Machine learning reveals sleep quality and anxiety as major predictors of depression

By Vijay Kumar MalesuReviewed by Lily Ramsey, LLMSep 12 2024

In a recent study published in the eBioMedicine, a group of researchers predicted depressive symptom severity (DSS) (intensity or degree of depressive symptoms an individual is experiencing) using anxiety, sleep quality, and brain structural measurements while testing the generalizability of the findings across three independent datasets.

​​​​​​​Study: Prediction of depressive symptoms severity based on sleep quality, anxiety, and gray matter volume: a generalizable machine learning approach across three datasets. Image Credit: WPixz/Shutterstock.com

Background 

About 25% of the population in modern societies experiences depressive symptoms, which have increased over time and can lead to major depressive disorder (MDD) within 15 years. Early detection is crucial to reducing depression rates.

Sleep disturbances, particularly insomnia, are linked to depression, but the relationship is complex due to individual differences in genetics, anxiety, and stress responses.

Further research is needed to understand better the interactions between sleep, anxiety, brain structure, and depressive symptoms and to enhance the accuracy and generalizability of predictive models across diverse populations.

About the study 

The Human Connectome Project (HCP)-Young dataset, acquired by the Washington University-University of Minnesota (WU-Minn HCP) consortium, includes healthy young adults aged 22-35, excluding those with current neurological, psychiatric, cardiovascular conditions, substance abuse, or treatments.

The study selected 1,101 individuals with complete structural Magnetic Resonance Imaging (MRI), sleep quality, anxiety, and depressive symptoms data from 1,206 participants. A secondary analysis excluded those with a history of clinical depression. The HCP-Aging dataset recruited adults aged 36 to over 100, but due to the design of the DSS questionnaire, only participants aged 36-59 were included.

Lastly, the Enhanced Nathan Kline Institute (eNKI) dataset, a community-representative sample, provided both cross-sectional and longitudinal data for participants aged 18-59. This allowed researchers to assess the generalizability of machine learning models and predict future DSS based on baseline sleep and anxiety measures.

Ethical approvals for each cohort are available online, with analysis permission granted by the University Hospital of the Heinrich-Heine University Düsseldorf.

Data were processed through advanced neuroimaging and machine learning techniques, focusing on gray matter volume and neurobehavioral predictors. Predictions were tested on independent datasets to validate the models.

Mediation analyses further explored the relationships between sleep quality, anxiety, and DSS, helping to clarify how brain and behavioral factors contribute to depressive symptoms across populations.

Study results 

The primary dataset used in this investigation, HCP-Young, consisted of 1,101 participants aged 22-35 years (mean age = 28.79 ± 3.69), with 54.3% female participants. Notably, 9% (103 individuals) had a history of Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV)-based depression episodes.

The study used two additional datasets for out-of-cohort validation: HCP-Aging and eNKI. The HCP-Aging dataset included 378 participants aged 36-59 (mean age = 47.3 ± 7), with 57.9% female. The eNKI dataset had 334 participants with cross-sectional data, aged 18-59 (mean age = 37 ± 13.8), and 62% were female.

Additionally, 66 participants from eNKI had longitudinal records, with a gap of 1-5 years between their two visits. Among these, 26 participants received neurofeedback therapy between their visits, with a gap of 653 days on average, while 40 participants without therapy had a gap of 847 days.

In the HCP-Young dataset, machine learning (ML) models using sleep quality could predict DSS (r = 0.43, R2 = 0.18, rMSE = 2.73). Adding anxiety to sleep quality improved the prediction significantly (r = 0.67, R2 = 0.45, rMSE = 2.25), whereas adding gray matter volume (GMV) did not enhance the results.

The study found no feature redundancy among sleep quality scores, and ensemble learning (LS-boost) was automatically selected for all models. Removing participants with a history of depression yielded similarly robust results (r = 0.61, R2 = 0.37, rMSE = 2.18).

Sleep-related components such as daytime dysfunction, sleep disturbance, and subjective sleep quality were the most predictive of DSS. Importantly, reverse predictions (DSS predicting sleep quality) were weaker, suggesting that sleep quality is a better predictor of DSS than vice versa.

The study also replicated its findings in the HCP-Aging and eNKI datasets, demonstrating the generalizability of ML models trained on the HCP-Young dataset.

In the HCP-Aging dataset, sleep quality predicted DSS (r = 0.57, R2 = 0.27, rMSE = 2.64), with the addition of anxiety further improving the prediction (r = 0.72, R2 = 0.50, rMSE = 2.19). Similar results were found in the eNKI dataset.

In the longitudinal eNKI subsample, ML models predicted future DSS based on baseline sleep quality and anxiety, but neurofeedback therapy appeared to diminish this predictability. This suggests that interventions like neurofeedback therapy could alter the relationship between sleep quality and DSS over time.

Conclusions 

To summarize, the findings demonstrated that sleep quality could reliably predict DSS across three independent datasets. Adding anxiety to the model further enhanced predictive accuracy. However, structural and functional brain measurements did not significantly predict DSS or mediate the relationship between sleep quality and DSS.

The ML models yielded similar results across independent datasets, underscoring the generalizability of the approach.

In a longitudinal subsample, the models also successfully predicted future DSS based on baseline sleep quality and anxiety. Complementary analyses confirmed the robustness of these findings, considering variables such as depression history, sleep-related confounders, and anxiety's mediating role.