Advancing Brain Age Estimation: Normative Deviation Mapping (NDM) for Sensitive Detection of Pathological Aging

• We need to reconsider whether it is appropriate to use chronological age as ground truth in machine learning (ML). This approach creates a "paradox" because a perfect prediction of chronological age would leave no residual "age gap", yet the age gap is the very metric thought to capture the individual's actual biological health and risk for age-related diseases. • There is no clear boundary between normal aging and pathological aging. Significant individual variation exists in the rate of brain maturation and aging, even among healthy individuals. Brain age should more accurately reflect aging as a biological marker even in healthy individuals, enabling more precise assessment of cognitive decline and disease risk. • There is a risk that over-optimization of MAE in machine learning-based brain age estimation suppresses the intrinsic biological changes associated with healthy aging. • To avoid these issues in ML, we propose brain age prediction using a Normative Deviation Mapping (NDM) model for brain age prediction as a robust alternative to conventional ML models to address these issues. • Compared to ML models, the NDM model demonstrates favorable performance in reflecting pathological changes and the effect of aging on cognitive function in the elderly. Brain age is a valuable neuroimaging-based biomarker for assessing brain health, typically estimated using machine learning (ML) models. However, ML approaches suffer from inherent bias, requiring post-hoc correction, and may mask age-related biological variation, limiting their sensitivity to detect subtle biological aging. To overcome these limitations, we proposed a normative deviation mapping (NDM) model as an alternative to conventional ML. We analyzed MRI-derived volumes of 223 brain regions from 10,539 participants (aged 4–98 years). The NDM model assumes a normal distribution for age-specific volumes to calculate regional deviations, which are then aggregated across the brain to determine the final brain age. Compared to standard ML models (e.g., neural networks, extreme gradient boosting), the NDM model effectively eliminated regression bias. Consequently, the NDM model mitigated the underestimation of brain age in older adults, significantly enhancing the detection of pathological changes associated with neurodegenerative diseases, such as Alzheimer's disease. Furthermore, in healthy individuals, the NDM model showed a stronger correlation with cognitive function than chronological age. Our findings indicate that the use of ComBat-GAM for data harmonization could unintentionally mitigate the pathological associations of the brain age gap, suggesting a need for caution to preserve vital biological information. Overall, our model outperforms conventional ML in detecting pathological changes and reflecting biological brain age, while revealing the effects of amyloid accumulation and lifestyle habits on brain health, offering a more robust and biologically meaningful biomarker. The normative deviation mapping model circumvents the bias correction typically required for machine learning. By identifying pathological brain abnormalities, age-related cognitive decline, and latent brain aging in healthy individuals, it serves as a robust biomarker with competitive performance compared to conventional machine learning models.