Exploring the Impact of Physical Activity Metrics on Calorie Consumption: A Machine Learning Approach Combined with SHAP Analysis

Abstract

Rising global obesity necessitates precise, personalized management of energy balance. While machine learning effectively handles complex physiological data, traditional “black-box” models lack the transparency needed to understand how specific behaviors impact metabolic outcomes. This opacity limits trust and the practical utility of these models in guiding actionable health interventions. To address this challenge, this study evaluates the effectiveness of activity features in predicting energy expenditure using four distinct machine learning paradigms: Support Vector Regression (SVR), Random Forest (RF), XGBoost, and Radial Basis Function Neural Network (RBFNN). These models were selected to compare performance across different algorithmic mechanisms (kernel-based, ensemble, and neural networks) on highly correlated datasets. Furthermore, SHAP (SHapley Additive exPlanations) analysis was integrated to visualize feature contributions and resolve the interpretability issue. Comparative analysis revealed that SVR demonstrated the strongest generalization ability, achieving an R² of 0.78 on the test set, whereas XGBoost, despite superior training performance, suffered from overfitting. SHAP analysis identified “Total Distance” and “Total Steps” as the most critical predictive features. Crucially, the results highlighted that “Very Active Minutes” significantly amplifies energy expenditure, whereas light activity and sedentary behavior contribute minimally. This study validates the robustness of SVR for small-scale, physiological datasets and demonstrates that integrating SHAP enhances model transparency8. These findings provide a data-driven theoretical foundation for optimizing exercise plans, suggesting that personalized health interventions should prioritize increasing the intensity and distance of physical activity to maximize health outcomes.

Introduction

The importance of healthy physical activity for individual health has been widely recognized globally, especially in the context of the ongoing rise in global obesity rates. Achieving a scientific balance between energy intake and expenditure has become one of the core issues in the field of public health¹. Studies have shown that increasing weekly physical activity from none to 150 min of moderate-intensity aerobic exercise can reduce the risk of cardiovascular disease-related mortality by 23%, the incidence of cardiovascular disease by 17%, and the incidence of type 2 diabetes by 26%². In this process, calorie expenditure plays a crucial role as an important indicator of human energy metabolism, encompassing various areas such as weight management, exercise performance, and the prevention of chronic diseases^3–5. According to studies addressing the role of caloric restriction in cancer treatment and prevention, it is hypothesized that at least a 20% caloric restriction in overweight or obese individuals contributes to cancer treatment and prevention. This approach helps reduce glucose, C-reactive protein, and insulin levels, all of which are associated with a lower risk of developing cancer⁶. Against the backdrop of rising global obesity rates, achieving a scientific balance between energy intake and expenditure has become one of the core issues in the field of public health.

For fitness enthusiasts and groups requiring special health management, such as individuals with obesity or chronic diseases, accurately predicting calorie expenditure not only aids in developing personalized fitness plans but also plays a critical role in preventing metabolic syndrome and cardiovascular diseases^7,8. However, due to significant differences in physiological characteristics and activity patterns among individuals, traditional estimation methods often fail to meet the demands for accuracy and personalization in practical applications. In conventional calorie expenditure estimation, the Metabolic Equivalent of Task (MET) method is widely used. This method provides a foundation for energy expenditure estimation by employing simple activity classification and average energy consumption values⁹. Nevertheless, the assumption of average values overlooks individual differences, fluctuations in activity intensity, and the impact of complex activity patterns. For the same type of running activity, variations in body weight, heart rate, and muscle efficiency among individuals can result in significant differences in actual energy expenditure^10,11. Consequently, while traditional statistical methods provide a baseline, they lack the granularity required to capture the non-linear relationship between dynamic physical behaviors and energy metabolism.

In recent years, with the widespread adoption of wearable devices, researchers have been able to obtain multidimensional dynamic data, including step count, heart rate, activity duration, and intensity, enabling more precise energy expenditure modeling^12,13. The data recorded by these devices reveal the differences in total energy expenditure between high-intensity activities (e.g., sprinting) and low-intensity activities (e.g., walking)¹⁴. However, integrating multiple activity characteristics to accurately predict calorie consumption remains a significant challenge in current research.

With the advancement of data science, machine learning has gradually become an effective tool for handling high-dimensional nonlinear data, demonstrating significant potential, particularly in the field of health data analysis¹⁵. Studies have shown that tree-based machine learning models, such as XGBoost, Random Forest, and LightGBM, can automatically learn feature interactions in multidimensional data, thereby significantly improving the accuracy of calorie expenditure predictions¹⁶. In addition, these models can adapt to dynamic and diverse activity patterns, overcoming the limitations of traditional methods in nonlinear modeling. However, machine learning models are often considered “black boxes,” with their prediction mechanisms lacking transparency. This makes it challenging for users to understand how the models utilize input features to generate predictions¹⁷. This issue is particularly prominent in the field of health management, as the interpretability of models directly affects the trust of users and professionals in the prediction results. Existing research has largely focused on minimizing prediction error metrics (e.g., RMSE) while neglecting the explanatory mechanisms of the models. This lack of transparency creates a barrier to practical application: users are presented with a calorie number but are not informed which specific behaviors (e.g., increasing step frequency vs. extending distance) drove that result.

Although machine learning techniques excel in improving prediction performance, their “black-box” nature remains a significant obstacle in practical applications. Users not only demand highly accurate predictions but also wish to understand how specific activity features influence calorie expenditure to make informed decisions in daily life. In recent years, the rise of Explainable Artificial Intelligence (XAI) has provided a new pathway to address this challenge¹⁸. Among them, the SHAP (SHapley Additive exPlanations) method, due to its game-theory-based properties, can intuitively quantify the contribution of each input feature to the prediction outcome, thereby providing clear explanations for the outputs of complex machine learning models¹⁹. The potential of SHAP analysis in health management applications has gradually become evident; however, its specific mechanisms in calorie expenditure prediction remain underexplored. This opens up new research directions for balancing high accuracy and interpretability.

To address the aforementioned challenges, this study aims to integrate machine learning techniques with SHAP analysis to explore the complex effects of activity features on calorie expenditure. The specific objectives of the study include: (1) comparing the performance of different machine learning models in calorie expenditure prediction to validate their applicability. (2) utilizing SHAP analysis to uncover the importance of activity features and quantify their specific impacts on prediction outcomes. And (3) providing data-driven scientific evidence for personalized health management and activity planning. This study differentiates itself from existing literature by not only benchmarking prediction accuracy but also systematically evaluating four distinct algorithmic paradigms—Support Vector Regression (SVR), Radial Basis Function Neural Network (RBFNN), Random Forest (RF), and XGBoost. The rationale for selecting these specific models is to assess how different learning mechanisms (kernel-based, neural, and ensemble) handle highly correlated physiological features, ensuring a robust framework for interpretability analysis.This study seeks to combine high-accuracy prediction with interpretability analysis to offer a more comprehensive understanding of the mechanisms underlying activity features.

The remainder of this paper is organized as follows: Section Methods outlines the methodology, detailing the data sources, preprocessing procedures, and the theoretical frameworks of the selected machine learning models and SHAP analysis. Section Results presents the results, covering the correlation analysis, comparative model performance, and visual interpretations of feature contributions. Section Discussion provides a discussion on the biological implications of these findings, addresses the issue of feature multicollinearity, and offers actionable health management advice. Finally, Section Conclusions summarizes the conclusions and limitations of the study.

Methods

Machine learning model

Support vector regression

SVR is an extension of Support Vector Machines (SVM) for regression tasks. Its core principle lies in mapping data into a high-dimensional space using a kernel function and finding a regression hyperplane that minimizes the error²⁰. SVR employs an epsilon-insensitive loss function to define the range of support vectors, calculating loss only for points that exceed the error margin. This enhances the model’s robustness. The objective function is:

1

with the constraints:

2

Where represents the features mapped by the kernel function to a high-dimensional space. is The regularization parameter, which controls the trade-off between model complexity and tolerance for error. is The epsilon-insensitive loss, defining a margin of tolerance within which prediction errors are ignored.

Radial basis function neural network

RBFNN is a feedforward neural network that uses radial basis functions (such as Gaussian functions) as activation functions in the hidden layer to capture nonlinear features²¹. RBFNN consists of an input layer, a hidden layer, and an output layer, where the number of nodes in the hidden layer and the kernel width parameter significantly impact model performance. The key advantage of RBFNN lies in its sensitivity to nonlinear relationships, making it suitable for capturing complex data patterns. However, its performance depends on appropriate parameter selection and sufficient data support. RBFNN is a three-layer neural network where the hidden layer uses radial basis functions as activation functions. The output of the network is:

3

Where represents the number of nodes in the hidden layer and is the radial basis function.

4

Where represents the center of the hidden layer node, is the kernel width parameter, is the weight from the hidden layer to the output layer and is the bias term.

Random forest

RF is an ensemble learning model that constructs multiple decision trees and averages their predictions to reduce the risk of overfitting from a single tree²². RF is an ensemble method based on decision trees, and its predictive output is the average of all individual decision trees:

5

Where is the total number of decision trees in the forest; is the prediction from the t-th decision tree.

eXtreme gradient boosting

XGBoost is an efficient gradient boosting framework that incrementally improves the performance of weak learners (decision trees) through a weighted optimization strategy²³. It supports various regularization methods (L1 and L2) to control model complexity, prevent overfitting, and effectively handle complex feature interactions. XGBoost is a gradient boosting algorithm designed to minimize an objective function, which is defined as:

6

Where is the loss function measuring the difference between the predicted and actual values. is the regularization term to control model complexity.

7

Where T is the number of leaves in the tree and is the weight of the leaf.

Evaluation indicators

This study comprehensively evaluates the predictive performance of machine learning models by utilizing a combination of R², MAE (Mean Absolute Error), MBE (Mean Bias Error), and RMSE (Root Mean Square Error).

R² reflects the model’s ability to explain the variance in the target variable, with values ranging from [0,1]. A value closer to 1 indicates stronger explanatory power of the model for the target variable. Negative values may suggest that the model’s predictive performance is even worse than a simple mean prediction²⁴.

MAE is used to measure the average absolute difference between predicted and actual values, serving as an intuitive indicator of error magnitude²⁵. A smaller MAE indicates that the model’s predictions are closer to the true values.

MBE is used to measure the systematic bias in the model’s predictions, reflecting the average deviation and direction of the predicted values relative to the actual values. The closer the absolute value of MBE is to zero, the smaller the bias between the model’s predictions and the actual values²⁶.

RMSE is an important metric for measuring the overall level of prediction error, reflecting the standard deviation of the differences between the predicted values and the actual values²⁷.

The formulas for these metrics are given below.

8

9

10

11

Where represents the actual value of the sample, represents the predicted value of the sample and represents the mean of the actual values.

SHAP analysis

Previous studies have typically focused on improving the predictive accuracy of models and comparing the performance of different models, while paying less attention to the interpretability of model outputs. However, in recent years, the demand for interpretable machine learning techniques has been increasing in the fields of sports science and health management. Such techniques not only enhance the transparency of model predictions but also provide deeper insights into how feature variables influence the outcomes. Against this backdrop, SHAP has emerged as a widely used method for interpretive analysis, garnering significant attention.

SHAP is a model interpretability method based on cooperative game theory, used to quantify the contribution of each feature to the model’s predictions. Its core lies in calculating the SHAP value for each feature, representing the marginal contribution of that feature to the model’s output²⁸. This method ensures a fair allocation of prediction values by considering all possible combinations of features. The specific formula is:

12

Where represents the SHAP value of feature , is the set of all features, is a subset of features excluding , and is the model output using only the subset . SHAP values satisfy the additive property, meaning that the model prediction can be expressed as the sum of the baseline prediction and the SHAP values of all features, that is,

13

This study employs SHAP analysis to interpret a predictive model for energy expenditure in physical activities, using both global and local analyses to uncover the contributions of key features to energy expenditure predictions. To comprehensively evaluate model performance and select the most suitable model for SHAP analysis, four machine learning methods (SVR, RBFNN, RF, and XGBoost) were compared in predicting calorie consumption. Based on the optimal model, SHAP analysis was conducted to quantify the contributions of activity-related features to energy expenditure predictions and to reveal the roles of these features in activities of varying intensity.

Data sources

The data used in this study originates from the FitBit Fitness Tracker Data dataset, hosted on Kaggle and provided by Mobius. The data was collected through Amazon Mechanical Turk surveys, covering the period from March 12, 2016, to May 12, 2016. A total of 30 FitBit users voluntarily contributed their personal tracking data, which includes daily activity metrics, calorie expenditure, and other features. The dataset is stored in a long format, with unique user IDs and timestamps indicating dates and times. This dataset is open-source and explicitly authorized for public domain use without requiring additional permissions²⁹.

Data preprocessing

The raw data used in this study originates from the “daily_activity dataset”, which includes daily activity metrics such as step count, total distance, activity intensity, and calorie expenditure. The specific model construction framework is shown in Fig. 1. Initially, data cleaning and formatting operations were conducted. To improve the efficiency of model prediction, we selected features highly correlated with calorie expenditure from the “daily_activity dataset”. These features include total steps, total distance, very active distance, light active distance, and calories (calorie expenditure, the target variable).

Fig. 1.

Construction principle of an interpretable machine learning framework.

Subsequently, numerical features were standardized to ensure consistent value ranges across all features, providing a compatible input format for models such as SVR and RF. Finally, the data was split into training and testing sets in a 7:3 ratio, which were used for model training and performance evaluation.

Results

Correlation analysis of physical activity features

To assess the issue of multicollinearity among features, we calculated the Pearson correlation coefficients between variables (Fig. 2). A correlation coefficient close to + 1 indicates a significant positive correlation, while a value close to −1 represents a significant negative correlation. Values near 0 suggest weak or no correlation between variables. In the Fig. 2, the color gradient from blue to red represents correlations ranging from low to high. The results indicate strong correlations among variables such as total steps, total distance, and very active distance (with coefficients exceeding 0.8). This high degree of correlation suggests severe multicollinearity, which could negatively impact the accuracy and interpretability of subsequent models. Therefore, special attention is needed when handling these variables during feature selection and modeling.

Fig. 2.

Heatmap of correlations between activities and calorie expenditure.

Although the Pearson correlation coefficient between total steps and total distance is relatively high (> 0.9), these two variables reflect distinct aspects of users’ physical activity characteristics and hold unique explanatory significance. Total steps primarily focuses on the frequency of movements, while total distance provides a comprehensive reflection of the spatial extent and intensity of the activity. Both variables may uniquely contribute to the model’s outcomes in practical application scenarios. For example, total distance offers a more comprehensive estimate of energy expenditure, whereas total steps directly indicates behavioral patterns such as walking frequency.

Moreover, this study emphasizes a multidimensional analysis of physical activity rather than the independent contribution of a single variable. Therefore, despite the high correlation between these two variables, they are not directly removed from the model but are instead treated as complementary information for modeling purposes.

Descriptive statistical analysis

Figure 3 illustrates the relationship between “Very Active Minutes” and “Moderately Active Minutes” with calorie expenditure. A strong positive correlation is observed between “Very Active Minutes” and calorie expenditure, indicating that an increase in high-intensity activity significantly boosts calorie burn. In contrast, “Moderately Active Minutes” shows a weaker positive correlation with calorie expenditure, though still exhibiting an upward trend. Most participants have active minutes below the median, but a noticeable increase in calorie expenditure is evident with higher active minutes, highlighting the greater impact of high-intensity activity on energy expenditure.

Fig. 3.

Comparison of the impact of “Very Active Time” and “Moderately Active Time” with calorie consumption.

As shown in Fig. 4, “Lightly Active Distance” exhibits a weak positive correlation with calorie expenditure. Calorie burn tends to increase with longer lightly active distances, but the correlation is relatively weak, and the data points are widely dispersed, indicating that light activity has a limited impact on calorie expenditure. In contrast, “Sedentary Active Distance” shows almost no significant correlation with calorie expenditure. Most data points are concentrated in the very low activity distance range, suggesting that sedentary activity contributes minimally to energy expenditure. The red dashed line and blue dashed line mark the medians of the respective variables and calorie expenditure, further highlighting the differing impacts of various activity types on energy consumption.

Fig. 4.

Comparison of the impact of “Lightly Active Distance” and “Sedentary Active Distance” on calorie consumption.

As shown in Fig. 5, “Very Active Distance” demonstrates a significant positive correlation with calorie expenditure. As the activity distance increases, calorie burn rises notably, especially in the range above the median, where calorie expenditure is more widely distributed. This highlights the substantial contribution of high-intensity activities to energy expenditure. In comparison, “Moderately Active Distance” exhibits a relatively weaker correlation with calorie expenditure. Although a positive trend is present, the data points are more densely concentrated below the median, indicating a relatively limited impact of moderate-intensity activities on calorie burn. The red and blue dashed lines mark the medians of activity distance and calorie expenditure, respectively, further illustrating the differing effects of these two activity types on energy consumption.

Fig. 5.

Comparison of the “Very Active Distance” and “Moderately Active Distance” on calorie consumption.

As shown in Fig. 6, “Total Steps” exhibits a strong positive correlation with calorie expenditure. As the step count increases, calorie burn rises significantly, with a more pronounced upward trend observed in the range above the median. Similarly, “Total Distance” also demonstrates a significant positive relationship with calorie expenditure, indicating that increased movement distance directly impacts energy consumption. In both graphs, the red and blue dashed lines mark the medians for step count, distance, and calorie expenditure. Most data points are concentrated below the median; however, when step count and distance exceed the median, calorie expenditure becomes more widely distributed. This further underscores the critical role of total steps and total distance in energy consumption.

Fig. 6.

Comparison of the impact of total steps and total distance on calorie consumption.

As shown in Fig. 7, “Lightly Active Minutes” exhibits a weak positive correlation with calorie expenditure. Although calorie burn tends to increase with longer light activity durations, the correlation is relatively weak, and the data points are widely dispersed, with most concentrated below the median. This suggests that light activity has a limited impact on calorie expenditure. The relationship between “Sedentary Minutes” and calorie expenditure is relatively complex. While some participants with longer sedentary durations exhibit higher calorie burn, there is no clear linear trend overall. The red and blue dashed lines indicate the medians of activity time and calorie expenditure, respectively, highlighting the differing impacts of sedentary and light activity on energy consumption. Additionally, the results emphasize the potential negative effects of prolonged sedentary time on total energy expenditure.

Fig. 7.

Comparison of the impact of “Lightly Active Minutes” and “Sedentary Minutes” on calorie consumption.

Performance of machine learning approaches

Table 1 summarizes the performance of four machine learning models—SVR, RBFNN, RF, and XGBoost—in predicting calorie expenditure. Model performance was evaluated using RMSE, MAE, MBE, and R² based on 5-fold cross-validation, where the reported results represent the average performance across all folds. This evaluation strategy was adopted to reduce the influence of data partitioning and to provide a more reliable assessment of model generalization ability.

Table 1.

Performance comparison of different models on training and test sets.

Model	Training				Testing
	RMSE	MAE	MBE	R 2	RMSE	MAE	MBE	R 2
SVR	300.79	197.32	−4.09	0.83	329.22	229.58	7.48	0.78
RBFNN	385.68	209.26	0.07	0.83	278.83	243.86	10.16	0.75
RF	277.73	208.437	−2.26	0.84	438.34	336.54	5.51	0.67
XGboost	162.09	98.69	1.07	0.94	458.89	353.2	187.92	0.63

Across the evaluated models, SVR demonstrates comparatively favorable predictive performance on the testing folds. On the training folds, the SVR model achieves an average R² of 0.83, an RMSE of 300.79, and an MAE of 197.32, indicating a good fit between predicted and observed values. Although XGBoost attains the highest training R² (0.94) and the lowest RMSE (162.09), its performance shows a notable decline on the corresponding testing folds, suggesting potential overfitting.

On the testing folds, SVR achieves an average R² of 0.78, reflecting favorable generalization performance relative to the other models. Its RMSE and MAE are 329.22 and 229.58, respectively, both lower than those of most competing approaches. In addition, the MBE of SVR is 7.48, suggesting a modest average bias in the predictions. While RBFNN yields a lower RMSE on the testing folds (278.83), its lower R² value (0.75) and higher MAE (243.86) suggest relatively larger overall prediction errors. RF and XGBoost exhibit weaker testing performance, with average R² values of 0.67 and 0.63, respectively. Notably, XGBoost shows a pronounced discrepancy between training and testing performance across folds, further indicating limited generalization capability.

The results in Table 2 suggest that SVR achieves a favorable balance between training performance and testing generalization in the calorie expenditure prediction task. Compared with more complex models, SVR demonstrates relatively stable predictive behavior across datasets while maintaining moderate error and bias levels.

Table 2.

Test results of different models on training and test sets.

Model	Training	Testing
SVR
RBFNN
RF
XGboost

SHAP explanation

To further investigate the specific impact of activity features on calorie expenditure, this study conducted SHAP analysis based on the SVR model. The Fig. 8 illustrates the importance of each feature in the model’s predictions and their effects on the target variable. Features are ranked in descending order of importance, with Total Distance identified as the most significant feature, indicating that the total distance walked contributes the most to the model’s predictions. This is followed by “Total Steps”, showing that step count is also a key predictor. Other features, such as “Very Active Minutes” and “Fairly Active Minutes”, also have some influence but are less critical than the top two features. In contrast, features like “Sedentary Minutes” show minimal importance, suggesting that sedentary time has limited impact on the target variable. The feature importance highlights the primary basis on which the model makes predictions according to the data patterns, providing interpretability and supporting the research findings.

Fig. 8.

Feature importance analysis based on SHAP values.

The SHAP value distribution for each feature illustrates how the magnitude of the feature values (distinguished by color) affects the direction and extent of the model’s predictions. For example, for the most important feature “Total Distance”, higher feature values (red points) correspond to positive SHAP values, indicating that larger walking distances significantly increase the predicted value. Conversely, lower feature values (blue points) correspond to negative SHAP values, indicating that smaller walking distances reduce the predicted value. Similarly, higher values of “Total Steps” show a similar positive influence on the predictions. In contrast, “Sedentary Minutes” is primarily concentrated in the negative region, suggesting that longer sedentary durations may decrease the predicted value of the target variable. These relationships clarify how changes in specific feature values influence the model’s output, providing intuitive explanations of the model’s decision-making process.

As shown in Figs. 9 and 10, the SHAP values for “Total Steps” decrease as “Total Steps” increases, indicating a non-linear relationship. Within a certain range, lower step counts (smaller Total Steps values) positively contribute to the prediction, while higher step counts (larger Total Steps values) negatively impact the prediction. This non-linearity reveals the model’s sensitivity to “Total Steps”, suggesting that an increase in step count does not always lead to positive outcomes. This may be due to higher step counts reflecting other underlying behavioral patterns or conditions that suppress the prediction.

Fig. 9.

Relationship between total steps and SHAP values.

Fig. 10.

Relationship between total distance and SHAP values.

Additionally, the color dimension represents the distribution of another feature, “Total Distance”, revealing the interaction between “Total Steps” and “Total Distance”. When “Total Distance” is high (red points), the negative impact of “Total Steps” is more pronounced, whereas when “Total Distance” is low (blue points), the influence of “Total Steps” is relatively weaker. This indicates that the contribution of “Total Steps” to the target variable is closely tied to “Total Distance”. Such interactions provide deeper interpretability for the model and suggest that researchers should focus on the combined effects of these two features to optimize the predictive performance in practical applications.

Based on the mean absolute SHAP values, “Total Distance” and “Total Step” are the two features that contribute the most to the model’s predictions (Fig. 11). This indicates that total walking distance and total step count have the most significant impact on predicting the target variable, likely energy expenditure or health indicators. The underlying reason is that total walking distance and step count directly reflect overall physical activity levels, both of which are positively correlated with energy expenditure. Studies have shown that greater walking distances and step counts are typically associated with higher energy consumption, aligning well with their strong contribution to the target variable in the model.

Fig. 11.

Ranking of features by average SHAP values impacting model output.

In addition to “Total Distance” and “Total Steps”, “Very Active Minutes” and “Fairly Active Minutes” also play significant roles in the model’s predictions, representing the duration of high-intensity and moderate-intensity activities, respectively. The relatively high SHAP values for these features indicate that engaging in high-intensity and moderate-intensity activities substantially increases energy expenditure, aligning with fundamental principles of exercise physiology: high-intensity activities demand greater energy supply compared to low-intensity activities.

Meanwhile, “Light Active Distance” and “Sedentary Minutes” contribute less to the model, suggesting that low-intensity activities and sedentary behavior have limited impact on the predictions, though they still provide some reference value. Extended sedentary durations may suppress energy expenditure, while light-intensity activities, though weaker in their immediate effects, could still contribute to long-term health benefits. The core features, “Total Distance” and “Total Steps”, highlight that the key to the model’s predictive power lies in capturing changes in overall activity levels. At the same time, “Very Active Minutes” and “Fairly Active Minutes” emphasize the critical role of high-intensity activities in predicting energy expenditure. In contrast, the weaker influence of low-intensity activities and sedentary behavior further underscores the importance of increasing activity intensity to boost energy consumption.

Discussion

Multicollinearity and feature selection

This study found a high correlation (Pearson correlation coefficient > 0.9) between “Total Steps” and “Total Distance”, indicating substantial information redundancy shared by these two variables in users’ activity data. Such redundancy may lead to multicollinearity issues, which are common in regression modeling. Multicollinearity can result in unstable parameter estimates, thereby negatively impacting the model’s predictive accuracy and interpretability. Addressing this issue is critical to ensure that the model captures distinct contributions from each variable while maintaining robustness in predictions and insights³⁰. However, unlike traditional approaches, this study chose not to remove one of these highly correlated variables but instead retained both “Total Steps” and “Total Distance”. The rationale lies in the complementary roles these two variables play in capturing the characteristics of users’ activity behaviors. “Total Steps” primarily reflects the frequency of movement, serving as a time-accumulated variable that is well-suited for capturing short-term, high-frequency activities. In contrast, “Total Distance” focuses more on representing the intensity and spatial extent of activity, making it a valuable metric for long-duration exercise. By preserving both features, the model benefits from their distinct contributions, ensuring a more comprehensive understanding of user behavior and enhancing the interpretability and robustness of the predictions³¹.

From a theoretical perspective, multicollinearity is typically considered an issue that needs to be addressed. When the correlation coefficient between variables exceeds 0.8, multicollinearity can significantly reduce the stability and interpretability of the model. It is generally recommended to apply dimensionality reduction or variable selection methods to mitigate this problem³². However, this traditional approach may overlook the unique explanatory significance of variables in practical applications. While some variables may exhibit high correlations, retaining them can provide additional information for capturing the multidimensional characteristics of behaviors. In regression modeling, multicollinearity does not impact relative or absolute predictive accuracy, emphasizing the value of preserving such variables for enhanced model performance and interpretability³³. Based on this theory, this study chose to retain both “Total Steps” and “Total Distance”, aiming to fully leverage their complementarity in representing activity frequency and intensity. This approach seeks to better capture the multidimensional characteristics of energy expenditure.

Regarding dimensionality reduction and feature engineering, this study deliberately excluded Principal Component Analysis (PCA) and manual feature derivation (e.g., “Stride Length”) to prioritize interpretability and clinical utility. While PCA effectively resolves multicollinearity, its abstract components (e.g., “PC1”) lack the physical semantics required for actionable health advice, whereas retaining original features allows for direct behavioral guidance³⁴. Furthermore, although constructing a “Stride Length” feature is theoretically sound, the kernel-based SVR model inherently captures non-linear interactions between variables. Manually forcing a composite ratio could obscure the distinct marginal contributions of step frequency versus spatial coverage, while the model’s regularization effectively suppresses variance inflation to maintain predictive robustness despite feature correlations.

It is worth noting that, in addition to “Total Steps” and “Total Distance”, this study also identified a moderate correlation between “Very Active Distance” and “Moderately Active Distance” (Pearson correlation coefficient approximately 0.6–0.7). Such feature correlations are not uncommon in activity behavior data. Given the significant inter-individual variability in physiological responses to exercise, employing a multivariable and personalized intensity prescription approach can more accurately target specific training stimuli³⁵. Exercise features of different intensities may exhibit some degree of statistical overlap, as high-intensity and moderate-intensity activities are often closely linked in terms of time and space. However, this overlap is not always negative, as it can help the model capture user behavior characteristics more comprehensively.

Effect of activity characteristics on calorie consumption

This study conducted a descriptive statistical analysis of the relationship between activity features and calorie expenditure, revealing the varying impacts of different activity types on energy consumption. The results indicate that high-intensity activities, such as “Very Active Minutes” and “Very Active Distance,” have the most significant influence on calorie expenditure, while moderate- and low-intensity activities have relatively weaker effects. Additionally, sedentary behaviors, such as “Sedentary Minutes”, exhibit a more complex relationship with calorie expenditure, generally showing low correlation in most cases. These findings not only provide direct evidence for the importance of high-intensity activities but also highlight the relative roles of low-intensity activities and sedentary behaviors in energy consumption.

The significant promotion of energy expenditure by high-intensity activities aligns with findings from numerous classic studies^36–38. Compared to low-intensity activities, high-intensity activities can substantially increase cardiovascular load and metabolic rate, thereby significantly boosting energy expenditure^39–41. This study further revealed that energy expenditure from high-intensity activities exhibits a nonlinear growth trend once activity levels exceed the median. This phenomenon may be attributed to the increased oxygen demand, glycogen breakdown, and fat oxidation required for high-intensity activities, leading to a nonlinear metabolic response. High-intensity exercise not only enhances energy expenditure but also improves physical fitness, making low-intensity exercise more tolerable, which can contribute to effective weight management³⁷. During high-intensity exercise, the sympathetic nervous system induces rapid catabolism of energy reserves, followed by transcriptional reprogramming to guide anabolic changes for recovery and improved subsequent exercise performance⁴².

In contrast, moderate-intensity activities have a weaker impact on energy expenditure but still show a positive correlation. A study indicates that moderate-intensity exercise results in higher daily physical activity energy expenditure compared to non-exercise days, although it is lower than the energy expended during high-intensity exercise⁴³. However, the sample distribution in this study shows that most participants had relatively short durations of moderate-intensity activity, which may have limited its overall contribution to energy expenditure. Additionally, statistical analysis reveals that the contribution of moderate-intensity activity to calorie expenditure is primarily concentrated around the median of the sample distribution, while its effect is relatively weaker in samples dominated by high-intensity activity. This finding suggests that in exercise interventions targeting high-intensity activity, the marginal benefits of moderate-intensity activity may be less significant than expected. High-intensity interval training has been shown to enhance motor cortex plasticity more effectively than moderate-intensity exercise, with the effects of moderate-intensity exercise being intermediate⁴⁴.

For low-intensity activities and sedentary behavior, this study found their contribution to calorie expenditure to be limited, characterized by weak correlations and concentrated distributions. Low-intensity activities typically do not significantly increase total energy expenditure, while sedentary behavior may have potential adverse effects on energy metabolism and health indicators⁴⁵. However, unlike many studies that directly emphasize the negative effects of sedentary behavior, the results of this study suggest that its impact varies depending on the sample distribution. For instance, in a subset of participants with prolonged sedentary time but equally high durations of high-intensity activity, calorie expenditure did not significantly decline. This indicates that the effects of sedentary behavior may need to be assessed in conjunction with other activity characteristics rather than being viewed in isolation. Nonetheless, existing research has demonstrated that reallocating sedentary time to physical activities can lead to substantial energy expenditure. For individuals who find combining light- and moderate-intensity activities challenging, moderate-intensity physical activity offers an effective alternative⁴⁶.

Additionally, this study used distribution plots to reveal the complex relationships between various activity features and calorie expenditure. For example, the relationship between “Very Active Distance” and energy expenditure showed a strong positive correlation, which became even more pronounced when activity distances exceeded the median. These findings underscore the central role of high-intensity activity in energy metabolism and provide direct evidence to inform the design of exercise interventions⁴⁷. Additionally, the distributions of “Lightly Active Distance” and “Sedentary Active Distance” were more dispersed, further indicating that low-intensity activities and sedentary behaviors have a limited impact on calorie expenditure. However, they may still hold some reference value in the context of long-term health management.

Performance of machine learning models

This study compared the performance of four machine learning models—SVR, RBFNN, RF, and XGBoost—in the task of predicting calorie expenditure. The rationale for selecting these specific models was to evaluate the adaptability of distinct algorithmic paradigms in handling physiological data: SVR represents kernel-based learning suitable for small samples; RBFNN represents neural networks capturing global non-linearity; RF represents bagging ensembles for variance reduction; and XGBoost represents boosting ensembles for bias reduction.

The results indicate that the SVR model demonstrated the best overall performance, particularly showing high generalization ability on the test set (R² = 0.78, RMSE = 329.22). To contextualize this performance, we compared our findings with recent studies in the field. For instance, research evaluating gradient boosting frameworks (including LightGBM and XGBoost) for fitness tracking has consistently reported high predictive precision in data-rich environments⁴⁸. Similarly, investigations utilizing deep learning architectures, such as Long Short-Term Memory (LSTM) networks, have demonstrated strong predictive capabilities when processing temporal accelerometry data⁴⁹. While these complex models achieve superior fits on large-scale or temporal datasets, our SVR model differentiates itself by demonstrating competitive robustness specifically in scenarios with limited sample sizes and highly correlated physiological features, where complex deep learning or boosting models often incur high computational costs or overfitting risks.

The superior performance of SVR can be attributed to its regularization properties and kernel methods, which enhance its robustness in handling highly correlated features. Regularization effectively controls model complexity, preventing overfitting, while kernel methods capture nonlinear relationships between features, enabling more accurate predictions⁵⁰. In this study, highly correlated features (such as “Total Steps” and “Total Distance”) may have interfered with the performance of other models, such as RF and XGBoost. However, the regularization capability of SVR significantly mitigated this issue, allowing it to maintain robust performance.

In contrast, the decline in XGBoost’s performance on the test set may be attributed to its complexity and high sensitivity to feature interactions. In high-dimensional data, where samples are relatively sparse, XGBoost’s accuracy tends to decrease, making it more prone to overfitting⁵¹. Additionally, RF achieved the lowest test set R² value of 0.67, demonstrating the poorest performance among the models. While RF is relatively well-suited for handling high-dimensional data and datasets with missing values, it struggles with high-cardinality categorical variables and lacks strong interpretability, which may have contributed to its suboptimal results in this study⁵². RBFNN achieved a relatively low RMSE on the test set (278.83), but its R² value (0.75) and MAE (243.86) indicate that its overall explanatory power was inferior to that of SVR. This may be due to RBFNN’s limited ability to learn global patterns when the dataset size is small, which could hinder its overall performance⁵³.

This study experimentally validated the advantages of SVR in scenarios involving medium- to small-scale datasets and highly correlated features. SVR achieved a good balance between training accuracy and test generalization ability, providing a robust tool for analyzing exercise data in personalized health management. At the same time, the results also suggest that XGBoost and RF are better suited for scenarios with stronger feature independence or larger datasets, while RBFNN has potential for improvement through optimization of network parameters and node selection.

Regarding robustness, although this study utilized a single dataset, the rigorous 5-fold cross-validation procedure ensures that the SVR model maintains stability across different data partitions. This suggests that the proposed method possesses a degree of internal robustness, making it a viable tool for personalized health management even when large-scale data is unavailable.

Explaining analysis for calorie consumption using SHAP

To further explore the contribution of features to calorie expenditure prediction, this study conducted SHAP analysis based on the SVR model. SHAP is an interpretability method rooted in game theory that quantifies the impact of each feature on the model’s output, thereby revealing the relationships between features and the target variable⁵⁴. The SHAP feature importance ranking indicates that “Total Distance” has a significant advantage in capturing the spatial dimension of physical activity, making the most substantial contribution to calorie expenditure prediction. This suggests that the greater the distance covered, the higher the energy expenditure, highlighting that physical activity increases total energy consumption, particularly under conditions of positive energy balance⁵⁵. At the same time, “Total Steps”, as a direct reflection of movement frequency, is also an important input variable for the model.

Research has shown that step count can serve as an effective indicator of physical activity frequency, particularly when assessing daily activity levels, and it provides strong explanatory power for energy expenditure⁵⁶. While these features are highly correlated, the SHAP analysis implies a non-linear interaction; specifically, the observed distribution in the dependence plots is consistent with a ‘stride efficiency’ mechanism. When high step counts are not accompanied by a proportional increase in distance (indicative of low-intensity movement), the marginal contribution to calorie expenditure tends to plateau. This ‘saturation threshold’ suggests that metabolic cost is maximized only when frequency and spatial coverage are optimized synchronously.

Additionally, SHAP analysis revealed that high-intensity activity features, such as “Very Active Minutes” and “Fairly Active Minutes”, make significant positive contributions to the model’s predictions. Moderate- and high-intensity activity durations have a greater impact on energy metabolism compared to activity distance. This observation has not been effectively validated in existing studies. Research indicates that exercise intensity influences both the magnitude and duration of excess post-exercise oxygen consumption (EPOC), whereas exercise duration only affects the duration of EPOC⁵⁷. Our study found that engaging in moderate- to high-intensity exercise during effective workout periods is more effective in consuming energy. High-intensity exercise increases energy expenditure by reducing movement efficiency, enhancing work efficiency, and promoting muscle mass growth³⁷.

In contrast, low-intensity activity features (e.g., “Lightly Active Minutes” and “Lightly Active Distance”) and sedentary behaviors (e.g., “Sedentary Minutes”) have more dispersed SHAP value distributions and lower contributions. This indicates that these features have limited impact on the model’s output but still hold some reference value in specific contexts. While low-intensity activities and sedentary behaviors make minimal direct contributions to total energy expenditure, they may indirectly influence health outcomes by affecting long-term metabolic rates^58,59. Even if individuals meet current physical activity guidelines, sedentary behavior remains an independent predictor of metabolic risk. The SHAP analysis results of this study support this perspective, while also suggesting that greater emphasis should be placed on incorporating high-intensity activities when designing exercise interventions.

Personalized health recommendations

By translating the interpretability results from the SHAP analysis into practical applications, this study proposes a data-driven framework for personalized health management. Unlike traditional “one-size-fits-all” guidelines, our model enables dynamic exercise prescriptions tailored to individual behavioral patterns, as detailed below:

Optimizing Stride Efficiency and Movement Quality Based on the interaction analysis between “Total Steps” and “Total Distance,” the model identifies a saturation point where increasing step count without a corresponding increase in distance yields diminishing metabolic returns. For users exhibiting high step counts but disproportionately low distance—indicative of short strides or shuffling behavior—the intervention strategy should prioritize movement quality over quantity. Health management platforms should recommend increasing walking speed or stride length rather than simply accumulating more steps. This adjustment maximizes the distance-per-step ratio, thereby elevating the metabolic cost of the activity and helping users overcome the efficiency plateau associated with low-intensity stepping.

Prioritizing High-Intensity Intervals for Time Efficiency For users with limited exercise time, the model underscores the critical importance of intensity, as evidenced by the steep positive contribution of “Very Active Minutes” to energy expenditure. A personalized prescription for this demographic should emphasize increasing the intensity of activity rather than extending duration. For instance, engaging in short bouts of brisk walking or jogging is more metabolically effective than longer periods of slow walking. This aligns with the physiological principle that higher intensity triggers greater excess post-exercise oxygen consumption (EPOC), validating the recommendation to prioritize vigorous intervals to maximize calorie burn within constrained timeframes.

Strategic Interruption of Sedentary Behavior While “Sedentary Minutes” contributes less directly to total energy output compared to active features, its negative impact accelerates beyond specific thresholds, posing long-term metabolic risks. Consequently, health recommendations should focus on the management of sedentary behavior through frequency interruption. Users should be alerted when continuous sedentary time approaches the model’s negative inflection point. The recommended intervention is to incorporate frequent micro-movements, such as standing up or stretching every hour, rather than relying solely on a single workout session to offset the cumulative adverse effects of prolonged sitting.

Limitations

While this study has achieved certain results in predicting calorie expenditure and interpreting feature contributions, several limitations must be acknowledged. First, the relatively small and unevenly distributed sample size—particularly for moderate- and low-intensity activities—along with the exclusion of physiological (e.g., heart rate, age) and environmental factors, partially limits the comprehensiveness of the predictions and their precision in specific scenarios. Second, four representative models (SVR, RF, XGBoost, and RBFNN) were selected to cover mainstream regression mechanisms. While this selection does not exhaust all high-complexity variants (e.g., deep learning architectures), it provides sufficient algorithmic representation given the current data scale. Future research could further explore the potential of complex models based on larger-scale datasets. Finally, constrained by the data format, this study primarily relied on k-fold cross-validation to assess internal validity, without conducting time-series or cross-dataset validation. Future research should introduce independent external datasets to rigorously test the model’s external robustness and generalizability across different populations.

Conclusions

This study, based on the FitBit Fitness Tracker Data dataset, first conducted multicollinearity analysis and descriptive statistics to explore the relationship between activity features and calorie expenditure and to identify the impact of key variables. Subsequently, it compared the performance of different machine learning algorithms, including SVR, RF, and XGBoost, in predicting energy expenditure. Finally, SHAP analysis was applied to the model outputs to investigate feature importance and interactions, enhancing the interpretability of the models. The main conclusions are as follows:

1. Relationship Between Activity Features and Energy Expenditure High-intensity activities (“Very Active Minutes”) and greater activity distances (“Total Distance”) significantly increased energy expenditure, demonstrating a strong positive correlation. This effect was particularly notable in activity ranges above the median, where calorie expenditure increased markedly. “Total Steps” and “Total Distance” emerged as two key variables. Although highly correlated (Pearson correlation coefficient > 0.9), they uniquely reflect movement frequency and spatial extent, respectively, offering distinct explanatory value and complementary information for the model. In contrast, low-intensity activities (“Light Active Distance”, “Light Active Minutes”) had a smaller impact on energy expenditure, while sedentary behavior (“Sedentary Minutes”) showed no significant relationship with calorie expenditure.

2. Performance of Machine Learning Models In the task of predicting energy expenditure, the SVR model demonstrated the best overall performance, achieving an R² of 0.78, RMSE of 329.22, and MAE of 229.58 on the test set, outperforming other models such as RF and XGBoost. Although XGBoost performed exceptionally well on the training set (R² = 0.94, RMSE = 162.09), its performance dropped significantly on the test set (R² = 0.63), indicating overfitting. In contrast, the SVR model achieved a good balance between training accuracy and test generalization ability.

3. SHAP Analysis Results SHAP analysis further revealed the contributions of individual features to the model’s predictions. “Total Distance” and “Total Steps” were the two most influential features, with “Very Active Minutes” (high-intensity activity duration) and “Fairly Active Minutes” (moderate-intensity activity duration) also having significant impacts. SHAP value analysis showed that samples with higher total distances and step counts substantially increased the predicted values, while “Sedentary Minutes” (sedentary time) had minimal or even negative contributions. These findings indicate that moderate- to high-intensity activities and overall activity levels are the primary factors influencing energy expenditure.

This study highlights the importance of increasing moderate- to high-intensity activities, emphasizing that high-intensity activities and greater activity distances should be focal points in health management and exercise interventions. Through the application of machine learning models, particularly the robust performance of the SVR model and the interpretability provided by SHAP analysis, scientific evidence can be offered for personalized health management. The predictive results based on core activity features can assist in designing more effective exercise plans, optimizing energy expenditure, and improving overall health outcomes.

Author contributions

Authors’contributions: S.L. and Y.Z. conceptualized the study and designed the experiments. S.L. conducted the data collection and preprocessing and Y.Z. performed the data analysis and interpretation. S.L. and Y.Z. collaboratively wrote the main manuscript text. All authors reviewed the manuscript.

Data availability

Activity data that support the findings of this study have been deposited in the Kaggle repository under the title FitBit Fitness Tracker Data (https://www.kaggle.com/datasets/arashnic/fitbit).

Code availability

The machine learning framework was developed using MATLAB and Python. The complete source code is available under an open-source license at https://github.com/afkzy/A-Machine-Learning-Approach-Combined-with-SHAP-Analysis.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval and consent to participate

Not applicable. Our research is exempt from formal ethical review in China. According to the Ethical Review Measures for Life Science and Medical Research Involving Human Beings (Article 32), the use of anonymous data does not require review. All participants provided informed consent prior to their inclusion in the study. Participants were fully informed about the purpose, procedures, potential risks, and benefits of the research, and they voluntarily agreed to participate. Confidentiality and anonymity of participants were strictly maintained throughout the study.