Quantifying Uncertainty in Brain Age Predictions via Conformal Prediction
- 1Medical University of South Carolina
Abstract
Youth can exhibit signs of delayed or accelerated neurodevelopment, measurable via brain magnetic resonance imaging (MRI). Brain age prediction seeks to estimate brain age at an individual level using machine learning models fitted on healthy individuals. The brain age gap (BAG): the difference between brain age and chronological age, has been studied as a potential biomarker. However, BAG suffers from known limitations, including dependence on chronological age, regression to the mean, and challenges in interpretability. As an alternative framework, we introduce brain age intervals (BAIs) as a normative interval on an individual's chronological age, based on MRI measures. BAIs can be derived using a recent statistical framework called conformal prediction that yields prediction intervals with guaranteed coverage. We train several brain age prediction models on structural and functional MRI scans from the Reproducible Brain Charts dataset (aged 6-22) and estimated BAIs. The BAIs demonstrated stable empirical coverage close to the nominal 90% level (median coverage=91%, width=7.36 years, root mean square error (RMSE)=2.20 years across 100 repeated train-test splits). Associations between interval coverage and clinical measures (p-factor, parental education) were weak, with only a modest directional signal for parental education. Together, these findings demonstrate the feasibility of uncertainty-aware brain age modeling in youth populations while highlighting the need for larger and more diverse samples to detect meaningful clinical and environmental associations.
Related
- Learning from Error Predictions Rather Than Prediction Errors: A Perspective on Self-Supervised Learning in the Brain
- Representation-aware conformal prediction for reliable uncertainty quantification in drug-target affinity
- Neural representation of the subjective uncertainty of prediction in the human brain
- CaliPPer: quantifying, predicting and improving AI model performance for binding prediction
- Contributions of Site Heterogeneity in Multi-site Brain-Phenotype Predictions
BibTeX
@misc{samanta2026quantifying,
title = {{Quantifying Uncertainty in Brain Age Predictions via Conformal Prediction}},
author = {U. Samanta and G. Shafiei and H. T. Karim and T. D. Satterthwaite and A. Chen},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.04.749442},
url = {https://www.biorxiv.org/content/10.64898/2026.09.04.749442v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Quantifying Uncertainty in Brain Age Predictions via Conformal Prediction
AU - U. Samanta
AU - G. Shafiei
AU - H. T. Karim
AU - T. D. Satterthwaite
AU - A. Chen
PY - 2026
DA - 2026/09/04
PB - bioRxiv
DO - 10.64898/2026.09.04.749442
UR - https://www.biorxiv.org/content/10.64898/2026.09.04.749442v1
AB - Youth can exhibit signs of delayed or accelerated neurodevelopment, measurable via brain magnetic resonance imaging (MRI). Brain age prediction seeks to estimate brain age at an individual level using machine learning models fitted on healthy individuals. The brain age gap (BAG): the difference between brain age and chronological age, has been studied as a potential biomarker. However, BAG suffers from known limitations, including dependence on chronological age, regression to the mean, and challenges in interpretability. As an alternative framework, we introduce brain age intervals (BAIs) as a normative interval on an individual's chronological age, based on MRI measures. BAIs can be derived using a recent statistical framework called conformal prediction that yields prediction intervals with guaranteed coverage. We train several brain age prediction models on structural and functional MRI scans from the Reproducible Brain Charts dataset (aged 6-22) and estimated BAIs. The BAIs demonstrated stable empirical coverage close to the nominal 90% level (median coverage=91%, width=7.36 years, root mean square error (RMSE)=2.20 years across 100 repeated train-test splits). Associations between interval coverage and clinical measures (p-factor, parental education) were weak, with only a modest directional signal for parental education. Together, these findings demonstrate the feasibility of uncertainty-aware brain age modeling in youth populations while highlighting the need for larger and more diverse samples to detect meaningful clinical and environmental associations.
ER -