A Transfer Learning Protocol for Hybrid Deep–Machine Learning Models to Predict and Interpret Weight Loss

Abstract

The objective of this study is to develop a three-stage transfer learning protocol for a hybrid CNN–Attention–XGBoost architecture to improve the accuracy and interpretability of weight-loss prediction in Iranian clinical data. The proposed model consisted of a convolutional neural network, an attention mechanism, and an XGBoost regressor. The CNN–Attention module was first pre-trained on an international dataset to learn general temporal weight-change patterns. The learned parameters were then fine-tuned on an Iranian dataset to adapt the model to local characteristics. Finally, the extracted features were passed to an XGBoost regressor for interpretable prediction. Model performance was evaluated using MAE, RMSE, and R², and statistical tests were applied to compare model effectiveness. The hybrid transfer-learning model showed significantly higher accuracy and lower error compared with the non-transfer baseline model. Temporal weight patterns and deep learned features contributed most strongly to predictive performance. Statistical testing confirmed that the improvement in accuracy was significant, indicating that transferred knowledge effectively enhanced model generalization. The proposed transfer learning protocol successfully enhanced both predictive performance and interpretability. Integrating deep feature extraction with an explainable decision layer provides a powerful and reliable approach for developing AI-based health systems that require both accuracy and transparency.