SYNERGY-VAE: An explainable generative deep learning framework for discovering depression subgroups from multimodal population health data
6 affiliations
- 1Department of Mathematics and Statistics, York University, Toronto, Ontario, Canada
- 2Biostatistics Department, Princess Margaret Cancer Centre, University Health Network, Toronto, Ontario, Canada
- 3Biostatistics Department, Dalla Lana School of Public Health, University of Toronto, Toronto, Ontario, Canada
- 4AI for Mental Health Program, St. Michael’s Hospital, Unity Health Toronto, Toronto, Ontario, Canada
- 5Department of Electrical and Computer Engineering, Toronto Metropolitan University, Toronto, Ontario, Canada
- 6Department of Psychiatry, University of Toronto, Toronto, Ontario, Canada
Abstract
Depression represents a complex, multifactorial disorder influenced by the interplay of biological, behavioral, and social determinants. While extensive health surveys, including the National Health and Nutrition Examination Survey (NHANES), gather comprehensive multidomain data, existing studies tend to analyze these domains independently, overlooking the potential to identify integrated, data-driven patterns of depression risk. We have developed SYNERGY-VAE, an interpretable and multimodal deep learning framework that is designed to identify latent health subgroups and facilitate the transparent prediction of depression risk, utilizing NHANES-derived multimodal data from a large analytic cohort that spans five distinct domains from 2005 to 2018. SYNERGY-VAE utilizes a variational autoencoder to learn a shared latent representation derived from five modalities of NHANES, specifically demographic, dietary, examination, laboratory, and questionnaire data. Clustering within this latent space revealed subpopulations exhibiting distinct health signatures. To enhance interpretability, we triangulated insights utilizing encoder weights, standardized mean differences, and permutation feature importance (PFI). Machine learning classifiers were trained within each cluster employing the top 30 PFI-ranked features, with performance evaluated through the receiver operating characteristic area under the curve (AUC) in a 70/30 train-test split. Three latent clusters emerged, each demonstrating markedly different observed prevalence rates of depression within the analytic sample, ranging from 6.8% to 10.9%. Cluster-specific models consistently surpassed pooled models in performance, with the highest predictive accuracy identified in Cluster 2 utilizing XGBoost (AUC = 0.839, 95% CI:0.804–0.874). The importance of features varied across clusters, indicating unique depression risk profiles specific to each subgroup. SYNERGY-VAE demonstrates the power of generative, explainable deep learning for the identification of latent health phenotypes, thereby furthering the objectives of precision mental health. By integrating high-dimensional, multimodal data and facilitating transparent subgroup discovery, our model may inform future stratified and context-aware screening research and contribute to precision psychiatry-oriented research in large health survey datasets. Author summary: Depression rarely has a single cause. It reflects a mix of biology, behavior, and social context that varies from person to person. We asked whether large, routinely collected health surveys could reveal hidden subgroups of people who share meaningful patterns linked to depression. Using a NHANES-derived health survey data, we combined five kinds of information: demographic details, diet, physical measures, laboratory tests, and questionnaire responses. We then used a generative deep learning approach to learn a shared, low-dimensional picture of each person’s overall health profile. Within this shared space, we found three subgroups with clearly different rates of depression and distinct health signatures. When we trained simple prediction models inside each subgroup, performance improved compared with using a single model for everyone. To keep our approach transparent, we used multiple methods to show which features shaped the subgroups and which features were most important for prediction. Our results suggest that multimodal health survey data can uncover useful depression phenotypes and may inform more tailored screening and prevention strategies. All data are public, and our code is openly available for reuse and extension.
Related
- An explainable multimodal framework for exploring depression-associated speech representations in Alzheimer's disease
- Hybrid Explainable Deep Ensemble Framework with self-attention mechanism for early prediction of depression severity using PHQ-9 clinical and behavioral data
- Explainable AI framework for Alzheimer’s disease prediction using multimodal data
- Multimodal fusion of SpO₂ time series and clinical data for explainable classification of obstructive sleep apnea severity
- Optical compression of multimodal clinical data into a unified visual memory for generative trajectory forecasting
BibTeX
@article{sharma2026synergyvae,
title = {{SYNERGY-VAE: An explainable generative deep learning framework for discovering depression subgroups from multimodal population health data}},
author = {Divya Sharma and Alice Rueda and Qiaowei Lin and Shakila Meshkat and Argyrios Perivolaris and Venkat Bhat},
journal = {PLOS Digital Health},
year = {2026},
month = sep,
doi = {10.1371/journal.pdig.0001719},
url = {https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001719},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - SYNERGY-VAE: An explainable generative deep learning framework for discovering depression subgroups from multimodal population health data
AU - Divya Sharma
AU - Alice Rueda
AU - Qiaowei Lin
AU - Shakila Meshkat
AU - Argyrios Perivolaris
AU - Venkat Bhat
PY - 2026
DA - 2026/09/24
JO - PLOS Digital Health
DO - 10.1371/journal.pdig.0001719
UR - https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001719
AB - Depression represents a complex, multifactorial disorder influenced by the interplay of biological, behavioral, and social determinants. While extensive health surveys, including the National Health and Nutrition Examination Survey (NHANES), gather comprehensive multidomain data, existing studies tend to analyze these domains independently, overlooking the potential to identify integrated, data-driven patterns of depression risk. We have developed SYNERGY-VAE, an interpretable and multimodal deep learning framework that is designed to identify latent health subgroups and facilitate the transparent prediction of depression risk, utilizing NHANES-derived multimodal data from a large analytic cohort that spans five distinct domains from 2005 to 2018. SYNERGY-VAE utilizes a variational autoencoder to learn a shared latent representation derived from five modalities of NHANES, specifically demographic, dietary, examination, laboratory, and questionnaire data. Clustering within this latent space revealed subpopulations exhibiting distinct health signatures. To enhance interpretability, we triangulated insights utilizing encoder weights, standardized mean differences, and permutation feature importance (PFI). Machine learning classifiers were trained within each cluster employing the top 30 PFI-ranked features, with performance evaluated through the receiver operating characteristic area under the curve (AUC) in a 70/30 train-test split. Three latent clusters emerged, each demonstrating markedly different observed prevalence rates of depression within the analytic sample, ranging from 6.8% to 10.9%. Cluster-specific models consistently surpassed pooled models in performance, with the highest predictive accuracy identified in Cluster 2 utilizing XGBoost (AUC = 0.839, 95% CI:0.804–0.874). The importance of features varied across clusters, indicating unique depression risk profiles specific to each subgroup. SYNERGY-VAE demonstrates the power of generative, explainable deep learning for the identification of latent health phenotypes, thereby furthering the objectives of precision mental health. By integrating high-dimensional, multimodal data and facilitating transparent subgroup discovery, our model may inform future stratified and context-aware screening research and contribute to precision psychiatry-oriented research in large health survey datasets. Author summary: Depression rarely has a single cause. It reflects a mix of biology, behavior, and social context that varies from person to person. We asked whether large, routinely collected health surveys could reveal hidden subgroups of people who share meaningful patterns linked to depression. Using a NHANES-derived health survey data, we combined five kinds of information: demographic details, diet, physical measures, laboratory tests, and questionnaire responses. We then used a generative deep learning approach to learn a shared, low-dimensional picture of each person’s overall health profile. Within this shared space, we found three subgroups with clearly different rates of depression and distinct health signatures. When we trained simple prediction models inside each subgroup, performance improved compared with using a single model for everyone. To keep our approach transparent, we used multiple methods to show which features shaped the subgroups and which features were most important for prediction. Our results suggest that multimodal health survey data can uncover useful depression phenotypes and may inform more tailored screening and prevention strategies. All data are public, and our code is openly available for reuse and extension.
SN - 2767-3170
ER -