Development and validation of a machine learning-based diagnostic model for depression in young populations: a retrospective cohort study

Depression in 10–26 years is often subtle but can greatly affect their physical and mental development. This study aims to develop a machine learning model for the diagnosis of depression. Data were collected from 600 patients aged 10–26 years between 2019 and 2025, including 350 non-depressed patients and 250 depressed patients. A gradient boosting machine (GBM) was used to build the model, shapley additive explanations (SHAP) plots employed for model interpretation and visualization. Based on Boruta feature selection, nine variables were identified for machine learning model construction: S100β, platelet(PLT), lactate dehydrogenase(LDH), white blood cell(WBC), creatinine kinase-mb(CK-Mb), gender, and LDH/PLT Ratio, LDH/WBC Ratio, LDH/HGB Ratio. The GBM showed the best performance with a threshold of 0.591, Accuracy of 0.878, Specificity of 0.941, and Precision of 0.913. It also demonstrated high levels of Sensitivity (0.797) and F1 Score (0.851). The decision curve analysis and calibration curve assessed the clinical validity and consistency of the model, indicating that the GBM model outperformed others. The most significant features identified were S100β, PLT, and LDH/PLT. By utilizing machine learning, key factors such as S100β, PLT, and LDH/PLT were identified, providing preliminary evidence for the diagnosis of depression in patients aged 10–26 years.