Prediction of cerebral white matter lesions using a multilayer perceptron applied to routine clinical data

Introduction: The detection of white matter lesions is limited by restricted access to magnetic resonance imaging, highlighting the need for predictive models based on accessible clinical data. Objective: To develop and compare a predictive model for cerebral white matter lesions using a multilayer perceptron (MLP) neural network trained with routine clinical and biochemical variables. Materials and methods: Cross-sectional study based on a secondary database of 1904 participants. Demographic, clinical, and metabolic variables were included, such as age, blood pressure, glucose, lipid profile, and body mass index. The MLP model was configured with one hidden layer of seven neurons and hyperbolic tangent and softmax activation functions. Backpropagation with cross-entropy loss and a 70/30% train-test split was applied. Its performance was compared with binary logistic regression and Random Forest models. Results: The MLP achieved an area under the curve (AUC) of 0.789 and an overall accuracy of 70.3%, outperforming logistic regression (AUC = 0.709) and Random Forest (AUC = 0.719). Sensitivity and specificity were 73.8% and 73.6%, respectively. Interpretability analysis using SHAP values identified age, antihypertensive use, diastolic blood pressure, glucose, and LDL as the most influential predictors. Conclusions: The neural network model demonstrated superior performance and greater explanatory capacity compared with classical methods, emphasising the role of vascular and metabolic factors as key modulators of cerebral white matter lesion risk.