Obesity prediction using an explainable deep learning framework based on LSTM-LIME with integrated visualization

Obesity is a major global health challenge requiring accurate and interpretable risk-assessment models to support early detection and prevention strategies. This study introduces a novel explainable deep learning framework for multiclass obesity prediction using a Saudi-specific dataset that integrates anthropometric, lifestyle, and dietary factors. Six models were evaluated including Long Short-Term Memory (LSTM), Bidirectional LSTM, Recurrent Neural Network (RNN), Deep Neural Network (DNN) specifically Multilayer Perceptron (MLP), TabNet, and Autoencoder. The Bi-LSTM model, with 96% accuracy, a macro recall of 0.96, a macro F1-score of 0.95, surpassed the other models in terms of predictive performance. Regression-style metrics such as the Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the Coefficient of Determination ([Formula: see text]) were also applied to assess ordinal misclassification and model calibration. The novelty of this work lies in the development of the first culturally specific Saudi multiclass obesity dataset and the integration of LSTM networks with Local Interpretable Model-Agnostic Explanations (LIME) within an interactive interface, enabling both predictive accuracy and transparent, user-centered visualization of obesity-risk factors. This approach advances current practice by combining explainable deep learning with region-specific health data for precision public-health applications.