Researchers from South Korea have built machine learning-based models that can predict mood episodes using only sleep and circadian rhythm data from wearable devices.
The team comprised researchers from the Institute for Basic Sciences (IBS), Korea Advanced Institute of Science and Technology (KAIST), and Korea University College of Medicine.
FINDINGS
In a study, which findings were published in Nature’s npj Digital Medicine journal, the research team first collected and analysed 429 days’ worth of sleep-wake data generated from Fitbit of 168 Korean patients with mood disorders, including major depression and bipolar disorders. It is said that mood disorders are closely associated with irregular sleep and circadian rhythms.
From this dataset, they extracted 36 sleep and circadian rhythm features, which were then applied to train models based on the machine learning library XGBoost to predict mood episodes.
Based on findings, the predictive models achieved 80%, 98%, and 95% accuracy in predicting depressive, manic, and hypomanic episodes, respectively.
Findings also suggested that daily changes in circadian rhythm may be a key predictor of mood episodes; delayed circadian rhythms can potentially lead to depressive episodes while advanced circadian rhythms can raise the possibility of manic episodes.
WHY IT MATTERS
While there already exist AI models for predicting the mood states of people with mood disorders, these typically require various data including sleep, heart rate, light exposure, phone usage, and GPS data, which may be costly to collect. It may also pose security and privacy risks and storage challenges, the research team emphasised.
“By developing a model that predicts mood episodes based solely on sleep-wake pattern data, we have reduced the cost of data collection and significantly improved clinical applicability,” claimed Kim Jae Kyoung, KAIST associate professor and IBS chief investigator.
The research team also point to the potential of complementing the prediction of mood states with the use of digital therapeutics to evaluate the daily risk of mood episode relapse and promote healthy sleep-wake cycles and circadian rhythms by providing reminders.
MARKET SNAPSHOT
A similar research in Singapore also utilised sleep patterns and circadian rhythms to predict depression risk. The AI-based Ycogni model by a research team from Nanyang Technology University demonstrated 80% accuracy in identifying those at high risk of depression using also sleep-wake data collected from Fitbit.
Early this year, research from Japan developed what could be the world’s first AI model that can predict Alzheimer’s disease using wearables data. Eisai and Oita University collected and analysed a range of biological and lifestyle data to develop a predictive model that can be used as a pre-screening tool for people suspected of having the disease.
