# Deep Learning–Based Structural Brain Age Estimation in Bipolar Disorder and Schizophrenia: A Single‐Site Pilot Study

## TL;DR

This study uses deep learning to estimate brain age in bipolar disorder and schizophrenia, finding accelerated aging early in life and key brain regions involved.

## Contribution

The study introduces a deep learning model for brain age estimation and identifies non-linear aging patterns in bipolar disorder and schizophrenia.

## Key findings

- Bipolar disorder and schizophrenia show elevated brain age early in adulthood, converging with healthy controls by midlife.
- Frontotemporal regions are key contributors to brain age predictions, especially in schizophrenia.
- Greater variability in brain age is observed in bipolar disorder and schizophrenia compared to healthy controls.

## Abstract

Accelerated brain aging has been implicated in severe mental illnesses, particularly schizophrenia (SZ) and bipolar disorder (BD). Brain–PAD, derived from structural MRI, offers a promising biomarker of neurobiological aging, but its developmental course, within‐group variability, and regional drivers remain incompletely understood. A three‐dimensional convolutional neural network (3D‐CNN) was trained exclusively on healthy controls (HC; n = 155) and then applied to independent BD (n = 122) and SZ (n = 161) groups. Brain–PAD was computed as predicted brain age minus chronological age. Age‐by‐group interactions, within‐group dispersion, and sensitivity analyses (e.g., piecewise regression, inverse probability weighting) were conducted. Gradient‐weighted Class Activation Mapping (Grad–CAM) was used to identify regional contributions to brain age predictions. The 3D‐CNN achieved high accuracy in HC (MAE = 3.05 years, r = 0.96), with reduced accuracy in BD (MAE = 8.86, r = 0.51) and SZ (MAE = 9.01, r = 0.48). Mean Brain–PAD was significantly elevated in BD (+4.2 ± 10.2 years) and SZ (+6.7 ± 8.7 years) relative to HC (+0.7 ± 3.5 years). Age‐by‐group analyses revealed that BD and SZ exhibited elevated Brain–PAD at younger ages, converging toward HC trajectories by midlife, followed by renewed divergence beyond age 40. This pattern was supported by piecewise and spline models showing steeper negative slopes in BD and SZ compared with HC. Variance and quantile regression indicated greater heterogeneity in BD and SZ across the Brain–PAD distribution. Grad–CAM highlighted temporal and frontal regions as central contributors across all groups; in SZ, Brain–PAD correlated positively with whole‐brain (r = 0.23, p = 0.004), frontal (r = 0.21, p = 0.009), and temporal (r = 0.20, p = 0.012) activations, whereas BD showed weaker and more diffuse associations. SZ and BD exhibit elevated Brain–PAD early in adulthood with greater heterogeneity than healthy controls. Frontotemporal regions contribute prominently to brain‐age predictions, reflecting model sensitivity to age‐informative structure. These findings support Brain–PAD as a group‐level marker of apparent brain aging and motivate longitudinal study of midlife divergence.

A deep–learning model (CNN) trained on healthy controls estimated structural brain age in bipolar disorder and schizophrenia. Both disorders showed elevated brain‐predicted age early in adulthood, convergence by midlife, and renewed deviation later, highlighting non‐linear brain aging and frontotemporal regions as key contributors identified by Grad–CAM.