Pace is not the same as a brain age gap
A conventional brain age model predicts how old a brain looks from a single scan. The difference between that prediction and chronological age is the brain predicted age difference, or brain-PAD, and it describes where an individual sits relative to a reference population at one time point. The paper argues that this position does not necessarily quantify how fast the brain is currently changing: someone could show a consistently positive brain-PAD across several scans without any ongoing structural change.
Brain-PACE targets a different quantity. It takes a baseline and a follow-up T1-weighted scan as a pair and predicts the interval between them, then expresses the result as a pace ratio, P, equal to the predicted interval divided by the observed interval. A value of 1 means the model's predicted structural change matches chronological ageing; values above 1 indicate faster predicted change. This is a ratio of predicted to observed time, not a difference in years, so it should not be read as a brain age gap.
The paper also builds an indirect comparison: a cross-sectional brain age model with the same backbone, spatial attention pooling, soft classification and Cramér distance loss, whose predictions at follow-up and baseline are subtracted to give a pace. That design keeps the architecture similar so the comparison isolates direct versus indirect estimation of change.
Paired scans and a distribution instead of a scalar
The framework is Siamese: two scans pass through shared network branches, and the model compares their embeddings rather than treating each scan independently. Spatial attention weights which embedding locations matter for the interval prediction. The output is a probability distribution over 80 bins of possible intervals rather than a single number, which lets the model report predictive uncertainty alongside its estimate.
Two additions distinguish Brain-PACE from the LILAC framework it extends. Soft label distribution learning lets the model learn from neighbouring bins instead of treating the true interval as a single hard target, and a Cramér distance objective scores the whole predicted distribution against the observed interval. The authors report that this combination improved probabilistic performance and reduced prediction bias relative to the KL-attention variant, and that Cramér-attention generally improved entropy and empirical coverage.
Performance is scored with the Continuous Ranked Probability Score, which compares the full predicted cumulative distribution with the observed interval, so lower values mean predictions that are both accurate and concentrated. The paper also reports mean absolute error, mean squared error, signed error and the pace ratio, and measures uncertainty as normalised Shannon entropy of the predicted distribution, where 0% is complete certainty and 100% is the maximum spread across all 80 bins.
Interval accuracy is similar to the indirect route
Against the LILAC+ baseline, Brain-PACE reduced pair-level mean absolute error from 0.779 to 0.663 years and mean squared error from 1.059 to 0.756 years squared. Those are interval-prediction errors in years, not brain age errors.
The head-to-head comparison with indirectly derived pace is more mixed. In the internal group, Brain-PACE showed numerically lower mean absolute error (0.663 versus 0.806 years) and lower signed error (0.017 versus 0.307 years), but neither difference survived false discovery rate correction. In the MCSA healthy cohort the two approaches performed similarly, and in WRAP Brain-PACE showed significantly greater underprediction (signed error −0.402 versus −0.029 years). In the ADNI mild cognitive impairment group, the indirect approach had numerically lower mean absolute error (0.766 versus 0.809 years), while Brain-PACE produced a significantly higher pace ratio (1.434 versus 1.220).
The authors interpret the larger deviations in the impairment group as possible sensitivity to disease-related accelerated ageing rather than simply worse accuracy. That interpretation is tested by asking whether the pace measure tracks clinical and biological markers, which is a separate question from interval error.
Faster pace tracks tau burden and clinical severity
Within the 68 ADNI participants with mild cognitive impairment, and after controlling for baseline age and median temporal gap, faster Brain-PACE was associated with greater tau burden measured as SUVR in the posterior cingulate (r = 0.59), precuneus (r = 0.47) and entorhinal cortex (r = 0.37). It was also associated with greater functional impairment on the FAQ (r = 0.35), poorer performance on the ADAS13 (r = 0.30) and greater impairment on the CDR sum of boxes (r = 0.32). The MMSE association was negative (r = −0.26) but did not survive correction.
The comparison measures behaved differently. Cross-sectional brain-PAD was associated mainly with structural features: lower hippocampal volume (r = −0.43), greater CDR score (r = 0.35), and lower precuneus and posterior cingulate volume, while its tau associations did not survive correction. Pace derived indirectly from repeated cross-sectional estimates showed weaker associations, with no feature correlation surviving correction; the strongest trends were entorhinal tau, posterior cingulate tau and FAQ score, each at r = 0.27 or 0.25 with q = 0.201.
These are partial correlations in one impaired cohort, so they describe co-variation rather than a causal chain. The authors suggest that accelerated Brain-PACE may reflect structural change occurring alongside active tau-related neurodegeneration rather than isolated volume loss, and they note that longitudinal tau PET would be needed to test whether the two change in parallel.
Predictions rest on distributed change, with calibration caveats
An Integrated Gradients analysis held the baseline scan fixed and interpolated toward the follow-up scan, so each voxel's attribution reflects the actual intensity change at that location multiplied by the model's average sensitivity along that path. Regions with no morphological change therefore receive zero attribution.
The resulting group map was distributed rather than focal, with the strongest contributions in the frontal and temporal poles, lateral occipital cortex, precentral and postcentral gyri, and brainstem, plus involvement of the precuneus, anterior cingulate, bilateral thalami, right caudate and posterior cingulate cortex. The authors read this as evidence that predicted intervals are informed by change across many cortical, subcortical and brainstem structures rather than a single anatomical region.
Uncertainty calibration did not hold equally well everywhere. Cramér-attention generally improved entropy and empirical coverage relative to KL-attention, but the paper reports reduced coverage in the mild cognitive impairment and WRAP cohorts, indicating that calibration remains sensitive to cohort and pathological differences. Predictive uncertainty from this model should therefore be read as cohort-dependent rather than uniformly reliable.
Direct pace adds a complementary, not replacement, phenotype
The paper's central claim is comparative: modelling the difference between two time points captures information that repeated cross-sectional brain age estimates do not recover. The support is that direct Brain-PACE classified more impaired participants as accelerated and showed associations with regional tau and clinical severity that the indirect pace and cross-sectional brain-PAD did not, while interval accuracy was broadly similar across approaches and sometimes favoured the indirect route.
That pattern matters for how the measure should be used. Brain-PACE is presented as a complementary longitudinal imaging phenotype sensitive to clinical and biological change in early neurodegeneration, not as a more accurate clock. Its acceleration labels depend on percentiles from an internal healthy reference cohort, its strongest biomarker evidence comes from 68 participants with mild cognitive impairment in one cohort, and its uncertainty estimates were less well calibrated in that same impaired group.
The practical priority the evidence supports is therefore a split by question. When the question is how old a brain appears at one time point, cross-sectional brain-PAD remains the relevant measure and is tied to structural volumes. When the question is how quickly structure is changing in early neurodegeneration, direct paired-scan pace is the measure that carried the tau and clinical associations here, with the caveat that its thresholds are reference-relative and its calibration in pathological groups still needs work.
Sources used in this answer
Brain-PACE: A Deep Siamese MRI Framework for Modelling Longitudinal Brain Acceleration
Directly estimated longitudinal brain pace classified 42.6% of 68 participants with mild cognitive impairment as accelerated versus 20.6% for indirectly derived pace, and correlated with posterior cingulate, precuneus and entorhinal tau burden where indirect and cross-sectional measures did not.