Deep Learning-Based Structural Brain Age Estimation in Bipolar Disorder and Schizophrenia: A Single-Site Pilot Study.

A 3D-CNN trained on healthy controls achieved high prediction accuracy in HC but substantially reduced accuracy when applied to BD and SZ groups.

• In healthy controls (n=155), the model achieved MAE=3.05 years and r=0.96.
• In bipolar disorder (n=122), accuracy was reduced to MAE=8.86 years and r=0.51.
• In schizophrenia (n=161), accuracy was further reduced to MAE=9.01 years and r=0.48.
• The model was trained exclusively on healthy controls and then applied to independent BD and SZ groups.

Mean Brain-PAD was significantly elevated in both BD and SZ relative to healthy controls.

• Mean Brain-PAD was +4.2 ± 10.2 years in BD.
• Mean Brain-PAD was +6.7 ± 8.7 years in SZ.
• Mean Brain-PAD was +0.7 ± 3.5 years in HC.
• Brain-PAD was computed as predicted brain age minus chronological age.

BD and SZ exhibited elevated Brain-PAD at younger ages that converged toward HC trajectories by midlife, followed by renewed divergence beyond age 40.

• Age-by-group interaction analyses revealed non-linear developmental trajectories in BD and SZ.
• Piecewise and spline models showed steeper negative slopes in BD and SZ compared with HC.
• The pattern included initial elevation at younger ages, convergence toward HC by midlife, and then renewed divergence beyond age 40.
• Sensitivity analyses including piecewise regression and inverse probability weighting were conducted to support these findings.

BD and SZ showed greater within-group heterogeneity in Brain-PAD compared to healthy controls.

• Variance and quantile regression analyses indicated greater heterogeneity in BD and SZ across the Brain-PAD distribution.
• Standard deviation of Brain-PAD was 10.2 years in BD and 8.7 years in SZ, compared to 3.5 years in HC.
• Within-group dispersion was explicitly examined as part of the analytical framework.

Grad-CAM analyses identified temporal and frontal regions as the central contributors to brain age predictions across all groups.

• Gradient-weighted Class Activation Mapping (Grad-CAM) was used to identify regional contributions to brain age predictions.
• In SZ, Brain-PAD correlated positively with whole-brain activation (r=0.23, p=0.004), frontal activation (r=0.21, p=0.009), and temporal activation (r=0.20, p=0.012).
• BD showed weaker and more diffuse associations between Brain-PAD and regional Grad-CAM activations compared to SZ.
• Frontotemporal regional contributions were described as reflecting 'model sensitivity to age-informative structure.'

The study employed inverse probability weighting as a sensitivity analysis to address potential confounding in group comparisons.

• Sensitivity analyses included piecewise regression, spline models, and inverse probability weighting.
• The 3D-CNN was a three-dimensional convolutional neural network trained on structural MRI data.
• The study was characterized as a single-site pilot study.
• Sample sizes were HC n=155, BD n=122, and SZ n=161.