Structural, Functional and Cognitive Validation of a Deep-Learning MCI-to-AD Conversion Model in OASIS-3
- 1Swinburne University of Technology
Abstract
BackgroundDeep-learning models predict mild cognitive impairment (MCI)-to-Alzheimers disease (AD) conversion from structural MRI with high accuracy. However, they are typically validated within a single cohort and judged on discrimination alone. Whether their risk scores are biologically grounded, and whether they generalise to independent data, remains unclear.
MethodsWe applied the ADNI-trained Temporal Adaptive Fusion Network (TAF-Net), without retraining, to 101 MCI participants from OASIS-3 (25 converters, 76 stable) and tested whether its conversion-risk scores were grounded in independent structural, functional, and cognitive markers of AD. Regional atrophy rates were derived from longitudinal FreeSurfer, cognition from longitudinal MMSE and CDR-Sum-of-Boxes, and baseline resting-state functional connectivity from fMRIPrep, in structural and functional subsamples of 41 and 40 participants. Associations used rank-based statistics with false-discovery-rate correction.
ResultsHigher TAF-Net risk tracked faster atrophy in medial-temporal AD-signature regions but not in AD-spared cortex -- an anatomically specific coupling that survived adjustment for global atrophy. In external validation, risk discriminated converters (AUC = 0.72), comparable to native atrophy and strongly concordant with it; atrophy statistically accounted for the models prognostic signal. Discrimination transferred but calibration did not: the two lowest tertiles of risk were assigned near-zero probability yet converted at 15%. Risk also tracked the multi-year rate of cognitive decline and, at baseline, was associated with reduced within-network functional-connectivity integrity, concentrated in salience and default-mode hubs; longitudinal functional analyses were underpowered.
ConclusionsA conversion model trained on one cohort produced risk scores that, in an independent cohort, were grounded in the structural, functional, and cognitive hallmarks of Alzheimers disease -- supporting biological validity and external generalisation. The score, however, largely re-expresses the neurodegenerative substrate captured by structural atrophy. Rank ordering transferred across cohorts but absolute risk did not, so the score requires recalibration before its values can be interpreted as individual probabilities.
Related
- Modelling the risk of conversion from mild cognitive impairment to Alzheimer's disease using machine learning.
- Longitudinal Changes in Structural and Functional Brain Imaging and Their Association with Cognitive and Neuropsychiatric Symptoms in Hashimoto Encephalopathy with Anti-NAE Antibody
- Functional and Structural Imaging Features of Cognitive Profiles in Multiple Sclerosis.
- Identifying Mild Cognitive Impairment from Routinely Collected Physical, Functional and Psychosocial Measures in Community-Dwelling Older Adults: A Comparative Machine-Learning and Regression Modelling Study in the Northern Japanese ORANGE Registry.
- Machine learning-derived Alzheimer's disease dimensions: neuroanatomical, cognitive, clinical, functional, and genetic risk heterogeneity signatures
BibTeX
@misc{fin2026structural,
title = {{Structural, Functional and Cognitive Validation of a Deep-Learning MCI-to-AD Conversion Model in OASIS-3}},
author = {S. Fin and A. Moayedikia},
howpublished = {medRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.05.26362318},
url = {https://www.medrxiv.org/content/10.64898/2026.09.05.26362318v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Structural, Functional and Cognitive Validation of a Deep-Learning MCI-to-AD Conversion Model in OASIS-3
AU - S. Fin
AU - A. Moayedikia
PY - 2026
DA - 2026/09/05
PB - medRxiv
DO - 10.64898/2026.09.05.26362318
UR - https://www.medrxiv.org/content/10.64898/2026.09.05.26362318v1
AB - BackgroundDeep-learning models predict mild cognitive impairment (MCI)-to-Alzheimers disease (AD) conversion from structural MRI with high accuracy. However, they are typically validated within a single cohort and judged on discrimination alone. Whether their risk scores are biologically grounded, and whether they generalise to independent data, remains unclear. MethodsWe applied the ADNI-trained Temporal Adaptive Fusion Network (TAF-Net), without retraining, to 101 MCI participants from OASIS-3 (25 converters, 76 stable) and tested whether its conversion-risk scores were grounded in independent structural, functional, and cognitive markers of AD. Regional atrophy rates were derived from longitudinal FreeSurfer, cognition from longitudinal MMSE and CDR-Sum-of-Boxes, and baseline resting-state functional connectivity from fMRIPrep, in structural and functional subsamples of 41 and 40 participants. Associations used rank-based statistics with false-discovery-rate correction. ResultsHigher TAF-Net risk tracked faster atrophy in medial-temporal AD-signature regions but not in AD-spared cortex -- an anatomically specific coupling that survived adjustment for global atrophy. In external validation, risk discriminated converters (AUC = 0.72), comparable to native atrophy and strongly concordant with it; atrophy statistically accounted for the models prognostic signal. Discrimination transferred but calibration did not: the two lowest tertiles of risk were assigned near-zero probability yet converted at 15%. Risk also tracked the multi-year rate of cognitive decline and, at baseline, was associated with reduced within-network functional-connectivity integrity, concentrated in salience and default-mode hubs; longitudinal functional analyses were underpowered. ConclusionsA conversion model trained on one cohort produced risk scores that, in an independent cohort, were grounded in the structural, functional, and cognitive hallmarks of Alzheimers disease -- supporting biological validity and external generalisation. The score, however, largely re-expresses the neurodegenerative substrate captured by structural atrophy. Rank ordering transferred across cohorts but absolute risk did not, so the score requires recalibration before its values can be interpreted as individual probabilities.
ER -