Robust brain structural normative modeling method based on multi-level data harmonization

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder characterized by cognitive decline, which substantially affects daily functioning and quality of life. Early detection and intervention are crucial for disease management, particularly during the initial stages. Mild cognitive impairment (MCI), often considered a precursor to AD, presents an opportunity for timely interventions to slow disease progression. Traditional diagnostic methods rely primarily on clinical assessments and subjective evaluations, which may not capture the subtle structural changes in the brain associated with cognitive decline. Magnetic resonance imaging is a valuable tool for assessing brain morphology and providing quantitative measures of structural characteristics that can aid in cognitive impairment diagnosis. Recent advancements in normative modeling have provided innovative approaches for quantifying deviations in brain structure in individuals with cognitive impairment. Normative methods establish a baseline for healthy brain morphology, allowing for the identification of deviations in patients. However, traditional normative approaches often face challenges related to multicenter datasets, including batch effects resulting from variability in imaging protocols and instruments across different research sites. Moreover, conventional linear regression models used to assess deviations typically do not account for age-related nonlinear changes, and may not adequately consider the influence of the standard deviation on the population distribution. Therefore, there is an urgent need for robust methods to improve the accuracy and reliability of brain structural assessments in diverse populations.To address these challenges, this study proposes a robust construction method for normative brain structural models based on multilevel data harmonization. This method aims to enhance deviation quantification in individuals with cognitive impairments across multicenter datasets with varying demographic factors, such as age and gender. The research began by employing the ComBat method, which is a recognized technique for bias correction in multicenter data. ComBat effectively reduces batch effects induced by differences in imaging equipment and acquisition protocols, ensuring that data from various centers can be accurately compared. Following this, the study established a normative model for a healthy population aged 55–95. This model uses log transformation in conjunction with linear regression to correct for age- and sex-related variations in morphological features. From this normative model, Z-scores were extracted as individual deviation indicators from the healthy population. To classify individuals, an autogluon machine-learning model was employed and trained to differentiate between healthy controls (HC), patients with MCI, and patients with AD using Z-score features derived from the normative model. The results of this study demonstrate the efficacy of the proposed method for classifying individuals with cognitive impairment. By leveraging the Autogluon automated machine learning framework, the robust construction of the brain structural normative model achieved a classification accuracy of 79.87%±2.61% between AD and HC, and 75.43%±3.39% between MCI and HC. These results are notably higher than those obtained under single-center conditions, which reported accuracies of 75.40%±4.24% and 72.82%±6.08%, respectively. The proposed method improves classification accuracy and exhibits greater robustness against variations in data, as evidenced by lower standard deviations across multicenter datasets. Additionally, a comparative analysis reveals that the performance of the proposed method in classifying AD and HC significantly surpasses that of traditional residual-based indicators (72.48% ±3.84%), and in MCI vs. HC classification, it demonstrates improved accuracy and consistency compared to residual-based approaches (74.81%±5.19%), with reduced variability (e.g., ±3.39% vs. ±5.19% in MCI classification) highlighting its enhanced reliability.This study underscores the potential value of the robust brain structural normative model construction method based on multilevel data harmonization in clinical applications, paving the way for improved early diagnosis and timely intervention strategies for AD and related cognitive disorders. As the healthcare landscape continues to evolve, such advancements will be critical in enhancing our understanding and management of neurodegenerative diseases, ultimately contributing to better patient outcomes.