A Comparative Analysis of Deep Learning Architectures for Obesity Classification Using Structured Data

Obesity is a significant global health concern, necessitating accurate and efficient diagnostic tools to classify individuals based on obesity levels. This study investigates the performance of five deep learning architectures: Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional LSTM (BiLSTM) in classifying obesity levels using structured data. The dataset comprises clinical, demographic, and lifestyle features, and is preprocessed through normalization, label encoding, and Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance. Each model was evaluated using accuracy, precision, recall, and F1-score metrics under stratified 10-fold cross-validation. The results indicate that MLP achieved the highest performance across all metrics, with an accuracy of 99.05%, followed closely by CNN at 98.77%. Sequential models, including LSTM, GRU, and BiLSTM, exhibited comparatively lower performance, achieving accuracies of 83.80%, 86.59%, and 86.78%, respectively. The superior performance of MLP and CNN underscores their suitability for structured datasets with static features, while the sequential models struggled due to the lack of temporal dependencies in the data. This study highlights the importance of aligning model architecture with dataset characteristics for optimal performance. The findings suggest that MLP and CNN are effective choices for obesity classification tasks, providing robust and computationally efficient solutions. Future work could explore hybrid models and incorporate temporal features to enhance the performance of sequential architecture.