Diagnosing Schizophrenia with Routine Blood Tests: A Comparative Analysis of Machine Learning Algorithms

Yavuz Selim Ogur¹Ayse Erdogan Kaya²Nur Banu Oğur^3*Beyza Erdogan Akturk⁴

• ¹Serdivan State Hospital, Sakarya, Türkiye
• ²Hitit University Çorum Erol Olçok Training and Research Hospital, Çorum, Türkiye
• ³Faculty of Computer and Information Sciences, Sakarya University, Sakarya, Türkiye
• ⁴Tarsus Devlet Hastanesi, Mersin, Türkiye

Schizophrenia is a severe mental disorder with a prevalence of approximately 1% in the general population, diagnosed primarily based on clinical criteria. The absence of objective diagnostic methods and reliable biomarkers presents significant challenges for accurate diagnosis and effective treatment. Recently, peripheral blood biomarkers have gained attention for their potential in understanding schizophrenia's biological basis. Concurrently, machine learning algorithms are increasingly used to analyze high-dimensional data to improve diagnostic accuracy and predict treatment responses. This retrospective case-control study aimed to develop a practical, costeffective, and high-performance classification model differentiating schizophrenia patients from healthy individuals using routine hematological and biochemical parameters. Conducted at a tertiary care hospital, the study included 203 schizophrenia patients treated within five years and 192 age-and sex-matched healthy controls. Demographic data and routine blood parameters (complete blood count and biochemical panels) were extracted from medical records. Missing data exceeding 85% per variable were excluded, and remaining missing values were imputed after train-test splitting to prevent data leakage. Optimal biomarker subsets were identified using the Grey Wolf Optimization (GWO) algorithm. Subsequently, five machine learning models-Random Forest (RF), XGBoost, Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Logistic Regression (LR)-were trained and evaluated via stratified 10-fold cross-validation. Groups were homogeneous regarding age and sex distribution. Prior to GWO optimization, the highest accuracies were obtained by XGBoost (95.55%) and Random Forest (94.63%). Postoptimization, both models maintained robust performance: Random Forest improved accuracy (94.95%) and recall (96.25%), whereas XGBoost achieved the highest accuracy (95.90%) and strong specificity (95.54%). The highest Area Under the Curve (AUC) values post-optimization were achieved by XGBoost (0.96) and Random Forest (0.95), indicating strong diagnostic capability. Key biomarkers distinguishing schizophrenia included total protein, glucose, iron, creatine kinase, total bilirubin, uric acid, calcium, and sodium, with schizophrenia patients showing notably lower glucose levels compared to controls, contrary to typical findings. Differences in triglycerides, liver enzymes, sodium, and potassium lacked clear clinical significance. The developed models achieved diagnostic accuracy comparable to studies using more costly biomarkers, highlighting their potential clinical practicality and economic advantages. External validation is recommended to confirm the generalizability of these results.