Elucidating salient site-specific functional connectivity features and site-invariant biomarkers in schizophrenia via deep neural networks

Yi Hao Chan, Wei Chee Yew, Qian Hui Chew, Kang Sim, Jagath C. RajapakseScientific 2023 · DOI

TL;DR

This paper introduces SHRED-III, a novel deep learning framework that simultaneously harmonizes multi-site neuroimaging data and leverages semi-supervised learning to identify both generalizable biomarkers and site-specific features of schizophrenia. Their approach effectively disentangles these feature types, revealing a key site-invariant connection (transverse temporal gyrus-paracentral lobule) and highlighting the importance of site-specific findings.

What Problem Does This Paper Address?

Schizophrenia is a complex disorder with significant heterogeneity in its presentation and underlying neurobiology. Functional connectivity (FC) patterns, derived from fMRI, are promising biomarkers, but their discovery is hampered by variations across different study sites (e.g., different scanners, protocols, populations). This "site effect" can obscure true biomarkers and lead to findings that aren't generalizable. Existing methods struggle to disentangle these site-specific variations from true, site-invariant biological signals. Furthermore, limited labeled data in neuroimaging studies presents a challenge for traditional supervised learning, leading to potential overfitting.

Methods

The core of this work is the SHRED-III (Semi-supervised learning with data HaRmonisation via Encoder-Decoder-classifier) architecture. It's a deep neural network designed to tackle multi-site data challenges. Key components include:

  • Data Harmonization: This module aims to reduce site-specific biases in the data, making it more comparable across different sources. Unlike some methods that might just aggregate data, SHRED-III's harmonization is integrated into the learning process.
  • Encoder-Decoder-Classifier (EDC) Module: This part performs semi-supervised learning (SSL).
  • Encoder-Decoder: A variational auto-encoder (VAE) learns a compressed representation of the data and tries to reconstruct it. This reconstruction loss is crucial for learning even with limited labeled data.
  • Classifier: A deep neural network (DNN) classifier uses the learned representations to predict disease status.
  • End-to-End Training: SHRED-III is trained as a single, unified model, allowing the harmonization and classification components to learn together. This is an improvement over previous methods that might apply harmonization as a pre-processing step.
  • Biomarker Generation: After training, Integrated Gradients are used to compute feature attribution scores, highlighting which FC connections are most important for the model's predictions.

Datasets

The study utilized four resting-state fMRI datasets:

  • Local Dataset: Collected from the Institute of Mental Health (IMH), Singapore.
  • External Datasets:
    • UCLA Consortium for Neuropsychiatric Phenomics LA5c Study
    • COllaborative Informatics and Neuroimaging Suite Data Exchange tool (COINS)
    • Neuromorphometry by Computer Algorithm Chicago (NMorphCH) dataset

In total, 417 scans (249 normal controls, 168 schizophrenia patients) were analyzed. The paper notes that the local IMH dataset was smaller and had PANSS scores available, allowing for a proof-of-concept association study.

Key Insights and Ideas

  • SSL is Key for Performance: The reconstruction loss component of SSL significantly boosted model performance, especially on smaller datasets, outperforming models without SSL.
  • Harmonization's Double-Edged Sword: While data harmonization aims to reduce site effects, it can sometimes lead to lower model generalizability, particularly by decreasing sensitivity. SHRED-III mitigates this effect better than simpler harmonization approaches.
  • Disentangling Site-Specific vs. Site-Invariant: The SHRED-III framework successfully differentiates between biomarkers that are consistently found across all sites (site-invariant) and those that are unique to specific sites (site-specific). This is a critical step for reliable biomarker discovery.
  • Uncus-Fusiform Gyrus as a Site-Specific Example: The uncus-fusiform gyrus connection was identified as a prominent feature in the large UCLA dataset but was effectively downplayed by SHRED-III, suggesting it's a site-specific artifact rather than a universal biomarker.
  • Transverse Temporal Gyrus-Paracentral Lobule as a Robust Biomarker: This specific FC connection emerged as a strong, site-invariant biomarker for schizophrenia across multiple analyses.

Key Results

  • Classification Accuracy: While performance varied, the introduction of SSL generally led to statistically significant improvements in classification accuracy across individual sites. Data harmonization, when applied alone, sometimes decreased performance, especially on smaller sites. SHRED-III showed robust performance, with less degradation from harmonization compared to other SHRED variants.
  • Site-Invariant Biomarker: The connection between the transverse temporal gyrus and paracentral lobule was consistently identified as a salient, site-invariant FC feature across different models and datasets.
  • Site-Specific Features: The study successfully identified site-specific salient FC features, with the paracentral lobule being particularly implicated in the local IMH dataset.
  • PANSS Correlation: A significant negative correlation was found between the FC value of the inferior parietal lobule-parahippocampal gyrus connection and PANSS scores (both total and general psychopathology) in the local dataset, suggesting a link between this FC pattern and symptom severity.

Limitations and Future Work

  • Dataset Size: The study acknowledges that the benefits of combining data harmonization and SSL might be more pronounced with larger pooled datasets (e.g., ~800 subjects, as seen in prior work).
  • Loss Explosion: Training VAEs on small, high-dimensional datasets can lead to exploding loss, which was managed here with gradient clipping.
  • Interpreting Intermediate Saliency Scores: Distinguishing between truly site-invariant and site-specific features based on saliency scores can be challenging when scores fall in an intermediate range. Future work could focus on developing algorithms to better threshold these scores.
  • Federated Learning: Extending SHRED to a federated learning setting would allow for site-invariant biomarker discovery without pooling raw data, addressing privacy concerns.
  • Dynamic FC and Subtypes: Future work could incorporate dynamic FC analysis and clustering modules to capture temporal dynamics and identify schizophrenia subtypes.

Summary updated

Abstract

Schizophrenia is a highly heterogeneous disorder and salient functional connectivity (FC) features have been observed to vary across study sites, warranting the need for methods that can differentiate between site-invariant FC biomarkers and site-specific salient FC features. We propose a technique named Semi-supervised learning with data HaRmonisation via Encoder-Decoder-classifier (SHRED) to examine these features from resting state functional magnetic resonance imaging scans gathered from four sites. Our approach involves an encoder-decoder-classifier architecture that simultaneously performs data harmonisation and semi-supervised learning (SSL) to deal with site differences and labelling inconsistencies across sites respectively. The minimisation of reconstruction loss from SSL was shown to improve model performance even within small datasets whilst data harmonisation often led to lower model generalisability, which was unaffected using the SHRED technique. We show that our proposed model produces site-invariant biomarkers, most notably the connection between transverse temporal gyrus and paracentral lobule. Site-specific salient FC features were also elucidated, especially implicating the paracentral lobule for our local dataset. Our examination of these salient FC features demonstrates how site-specific features and site-invariant biomarkers can be differentiated, which can deepen our understanding of the neurobiology of schizophrenia.
BibTeX
@inproceedings{chan2023elucidating,
  author = {Yi Hao Chan and Wei Chee Yew and Qian Hui Chew and Kang Sim and Jagath C. Rajapakse},
  title = {{Elucidating salient site-specific functional connectivity features and site-invariant biomarkers in schizophrenia via deep neural networks}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-48548-w},
}
Figure