428. FEATURE SELECTION AND ALGORITHM OPTIMIZATION FOR BRAIN AGE PREDICTION IN SCHIZOPHRENIA PATIENTS

Abstract Background Schizophrenia is a severe mental disorder that significantly impacts daily life. Symptoms range from mild issues such as anxiety and insomnia to severe conditions such as hallucinations and delusions. Accelerated brain aging is a characteristic of schizophrenia, manifested as structural brain changes, including reductions in gray and white matter. Early diagnosis and treatment are critical to slowing disease progression and improving patients' quality of life. However, traditional diagnostic methods mainly rely on behavioral assessments, which are subjective and limited in accuracy. Recent advancements in neuroimaging and artificial intelligence have created new opportunities for early and more accurate diagnosis. Aims & Objectives This study aims to develop a machine learning model based on MRI images to predict the brain age of schizophrenia patients, enabling early diagnosis and monitoring of disease progression. The specific objectives are: Method MRI images were sourced from three public datasets (COBRE, MCIC, UCLA). The CAT12 toolbox was used to encode the images, generating structured numerical data, and Quantile Transformer uniform was employed for data normalization to reduce errors and improve model efficiency. For model training, various regression algorithms were tested, and Bayesian Ridge was identified as the most suitable for the data. Hyperparameters were optimized using GridSearchCV to enhance model performance, with key metrics including Mean Squared Error (MSE) and Pearson Correlation Coefficient (PCC). A genetic algorithm was then applied iteratively for feature selection, identifying critical features affecting prediction outcomes. Model performance was evaluated using MSE and PCC, comparing results before and after feature selection. Results The study found that combining QuantileTransformer normalization with the Bayesian Ridge regression algorithm achieved optimal model performance. After hyperparameter tuning and feature selection via the genetic algorithm, the model achieved a PCC of 0.95. Key features identified included changes in white matter volume, cavity size, and alterations in the amygdala. Discussion & Conclusions This study demonstrates that using a machine learning model to predict the brain age of schizophrenia patients is feasible. The results show that efficient preprocessing tools (e.g., CAT12) and feature selection techniques (e.g., genetic algorithms) are critical to enhancing model performance. Additionally, identifying specific brain structural changes related to accelerated brain aging provides new insights into the pathology of schizophrenia. This research lays the groundwork for integrating machine learning techniques into clinical applications, improving diagnostic efficiency, and facilitating early intervention. Future studies could explore larger datasets, additional features, and alternative algorithms to further enhance predictive accuracy and clinical applicability.