Brain morphometric features predict depression symptom phenotypes in late-life depression using a deep learning model

Bing Cao, Erkun Yang, Lihong Wang, Zhanhao Mo, David C. Steffens, Han Zhang, Mingxia Liu, Guy G. PotterFrontiers in Neuroscience 2023 · DOI

Abstract

We validated the effectiveness of using deep learning approaches on sMRI for predicting depression symptom phenotypes in LLD. We were able to identify deep embedded local morphological differences in symptom phenotypes in the brains of those with LLD, which is promising for symptom-targeted treatment of LLD. Future research with machine learning models integrating multimodal imaging and clinical data can provide additional discriminative information.

Introduction

Late-life depression (LLD) is a highly heterogeneous condition, characterized by a diverse array of clinical presentations. While two patients may share a diagnosis, their symptom profiles—ranging from severe anxiety and suicidal ideation to physical symptoms like sleep disturbance and appetite loss—can differ significantly. Traditional neuroimaging studies have often approached LLD as a monolithic entity, comparing depressed cohorts against healthy controls. However, this binary approach fails to account for the underlying neurobiological diversity of the disease. Mapping specific symptom phenotypes to distinct structural brain changes is essential for developing personalized, symptom-targeted interventions.

Applying deep learning to clinical populations like LLD presents unique challenges, primarily due to the limited availability of massive datasets typically required for training complex neural networks. This study addresses these challenges by developing a sample-efficient deep learning framework to predict symptom phenotypes from structural MRI (sMRI) data.

Methodology

The researchers developed a specialized "lightweight" 3D deep learning architecture, designated as 3D-ResNet-12, to mitigate the risk of overfitting on a modest sample size.

Model Architecture and ROI Selection

The standard ResNet-18 architecture was simplified into a 12-layer version. Rather than processing the entire brain volume, which increases computational complexity and noise, the model utilized a patch-based approach. Thirty-four a priori Regions of Interest (ROIs) were selected based on the AAL3 atlas, focusing on areas previously associated with LLD pathology.

Architecture of the 3D-ResNet-12 model for ROI-based symptom phenotype prediction.

Clinical Data and Factor Analysis

Data were obtained from the Neurocognitive Outcomes of Depression in the Elderly (NCODE) study, comprising 116 adults (age 60+) with LLD. Clinical severity was assessed using the Hamilton Rating Scale for Depression (HDRS-17) and the Montgomery-Åsberg Depression Rating Scale (MADRS). To capture the multidimensional nature of the disease, five symptom factors were derived: (1) Anhedonia, (2) Suicidality, (3) Appetite, (4) Sleep Disturbance, and (5) Anxiety.

Statistical Approach

The model predicted factor scores using Iteratively Reweighted Least Squares (IRLS) regression, a robust method designed to minimize the influence of outliers.

Results

The deep learning model demonstrated varying degrees of predictive success across the five symptom phenotypes, suggesting that different symptoms have distinct structural signatures in the brain.

Predictive Performance

The highest predictive accuracy was achieved for the Anxiety factor, which showed a significant Pearson correlation between estimated and real scores ($r = 0.2589, p < 0.001$). The Suicidality factor also yielded statistically significant results ($r = 0.1436, p < 0.001$). In contrast, predictions for Appetite, Anhedonia, and Sleep Disturbance did not reach statistical significance, suggesting these symptoms may be better characterized by functional connectivity or white matter integrity rather than gray matter morphology.

Correlation between real and predicted ROI-level scores for depression symptom phenotypes.

Neuroanatomical Hubs

Analysis of regional importance revealed that the Anterior Cingulate Cortex (ACC) and Orbitofrontal Cortex (OFC) were the most discriminative regions across nearly all symptom factors. Notably, the ACC was identified as the top predictive region for four out of the five phenotypes, reinforcing its role as a critical hub in the neurobiology of LLD.

Discussion and Conclusion

This study validates the use of lightweight deep learning models for identifying local morphological differences associated with specific depression symptoms in older adults. By focusing on ROI-based patches and a shallower architecture, the researchers achieved meaningful predictive power with a relatively small clinical dataset ($N=116$).

Despite these promising results, several limitations remain. The reliance on a single imaging modality (T1-weighted sMRI) may overlook important pathological features. Future research should aim to integrate multimodal data, such as T2-FLAIR for white matter lesion detection and functional MRI for network connectivity analysis. Additionally, while the a priori ROI selection was grounded in existing literature, whole-brain discovery methods in larger, multi-site cohorts will be necessary to fully map the structural landscape of LLD symptom phenotypes. Ultimately, these findings provide a potential roadmap for the development of symptom-targeted treatments in geriatric psychiatry.

Summary updated

Abstract

We validated the effectiveness of using deep learning approaches on sMRI for predicting depression symptom phenotypes in LLD. We were able to identify deep embedded local morphological differences in symptom phenotypes in the brains of those with LLD, which is promising for symptom-targeted treatment of LLD. Future research with machine learning models integrating multimodal imaging and clinical data can provide additional discriminative information.
BibTeX
@inproceedings{cao2023brain,
  author = {Bing Cao and Erkun Yang and Lihong Wang and Zhanhao Mo and David C. Steffens and Han Zhang and Mingxia Liu and Guy G. Potter},
  title = {{Brain morphometric features predict depression symptom phenotypes in late-life depression using a deep learning model}},
  booktitle = {Frontiers in Neuroscience},
  year = {2023},
  doi = {10.3389/fnins.2023.1209906},
}
Figure