages, as well as percentage characteristics of reaction direction, were used. The predictor set included nine variables: RT1_mean, RT1_SD, RT2_mean, RT2_SD, RT3_mean, RT3_SD, Premature_pct, Delayed_pct, and Accurate_pct. These indicators reflected the speed and variability of reaction at each test stage, as well as the ratio of premature, delayed, and accurate reactions. The Boosted Tree Classifier algorithm was used to train the model in the Create ML environment in Xcode. This algorithm was selected due to its ability to work with tabular data, identify nonlinear relationships between predictors, and generate classification decisions based on a set of features. To evaluate classification performance, accuracy, precision, recall, and F1-score were used. Accuracy reflected the overall proportion of correctly classified results. Within the training-validation dataset (n = 264), Create ML reported 100% classification accuracy during model development. When evaluated on the independent test sample (n = 48), which was not used during model development, the model achieved an overall accuracy of 94%.

Precision characterized the proportion of correct predictions among all cases that the model assigned to a particular class. Recall showed the model’s ability to identify all observations that actually belonged to the corresponding class. F1-score was used as a generalized indicator that combines precision and recall and allows the quality of classification to be assessed separately for each RMO level (Table 4).

The classification performance of the model by RMO levels indicates that the model identified the High class most accurately. The Medium class was the most difficult to classify, as it lies between the extreme levels. The trained model was exported in Core ML format, which enabled its subsequent integration into the “Reaction RMO Pro” mobile application on iPadOS (Figure 1).

4. Discussion

The results of the present study confirm the feasibility of a comprehensive approach to assessing reaction to a moving object in combat sports athletes. The conditions of competitive activity in combat sports require athletes not only to produce a fast motor response but also to accurately anticipate the moment of interaction with a moving object, maintain reaction stability, and control the timeliness of action execution. This is consistent with data from systematic reviews, which emphasize that combat sports athletes must quickly identify relevant visual cues, anticipate the opponent’s actions, and make decisions under time constraints [11,23]. Therefore, assessing the RMO function only by mean reaction time is insufficient, as it does not fully reflect individual characteristics of sensorimotor control.

In the present study, reaction to a moving object was considered an integral characteristic combining reaction speed and stability. For this reason, the RMO_Index was formed to summarize the results and included standardized values of mean reaction time and reaction variability. This approach made it possible to move from the analysis of separate indicators to a comprehensive assessment of the RMO function level and to form balanced High, Medium, and Low classes for subsequent machine learning. The feasibility of using computerized testing of reaction to a moving object in combat sports athletes is supported by previous studies, which have shown that the RMO method allows quantitative assessment of response characteristics under conditions of gradually increasing task complexity [8]. Another important finding is that most athletes demonstrated a predominance of premature or delayed reactions, whereas the proportion of accurate reactions was relatively small. This may indicate that, under RMO test conditions, not only the absolute accuracy of the response is important, but also the individual response strategy. The predominance of premature reactions may reflect a tendency toward anticipatory responding, whereas the predominance of delayed reactions may be associated with a more cautious or delayed response to a moving stimulus.

The obtained results provided a methodological basis for developing a machine learning model that performed automated classification of the level of reaction to a moving object. The high accuracy of the model on the independent test sample indicates that the indicators of individual stages of the RMO test and the characteristics of reaction direction contain sufficient information to distinguish between the High, Medium, and Low levels. The use of machine learning in this context corresponds to current trends in sports analytics, where AI/ML approaches are applied to analyze multidimensional data, support the monitoring of athletes’ preparedness, and assist coaching decision-making. However, researchers emphasize the need for model interpretability, external validation, and cautious use of automated conclusions [24,25].

One of the key results of the study is the development of the integral RMO_Index, which combines two basic components of reaction quality: speed and stability. This approach is methodologically appropriate because mean reaction time alone does not allow a full characterization of an athlete’s level of sensorimotor control. With the same mean reaction time, athletes may differ substantially in response variability, which is important for combat sports, where the effectiveness of an action depends not only on speed but also on the repeatability and controllability of the response. This is consistent with current views on perceptual-cognitive activity in combat sports, according to which an athlete’s success is determined not only by reaction speed but also by the ability to effectively use visual information, anticipate the opponent’s actions, and implement a motor response in a timely manner [11,23,26]. The results of the study also indicate the importance of reaction direction in the RMO test. The predominance of premature or delayed reactions reflects an athlete’s individual response style to a moving object. In combat sports practice, a premature reaction may be associated with a tendency toward anticipatory responding or risky decision-making, whereas a delayed reaction may indicate a more cautious strategy, a delay in motor response, or insufficient accuracy in movement prediction. This interpretation is consistent with evidence that elite combat sports athletes make better use of advance information about an opponent’s actions and demonstrate advantages in perceptual anticipation compared with lower-level athletes [10,11]. At the same time, the scientific literature emphasizes that isolated reaction time is not always a sufficient predictor of success in combat sports; therefore, it should be considered together with other perceptual-cognitive indicators [6,27].

The division of athletes into High, Medium, and Low levels based on the tertile distribution of the RMO_Index made it possible to move from a purely quantitative assessment to a classification that is more convenient for practical use. This approach is appropriate in sports monitoring, since it is easier for a coach or specialist to interpret the result as a level of functional manifestation than to analyze separate numerical indicators. At the same time, it is important to emphasize that these levels are relative and were formed based on a specific reference sample of combat sports athletes. The Medium level was the most difficult to classify, which is methodologically expected. This class is located between the extreme High and Low levels; therefore, its boundaries are less distinct. Medium-level athletes may have mixed profiles, that is, a sufficiently fast but unstable reaction or, conversely, a stable but slower response. For this reason, distinguishing the Medium level is more difficult than identifying the High and Low levels. A similar problem is typical of classification tasks in sports analytics, where intermediate classes often have less pronounced features than extreme categories, and model quality largely depends on the size, balance, and representativeness of the data [28].

Previous studies on the computerized assessment of reaction to a moving object in combat sports athletes also emphasize the appropriateness of using quantitative indicators to characterize psychophysiological functions under conditions of gradually increasing task complexity [8]. The high accuracy of the classification model on the independent test sample indicates that stage-specific indicators of the RMO test and characteristics of reaction direction contain sufficient information for the automated determination of the level of reaction to a moving object. It is particularly important that RT_avg, RT_SD, and RMO_Index were not included in the predictor set, that is, the indicators that directly formed the target variable. This reduces the risk of target information leakage and increases the methodological correctness of model development. The use of machine learning for the analysis of sports data corresponds to current trends, as ML approaches make it possible to process multidimensional indicators, identify complex relationships between variables, and support the monitoring of athletes’ preparedness [28,29].

Recent studies indicate that machine learning methods are increasingly being used specifically in combat sports to analyze technical-tactical actions, predict performance, assess athletes’ functional state, and support the training process. Studies devoted to sports analytics emphasize that ML algorithms make it possible to identify hidden relationships between athletes’ psychophysiological and technical characteristics that are difficult to detect using traditional statistical methods [19,30]. In particular, recent studies use artificial intelligence for the automated analysis of athletes’ motor actions, prediction of the effectiveness of technical-tactical actions, assessment of the risk of overfatigue, and individualization of training loads [21,31]. In this context, the proposed classification model for the level of reaction to a moving object expands the possibilities for the practical application of ML in combat sports, as it enables the automated assessment of one of the key psychophysiological functions directly related to decision-making speed, anticipation, and sensorimotor control in athletes. It is also important that the use of an on-device Core ML model enables the mobile application to operate autonomously without the need to transmit data to external servers, which increases the efficiency of sports monitoring and simplifies the integration of digital technologies into everyday training practice.

The practical significance of the obtained results lies in the possibility of using the integral RMO_Index and the machine learning model for the rapid assessment of reaction to a moving object in combat sports athletes. The proposed approach makes it possible not only to determine separate indicators of reaction speed and variability, but also to form a generalized level of manifestation of the RMO function. This may be useful for monitoring psychophysiological functions, identifying individual response characteristics, and controlling changes during the training process. The integration of the model into a mobile application creates conditions for the automated interpretation of testing results immediately after its completion. This approach may simplify the use of the RMO test in coaching practice, since the result is presented not only as numerical indicators but also as a classification level - High, Medium, or Low - with corresponding probabilities of belonging to each class. At the same time, recent reviews on the use of artificial intelligence in sports science emphasize that such models require transparent description, testing on independent samples, and cautious interpretation of results in the practice of the training process [32,33,34]. The obtained results also confirm the promise of using the concept of smart monitoring in combat sports. Unlike the traditional approach, in which the results of psychophysiological testing are analyzed mainly after the examination has been completed, the integration of an ML model into a mobile application allows rapid data interpretation directly during the training process. This creates the possibility of quickly detecting changes in the athlete’s sensorimotor state, monitoring the individual dynamics of indicators, and personalizing training effects. Recent studies on artificial intelligence in sport emphasize that the combination of mobile platforms, wearable technologies, and ML algorithms forms the basis for the development of next-generation adaptive sports monitoring systems [20,35]. In the context of combat sports, this is especially important, since an athlete’s psychophysiological state may change rapidly under the influence of fatigue, emotional tension, training load intensity, and the specific demands of competitive activity.

At the same time, the study has certain limitations. First, the formed High, Medium, and Low levels are relative, as they are based on the tertile distribution of the integral RMO_Index in a specific sample of 312 combat sports athletes. Therefore, the obtained index thresholds should be considered as reference values for this sample rather than universal norms for all athletes. Second, the model requires further external validation on independent samples of athletes of different ages, sexes, sports qualifications, and specializations in various combat sports. Third, this study did not analyze the relationship between RMO levels and indicators of competitive performance, which limits the possibility of directly interpreting the obtained levels as predictors of sports performance. The need for external validation and data standardization is one of the key methodological issues in contemporary AI/ML research in sport [28,32].

The prospects for further research include expanding the sample, testing the stability of the integral RMO_Index in different groups of athletes, and conducting external validation of the ML model. It would also be appropriate to compare different machine learning algorithms for classifying the level of reaction to a moving object, including Random Forest, Support Vector Machine, Neural Network, and other approaches to tabular classification. A separate promising direction is the study of the relationship between the integral RMO_Index, the High, Medium, and Low levels, and indicators of competitive performance in athletes from different combat sports. This will make it possible to clarify the practical significance of the RMO function for bout effectiveness and to substantiate the use of this indicator in the system of comprehensive monitoring of athletes’ preparedness.

5. Conclusions

The results of the study confirmed the feasibility of a comprehensive assessment of reaction to a moving object in combat sports athletes, taking into account not only mean reaction time but also reaction variability and response direction. This approach makes it possible to more fully characterize the individual features of athletes’ sensorimotor control. For a generalized assessment of the RMO test results, an integral RMO_Index was formed, combining the standardized values of mean reaction time and the standard deviation of reaction time. Higher index values corresponded to a faster and more stable reaction to a moving object. Based on the tertile distribution of the RMO_Index, three levels of reaction to a moving object were formed: High, Medium, and Low. This made it possible to construct a balanced target variable, RMO_Level_Target, in which each class was represented by an equal number of observations. The developed machine learning model based on the Boosted Tree Classifier algorithm demonstrated high classification quality for the RMO level. On the independent test sample, the model achieved an accuracy of 94%, indicating that stage-specific indicators of the RMO test and characteristics of reaction direction are sufficiently informative for the automated determination of the level of reaction to a moving object. The best classification results were obtained for the High and Low levels, whereas the Medium level was more difficult to distinguish, which is methodologically expected due to its intermediate position between the High and Low levels. The proposed approach can be used in sports monitoring practice for the rapid assessment of the RMO function in combat sports athletes. The integration of the ML model into a mobile application based on Core ML creates the possibility of automated on-device classification of testing results without transmitting data to external servers.

Further research should focus on the external validation of the model in independent samples of athletes of different ages, qualification levels, and combat sport disciplines, as well as on studying the relationship between the RMO_Index and the High, Medium, Low levels with indicators of competitive performance.