<p>Diabetic nephropathy is one of the most important complications of type 1 and type 2 diabetes, and its early diagnosis is difficult due to the high dimensionality of medical data and the low transparency of current models. Artificial intelligence-based predictive approaches can play a crucial role in early intervention and reducing the risk of kidney failure; however, they often face challenges such as overfitting, the selection of uninterpretable features, and a lack of quantitative indicators for Explainability. In this study, an explainable hybrid model named DNP-MTC is presented, in which the Trees Social Relations algorithm is used to select clinically meaningful features from the original data, and these features are then transferred to a multilayer perceptron. To achieve the optimal architecture and parameters of the network, the Coral Reefs Optimization algorithm is used. Using a dataset of over 101,000 samples and 50 features from 10 years of clinical data from 130 US hospitals, the model achieved 95% accuracy, 95% Recall, 96% F-Measure, and 98% precision. This framework offers an accurate, efficient, and interpretable approach to predicting diabetic nephropathy. By reducing the risk of overfitting and utilizing selected features relevant to clinical realities, it can serve as a practical tool in medical decision support systems. Additionally, in terms of Explainability, this study introduces a novel metric called the Explainable Feature Selection Metric (XFSM), which measures the importance of features selected by the model compared to the assessment of clinical experts. This approach enhances the transparency and reliability of predictions for use in sensitive medical settings, such as the early diagnosis of diabetic nephropathy.</p>