DeepHealthNet: Adolescent Obesity Prediction System Based on a Deep Learning Framework

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Aug 31 2023

In an article recently submitted to the ArXiv* server, researchers addressed global concerns regarding childhood and adolescent obesity by utilizing artificial intelligence (AI) technology to achieve accurate prediction and deliver personalized feedback. This approach used a deep learning model, DeepHealthNet, that was trained with limited health data and achieved a higher prediction accuracy. Many gender-specific variations were also identified, with boys reaching an accuracy of 0.9320 and girls at 0.9163. This innovative system demonstrated promise in addressing obesity and promoting improved health outcomes.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information.

Background

The surge in childhood and adolescent obesity rates on a global scale has sparked considerable alarm, mainly because of the linked chronic diseases and enduring health hazards. To address this issue, AI technology is being leveraged as a potent tool for precisely forecasting obesity rates and furnishing tailored feedback to adolescents that aims to combat this pressing health challenge effectively. The intersection of AI and public health is paving the way for innovative interventions in obesity management.

Proposed method

In the present paper, a fundamental aspect of the methodology revolves around meticulously evaluating the predictive models. The significance of obtaining reliable performance estimates cannot be overstated, as it underpins the credibility of the findings. To achieve this, the 10-fold cross-validation approach was employed. This is a widely recognized technique in machine learning evaluation.

The crux of the 10-fold cross-validation lies in its ability to comprehensively assess the performance of the model and its potential to generalize to unseen data. This process involves systematically partitioning the dataset into ten subsets. During each iteration, the models are trained on 90% of the data, and the remaining 10% are held aside for rigorous testing. This rigorous testing set allows us to gauge how well the models can predict previously unseen instances.

The distinctive feature of the 10-fold cross-validation lies in its iterative utilization across 10 separate data partitions. This repetitive process provides a robust understanding of how the models perform on various dataset segments. The resulting evaluation metrics are then averaged over these 10 folds, providing a more stable estimate of performance and reducing the impact of data variability.

One of the central advantages of this approach is its ability to mitigate the risk of overfitting. The performance is not skewed by idiosyncrasies present in a single partition. It is achieved by validating the models on multiple distinct subsets of the data. Consequently, the evaluation becomes more immune to the noise and fluctuations inherent in any dataset.

A clearer and more accurate representation of the capabilities of the predictive models is achieved through this rigorous 10-fold cross-validation process. This methodology significantly bolsters the reliability and generalizability of the findings, and also the robustness of the proposed approach in predicting obesity rates among adolescents.

Experimental results

The outcomes displayed the impressive accomplishments of the proposed deep learning framework. It achieved an average accuracy of 0.8837, surpassing all the compared models. The F1-score, which gauges the equilibrium between precision and recall, also demonstrated a remarkable value of 0.8797. The recall value, indicating the ratio of true positives detected, was 0.8958, while the precision value, representing the accuracy of positive predictions, matched the overall accuracy at 0.8837.

Conversely, the LSTM model exhibited a relatively higher accuracy of 0.7008, and the support vector machine (SVM) classifier from traditional machine learning (ML) models achieved a respectable accuracy of 0.6846. However, the Naive Bayes (NB) model exhibited the lowest accuracy of 0.3727, similar to random chance. Consequently, DeepHealthNet displayed superior performance across accuracy, F1-score, recall, and precision compared to all models. The standard deviation remained consistently below 0.02 for all models across various folds, indicating stable training without significant variance.

Moreover, a statistical analysis indicated significant differences between the proposed and compared models concerning performance metrics. The obtained p-values signified substantial statistical distinctions, underscoring the superior performance of the proposed framework in accuracy, F1 score, recall, and precision. Notably, even when assessing performance separately for boys and girls, the proposed model's statistical significance and higher performance were maintained. Convergence analysis revealed that the model achieved stability around 100 epochs for training accuracy and 150 epochs for loss, showcasing effective learning and dependable predictions.

The study demonstrated the proposed deep learning framework's strong ability to predict adolescent obesity levels, outperforming other models in various metrics. It achieved a high average accuracy of 0.8837, with excellent F1-score, recall, and precision values. Statistical analysis confirmed its significant superiority. The model's efficacy was consistent across genders, and its performance improved gradually over time, peaking at the 162nd day session.

Future Work

Future research should focus on enhancing the generalizability of the model by testing it on diverse datasets. The prediction accuracy can be enhanced by incorporating additional influencing factors such as socioeconomic status and genetic predisposition. A longitudinal analysis could provide insights into the performance of the model over extended timeframes. Additionally, efforts should be directed toward streamlining the implementation of the model in real-world settings, which can make it more accessible and applicable in practical scenarios.

Conclusion

To summarize, the study presented a deep-learning framework designed to predict obesity levels in adolescents. The obtained results highlighted the remarkable performance of the model compared to alternative methods. It showed elevated accuracy, F1-score, recall, and precision metrics. The significance of the model was confirmed through statistical analysis. Importantly, its effectiveness extended consistently across different genders by strengthening its reliability. Visualizations also indicated the competence of the model in delivering dependable long-term predictions and surpassing competing approaches.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information.