Transcranial Sonography and Machine Learning for Schizophrenia Identification (TCS-ML-SZ)

This is a prospective observational case-control study designed to develop and evaluate a machine-learning model for identifying schizophrenia using transcranial sonography (TCS) structural imaging features and clinical information.

Schizophrenia is clinically heterogeneous, and diagnosis depends mainly on clinical symptoms and psychiatric assessment. TCS is a non-invasive imaging method that can assess selected deep brain structures through the temporal acoustic window. Previous studies suggest that ultrasound features of structures such as the substantia nigra, raphe nuclei, and third ventricle may be related to neuropsychiatric disorders. This study will investigate whether TCS-derived structural imaging features, combined with clinical variables, can support auxiliary identification of schizophrenia.

Adults aged 18 to 65 years with schizophrenia diagnosed according to ICD-10 criteria and matched adults without a personal or family history of psychiatric disorders will be enrolled. The planned enrollment is 200 participants, including approximately 100 participants with schizophrenia and 100 healthy controls. Participants will undergo baseline TCS assessment and clinical data collection. No therapeutic intervention will be assigned by the investigators, and participation will not replace or alter usual clinical care.

TCS assessments will focus on selected brain structural imaging features, including substantia nigra echogenicity, raphe nuclei echogenicity, and third-ventricle width. Clinical information may include demographic characteristics, medical history, family history, disease course, medication history, and symptom assessment data when available. TCS measurements will be performed according to a standardized procedure, and image quality control will be conducted to reduce measurement variability.

The collected TCS and clinical variables will be integrated into a structured dataset for model development. Candidate machine-learning methods may include logistic regression, random forest, support vector machine, and XGBoost. Feature selection and model optimization will be performed within the model development process. Internal validation will be used to assess model performance, and additional independent data may be used for external validation if available.

Model performance will be evaluated using discrimination, calibration, and clinical utility metrics, including the area under the receiver operating characteristic curve, sensitivity, specificity, accuracy, F1 score, calibration assessment, and decision curve analysis where appropriate. Model interpretability will be explored using SHAP to assess the relative contribution of TCS imaging features and clinical variables.

The resulting model is intended as an auxiliary research tool for schizophrenia identification. It is not intended to make a definitive diagnosis, replace psychiatric assessment, or guide treatment decisions independently.