Sequential Deep Learning Model for Obesity Prediction Based on Physical Fitness Factors: An Analysis of Data from the 2010–2023 Korean National Physical Fitness Data 

Sequential Deep Learning Model for Obesity Prediction Based on Physical Fitness Factors: An Analysis of Data from the 2010–2023 Korean National Physical Fitness Data | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Sequential Deep Learning Model for Obesity Prediction Based on Physical Fitness Factors: An Analysis of Data from the 2010–2023 Korean National Physical Fitness Data Jun-Hyun Bae, Yunho Sung, Xinxing Li, Wook Song This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4782187/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Obesity, a "global syndemic," increases the risk of noncommunicable diseases; therefore, the prediction and management of obesity is crucial. Regular physical activity and cardiorespiratory fitness are inversely correlated with obesity, highlighting the need for effective models for predicting obesity. Aim This study aimed to predict obesity using physical fitness factors, including those related to cardiorespiratory fitness, determined via deep neural network analysis of data obtained from the 2010–2023 Korean National Physical Fitness Award. Methods A deep learning approach was implemented to analyze the data obtained from 108,304 participants, and variables such as exercise-induced oxygen consumption during a 20-m shuttle run test (20-m VO 2 max), gender, and relative grip strength were considered. Stratified K-fold cross-validation, SHapley Additive exPlanations (SHAP), and local interpretable model-agnostic description analyses were employed to evaluate the model performance and feature importance. Results The neural network yielded a high accuracy score (0.87–0.88), with Fold 4 providing the optimized model for obesity classifications. Features such as 20-m VO 2 max, gender, and relative grip strength significantly influenced the obesity predictions, and low 20-m VO 2 max levels were key predictors of obesity. Discussion This study confirmed the efficacy of the proposed deep neural network in predicting obesity based on physical fitness factors and clarified the significant predictors of obesity. Conclusion The results of this study may potentially be used for devising personalized obesity-management strategies that emphasize the importance of cardiorespiratory fitness. Health sciences/Health care/Weight management Health sciences/Risk factors Obesity Physical Fitness Hyperparameter Deep Learning 20-m Shuttle Run Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction The prevalence of obesity is increasing rapidly worldwide. According to the recent epidemic-related definition of obesity, this disorder has been labeled as a "global syndemic." Moreover, obesity is one of the most significant risk factors related to physical and psychological health problems [1] and increases the risks of noncommunicable diseases, cardiovascular diseases, cancer, chronic respiratory diseases, and diabetes [2–5]. Regular physical activity and improving physical fitness reduce the risk of obesity [6]. In particular, cardiorespiratory fitness is negatively correlated with obesity [7–9], and increased exercise-induced oxygen consumption (VO 2 max) improves the physical fitness of obese individuals [10, 11]. Improving physical fitness, especially cardiorespiratory fitness (CRF), is a more effective strategy for treating obesity than focusing only on exercise-induced weight loss [12]. Moreover, the results of a 20-m shuttle run test can be used to determine CRF [13]; this test can be considered a relevant field test for assessing the CRF of obese adolescents [14] and youth [15]. Moreover, a systematic review and meta-analysis revealed that the results of a 20-m shuttle run test were negatively correlated with body mass index (BMI), body fat percentage, and waist circumference [16]. A previous study focused on youths and adolescents also revealed that this test was a major indicator of CRF. Another study indicated that the results of a 20-m shuttle run test were significant indicators of VO 2 max, and the results exhibited acceptable accuracy when a large sample of adults was tested [17]. Recently, artificial intelligence-based methods involving machine and deep learning (neural networks) have been developed for predicting obesity [18, 19]. A previous study showed that support vector machine, simple K-means, and decision trees are suitable algorithms for detecting obesity levels using big data sources. A neural network model constructed using data on physical fitness factors showed that an increased risk of obesity was associated with low physical fitness levels [20]. Many studies have considered national health and nutrition examination surveys that collected physical activity and basic clinical information via machine learning methods. Most of the results revealed that doing a physical activity without being physically fit is a key factor in determining obesity levels [21, 22]. The results obtained from a model for predicting physical exercise adherence using deep neural networks indicated that the analysis was a good indicator of physical exercise adherence without the physical fitness factor [23]. However, many studies have evaluated the performances (e.g., accuracy, precision, recall, and F1 scores) of such models without conducting feature importance analyses. As obesity can be predicted using physical fitness factors, these factors should be analyzed and ranked. This study used a deep neural network and data obtained from the 2010–2023 Korean National Physical Fitness Award (NFA) to predict obesity based on physical fitness factors. Additionally, this study hypothesized that the result of a 20-m shuttle run test is a physical factor that can be used to predict obesity. 2. Materials and Methods 2.1. Dataset and Measurements All participants voluntarily participated in the Korean NFA project at 75 locations nationwide [24]. In this study, various physical fitness factors were investigated for predicting obesity by incorporating variables such as age (years), diastolic blood pressure (DBP, mmHg), systolic blood pressure (SBP, mmHg), relative grip strength (%), sit and reach distance (cm), sit-up repetitions (count), VO 2 max during the 20-m shuttle run (ml/kg/min), the time required for a 10-m round trip (s), height achieved when jumping in place (cm), and gender (male = 1, female = 2) [25]. Additionally, muscular strength was measured using a handheld dynamometer (Takei, Niigata, Japan), with the participants gripping the handle and pulling with maximum force for 5 s at a 15° angle. The largest of the measurements obtained from both hands was recorded. Relative handgrip strength was calculated by dividing the absolute handgrip strength (kg) by body weight (kg) and multiplying by 100. CRF was assessed using the results of the 20-m shuttle run test in which participants ran back and forth on a 20 m course at increasing speeds, as signaled by beeps. The maximum number of runs was recorded for each participant, and the data obtained from the participants who failed to reach the line before the warning beep were excluded. VO 2 max was calculated using the following estimation formulae based on the number of completions of the 20-m shuttle run test: VO 2 max of a male = 48.550 − 0.080 \(\:\times\:\) (age in years) − 0.021 \(\:\times\:\) (height in cm) − 0.146 \(\:\times\:\) (weight in kg) + 0.234 \(\:\times\:\) (laps), and VO 2 max of a female = 34.305 − 0.039 \(\:\times\:\) (age in years) + 0.018 \(\:\times\:\) (height in cm) − 0.161 \(\:\times\:\) (weight in kg) + 0.305 \(\:\times\:\) (laps) [26, 27]. All measurements on Korean subjects aged between 19–64 years were collected by trained physical fitness instructors [28]. 2.2. Data Collection Formal consent was not required for this study. The Research Ethics Committee of Hyupsung University approved the dataset (IRB no. 7002320-202303-HR-001), and all methods were implemented in accordance with the relevant guidelines. The data comprised exercise and health-related measurements on Koreans aged 19–64 years collected from 19 national fitness centers. The original Korean NFA data were collected between Jan 2010–Mar 2023 (n = 1,810,208), and the first stage excluded data from persons aged \(\:\ge\:\) 65 years (n = 948,009). In the second stage, data with > 20% missing values collected from individuals were excluded (n = 246,842), along with values > Q3 + 1.5 \(\:\times\:\) IQR or < Q1–1.5 \(\:\times\:\) IQR (Q: quartile, IQR: interquartile range. n = 374,943). The final sample size comprised 204,334 participants (Fig. 1 ). The data analysis environment was implemented on Windows 11 (x64 version), with a 13th Gen Intel(R) Core (TM) i9-13900HX processor (2.20 GHz), 32 GB RAM, NVIDIA GeForce RTX 4060 graphics processing unit, and Python (version 3.11.7) with TensorFlow (version 2.15.0). 2.3. Data Independent and Dependent Variables This study used independent variables (features) including age, DBP, SBP, relative grip strength, sit and reach distance, sit-up repetitions (count), VO 2 max during the 20-m shuttle run, time required for a 10-m round trip (s), height achieved when jumping in place, and gender. The dependent variable (target) was obesity classification (obesity \(\:\ge\:\) 30 kg/m 2 ), as per the World Health Organization [29], and was represented in a numeric format (0 = normal, 1 = obese). 2.4. Statistical Modeling 2.4.1. Variance Inflation Factor, Tolerance, and Correlation between Variables The dataset included both independent and dependent variables. A constant term was added to the dataset to include an intercept in the regression model. The variance inflation factor (VIF) corresponding to each variable was calculated using the formula (1/1 − R 2 ), where R 2 is the regression coefficient of the variable with respect to all the other variables. VIF assesses the extent to which the variance of an estimated regression coefficient is inflated owing to multicollinearity. A VIF threshold > 10 was used to identify high multicollinearity [30]. Variables with VIF values > 10 were considered to have significant multicollinearity. Tolerance was calculated as the reciprocal of VIF using the formula (1/VIF). A tolerance of < 0.1 indicated high multicollinearity in the dataset [31]. VIF and tolerance measure the extent to which a variable is unexplained by other predictors in the model. 2.4.2. Data Normalization and Sampling The data were normalized using MinMaxScaler to avoid overreliance on certain features during speed learning by restricting all variables to a range of 0–1. Additionally, the datasets were balanced via undersampling using RandomUnderSampler (random state = 42) by reducing oversampling between "normal" and "obese." The data were split into training and test sets using a train-test split function. An 80 (training):20 (test) split was used in this study. 2.4.3. Analysis Using Neural Network 2.4.3.1. Model Setup and Stratified K-fold Cross-validation The dataset was split into training and validation sets using stratified K-fold (n = 5). The model setup ensured that each fold had the same proportion of "obese" classification labels (normal versus obese) as that in the original dataset. Additionally, the training and validation sets were defined for each fold [32–34]. 2.4.3.2. Model Structure and Compilation This study used a sequential neural network model in which the first layer contained 64 units. The ReLU activation function was used, and the input data had a shape of 10. A dropout layer with a rate of 0.2 was added to prevent overfitting. Another dense layer with 32 units and ReLU activation was present. The final layer was a dense layer with a single unit and a sigmoid activation function, suitable for binary classification (normal vs. obese). The model was compiled with a binary cross-entropy loss function, the Adam optimizer with a learning rate of 0.001 was used, and accuracy was used to monitor the training performance of the model [35, 36]. 2.4.3.3. EarlyStopping and ModelCheckpoint Callbacks Two callbacks were used to monitor and control the training dataset. EarlyStopping monitored the validation loss, terminated the training (patience = 20), and set the weights of the model to those of the optimized model. ModelCheckpoint monitored the validation loss and saved the model with the lowest validation loss during training [37–39]. 2.4.3.4. Model Training and Loading the Optimized Model The model was trained using the training data for a maximum of 200 epochs with a batch size of 16. The training process aimed to minimize binary cross-entropy loss. If the validation loss did not improve for 20 consecutive epochs, the training was terminated and the model weights were set to the optimized state using the EarlyStopping callback. After the training, the model with the optimized performance on the validation set loaded the “TensorFlow and Keras model” function. These functions were stored in the optimized neural network model [39, 40]. 2.4.3.5. Performance Evaluation and Prediction Using the Optimized Model The optimized model was used to predict the validation set. Predictions were obtained for each sample in the validation set. To convert the probabilities generated by the model into binary predictions, a threshold of 0.5 was applied. If a predicted probability was greater than 0.5, the sample was classified as positive (obese); otherwise, the sample was classified as negative (normal). The performance of the model was evaluated on the validation set in terms of accuracy, precision, recall, F1 score, mean absolute error (MAE), and mean squared error (MSE). The evaluation of performance metrics (accuracy, precision, recall, and F1 score) calculated the validation set of each fold. A confusion matrix was prepared using true negatives (TNs), false positives (FPs), false negatives (FNs), and true positives (TPs) for each fold [41–43]. After training, the precision-recall curve was plotted using the probabilities predicted using the validation dataset for each fold. The area under the precision-recall curve (AUPRC) [44], a summary metric that measures the performance of the model across all classification thresholds, was calculated. Finally, the result was used to visualize the precision-recall curve and AUPRC corresponding to each fold to determine the ability of the model to distinguish between the normal and obese classes (Supplement 2(a)). 2.5. Model-agnostic Algorithms Used in the Optimized Model Subsequently, we evaluated the feature importance of each variable used in the model by applying SHapley Additive exPlanations (SHAP) [45] and local interpretable model-agnostic explanations (LIME) analysis [46]. The SHAP analysis used DeepExplainer, which was designed for deep learning models, and the SHAP values were calculated using the scaled training dataset. The SHAP values obtained after the analysis represent the contribution of each input feature to the prediction obtained from the model. The dependence of the SHAP plot uses the relationship between the SHAP and actual values of the feature corresponding to each observation in the dataset. In LIME analysis, LimeTabularExplainer was used, and the probabilities corresponding to the normal and obese classes were generated for specific datapoints from the test dataset. This method evaluates the importance of each variable based on its contribution to the predictions. 3. Results 3.1. VIFs and Tolerances Corresponding to the Variables As shown in Fig. 2 , high multicollinearity was identified when the VIF values were > 10, which corresponds to a tolerance value of less than 0.1. No variables exceeded the VIF threshold of 10 or had a tolerance of < 0.1, indicating that no severe multicollinearity issues existed based on the set VIF and tolerance threshold. 3.2. Neural Network Analysis Using the Stratified K-fold Model In Fig. 3 (a), the confusion matrix corresponding to Fold 1 had TN (n = 1,211), FN (n = 177), FP (n = 171), and TP (n = 1,200) of 43.89%, 6.42%, 6.20%, and 43.49%, respectively. The receiver operating characteristic-area under curve (ROC-AUC) corresponding to Fold 1 was 0.95. The accuracy and precision scores were 0.87 and 0.875, respectively, recall score was 0.871, and F1 score was 0.873. Training was terminated after 51 epochs using callbacks and the early stopping method. The confusion matrix corresponding to Fold 2 had TN (n = 1,188), FN (n = 152), FP (n = 194), and TP (n = 1,225) of 43.06%, 5.51%, 7.03%, and 44.40%, respectively. The ROC-AUC of Fold 2 was 0.95. The accuracy and precision scores were 0.875 and 0.863, respectively, recall score was 0.890, and F1 score was 0.876. Training was terminated after 64 epochs using callbacks and the early stopping method. The confusion matrix corresponding to Fold 3 had TN (n = 1,198), FN (n = 153), FP (n = 184), and TP (n = 1,224) of 43.42%, 5.55%, 6.67%, and 44.36%, respectively. The ROC-AUC of Fold 3 was 0.95. The accuracy and precision scores were 0.878 and 0.869, respectively, recall score was 0.889, and F1 score was 0.879. T raining was terminated after 60 epochs using callbacks and the early stopping method. The confusion matrix corresponding to Fold 4 had TN (n = 1,200), FN (n = 151), FP (n = 182), and TP (n = 1,225) of 43.51%, 5.47%, 6.60%, and 44.42%, respectively. The ROC-AUC of Fold 4 was 0.95. The accuracy and precision scores were 0.879 and 0.871, respectively, recall score was 0.890, and F1 score was 0.880. Training was terminated after 106 epochs using callbacks and the early stopping method. The confusion matrix corresponding to Fold 5 had a TN (n = 1,188), FN (n = 146), FP (n = 194), and TP (n = 1,230) of 43.07%, 5.29%, 7.03%, and 44.60%, respectively. The ROC-AUC of Fold 5 was 0.95. The accuracy and precision scores were 0.877 and 0.864, respectively, recall score was 0.894, and F1 score was 0.879. Training was terminated after 57 epochs using callbacks and the early stopping method. The overall stratified 5-fold analysis (Fig. 3 (b)) showed that the average of the ROC curves was 0.949 and the MAE and MSE were 0.178 and 0.089, respectively. The average accuracy, precision, recall, and F1 scores were 0.877, 0.869, 0.887, and 0.878, respectively. 3.3. Optimized Neural Network Model Used for Obesity Classification Among the five stratified K-folds, the Fold 4 model was the optimal neural network for predicting obesity using callbacks, and the early stopping method yielded optimized training and validation results based on loss and accuracy (Fig. 4 ). In Fig. 4 (a), the confusion matrix corresponding to the optimal neural network model had TN (n = 1,462), FN (n = 184), FP (n = 249), and TP (n = 1,554) of 42.39%, 5.53%, 7.22%, and 45.06%, respectively. The ROC-AUC was 0.95. The accuracy, precision and, recall, and F1 scores were 0.873, 0.866, 0.855, and 0.876, respectively. As shown in Fig. 4 (b), the neural network optimized using callbacks and the early stopping method achieved optimal training and validation results based on loss and accuracy after 20 epochs. For the first 50 datapoints, the MAE and MSE were 0.122 and 0.041, respectively, between the actual and predicted values. The R 2 score of the representative 1–50 datapoints was 0.833, and the average prediction error for the first 50 datapoints was 0.171 (Fig. 4 (c)). 3.4. SHAP and LIME Analysis Using the Optimized Model for Obesity Classifications The SHAP feature-importance analysis showed that the top three variables were 20-m VO 2 max (importance: 0.339, prediction: 0.586), gender (importance: 0.2481, prediction: 0.004), and relative grip strength (importance: 0.135, prediction: 0.76) (Fig. 5 (a)). The SHAP dependence plot in Fig. 5 (c) shows that SHAP values decreased as 20-m VO 2 max and relative grip strength increased; therefore, as these variables increase in value, their impact on model prediction decreases, thereby increasing the number of normal classifications. Low 20-m VO 2 max and relative grip strength resulted in the aggregation of red points (female) at the high end of the SHAP value range, indicating that for the same level of fitness, being female had a higher impact on the predictions than being male. Conversely, the predictions obtained under high 20-m VO 2 max and relative grip strength were less differentiated by gender. These results imply that for low 20-m VO 2 max (Spearman correlation r = −0.642, p = 0.000) and relative grip strength (Spearman correlation r = −0.549, p = 0.000), gender predicted more accurate obesity classification; however, this effect reduced when these two variables increased. As shown in Fig. 5 (b), LIME analysis reveals that the number of instances in which obesity is predicted is 2,608. The intercept and predicted local values (LIME value) obtained from LIME analysis were 0.511 and 0.668, respectively, which indicate obesity predictions. The confidence in the LIME prediction indicates that the prediction probability for the instance obtained from the original model corresponds to the obese class (R 2 = 0.995). LIME feature importance analysis yielded two values, i.e., 20-m VO 2 max less than 0.30 (importance = 0.544) and relative grip strength less than 0.36 (importance = 0.171). The original 20-m VO 2 max and relative grip strength obtained from the Minmax scaler were 28.71 ml/kg/min and 38.17%, respectively. The strong positive contributions of these features indicate that low 20-m VO 2 max and relative grip strength significantly increased obesity classification. The interaction between these two variables is shown in Fig. 5 (d) (Pearson correlation coefficient = 0.727, p = 0.000). Increases in 20-m VO 2 max decreased the SHAP values. This result indicates that a high 20-m VO 2 max corresponds to a low probability of obesity according to the predictions obtained from the model. Furthermore, relative grip strength interacted with 20-m VO 2 max and affected the SHAP values and, in turn, the obesity predictions. However, the increase in relative grip strength (toward the red color) decreased the SHAP values, reducing the probability of obesity classification. 4. Discussion This study used deep neural networks to evaluate the physical fitness factors that affect obesity predictions. This study is a follow-up to our previous study [47], utilizing the same analysis methodology and dataset. The results revealed that neural network analysis across five stratified folds yielded consistent performances with TN, FN, FP, TP, and ROC curves of over 0.95. The accuracy, precision, recall, and F1 scores were in the range of 0.87–0.88% (Fig. 3 ). The results obtained from Fold 4 were the best for determining the optimized neural network for obesity classification. The TN and FN percentages were balanced, with an ROC value of 0.95. Performance metrics, such as accuracy, precision, and recall, were approximately 0.87 (Fig. 4 ). SHAP and LIME analyses clarified the influence of features such as 20-m VO 2 max, gender, and relative grip strength on obesity classifications (Figs. 5 (a) and (b)). Gender and relative grip strength were important factors influencing obesity prediction; however, the interaction of 20-m VO 2 max decreased the probability of obesity classification (Fig. 5 (c)). Therefore, 20-m VO 2 max (original value of < 28.71 ml/kg/min) was an important physical factor in predicting obesity. A previous study proposed a deep neural network model using stratified five k-folds with good accuracy (71%) and explained that low physical fitness, which included the moderate effect-size correlations of aerobic fitness, upper limb strength, and sprint time, in adolescents is correlated with increased obesity risk [20]. Our model exhibited higher accuracy levels than that of the previously reported model; the average accuracy obtained from the stratified five K-folds was 87.7%. The 20-m VO 2 max greatly predicted the obesity classifications, similar to the observations reported in previous studies. The obesity classifications (obesity = 1, normal = 0, degree of prediction to towards 1) increased with the decrease in 20-m VO 2 max. These results are supported by the fact that high BMI is associated with relatively low O 2 uptake [48]. Our results showed that 20-m VO 2 max is the highest feature importance variable for predicting obesity (Figs. 5 -(a) and (b)). The average deep learning modeling error between aerobic exercise and obesity reduction obtained in a previous study was 0.053%, whereas the average performance accuracy error (%) was approximately 0.186% [49]. Moreover, neural network models provided better obesity classifications than the logistic regression model [50]. Both the previous studies support the results obtained from the deep neural network analysis in this study, i.e., a reduced 20-m VO 2 max increased the probability of obesity prediction. Our SHAP analysis showed that gender was the second most important factor for predicting obesity (Fig. 5 (a)). The results indicate that increased 20-m VO 2 max in males decreased the corresponding SHAP importance values (Fig. 5 (c)). This result may correspond to males who attempt to lose weight and, therefore, may increase exercise and reduce fat intake. By contrast, females are likely to participate in weight-loss programs, administer prescription diet pills, follow special diets, and eat more vegetables and fruits [50]. Males who attempt weight loss are more likely to lose weight successfully than females [51]. Our results are supported by males having high aerobic capacity (VO 2 max /lean body mass (kg)) and BMI (> 1.37 kg/m 2 ) and low VO 2 max/lean body mass by 1 ml/kg/min [52]. In addition, the results of SHAP and LIME analyses showed that relative grip strength was an important factor in predicting obesity (Figs. 5 (a) and (b)) in females (Fig. 5 (c)). However, Fig. 5 (d) indicates that a decrease in relative grip strength decreases 20-m VO 2 max and, in turn, the SHAP value, thereby yielding a positive prediction of obesity. Therefore, this result indicates that 20-m VO 2 max was a more important factor in predicting obesity than relative grip strength (Fig. 5 (b)). The results of this study indicate that females (BMI > 30 kg/m 2 , aged 35–45 years) should engage in aerobic exercises, which are more beneficial for preventing cardiovascular diseases due to obesity than resistance exercise training [53]. In addition, engaging in aerobic exercises can effectively decrease obesity risk [54–56]. 5. Conclusion A predictive neural network model was proposed, and data obtained from the Korean NFA program (2010–2023) were analyzed. The model demonstrated high reliability in terms of obesity classification and included significant physical fitness factors such as 20-m VO 2 max (with values < 28.71 ml/kg/min indicating high obesity). The consistency of the model across various metrics (accuracy, precision, recall, and F1 scores) reaffirmed that aerobic capacity can be used for obesity-risk assessment. This study also highlighted the potential gender-specific interventions to reduce the risk of obesity. Additionally, this study highlighted the potential of using neural network models to devise public health strategies for early obesity detection and personalized treatment plans. The limitations of this study include not incorporating dietary intake and physical activity levels, potentially overlooking crucial obesity influencers, and limiting the applicability of the model. Overfitting owing to the use of various sequential neural network architectures may restrict the generalizability and real-world prediction accuracy of the model. Additionally, the adoption of a general obesity classification threshold may be unsuitable for several populations, especially those of Asian countries, owing to different body compositions and risk profiles. Finally, relying on indirect VO 2 max estimates from the 20-m shuttle run test, instead of direct measurements, may reduce the precision of the predictions related to physical fitness factors, thereby affecting the overall robustness and applicability of the model when attempting to accurately predict obesity. Declarations Data Availability Statement All data generated during the current study have been included in this published article and its original dataset and coding file are available from the corresponding and first author (Jun-Hyun Bae & Wook Song) on reasonable request. Ethics Approval Approval was obtained by the Research Ethics Committee of Hyupsung University (approval no. 7002320-202303-HR-001). The procedures used in this study adhere to the tenets of the Declaration of Helsinki. The studies were conducted in accordance with local legislation and institutional requirements. Written informed consent for participants was not required from the participants, their legal guardians, or next of kin in accordance with the national legislation and institutional requirements. The informed consent was waived. Author Contributions J-HB and YHS contributed to data collection, analysis, and writing of the manuscript. J-HB, YHS, and WS were involved in data collection and reviewing the manuscript. All the authors contributed to the manuscript and approved the submitted version. Funding This work was supported by the Basic Science Research Program of the National Research Foundation of Korea (NRF) funded by the Ministry of Education (RS-2023-00247075) and Kyungil University research fund. Acknowledgements None Competing Interests & Disclosure of Potential Conflicts of Interest The authors declare that this study was conducted in the absence of any commercial or financial relationships that could be construed as potential conflicts of interest. References Hong, I., et al., Relationship Between Physical Activity and Overweight and Obesity in Children: Findings From the 2012 National Health and Nutrition Examination Survey National Youth Fitness Survey. Am J Occup Ther, 2016. 70 (5): p. 7005180060p1-8. Yang, Y.S., et al., Obesity Fact Sheet in Korea, 2021: Trends in Obesity Prevalence and Obesity-Related Comorbidity Incidence Stratified by Age from 2009 to 2019. J Obes Metab Syndr, 2022. 31 (2): p. 169-177. Poirier, P., et al., Obesity and cardiovascular disease: pathophysiology, evaluation, and effect of weight loss: an update of the 1997 American Heart Association Scientific Statement on Obesity and Heart Disease from the Obesity Committee of the Council on Nutrition, Physical Activity, and Metabolism. Circulation, 2006. 113 (6): p. 898-918. Kim, D.-S. and P.E. Scherer, Obesity, diabetes, and increased cancer progression. Diabetes & metabolism journal, 2021. 45 (6): p. 799-812. Nyberg, S.T., et al., Obesity and loss of disease-free years owing to major non-communicable diseases: a multicohort study. The lancet Public health, 2018. 3 (10): p. e490-e497. Kyröläinen, H., et al., Physical Fitness Profiles of Young Men. Sports Medicine, 2010. 40 (11): p. 907-920. Kim, H.J., et al., Relationships of physical fitness and obesity with metabolic risk factors in children and adolescents: Chungju city cohort study. Ann Pediatr Endocrinol Metab, 2016. 21 (1): p. 31-8. Kim, J.-W., et al., Association between obesity and various parameters of physical fitness in Korean students. Obesity Research & Clinical Practice, 2013. 7 (1): p. e67-e74. Pojskic, H. and B. Eslami, Relationship Between Obesity, Physical Activity, and Cardiorespiratory Fitness Levels in Children and Adolescents in Bosnia and Herzegovina: An Analysis of Gender Differences. Frontiers in Physiology, 2018. 9 . Castro, E.A., et al., What is the most effective exercise protocol to improve cardiovascular fitness in overweight and obese subjects? Journal of Sport and Health Science, 2017. 6 (4): p. 454-461. Zhou, N., Assessment of aerobic exercise capacity in obesity, which expression of oxygen uptake is the best? Sports Medicine and Health Science, 2021. 3 (3): p. 138-147. Gaesser, G.A. and S.S. Angadi, Obesity treatment: Weight loss versus increasing fitness and physical activity for reducing health risks. iScience, 2021. 24 (10): p. 102995. Lang, J., et al., Systematic review of the relationship between 20 m shuttle run performance and health indicators among children and youth. Journal of Science and Medicine in Sport, 2017. 21 . Rey, O., et al., Psycho-Physiological Responses of Obese Adolescents to an Intermittent Run Test Compared with a 20-M Shuttle Run. J Sports Sci Med, 2016. 15 (3): p. 451-459. Moran, C.A., et al., Performance and reproducibility on shuttle run test between obese and non-obese children: a cross-sectional study. BMC Pediatrics, 2017. 17 (1): p. 68. Przednowek, K., et al., Predictive modeling of VO2max based on 20 m shuttle run test for young healthy people. Applied Sciences, 2018. 8 (11): p. 2213. Chung, J.W., O. Lee, and K.H. Lee, Estimation of maximal oxygen consumption using the 20m shuttle run test in Korean adults aged 19-64 years. Science & Sports, 2023. 38 (1): p. 68-74. DeGregory, K.W., et al., A review of machine learning in obesity. Obesity Reviews, 2018. 19 (5): p. 668-685. Ferdowsy, F., et al., A machine learning approach for obesity risk prediction. Current Research in Behavioral Sciences, 2021. 2 : p. 100053. Forte, P., et al., A Deep Learning Neural Network to Classify Obesity Risk in Portuguese Adolescents Based on Physical Fitness Levels and Body Mass Index Percentiles: Insights for National Health Policies. Behavioral Sciences, 2023. 13 (7): p. 522. Cheng, X., et al., Does Physical Activity Predict Obesity-A Machine Learning and Statistical Method-Based Analysis. Int J Environ Res Public Health, 2021. 18 (8). Thamrin, S.A., et al., Predicting Obesity in Adults Using Machine Learning Techniques: An Analysis of Indonesian Basic Health Research 2018. Frontiers in Nutrition, 2021. 8 . Jossa-Bastidas, O., et al., Predicting Physical Exercise Adherence in Fitness Apps Using a Deep Learning Approach. Int J Environ Res Public Health, 2021. 18 (20). Lee, S.-H., et al., Estimation of health-related physical fitness (hrpf) levels of the general public using artificial neural network with the national fitness award (nfa) datasets. International journal of environmental research and public health, 2021. 18 (19): p. 10391. Ko, B.-g., et al., Prediction equations of physical fitness age for Korean adults. Exercise Science, 2021. 30 (3): p. 352-360. Park, S., et al., Normative Reference Values of Physical Fitness Levels in Koreans: Results from the National Fitness Award Project (2017-2019). Exercise Science, 2022. 31 (4): p. 511-526. Jung, J., et al., Im S. Model development of fitness cer-tification center . 2014, Research report, Ministry of Culture, Sports and Tourism. Bae, J.-H., et al., Prediction models of grip strength in adults above 65 years using Korean National Physical Fitness Award Data from 2009 to 2019. European Geriatric Medicine, 2023. 14 (5): p. 1059-1064. Weir, C.B. and A. Jan, BMI classification percentile and cut off points. 2019. Kalnins, A. and K. Praitis Hill, The VIF Score. What is it Good For? Absolutely Nothing. Organizational research methods, 2023: p. 10944281231216381. Krishnamoorthy, K. and T. Mathew, Statistical tolerance regions: theory, applications, and computation . 2009: John Wiley & Sons. Maqsood, S. and R. Damaševičius, Multiclass skin lesion localization and classification using deep learning based features fusion and selection framework for smart healthcare. Neural Networks, 2023. 160 : p. 238-258. Yan, Y., et al. Deep learning for imbalanced multimedia data classification . in 2015 IEEE international symposium on multimedia (ISM) . 2015. IEEE. Refaeilzadeh, P., L. Tang, and H. Liu, Cross-validation. Encyclopedia of database systems, 2009: p. 532-538. Peng, D., et al., Addressing the multi-label imbalance for neural networks: An approach based on stratified mini-batches. Neurocomputing, 2021. 435 : p. 91-102. Piñeyro, L., A. Pardo, and M. Viera. Structure verification of deep neural networks at compilation time using dependent types . in Proceedings of the XXIII Brazilian Symposium on Programming Languages . 2019. Nehra, N., P. Sangwan, and D. Kumar, Artificial Neural Networks: A Comprehensive Review. Handbook of Machine Learning for Computational Optimization, 2021: p. 203-227. Sabiri, B., B. El Asri, and M. Rhanoui. Mechanism of Overfitting Avoidance Techniques for Training Deep Neural Networks . in ICEIS (1) . 2022. Bae, J.-H., J.-w. Seo, and D.Y. Kim, Deep-learning model for predicting physical fitness in possible sarcopenia: analysis of the Korean physical fitness award from 2010 to 2023. Frontiers in Public Health, 2023. 11 . Lee, S., et al., Background information of deep learning for structural engineering. Archives of Computational Methods in Engineering, 2018. 25 : p. 121-129. Li, Y. and Z. Chen, Performance evaluation of machine learning methods for breast cancer prediction. Appl Comput Math, 2018. 7 (4): p. 212-216. Pham, B.T., et al., Performance evaluation of machine learning methods for forest fire modeling and prediction. Symmetry, 2020. 12 (6): p. 1022. Luque, A., et al., The impact of class imbalance in classification performance metrics based on the binary confusion matrix. Pattern Recognition, 2019. 91 : p. 216-231. Khan, S.A. and Z.A. Rana. Evaluating performance of software defect prediction models using area under precision-Recall curve (AUC-PR) . in 2019 2nd International Conference on Advancements in Computational Sciences (ICACS) . 2019. IEEE. Mosca, E., et al. SHAP-based explanation methods: a review for NLP interpretability . in Proceedings of the 29th International Conference on Computational Linguistics . 2022. Kumarakulasinghe, N.B., et al. Evaluating local interpretable model-agnostic explanations on clinical machine learning classification models . in 2020 IEEE 33rd International Symposium on Computer-Based Medical Systems (CBMS) . 2020. IEEE. Bae, J.-H., et al., Neural network model for prediction of possible sarcopenic obesity using Korean national fitness award data (2010–2023). Scientific Reports, 2024. 14 (1): p. 14565. Umamaheswari, K., et al., VO2 Max and Body Mass in Overweight and Obese Young Adults. International Journal of Physiology, 2017. 5 (2): p. 23-27. Mu, P., Modeling Analysis of the Relationship between Adolescent Aerobic Exercise and Obesity Reduction Based on Deep Learning. Advances in Multimedia, 2022. 2022 : p. 4112169. Ergün, U., The classification of obesity disease in logistic regression and neural network methods. J Med Syst, 2009. 33 (1): p. 67-72. Tsai, S.A., et al., Gender Differences in Weight-Related Attitudes and Behaviors Among Overweight and Obese Adults in the United States. American Journal of Men's Health, 2016. 10 (5): p. 389-398. Sharma, H.B. and J. Kailashiya, Gender Difference in Aerobic Capacity and the Contribution by Body Composition and Haemoglobin Concentration: A Study in Young Indian National Hockey Players. J Clin Diagn Res, 2016. 10 (11): p. Cc09-cc13. Chaudhary, S., M.K. Kang, and J.S. Sandhu, The effects of aerobic versus resistance training on cardiovascular fitness in obese sedentary females. Asian J Sports Med, 2010. 1 (4): p. 177-84. Bellicha, A., et al., Effect of exercise training on weight loss, body composition changes, and weight maintenance in adults with overweight or obesity: An overview of 12 systematic reviews and 149 studies. Obesity Reviews, 2021. 22 : p. e13256. Skrypnik, D., et al., Effects of endurance and endurance strength training on body composition and physical capacity in women with abdominal obesity. Obesity facts, 2015. 8 (3): p. 175-187. Jamka, M., et al., Comparison of the Effect of Endurance, Strength and Endurance-Strength Training on Glucose and Insulin Homeostasis and the Lipid Profile of Overweight and Obese Subjects: A Systematic Review and Meta-Analysis. International Journal of Environmental Research and Public Health, 2022. 19 (22): p. 14928. Additional Declarations No competing interests reported. Supplementary Files supplement.pptx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4782187","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":339468744,"identity":"040aa062-8f2b-4fc6-8cc1-99113d6c3b9b","order_by":0,"name":"Jun-Hyun Bae","email":"","orcid":"","institution":"Seoul National University","correspondingAuthor":false,"prefix":"","firstName":"Jun-Hyun","middleName":"","lastName":"Bae","suffix":""},{"id":339468745,"identity":"23ad2f6d-306c-4b9d-86b9-fd212ce931d7","order_by":1,"name":"Yunho Sung","email":"","orcid":"","institution":"Seoul National University","correspondingAuthor":false,"prefix":"","firstName":"Yunho","middleName":"","lastName":"Sung","suffix":""},{"id":339468746,"identity":"83e24128-1723-4e02-a5a8-0c2121a4bfd6","order_by":2,"name":"Xinxing Li","email":"","orcid":"","institution":"Seoul National University","correspondingAuthor":false,"prefix":"","firstName":"Xinxing","middleName":"","lastName":"Li","suffix":""},{"id":339468747,"identity":"1200a27e-0dba-41d7-9cc0-cef8647cadeb","order_by":3,"name":"Wook Song","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAwklEQVRIiWNgGAWjYNCCHzYwVgKROhh70kjVwsB2mAQt8u29h1/z8Jy3N5dIYPzwgyEtn6AWgzPn0qx5LG4n7pyRwCzZw5Bj2UBQi0SOmXEOz+0EgxsJDNIMDBUGhB02A6SF7Zw9UAvzb6K0MNzIMX6cw3aAccONBDagLTmEtRicOWPG/LcnOXHDmYdtlj0GaUQ4rL3H+OOMH3b2BseTD9/4UZFMhMOAkSIBoRkbgJYSo4GBgfkDcepGwSgYBaNgxAIApxU4DcC0uokAAAAASUVORK5CYII=","orcid":"","institution":"Seoul National University","correspondingAuthor":true,"prefix":"","firstName":"Wook","middleName":"","lastName":"Song","suffix":""}],"badges":[],"createdAt":"2024-07-22 13:21:00","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4782187/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4782187/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":63157020,"identity":"db62e54b-adb7-44cc-9f74-279ffcbe0365","added_by":"auto","created_at":"2024-08-23 21:27:56","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":65378,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eData collection process. \u003c/strong\u003eData collection process. The original data collected from January 2010 to March 2023 included 1,810,208 participants. Data were excluded in three steps. The outlier in each variable was detected via the IQR method using the exclusion criterion (\u0026gt; Q3 + 1.5 × IQR \u0026amp; \u0026lt; Q1 – 1.5 × IQR). The final number of study subjects was 204,334.\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4782187/v1/bf974b500b5876e705ba4802.jpg"},{"id":63157652,"identity":"de27c233-3578-4ac4-94ce-b06c22ae3a8b","added_by":"auto","created_at":"2024-08-23 21:35:56","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":105351,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eVIF, tolerance, and correlation corresponding to each variable. a.\u003c/strong\u003eMulticollinearity between variables. The variance inflation factor (VIF) and tolerance identify the interdependence between independent variables in the model. The VIF threshold is \u0026gt;10, and the tolerance threshold is \u0026lt; 0.1. \u003cstrong\u003eb.\u003c/strong\u003ePearson correlation between variables. The correlation matrix is visualized using the Heatmap function from the Seaborn library.\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4782187/v1/77d32023d2e33808b1b4aba9.jpg"},{"id":63157025,"identity":"20512753-8f2a-4e46-92cd-643998ae1493","added_by":"auto","created_at":"2024-08-23 21:27:56","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":391619,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eBMI classification using the neural network model. a.\u003c/strong\u003e Stratified K-Fold cross-validation as a binary classification (0 = normal, 1 = obese) dataset. The neural network model is trained for each fold (n = 5), and the results are presented along with confusion matrices and ROC curves. \u003cstrong\u003eb.\u003c/strong\u003e Performance metrics (ROC-AUC, accuracy, precision, recall, F1-score, MAE, and MSE) averaged across all seven folds to assess the model’s effectiveness. Abbreviations: TN = true negatives, FP = false positives, FN = false negatives, TP= true positives, Ave = average, ROC-AUC = receiver operating characteristic-area under curve, MAE = mean absolute error, and MSE = mean squared error.\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4782187/v1/b1e5b92a82e60558b0f106b0.jpg"},{"id":63157021,"identity":"60581e8d-32db-4e87-886b-c9103af4d5fb","added_by":"auto","created_at":"2024-08-23 21:27:56","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":200962,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eBest neural network model for BMI classification. a.\u003c/strong\u003e Evaluation of the best neural network model using confusion matrix, accuracy, precision, recall, F1-score, MAE, MSE, and ROC curve. \u003cstrong\u003eb.\u003c/strong\u003e Overfitting and underfitting are identified by comparing the loss and accuracy trends obtained using the training and validation datasets and the best neural network model. \u003cstrong\u003ec.\u003c/strong\u003eData are generated to compare the predicted and actual values, and a subset of these values is used to visualize the model’s prediction accuracy using the actual data for the first 50 samples. The prediction error is calculated by considering the absolute difference between the predicted and actual values for the first 50 datapoints. The actual value with an associated error bar indicates the magnitude of the prediction error.\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4782187/v1/e3df19d45967cff975dc3adf.jpg"},{"id":63157023,"identity":"96f7555a-d07c-4e65-bf62-d0954ced169f","added_by":"auto","created_at":"2024-08-23 21:27:56","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":78333,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eResults obtained from the model-agnostic explainable algorithms using the best neural network model. a.\u003c/strong\u003e Detailed analysis of the SHAP feature importance obtained using the best neural network model, which computed SHAP values to measure the impact of each feature on model output using the training dataset. \u003cstrong\u003eb.\u003c/strong\u003e The LIME results show the feature importance for distinguishing the positive and negative of mean importance on the best neural network model. \u003cstrong\u003ec.\u003c/strong\u003e Impact of 20-m VO\u003csub\u003e2\u003c/sub\u003e max (ml/kg/min) and relative grip strength (%) on obesity predictions yielded by the best model. A SHAP dependence plot showing the relationship between 20-m VO\u003csub\u003e2\u003c/sub\u003e max (ml/kg/min) and relative grip strength (%) and the SHAP values obtained from the best neural network model. \u003cstrong\u003ed.\u003c/strong\u003e Feature interactions determined by SHAP value analysis (1–100 samples) using a dependence plot (blue: low, and red: high). All values were based on the Minmax scaler dataset.\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4782187/v1/3295faa5071c62fb4d060f4e.jpg"},{"id":68224497,"identity":"8041d6a5-ff97-4748-b804-67d7a691b1ad","added_by":"auto","created_at":"2024-11-05 03:01:55","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1624121,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4782187/v1/97c7c1bb-59b6-4055-a947-02bce705aef6.pdf"},{"id":63157651,"identity":"bf78a96f-d318-40f3-9304-6ef737a0c3b0","added_by":"auto","created_at":"2024-08-23 21:35:56","extension":"pptx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":343162,"visible":true,"origin":"","legend":"","description":"","filename":"supplement.pptx","url":"https://assets-eu.researchsquare.com/files/rs-4782187/v1/28db1c5b639fdc970a00f646.pptx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Sequential Deep Learning Model for Obesity Prediction Based on Physical Fitness Factors: An Analysis of Data from the 2010–2023 Korean National Physical Fitness Data ","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eThe prevalence of obesity is increasing rapidly worldwide. According to the recent epidemic-related definition of obesity, this disorder has been labeled as a \"global syndemic.\" Moreover, obesity is one of the most significant risk factors related to physical and psychological health problems [1] and increases the risks of noncommunicable diseases, cardiovascular diseases, cancer, chronic respiratory diseases, and diabetes [2\u0026ndash;5]. Regular physical activity and improving physical fitness reduce the risk of obesity [6]. In particular, cardiorespiratory fitness is negatively correlated with obesity [7\u0026ndash;9], and increased exercise-induced oxygen consumption (VO\u003csub\u003e2\u003c/sub\u003e max) improves the physical fitness of obese individuals [10, 11].\u003c/p\u003e \u003cp\u003eImproving physical fitness, especially cardiorespiratory fitness (CRF), is a more effective strategy for treating obesity than focusing only on exercise-induced weight loss [12]. Moreover, the results of a 20-m shuttle run test can be used to determine CRF [13]; this test can be considered a relevant field test for assessing the CRF of obese adolescents [14] and youth [15]. Moreover, a systematic review and meta-analysis revealed that the results of a 20-m shuttle run test were negatively correlated with body mass index (BMI), body fat percentage, and waist circumference [16]. A previous study focused on youths and adolescents also revealed that this test was a major indicator of CRF. Another study indicated that the results of a 20-m shuttle run test were significant indicators of VO\u003csub\u003e2\u003c/sub\u003e max, and the results exhibited acceptable accuracy when a large sample of adults was tested [17].\u003c/p\u003e \u003cp\u003eRecently, artificial intelligence-based methods involving machine and deep learning (neural networks) have been developed for predicting obesity [18, 19]. A previous study showed that support vector machine, simple K-means, and decision trees are suitable algorithms for detecting obesity levels using big data sources. A neural network model constructed using data on physical fitness factors showed that an increased risk of obesity was associated with low physical fitness levels [20]. Many studies have considered national health and nutrition examination surveys that collected physical activity and basic clinical information via machine learning methods. Most of the results revealed that doing a physical activity without being physically fit is a key factor in determining obesity levels [21, 22]. The results obtained from a model for predicting physical exercise adherence using deep neural networks indicated that the analysis was a good indicator of physical exercise adherence without the physical fitness factor [23]. However, many studies have evaluated the performances (e.g., accuracy, precision, recall, and F1 scores) of such models without conducting feature importance analyses. As obesity can be predicted using physical fitness factors, these factors should be analyzed and ranked. This study used a deep neural network and data obtained from the 2010\u0026ndash;2023 Korean National Physical Fitness Award (NFA) to predict obesity based on physical fitness factors. Additionally, this study hypothesized that the result of a 20-m shuttle run test is a physical factor that can be used to predict obesity.\u003c/p\u003e"},{"header":"2. Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Dataset and Measurements\u003c/h2\u003e \u003cp\u003eAll participants voluntarily participated in the Korean NFA project at 75 locations nationwide [24]. In this study, various physical fitness factors were investigated for predicting obesity by incorporating variables such as age (years), diastolic blood pressure (DBP, mmHg), systolic blood pressure (SBP, mmHg), relative grip strength (%), sit and reach distance (cm), sit-up repetitions (count), VO\u003csub\u003e2\u003c/sub\u003e max during the 20-m shuttle run (ml/kg/min), the time required for a 10-m round trip (s), height achieved when jumping in place (cm), and gender (male\u0026thinsp;=\u0026thinsp;1, female\u0026thinsp;=\u0026thinsp;2) [25]. Additionally, muscular strength was measured using a handheld dynamometer (Takei, Niigata, Japan), with the participants gripping the handle and pulling with maximum force for 5 s at a 15\u0026deg; angle. The largest of the measurements obtained from both hands was recorded. Relative handgrip strength was calculated by dividing the absolute handgrip strength (kg) by body weight (kg) and multiplying by 100. CRF was assessed using the results of the 20-m shuttle run test in which participants ran back and forth on a 20 m course at increasing speeds, as signaled by beeps. The maximum number of runs was recorded for each participant, and the data obtained from the participants who failed to reach the line before the warning beep were excluded. VO\u003csub\u003e2\u003c/sub\u003e max was calculated using the following estimation formulae based on the number of completions of the 20-m shuttle run test: VO\u003csub\u003e2\u003c/sub\u003e max of a male\u0026thinsp;=\u0026thinsp;48.550 \u0026minus; 0.080 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\times\\:\\)\u003c/span\u003e\u003c/span\u003e (age in years) \u0026minus; 0.021 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\times\\:\\)\u003c/span\u003e\u003c/span\u003e (height in cm) \u0026minus; 0.146 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\times\\:\\)\u003c/span\u003e\u003c/span\u003e (weight in kg)\u0026thinsp;+\u0026thinsp;0.234 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\times\\:\\)\u003c/span\u003e\u003c/span\u003e (laps), and VO\u003csub\u003e2\u003c/sub\u003e max of a female\u0026thinsp;=\u0026thinsp;34.305 \u0026minus; 0.039 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\times\\:\\)\u003c/span\u003e\u003c/span\u003e (age in years)\u0026thinsp;+\u0026thinsp;0.018 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\times\\:\\)\u003c/span\u003e\u003c/span\u003e (height in cm) \u0026minus; 0.161 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\times\\:\\)\u003c/span\u003e\u003c/span\u003e (weight in kg)\u0026thinsp;+\u0026thinsp;0.305 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\times\\:\\)\u003c/span\u003e\u003c/span\u003e (laps) [26, 27]. All measurements on Korean subjects aged between 19\u0026ndash;64 years were collected by trained physical fitness instructors [28].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Data Collection\u003c/h2\u003e \u003cp\u003eFormal consent was not required for this study. The Research Ethics Committee of Hyupsung University approved the dataset (IRB no. 7002320-202303-HR-001), and all methods were implemented in accordance with the relevant guidelines. The data comprised exercise and health-related measurements on Koreans aged 19\u0026ndash;64 years collected from 19 national fitness centers. The original Korean NFA data were collected between Jan 2010\u0026ndash;Mar 2023 (n\u0026thinsp;=\u0026thinsp;1,810,208), and the first stage excluded data from persons aged \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\ge\\:\\)\u003c/span\u003e\u003c/span\u003e 65 years (n = 948,009). In the second stage, data with \u0026gt; 20% missing values collected from individuals were excluded (n = 246,842), along with values \u0026gt; Q3 + 1.5 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\times\\:\\)\u003c/span\u003e\u003c/span\u003e IQR or \u0026lt; Q1\u0026ndash;1.5 \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\times\\:\\)\u003c/span\u003e\u003c/span\u003e IQR (Q: quartile, IQR: interquartile range. n = 374,943). The final sample size comprised 204,334 participants (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The data analysis environment was implemented on Windows 11 (x64 version), with a 13th Gen Intel(R) Core (TM) i9-13900HX processor (2.20 GHz), 32 GB RAM, NVIDIA GeForce RTX 4060 graphics processing unit, and Python (version 3.11.7) with TensorFlow (version 2.15.0).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Data Independent and Dependent Variables\u003c/h2\u003e \u003cp\u003eThis study used independent variables (features) including age, DBP, SBP, relative grip strength, sit and reach distance, sit-up repetitions (count), VO\u003csub\u003e2\u003c/sub\u003e max during the 20-m shuttle run, time required for a 10-m round trip (s), height achieved when jumping in place, and gender. The dependent variable (target) was obesity classification (obesity \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\ge\\:\\)\u003c/span\u003e\u003c/span\u003e 30 kg/m\u003csup\u003e2\u003c/sup\u003e), as per the World Health Organization [29], and was represented in a numeric format (0 = normal, 1 = obese).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4. Statistical Modeling\u003c/h2\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e2.4.1. Variance Inflation Factor, Tolerance, and Correlation between Variables\u003c/h2\u003e \u003cp\u003eThe dataset included both independent and dependent variables. A constant term was added to the dataset to include an intercept in the regression model. The variance inflation factor (VIF) corresponding to each variable was calculated using the formula (1/1 \u0026minus; R\u003csup\u003e2\u003c/sup\u003e), where R\u003csup\u003e2\u003c/sup\u003e is the regression coefficient of the variable with respect to all the other variables. VIF assesses the extent to which the variance of an estimated regression coefficient is inflated owing to multicollinearity. A VIF threshold\u0026thinsp;\u0026gt;\u0026thinsp;10 was used to identify high multicollinearity [30]. Variables with VIF values\u0026thinsp;\u0026gt;\u0026thinsp;10 were considered to have significant multicollinearity. Tolerance was calculated as the reciprocal of VIF using the formula (1/VIF). A tolerance of \u0026lt;\u0026thinsp;0.1 indicated high multicollinearity in the dataset [31]. VIF and tolerance measure the extent to which a variable is unexplained by other predictors in the model.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.4.2. Data Normalization and Sampling\u003c/h2\u003e \u003cp\u003eThe data were normalized using MinMaxScaler to avoid overreliance on certain features during speed learning by restricting all variables to a range of 0\u0026ndash;1. Additionally, the datasets were balanced via undersampling using RandomUnderSampler (random state\u0026thinsp;=\u0026thinsp;42) by reducing oversampling between \"normal\" and \"obese.\" The data were split into training and test sets using a train-test split function. An 80 (training):20 (test) split was used in this study.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e2.4.3. Analysis Using Neural Network\u003c/h2\u003e \u003cdiv id=\"Sec10\" class=\"Section4\"\u003e \u003ch2\u003e2.4.3.1. Model Setup and Stratified K-fold Cross-validation\u003c/h2\u003e \u003cp\u003eThe dataset was split into training and validation sets using stratified K-fold (n\u0026thinsp;=\u0026thinsp;5). The model setup ensured that each fold had the same proportion of \"obese\" classification labels (normal versus obese) as that in the original dataset. Additionally, the training and validation sets were defined for each fold [32\u0026ndash;34].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section4\"\u003e \u003ch2\u003e2.4.3.2. Model Structure and Compilation\u003c/h2\u003e \u003cp\u003eThis study used a sequential neural network model in which the first layer contained 64 units. The ReLU activation function was used, and the input data had a shape of 10. A dropout layer with a rate of 0.2 was added to prevent overfitting. Another dense layer with 32 units and ReLU activation was present. The final layer was a dense layer with a single unit and a sigmoid activation function, suitable for binary classification (normal vs. obese). The model was compiled with a binary cross-entropy loss function, the Adam optimizer with a learning rate of 0.001 was used, and accuracy was used to monitor the training performance of the model [35, 36].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section4\"\u003e \u003ch2\u003e2.4.3.3. EarlyStopping and ModelCheckpoint Callbacks\u003c/h2\u003e \u003cp\u003eTwo callbacks were used to monitor and control the training dataset. EarlyStopping monitored the validation loss, terminated the training (patience\u0026thinsp;=\u0026thinsp;20), and set the weights of the model to those of the optimized model. ModelCheckpoint monitored the validation loss and saved the model with the lowest validation loss during training [37\u0026ndash;39].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section4\"\u003e \u003ch2\u003e2.4.3.4. Model Training and Loading the Optimized Model\u003c/h2\u003e \u003cp\u003eThe model was trained using the training data for a maximum of 200 epochs with a batch size of 16. The training process aimed to minimize binary cross-entropy loss. If the validation loss did not improve for 20 consecutive epochs, the training was terminated and the model weights were set to the optimized state using the EarlyStopping callback. After the training, the model with the optimized performance on the validation set loaded the \u0026ldquo;TensorFlow and Keras model\u0026rdquo; function. These functions were stored in the optimized neural network model [39, 40].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section4\"\u003e \u003ch2\u003e2.4.3.5. Performance Evaluation and Prediction Using the Optimized Model\u003c/h2\u003e \u003cp\u003eThe optimized model was used to predict the validation set. Predictions were obtained for each sample in the validation set. To convert the probabilities generated by the model into binary predictions, a threshold of 0.5 was applied. If a predicted probability was greater than 0.5, the sample was classified as positive (obese); otherwise, the sample was classified as negative (normal). The performance of the model was evaluated on the validation set in terms of accuracy, precision, recall, F1 score, mean absolute error (MAE), and mean squared error (MSE). The evaluation of performance metrics (accuracy, precision, recall, and F1 score) calculated the validation set of each fold. A confusion matrix was prepared using true negatives (TNs), false positives (FPs), false negatives (FNs), and true positives (TPs) for each fold [41\u0026ndash;43]. After training, the precision-recall curve was plotted using the probabilities predicted using the validation dataset for each fold. The area under the precision-recall curve (AUPRC) [44], a summary metric that measures the performance of the model across all classification thresholds, was calculated. Finally, the result was used to visualize the precision-recall curve and AUPRC corresponding to each fold to determine the ability of the model to distinguish between the normal and obese classes (Supplement 2(a)).\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e2.5. Model-agnostic Algorithms Used in the Optimized Model\u003c/h2\u003e \u003cp\u003eSubsequently, we evaluated the feature importance of each variable used in the model by applying SHapley Additive exPlanations (SHAP) [45] and local interpretable model-agnostic explanations (LIME) analysis [46]. The SHAP analysis used DeepExplainer, which was designed for deep learning models, and the SHAP values were calculated using the scaled training dataset. The SHAP values obtained after the analysis represent the contribution of each input feature to the prediction obtained from the model. The dependence of the SHAP plot uses the relationship between the SHAP and actual values of the feature corresponding to each observation in the dataset. In LIME analysis, LimeTabularExplainer was used, and the probabilities corresponding to the normal and obese classes were generated for specific datapoints from the test dataset. This method evaluates the importance of each variable based on its contribution to the predictions.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e3.1. VIFs and Tolerances Corresponding to the Variables\u003c/h2\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, high multicollinearity was identified when the VIF values were \u0026gt;\u0026thinsp;10, which corresponds to a tolerance value of less than 0.1. No variables exceeded the VIF threshold of 10 or had a tolerance of \u0026lt;\u0026thinsp;0.1, indicating that no severe multicollinearity issues existed based on the set VIF and tolerance threshold.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Neural Network Analysis Using the Stratified K-fold Model\u003c/h2\u003e \u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(a), the confusion matrix corresponding to Fold 1 had TN (n\u0026thinsp;=\u0026thinsp;1,211), FN (n\u0026thinsp;=\u0026thinsp;177), FP (n\u0026thinsp;=\u0026thinsp;171), and TP (n\u0026thinsp;=\u0026thinsp;1,200) of 43.89%, 6.42%, 6.20%, and 43.49%, respectively. The receiver operating characteristic-area under curve (ROC-AUC) corresponding to Fold 1 was 0.95. The accuracy and precision scores were 0.87 and 0.875, respectively, recall score was 0.871, and F1 score was 0.873. Training was terminated after 51 epochs using callbacks and the early stopping method. The confusion matrix corresponding to Fold 2 had TN (n\u0026thinsp;=\u0026thinsp;1,188), FN (n\u0026thinsp;=\u0026thinsp;152), FP (n\u0026thinsp;=\u0026thinsp;194), and TP (n\u0026thinsp;=\u0026thinsp;1,225) of 43.06%, 5.51%, 7.03%, and 44.40%, respectively. The ROC-AUC of Fold 2 was 0.95. The accuracy and precision scores were 0.875 and 0.863, respectively, recall score was 0.890, and F1 score was 0.876. Training was terminated after 64 epochs using callbacks and the early stopping method. The confusion matrix corresponding to Fold 3 had TN (n\u0026thinsp;=\u0026thinsp;1,198), FN (n\u0026thinsp;=\u0026thinsp;153), FP (n\u0026thinsp;=\u0026thinsp;184), and TP (n\u0026thinsp;=\u0026thinsp;1,224) of 43.42%, 5.55%, 6.67%, and 44.36%, respectively. The ROC-AUC of Fold 3 was 0.95. The accuracy and precision scores were 0.878 and 0.869, respectively, recall score was 0.889, and F1 score was 0.879. T raining was terminated after 60 epochs using callbacks and the early stopping method. The confusion matrix corresponding to Fold 4 had TN (n\u0026thinsp;=\u0026thinsp;1,200), FN (n\u0026thinsp;=\u0026thinsp;151), FP (n\u0026thinsp;=\u0026thinsp;182), and TP (n\u0026thinsp;=\u0026thinsp;1,225) of 43.51%, 5.47%, 6.60%, and 44.42%, respectively. The ROC-AUC of Fold 4 was 0.95. The accuracy and precision scores were 0.879 and 0.871, respectively, recall score was 0.890, and F1 score was 0.880. Training was terminated after 106 epochs using callbacks and the early stopping method. The confusion matrix corresponding to Fold 5 had a TN (n\u0026thinsp;=\u0026thinsp;1,188), FN (n\u0026thinsp;=\u0026thinsp;146), FP (n\u0026thinsp;=\u0026thinsp;194), and TP (n\u0026thinsp;=\u0026thinsp;1,230) of 43.07%, 5.29%, 7.03%, and 44.60%, respectively. The ROC-AUC of Fold 5 was 0.95. The accuracy and precision scores were 0.877 and 0.864, respectively, recall score was 0.894, and F1 score was 0.879. Training was terminated after 57 epochs using callbacks and the early stopping method. The overall stratified 5-fold analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(b)) showed that the average of the ROC curves was 0.949 and the MAE and MSE were 0.178 and 0.089, respectively. The average accuracy, precision, recall, and F1 scores were 0.877, 0.869, 0.887, and 0.878, respectively.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Optimized Neural Network Model Used for Obesity Classification\u003c/h2\u003e \u003cp\u003eAmong the five stratified K-folds, the Fold 4 model was the optimal neural network for predicting obesity using callbacks, and the early stopping method yielded optimized training and validation results based on loss and accuracy (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). In Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e(a), the confusion matrix corresponding to the optimal neural network model had TN (n\u0026thinsp;=\u0026thinsp;1,462), FN (n\u0026thinsp;=\u0026thinsp;184), FP (n\u0026thinsp;=\u0026thinsp;249), and TP (n\u0026thinsp;=\u0026thinsp;1,554) of 42.39%, 5.53%, 7.22%, and 45.06%, respectively. The ROC-AUC was 0.95. The accuracy, precision and, recall, and F1 scores were 0.873, 0.866, 0.855, and 0.876, respectively. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e(b), the neural network optimized using callbacks and the early stopping method achieved optimal training and validation results based on loss and accuracy after 20 epochs. For the first 50 datapoints, the MAE and MSE were 0.122 and 0.041, respectively, between the actual and predicted values. The R\u003csup\u003e2\u003c/sup\u003e score of the representative 1\u0026ndash;50 datapoints was 0.833, and the average prediction error for the first 50 datapoints was 0.171 (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e(c)).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e3.4. SHAP and LIME Analysis Using the Optimized Model for Obesity Classifications\u003c/h2\u003e \u003cp\u003eThe SHAP feature-importance analysis showed that the top three variables were 20-m VO\u003csub\u003e2\u003c/sub\u003e max (importance: 0.339, prediction: 0.586), gender (importance: 0.2481, prediction: 0.004), and relative grip strength (importance: 0.135, prediction: 0.76) (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(a)). The SHAP dependence plot in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(c) shows that SHAP values decreased as 20-m VO\u003csub\u003e2\u003c/sub\u003e max and relative grip strength increased; therefore, as these variables increase in value, their impact on model prediction decreases, thereby increasing the number of normal classifications. Low 20-m VO\u003csub\u003e2\u003c/sub\u003e max and relative grip strength resulted in the aggregation of red points (female) at the high end of the SHAP value range, indicating that for the same level of fitness, being female had a higher impact on the predictions than being male. Conversely, the predictions obtained under high 20-m VO\u003csub\u003e2\u003c/sub\u003e max and relative grip strength were less differentiated by gender. These results imply that for low 20-m VO\u003csub\u003e2\u003c/sub\u003e max (Spearman correlation r\u0026thinsp;=\u0026thinsp;\u0026minus;0.642, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.000) and relative grip strength (Spearman correlation r\u0026thinsp;=\u0026thinsp;\u0026minus;0.549, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.000), gender predicted more accurate obesity classification; however, this effect reduced when these two variables increased.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(b), LIME analysis reveals that the number of instances in which obesity is predicted is 2,608. The intercept and predicted local values (LIME value) obtained from LIME analysis were 0.511 and 0.668, respectively, which indicate obesity predictions. The confidence in the LIME prediction indicates that the prediction probability for the instance obtained from the original model corresponds to the obese class (R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.995). LIME feature importance analysis yielded two values, i.e., 20-m VO\u003csub\u003e2\u003c/sub\u003e max less than 0.30 (importance\u0026thinsp;=\u0026thinsp;0.544) and relative grip strength less than 0.36 (importance\u0026thinsp;=\u0026thinsp;0.171). The original 20-m VO\u003csub\u003e2\u003c/sub\u003e max and relative grip strength obtained from the Minmax scaler were 28.71 ml/kg/min and 38.17%, respectively. The strong positive contributions of these features indicate that low 20-m VO\u003csub\u003e2\u003c/sub\u003e max and relative grip strength significantly increased obesity classification.\u003c/p\u003e \u003cp\u003eThe interaction between these two variables is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(d) (Pearson correlation coefficient\u0026thinsp;=\u0026thinsp;0.727, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.000). Increases in 20-m VO\u003csub\u003e2\u003c/sub\u003e max decreased the SHAP values. This result indicates that a high 20-m VO\u003csub\u003e2\u003c/sub\u003e max corresponds to a low probability of obesity according to the predictions obtained from the model. Furthermore, relative grip strength interacted with 20-m VO\u003csub\u003e2\u003c/sub\u003e max and affected the SHAP values and, in turn, the obesity predictions. However, the increase in relative grip strength (toward the red color) decreased the SHAP values, reducing the probability of obesity classification.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eThis study used deep neural networks to evaluate the physical fitness factors that affect obesity predictions. This study is a follow-up to our previous study [47], utilizing the same analysis methodology and dataset. The results revealed that neural network analysis across five stratified folds yielded consistent performances with TN, FN, FP, TP, and ROC curves of over 0.95. The accuracy, precision, recall, and F1 scores were in the range of 0.87\u0026ndash;0.88% (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). The results obtained from Fold 4 were the best for determining the optimized neural network for obesity classification. The TN and FN percentages were balanced, with an ROC value of 0.95. Performance metrics, such as accuracy, precision, and recall, were approximately 0.87 (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). SHAP and LIME analyses clarified the influence of features such as 20-m VO\u003csub\u003e2\u003c/sub\u003e max, gender, and relative grip strength on obesity classifications (Figs.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(a) and (b)). Gender and relative grip strength were important factors influencing obesity prediction; however, the interaction of 20-m VO\u003csub\u003e2\u003c/sub\u003e max decreased the probability of obesity classification (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(c)). Therefore, 20-m VO\u003csub\u003e2\u003c/sub\u003e max (original value of \u0026lt;\u0026thinsp;28.71 ml/kg/min) was an important physical factor in predicting obesity.\u003c/p\u003e \u003cp\u003eA previous study proposed a deep neural network model using stratified five k-folds with good accuracy (71%) and explained that low physical fitness, which included the moderate effect-size correlations of aerobic fitness, upper limb strength, and sprint time, in adolescents is correlated with increased obesity risk [20]. Our model exhibited higher accuracy levels than that of the previously reported model; the average accuracy obtained from the stratified five K-folds was 87.7%. The 20-m VO\u003csub\u003e2\u003c/sub\u003e max greatly predicted the obesity classifications, similar to the observations reported in previous studies. The obesity classifications (obesity\u0026thinsp;=\u0026thinsp;1, normal\u0026thinsp;=\u0026thinsp;0, degree of prediction to towards 1) increased with the decrease in 20-m VO\u003csub\u003e2\u003c/sub\u003e max. These results are supported by the fact that high BMI is associated with relatively low O\u003csub\u003e2\u003c/sub\u003e uptake [48].\u003c/p\u003e \u003cp\u003eOur results showed that 20-m VO\u003csub\u003e2\u003c/sub\u003e max is the highest feature importance variable for predicting obesity (Figs.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e-(a) and (b)). The average deep learning modeling error between aerobic exercise and obesity reduction obtained in a previous study was 0.053%, whereas the average performance accuracy error (%) was approximately 0.186% [49]. Moreover, neural network models provided better obesity classifications than the logistic regression model [50]. Both the previous studies support the results obtained from the deep neural network analysis in this study, i.e., a reduced 20-m VO\u003csub\u003e2\u003c/sub\u003e max increased the probability of obesity prediction.\u003c/p\u003e \u003cp\u003eOur SHAP analysis showed that gender was the second most important factor for predicting obesity (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(a)). The results indicate that increased 20-m VO\u003csub\u003e2\u003c/sub\u003e max in males decreased the corresponding SHAP importance values (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(c)). This result may correspond to males who attempt to lose weight and, therefore, may increase exercise and reduce fat intake. By contrast, females are likely to participate in weight-loss programs, administer prescription diet pills, follow special diets, and eat more vegetables and fruits [50]. Males who attempt weight loss are more likely to lose weight successfully than females [51]. Our results are supported by males having high aerobic capacity (VO\u003csub\u003e2\u003c/sub\u003e max /lean body mass (kg)) and BMI (\u0026gt;\u0026thinsp;1.37 kg/m\u003csup\u003e2\u003c/sup\u003e) and low VO\u003csub\u003e2\u003c/sub\u003e max/lean body mass by 1 ml/kg/min [52].\u003c/p\u003e \u003cp\u003eIn addition, the results of SHAP and LIME analyses showed that relative grip strength was an important factor in predicting obesity (Figs.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(a) and (b)) in females (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(c)). However, Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(d) indicates that a decrease in relative grip strength decreases 20-m VO\u003csub\u003e2\u003c/sub\u003e max and, in turn, the SHAP value, thereby yielding a positive prediction of obesity. Therefore, this result indicates that 20-m VO\u003csub\u003e2\u003c/sub\u003e max was a more important factor in predicting obesity than relative grip strength (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(b)). The results of this study indicate that females (BMI\u0026thinsp;\u0026gt;\u0026thinsp;30 kg/m\u003csup\u003e2\u003c/sup\u003e, aged 35\u0026ndash;45 years) should engage in aerobic exercises, which are more beneficial for preventing cardiovascular diseases due to obesity than resistance exercise training [53]. In addition, engaging in aerobic exercises can effectively decrease obesity risk [54\u0026ndash;56].\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eA predictive neural network model was proposed, and data obtained from the Korean NFA program (2010\u0026ndash;2023) were analyzed. The model demonstrated high reliability in terms of obesity classification and included significant physical fitness factors such as 20-m VO\u003csub\u003e2\u003c/sub\u003e max (with values\u0026thinsp;\u0026lt;\u0026thinsp;28.71 ml/kg/min indicating high obesity). The consistency of the model across various metrics (accuracy, precision, recall, and F1 scores) reaffirmed that aerobic capacity can be used for obesity-risk assessment. This study also highlighted the potential gender-specific interventions to reduce the risk of obesity. Additionally, this study highlighted the potential of using neural network models to devise public health strategies for early obesity detection and personalized treatment plans. The limitations of this study include not incorporating dietary intake and physical activity levels, potentially overlooking crucial obesity influencers, and limiting the applicability of the model. Overfitting owing to the use of various sequential neural network architectures may restrict the generalizability and real-world prediction accuracy of the model. Additionally, the adoption of a general obesity classification threshold may be unsuitable for several populations, especially those of Asian countries, owing to different body compositions and risk profiles. Finally, relying on indirect VO\u003csub\u003e2\u003c/sub\u003e max estimates from the 20-m shuttle run test, instead of direct measurements, may reduce the precision of the predictions related to physical fitness factors, thereby affecting the overall robustness and applicability of the model when attempting to accurately predict obesity.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eData Availability Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll data generated during the current study have been included in this published article and its original dataset and coding file are available from the corresponding and first author (Jun-Hyun Bae \u0026amp; Wook Song) on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics Approval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eApproval was obtained by the Research Ethics Committee of Hyupsung University (approval no. 7002320-202303-HR-001). The procedures used in this study adhere to the tenets of the Declaration of Helsinki. The studies were conducted in accordance with local legislation and institutional requirements. Written informed consent for participants was not required from the participants, their legal guardians, or next of kin in accordance with the national legislation and institutional requirements. The informed consent was waived.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eJ-HB and YHS contributed to data collection, analysis, and writing of the manuscript. J-HB, YHS, and WS were involved in data collection and reviewing the manuscript. All the authors contributed to the manuscript and approved the submitted version.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the Basic Science Research Program of the National Research Foundation of Korea (NRF) funded by the Ministry of Education (RS-2023-00247075) and Kyungil University research fund.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNone\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests \u0026amp; Disclosure of Potential Conflicts of Interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that this study was conducted in the absence of any commercial or financial relationships that could be construed as potential conflicts of interest.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eHong, I., et al., \u003cem\u003eRelationship Between Physical Activity and Overweight and Obesity in Children: Findings From the 2012 National Health and Nutrition Examination Survey National Youth Fitness Survey.\u003c/em\u003e Am J Occup Ther, 2016. \u003cstrong\u003e70\u003c/strong\u003e(5): p. 7005180060p1-8.\u003c/li\u003e\n\u003cli\u003eYang, Y.S., et al., \u003cem\u003eObesity Fact Sheet in Korea, 2021: Trends in Obesity Prevalence and Obesity-Related Comorbidity Incidence Stratified by Age from 2009 to 2019.\u003c/em\u003e J Obes Metab Syndr, 2022. \u003cstrong\u003e31\u003c/strong\u003e(2): p. 169-177.\u003c/li\u003e\n\u003cli\u003ePoirier, P., et al., \u003cem\u003eObesity and cardiovascular disease: pathophysiology, evaluation, and effect of weight loss: an update of the 1997 American Heart Association Scientific Statement on Obesity and Heart Disease from the Obesity Committee of the Council on Nutrition, Physical Activity, and Metabolism.\u003c/em\u003e Circulation, 2006. \u003cstrong\u003e113\u003c/strong\u003e(6): p. 898-918.\u003c/li\u003e\n\u003cli\u003eKim, D.-S. and P.E. Scherer, \u003cem\u003eObesity, diabetes, and increased cancer progression.\u003c/em\u003e Diabetes \u0026amp; metabolism journal, 2021. \u003cstrong\u003e45\u003c/strong\u003e(6): p. 799-812.\u003c/li\u003e\n\u003cli\u003eNyberg, S.T., et al., \u003cem\u003eObesity and loss of disease-free years owing to major non-communicable diseases: a multicohort study.\u003c/em\u003e The lancet Public health, 2018. \u003cstrong\u003e3\u003c/strong\u003e(10): p. e490-e497.\u003c/li\u003e\n\u003cli\u003eKyr\u0026ouml;l\u0026auml;inen, H., et al., \u003cem\u003ePhysical Fitness Profiles of Young Men.\u003c/em\u003e Sports Medicine, 2010. \u003cstrong\u003e40\u003c/strong\u003e(11): p. 907-920.\u003c/li\u003e\n\u003cli\u003eKim, H.J., et al., \u003cem\u003eRelationships of physical fitness and obesity with metabolic risk factors in children and adolescents: Chungju city cohort study.\u003c/em\u003e Ann Pediatr Endocrinol Metab, 2016. \u003cstrong\u003e21\u003c/strong\u003e(1): p. 31-8.\u003c/li\u003e\n\u003cli\u003eKim, J.-W., et al., \u003cem\u003eAssociation between obesity and various parameters of physical fitness in Korean students.\u003c/em\u003e Obesity Research \u0026amp; Clinical Practice, 2013. \u003cstrong\u003e7\u003c/strong\u003e(1): p. e67-e74.\u003c/li\u003e\n\u003cli\u003ePojskic, H. and B. Eslami, \u003cem\u003eRelationship Between Obesity, Physical Activity, and Cardiorespiratory Fitness Levels in Children and Adolescents in Bosnia and Herzegovina: An Analysis of Gender Differences.\u003c/em\u003e Frontiers in Physiology, 2018. \u003cstrong\u003e9\u003c/strong\u003e.\u003c/li\u003e\n\u003cli\u003eCastro, E.A., et al., \u003cem\u003eWhat is the most effective exercise protocol to improve cardiovascular fitness in overweight and obese subjects?\u003c/em\u003e Journal of Sport and Health Science, 2017. \u003cstrong\u003e6\u003c/strong\u003e(4): p. 454-461.\u003c/li\u003e\n\u003cli\u003eZhou, N., \u003cem\u003eAssessment of aerobic exercise capacity in obesity, which expression of oxygen uptake is the best?\u003c/em\u003e Sports Medicine and Health Science, 2021. \u003cstrong\u003e3\u003c/strong\u003e(3): p. 138-147.\u003c/li\u003e\n\u003cli\u003eGaesser, G.A. and S.S. Angadi, \u003cem\u003eObesity treatment: Weight loss versus increasing fitness and physical activity for reducing health risks.\u003c/em\u003e iScience, 2021. \u003cstrong\u003e24\u003c/strong\u003e(10): p. 102995.\u003c/li\u003e\n\u003cli\u003eLang, J., et al., \u003cem\u003eSystematic review of the relationship between 20 m shuttle run performance and health indicators among children and youth.\u003c/em\u003e Journal of Science and Medicine in Sport, 2017. \u003cstrong\u003e21\u003c/strong\u003e.\u003c/li\u003e\n\u003cli\u003eRey, O., et al., \u003cem\u003ePsycho-Physiological Responses of Obese Adolescents to an Intermittent Run Test Compared with a 20-M Shuttle Run.\u003c/em\u003e J Sports Sci Med, 2016. \u003cstrong\u003e15\u003c/strong\u003e(3): p. 451-459.\u003c/li\u003e\n\u003cli\u003eMoran, C.A., et al., \u003cem\u003ePerformance and reproducibility on shuttle run test between obese and non-obese children: a cross-sectional study.\u003c/em\u003e BMC Pediatrics, 2017. \u003cstrong\u003e17\u003c/strong\u003e(1): p. 68.\u003c/li\u003e\n\u003cli\u003ePrzednowek, K., et al., \u003cem\u003ePredictive modeling of VO2max based on 20 m shuttle run test for young healthy people.\u003c/em\u003e Applied Sciences, 2018. \u003cstrong\u003e8\u003c/strong\u003e(11): p. 2213.\u003c/li\u003e\n\u003cli\u003eChung, J.W., O. Lee, and K.H. Lee, \u003cem\u003eEstimation of maximal oxygen consumption using the 20m shuttle run test in Korean adults aged 19-64 years.\u003c/em\u003e Science \u0026amp; Sports, 2023. \u003cstrong\u003e38\u003c/strong\u003e(1): p. 68-74.\u003c/li\u003e\n\u003cli\u003eDeGregory, K.W., et al., \u003cem\u003eA review of machine learning in obesity.\u003c/em\u003e Obesity Reviews, 2018. \u003cstrong\u003e19\u003c/strong\u003e(5): p. 668-685.\u003c/li\u003e\n\u003cli\u003eFerdowsy, F., et al., \u003cem\u003eA machine learning approach for obesity risk prediction.\u003c/em\u003e Current Research in Behavioral Sciences, 2021. \u003cstrong\u003e2\u003c/strong\u003e: p. 100053.\u003c/li\u003e\n\u003cli\u003eForte, P., et al., \u003cem\u003eA Deep Learning Neural Network to Classify Obesity Risk in Portuguese Adolescents Based on Physical Fitness Levels and Body Mass Index Percentiles: Insights for National Health Policies.\u003c/em\u003e Behavioral Sciences, 2023. \u003cstrong\u003e13\u003c/strong\u003e(7): p. 522.\u003c/li\u003e\n\u003cli\u003eCheng, X., et al., \u003cem\u003eDoes Physical Activity Predict Obesity-A Machine Learning and Statistical Method-Based Analysis.\u003c/em\u003e Int J Environ Res Public Health, 2021. \u003cstrong\u003e18\u003c/strong\u003e(8).\u003c/li\u003e\n\u003cli\u003eThamrin, S.A., et al., \u003cem\u003ePredicting Obesity in Adults Using Machine Learning Techniques: An Analysis of Indonesian Basic Health Research 2018.\u003c/em\u003e Frontiers in Nutrition, 2021. \u003cstrong\u003e8\u003c/strong\u003e.\u003c/li\u003e\n\u003cli\u003eJossa-Bastidas, O., et al., \u003cem\u003ePredicting Physical Exercise Adherence in Fitness Apps Using a Deep Learning Approach.\u003c/em\u003e Int J Environ Res Public Health, 2021. \u003cstrong\u003e18\u003c/strong\u003e(20).\u003c/li\u003e\n\u003cli\u003eLee, S.-H., et al., \u003cem\u003eEstimation of health-related physical fitness (hrpf) levels of the general public using artificial neural network with the national fitness award (nfa) datasets.\u003c/em\u003e International journal of environmental research and public health, 2021. \u003cstrong\u003e18\u003c/strong\u003e(19): p. 10391.\u003c/li\u003e\n\u003cli\u003eKo, B.-g., et al., \u003cem\u003ePrediction equations of physical fitness age for Korean adults.\u003c/em\u003e Exercise Science, 2021. \u003cstrong\u003e30\u003c/strong\u003e(3): p. 352-360.\u003c/li\u003e\n\u003cli\u003ePark, S., et al., \u003cem\u003eNormative Reference Values of Physical Fitness Levels in Koreans: Results from the National Fitness Award Project (2017-2019).\u003c/em\u003e Exercise Science, 2022. \u003cstrong\u003e31\u003c/strong\u003e(4): p. 511-526.\u003c/li\u003e\n\u003cli\u003eJung, J., et al., \u003cem\u003eIm S. Model development of fitness cer-tification center\u003c/em\u003e. 2014, Research report, Ministry of Culture, Sports and Tourism.\u003c/li\u003e\n\u003cli\u003eBae, J.-H., et al., \u003cem\u003ePrediction models of grip strength in adults above 65 years using Korean National Physical Fitness Award Data from 2009 to 2019.\u003c/em\u003e European Geriatric Medicine, 2023. \u003cstrong\u003e14\u003c/strong\u003e(5): p. 1059-1064.\u003c/li\u003e\n\u003cli\u003eWeir, C.B. and A. Jan, \u003cem\u003eBMI classification percentile and cut off points.\u003c/em\u003e 2019.\u003c/li\u003e\n\u003cli\u003eKalnins, A. and K. Praitis Hill, \u003cem\u003eThe VIF Score. What is it Good For? Absolutely Nothing.\u003c/em\u003e Organizational research methods, 2023: p. 10944281231216381.\u003c/li\u003e\n\u003cli\u003eKrishnamoorthy, K. and T. Mathew, \u003cem\u003eStatistical tolerance regions: theory, applications, and computation\u003c/em\u003e. 2009: John Wiley \u0026amp; Sons.\u003c/li\u003e\n\u003cli\u003eMaqsood, S. and R. Dama\u0026scaron;evičius, \u003cem\u003eMulticlass skin lesion localization and classification using deep learning based features fusion and selection framework for smart healthcare.\u003c/em\u003e Neural Networks, 2023. \u003cstrong\u003e160\u003c/strong\u003e: p. 238-258.\u003c/li\u003e\n\u003cli\u003eYan, Y., et al. \u003cem\u003eDeep learning for imbalanced multimedia data classification\u003c/em\u003e. in \u003cem\u003e2015 IEEE international symposium on multimedia (ISM)\u003c/em\u003e. 2015. IEEE.\u003c/li\u003e\n\u003cli\u003eRefaeilzadeh, P., L. Tang, and H. Liu, \u003cem\u003eCross-validation.\u003c/em\u003e Encyclopedia of database systems, 2009: p. 532-538.\u003c/li\u003e\n\u003cli\u003ePeng, D., et al., \u003cem\u003eAddressing the multi-label imbalance for neural networks: An approach based on stratified mini-batches.\u003c/em\u003e Neurocomputing, 2021. \u003cstrong\u003e435\u003c/strong\u003e: p. 91-102.\u003c/li\u003e\n\u003cli\u003ePi\u0026ntilde;eyro, L., A. Pardo, and M. Viera. \u003cem\u003eStructure verification of deep neural networks at compilation time using dependent types\u003c/em\u003e. in \u003cem\u003eProceedings of the XXIII Brazilian Symposium on Programming Languages\u003c/em\u003e. 2019.\u003c/li\u003e\n\u003cli\u003eNehra, N., P. Sangwan, and D. Kumar, \u003cem\u003eArtificial Neural Networks: A Comprehensive Review.\u003c/em\u003e Handbook of Machine Learning for Computational Optimization, 2021: p. 203-227.\u003c/li\u003e\n\u003cli\u003eSabiri, B., B. El Asri, and M. Rhanoui. \u003cem\u003eMechanism of Overfitting Avoidance Techniques for Training Deep Neural Networks\u003c/em\u003e. in \u003cem\u003eICEIS (1)\u003c/em\u003e. 2022.\u003c/li\u003e\n\u003cli\u003eBae, J.-H., J.-w. Seo, and D.Y. Kim, \u003cem\u003eDeep-learning model for predicting physical fitness in possible sarcopenia: analysis of the Korean physical fitness award from 2010 to 2023.\u003c/em\u003e Frontiers in Public Health, 2023. \u003cstrong\u003e11\u003c/strong\u003e.\u003c/li\u003e\n\u003cli\u003eLee, S., et al., \u003cem\u003eBackground information of deep learning for structural engineering.\u003c/em\u003e Archives of Computational Methods in Engineering, 2018. \u003cstrong\u003e25\u003c/strong\u003e: p. 121-129.\u003c/li\u003e\n\u003cli\u003eLi, Y. and Z. Chen, \u003cem\u003ePerformance evaluation of machine learning methods for breast cancer prediction.\u003c/em\u003e Appl Comput Math, 2018. \u003cstrong\u003e7\u003c/strong\u003e(4): p. 212-216.\u003c/li\u003e\n\u003cli\u003ePham, B.T., et al., \u003cem\u003ePerformance evaluation of machine learning methods for forest fire modeling and prediction.\u003c/em\u003e Symmetry, 2020. \u003cstrong\u003e12\u003c/strong\u003e(6): p. 1022.\u003c/li\u003e\n\u003cli\u003eLuque, A., et al., \u003cem\u003eThe impact of class imbalance in classification performance metrics based on the binary confusion matrix.\u003c/em\u003e Pattern Recognition, 2019. \u003cstrong\u003e91\u003c/strong\u003e: p. 216-231.\u003c/li\u003e\n\u003cli\u003eKhan, S.A. and Z.A. Rana. \u003cem\u003eEvaluating performance of software defect prediction models using area under precision-Recall curve (AUC-PR)\u003c/em\u003e. in \u003cem\u003e2019 2nd International Conference on Advancements in Computational Sciences (ICACS)\u003c/em\u003e. 2019. IEEE.\u003c/li\u003e\n\u003cli\u003eMosca, E., et al. \u003cem\u003eSHAP-based explanation methods: a review for NLP interpretability\u003c/em\u003e. in \u003cem\u003eProceedings of the 29th International Conference on Computational Linguistics\u003c/em\u003e. 2022.\u003c/li\u003e\n\u003cli\u003eKumarakulasinghe, N.B., et al. \u003cem\u003eEvaluating local interpretable model-agnostic explanations on clinical machine learning classification models\u003c/em\u003e. in \u003cem\u003e2020 IEEE 33rd International Symposium on Computer-Based Medical Systems (CBMS)\u003c/em\u003e. 2020. IEEE.\u003c/li\u003e\n\u003cli\u003eBae, J.-H., et al., \u003cem\u003eNeural network model for prediction of possible sarcopenic obesity using Korean national fitness award data (2010\u0026ndash;2023).\u003c/em\u003e Scientific Reports, 2024. \u003cstrong\u003e14\u003c/strong\u003e(1): p. 14565.\u003c/li\u003e\n\u003cli\u003eUmamaheswari, K., et al., \u003cem\u003eVO2 Max and Body Mass in Overweight and Obese Young Adults.\u003c/em\u003e International Journal of Physiology, 2017. \u003cstrong\u003e5\u003c/strong\u003e(2): p. 23-27.\u003c/li\u003e\n\u003cli\u003eMu, P., \u003cem\u003eModeling Analysis of the Relationship between Adolescent Aerobic Exercise and Obesity Reduction Based on Deep Learning.\u003c/em\u003e Advances in Multimedia, 2022. \u003cstrong\u003e2022\u003c/strong\u003e: p. 4112169.\u003c/li\u003e\n\u003cli\u003eErg\u0026uuml;n, U., \u003cem\u003eThe classification of obesity disease in logistic regression and neural network methods.\u003c/em\u003e J Med Syst, 2009. \u003cstrong\u003e33\u003c/strong\u003e(1): p. 67-72.\u003c/li\u003e\n\u003cli\u003eTsai, S.A., et al., \u003cem\u003eGender Differences in Weight-Related Attitudes and Behaviors Among Overweight and Obese Adults in the United States.\u003c/em\u003e American Journal of Men\u0026apos;s Health, 2016. \u003cstrong\u003e10\u003c/strong\u003e(5): p. 389-398.\u003c/li\u003e\n\u003cli\u003eSharma, H.B. and J. Kailashiya, \u003cem\u003eGender Difference in Aerobic Capacity and the Contribution by Body Composition and Haemoglobin Concentration: A Study in Young Indian National Hockey Players.\u003c/em\u003e J Clin Diagn Res, 2016. \u003cstrong\u003e10\u003c/strong\u003e(11): p. Cc09-cc13.\u003c/li\u003e\n\u003cli\u003eChaudhary, S., M.K. Kang, and J.S. Sandhu, \u003cem\u003eThe effects of aerobic versus resistance training on cardiovascular fitness in obese sedentary females.\u003c/em\u003e Asian J Sports Med, 2010. \u003cstrong\u003e1\u003c/strong\u003e(4): p. 177-84.\u003c/li\u003e\n\u003cli\u003eBellicha, A., et al., \u003cem\u003eEffect of exercise training on weight loss, body composition changes, and weight maintenance in adults with overweight or obesity: An overview of 12 systematic reviews and 149 studies.\u003c/em\u003e Obesity Reviews, 2021. \u003cstrong\u003e22\u003c/strong\u003e: p. e13256.\u003c/li\u003e\n\u003cli\u003eSkrypnik, D., et al., \u003cem\u003eEffects of endurance and endurance strength training on body composition and physical capacity in women with abdominal obesity.\u003c/em\u003e Obesity facts, 2015. \u003cstrong\u003e8\u003c/strong\u003e(3): p. 175-187.\u003c/li\u003e\n\u003cli\u003eJamka, M., et al., \u003cem\u003eComparison of the Effect of Endurance, Strength and Endurance-Strength Training on Glucose and Insulin Homeostasis and the Lipid Profile of Overweight and Obese Subjects: A Systematic Review and Meta-Analysis.\u003c/em\u003e International Journal of Environmental Research and Public Health, 2022. \u003cstrong\u003e19\u003c/strong\u003e(22): p. 14928.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Obesity, Physical Fitness, Hyperparameter, Deep Learning, 20-m Shuttle Run","lastPublishedDoi":"10.21203/rs.3.rs-4782187/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4782187/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eObesity, a \"global syndemic,\" increases the risk of noncommunicable diseases; therefore, the prediction and management of obesity is crucial. Regular physical activity and cardiorespiratory fitness are inversely correlated with obesity, highlighting the need for effective models for predicting obesity.\u003c/p\u003e\u003ch2\u003eAim\u003c/h2\u003e \u003cp\u003eThis study aimed to predict obesity using physical fitness factors, including those related to cardiorespiratory fitness, determined via deep neural network analysis of data obtained from the 2010\u0026ndash;2023 Korean National Physical Fitness Award.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eA deep learning approach was implemented to analyze the data obtained from 108,304 participants, and variables such as exercise-induced oxygen consumption during a 20-m shuttle run test (20-m VO\u003csub\u003e2\u003c/sub\u003e max), gender, and relative grip strength were considered. Stratified K-fold cross-validation, SHapley Additive exPlanations (SHAP), and local interpretable model-agnostic description analyses were employed to evaluate the model performance and feature importance.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe neural network yielded a high accuracy score (0.87\u0026ndash;0.88), with Fold 4 providing the optimized model for obesity classifications. Features such as 20-m VO\u003csub\u003e2\u003c/sub\u003e max, gender, and relative grip strength significantly influenced the obesity predictions, and low 20-m VO\u003csub\u003e2\u003c/sub\u003e max levels were key predictors of obesity.\u003c/p\u003e\u003ch2\u003eDiscussion\u003c/h2\u003e \u003cp\u003eThis study confirmed the efficacy of the proposed deep neural network in predicting obesity based on physical fitness factors and clarified the significant predictors of obesity.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eThe results of this study may potentially be used for devising personalized obesity-management strategies that emphasize the importance of cardiorespiratory fitness.\u003c/p\u003e","manuscriptTitle":"Sequential Deep Learning Model for Obesity Prediction Based on Physical Fitness Factors: An Analysis of Data from the 2010–2023 Korean National Physical Fitness Data ","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-08-23 21:27:51","doi":"10.21203/rs.3.rs-4782187/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"65150217-d0da-498f-b023-0bd6771b8a33","owner":[],"postedDate":"August 23rd, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":35962648,"name":"Health sciences/Health care/Weight management"},{"id":35962649,"name":"Health sciences/Risk factors"}],"tags":[],"updatedAt":"2024-11-05T02:53:47+00:00","versionOfRecord":[],"versionCreatedAt":"2024-08-23 21:27:51","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4782187","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4782187","identity":"rs-4782187","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}