A Clinical Risk Prediction Model for Depressive Disorders Based on Seven Machine Learning Algorithms

• Published in: International journal of general medicine Vol. 18; no. Issue 1; pp. 2461 - 2473
• Main Authors: Jin, Weifeng, Chen, Shuzi, Wang, Mengxia, Lin, Ping
• Format: Journal Article
• Language: English
• Published: New Zealand Dove Medical Press Limited 01.01.2025; Taylor & Francis Ltd; Dove; Dove Medical Press
• Subjects: Accuracy; Algorithms; Arginine; Biomarkers; Blood; Blood tests; Calibration; Clinical decision making; Data mining; Depressive disorders，Machine learn; Feature selection; Glucose; Health care; High density lipoprotein; Lipoproteins; Machine learning; Medical colleges; Medical examination; Medical research; Medicine, Experimental; Mental disorders; Mental health; Nomograms; Original Research; Phenylalanine; Phosphatase; Phosphatases; Regression analysis; Serotonin; Variables; machine learn; depressive disorders
• ISSN: 1178-7074; 1178-7074
• DOI: 10.2147/IJGM.S524016

To develop a clinical risk prediction model for depressive disorders using seven machine learning algorithms based on routine blood test indicators. A retrospective study was conducted, involving 284 patients with depressive disorders and 214 healthy controls recruited between January and October 2024. Clinical data, including age, sex, and routine blood test results, were collected. The dataset was randomly divided into a training set (70%; n=348) and a test set (30%; n=150). Univariate logistic regression analysis (p<0.1) was initially performed to identify potential predictors, followed by feature selection using the Boruta and LASSO algorithms. Seven machine learning algorithms were employed to construct predictive models, with their performance evaluated using metrics such as AUC, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), precision, recall, and F1 score. A multivariable logistic regression model was subsequently used to develop a nomogram, and its discrimination, calibration, and clinical utility were comprehensively assessed. Four significant predictors (alkaline phosphatase [AKP], serotonin, phenylalanine [Phe], and arginine [Arg]) were identified through univariate logistic regression combined with Boruta and LASSO feature selection. Among the seven algorithms, the random forest model exhibited the highest AUC, achieving an AUC of 1.000 (95% CI: 1.000-1.000) in the training set and 0.958 (95% CI: 0.931-0.985) in the test set. However, due to concerns about potential overfitting, the multivariable logistic regression model was selected as the final predictive model. A nomogram was constructed based on this model. This study successfully developed a clinically interpretable risk prediction model for depressive disorders by integrating machine learning algorithms and routine blood test indicators. The logistic regression model demonstrated robust performance across all metrics and holds potential as a reliable auxiliary tool for the diagnosis of depressive disorders.