A Multiverse Approach to Brain Age Prediction: Exploring the Impact of Modality, Bias Adjustment, and Atlas
- 1Tilburg University
Abstract
Brain age, an individual's age predicted from neuroimaging data, is a promising biomarker of brain health. Yet brain age estimates may depend on methodological choices within the prediction pipeline. We used a multiverse approach to systematically investigate effects of data modality, bias-adjustment method, and brain parcellation on brain-age prediction using data from 560 healthy adults (18.6-89 years) from the CamCAN dataset. Across 84 pipeline conditions, chronological age was predicted from connectivity features with elastic net regression, and performance was evaluated using mean absolute error (MAE). A multilevel model tested the effects of the pipeline factors and their interactions; moreover, cross-atlas consistency of the most predictive connections was assessed. MAEs varied substantially across conditions (5.26-18.30 years). Modality and bias adjustment had significant effects on MAE and interacted significantly. Concatenating structural and functional connectivity performed similarly to structural connectivity alone, and structure-function coupling produced the highest MAE in most conditions. Beheshti-based bias adjustment outperformed Cole-based adjustment, while combining adjustment procedures did not consistently improve performance. Atlas choice had no main effect on MAE, its interaction with modality was significant but negligible in size, and finer parcellations did not systematically improve performance. However, the network-level features driving predictions showed limited spatial overlap across atlases, and importance was distributed over many features. Brain-age prediction is thus sensitive to methodological choices, particularly modality and bias-adjustment strategy. Atlas choice appears less important for predictive accuracy but may affect which connections appear most predictive. These findings highlight the importance of systematically evaluating analytical choices in brain-age research.
Related
- Tri-Modality Representation Learning for Molecular Property Prediction
- Exploring the impact of intraspecific interaction between sibling colonies of Bacillus cereus MSM-S1 in the absence and presence of antibiotics
- Exploring Sleep Respiratory Endophenotypes in Prader-Willi Syndrome: Impact of Genetic Mutations
- Editorial: Exploring AI’s role in disease prediction and diagnosis through medical big data
- Exploring the feasibility, acceptability and impact of genomic newborn screening for rare diseases in England: A study protocol for the Generation Study - Process and Impact Evaluation (Version 2)
BibTeX
@misc{festor2026multiverse,
title = {{A Multiverse Approach to Brain Age Prediction: Exploring the Impact of Modality, Bias Adjustment, and Atlas}},
author = {J. Festor and L. Smolders and W. De Baene},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.12.758514},
url = {https://www.biorxiv.org/content/10.64898/2026.10.12.758514v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - A Multiverse Approach to Brain Age Prediction: Exploring the Impact of Modality, Bias Adjustment, and Atlas
AU - J. Festor
AU - L. Smolders
AU - W. De Baene
PY - 2026
DA - 2026/10/12
PB - bioRxiv
DO - 10.64898/2026.10.12.758514
UR - https://www.biorxiv.org/content/10.64898/2026.10.12.758514v1
AB - Brain age, an individual's age predicted from neuroimaging data, is a promising biomarker of brain health. Yet brain age estimates may depend on methodological choices within the prediction pipeline. We used a multiverse approach to systematically investigate effects of data modality, bias-adjustment method, and brain parcellation on brain-age prediction using data from 560 healthy adults (18.6-89 years) from the CamCAN dataset. Across 84 pipeline conditions, chronological age was predicted from connectivity features with elastic net regression, and performance was evaluated using mean absolute error (MAE). A multilevel model tested the effects of the pipeline factors and their interactions; moreover, cross-atlas consistency of the most predictive connections was assessed. MAEs varied substantially across conditions (5.26-18.30 years). Modality and bias adjustment had significant effects on MAE and interacted significantly. Concatenating structural and functional connectivity performed similarly to structural connectivity alone, and structure-function coupling produced the highest MAE in most conditions. Beheshti-based bias adjustment outperformed Cole-based adjustment, while combining adjustment procedures did not consistently improve performance. Atlas choice had no main effect on MAE, its interaction with modality was significant but negligible in size, and finer parcellations did not systematically improve performance. However, the network-level features driving predictions showed limited spatial overlap across atlases, and importance was distributed over many features. Brain-age prediction is thus sensitive to methodological choices, particularly modality and bias-adjustment strategy. Atlas choice appears less important for predictive accuracy but may affect which connections appear most predictive. These findings highlight the importance of systematically evaluating analytical choices in brain-age research.
ER -