Predicting mood episodes with sleep data: A breakthrough for mental health care

By Dr. Sanchari Sinha Dutta, Ph.D.Nov 21 2024

Discover how analyzing simple sleep patterns can revolutionize mental health treatment by predicting mood episodes with cutting-edge accuracy.

Study: Accurately predicting mood episodes in mood disorder patients using wearable sleep and circadian rhythm features. Image Credit: Chay_Tee / Shutterstock

A study published in the journal NPJ Digital Medicine describes the development of mathematical models that can accurately predict future mood episodes using only sleep-wake history and past mood episodes of patients with mood disorders.

Background

Sleep disturbance has a strong association with various mood disorders, including major depressive disorder and bipolar disorder. Monitoring mood-related sleep and activity patterns is generally conducted using wearable devices and smartphone sensors that noninvasively and passively collect physiological and behavioral data from patients in real-life settings.

Previous studies utilizing physiological data from wearable devices to develop machine-learning models have shown promising results in identifying individuals at risk of depression. A combination of machine learning and wearable technology has been used to predict daily mood status and control mood episodes in patients with psychological disorders. However, these models require several types of data, including sleep, heart rate, light exposure, phone usage, and GPS mobility, which restricts their application in real life.

In this study, scientists have developed a mathematical model for predicting mood episodes that requires only patients' histories of binary sleep-wake patterns and past mood episodes.

Development of the Model

Scientists collected sleep-wake pattern data and previous mood episode period data from 168 patients with major depressive disorder or bipolar disorder, aged 18–35 years, and of Korean ethnicity. The sleep dataset included complete sleep records for a minimum of 30 days.

They processed sleep-wake pattern data and derived a total of 36 sleep and circadian rhythm features, which were used as inputs for the machine learning classification algorithm that aimed to predict future depressive, manic, and hypomanic episodes in patients with mood disorders. By analyzing the significance of these features, they found strong associations between sleep, circadian features, and mood episodes.

Key features included circadian phase and amplitude Z-scores as well as wake time during long sleep windows, which emerged as the most important predictors of mood episodes.

Model Validation

Scientists validated the model's predictive efficacy using patients’ previous sleep-wake patterns and circadian rhythm data. They selected a specific 60-day range for each patient, half of which represented episodic days. They included data spanning this range in the model training set and used the data following the training range for validation.

Using training data from the 60-day range, the model accurately predicted next-day depressive, manic, and hypomanic episodes, with AUC (Area Under the Curve) values of 0.80, 0.98, and 0.95, respectively. The accuracy level of predicting manic and hypomanic episodes remained high when a 30-day range of training data was used. However, with reduced training data, the model's predictive accuracy for depressive episodes decreased significantly. Given these observations, scientists mentioned that sufficient training data is required for the model to maintain its high predictive accuracy.

The study noted challenges in predicting hypomanic episodes due to potential non-monotonic relationships between the circadian phase and mood states. This variability highlights the need for further research into distinct circadian profiles during hypomania.

Study Significance

Collectively, the study findings indicate that the model can accurately predict subsequent mood episodes by analyzing sufficient sleep-wake patterns around the time of a patient’s first mood episode. Since any changes in medication types or dose levels can affect patients’ circadian phases, scientists used a new testing set to determine the impact of medication changes on the model's predictive accuracy. They constructed testing sets containing data from the patients who did not change their medication types and dosages after the episode onset. Using these testing sets, they confirmed that the model’s predictive accuracy is not associated with the medication-induced changes in the circadian phase.

The study describes developing and validating a mathematical model that can accurately predict future mood episodes only from binary sleep-wake pattern data. The study also finds that key circadian features, including phase and amplitude shifts, are the most significant predictors of mood episodes, with the delayed phase linked to depressive episodes and the advanced phase linked to manic episodes.

Previous studies conducted at the molecular level have identified the association between circadian rhythm and mood disorders. These studies have shown that changes in the circadian rhythm may lead to abnormalities in serotonergic and dopaminergic circuits through alterations in rhythms of the circadian nuclear receptor REV-ERB alpha, which plays a pivotal role in the development of depressive and manic episodes.

Advantages and Limitations

Overall, the prediction model developed in this study opens up a new path toward diagnosing and treating mood episodes. The model's main advantage is that it requires only sleep-wake pattern data, which can be collected passively and easily through smartphones or wearable devices. Unlike previous studies that relied on simple sleep metrics like length and efficiency, this study incorporated comprehensive sleep features and a mathematical model to estimate circadian rhythms.

However, the study has some limitations. It included only patients who were compliant with wearable devices, and the sample was restricted to early-stage mood disorder patients in South Korea, limiting generalizability. Additionally, the observational nature of the study and reliance on wearable devices, which can be less precise than laboratory-grade measurements, were noted as constraints.

The authors suggest that future developments could focus on individualized prediction models tailored to patients’ specific circadian profiles and sleep patterns. This approach could enhance the accuracy and applicability of these tools for personalized mental health management.