Researchers have used everyday Fitbit data to train a machine learning algorithm to accurately predict mood episodes associated with bipolar disorder. It opens the door to using a personalized algorithm to drive treatment of the life-impacting condition.
Bipolar disorder’s (BD) characteristic mood episodes – the extreme swings between depression and mania, followed by a period of remission – can have a huge impact on a person’s work, relationships, and health. Treatment of BD is focused on limiting that impact, which requires the prompt identification and treatment of mood episodes.
Leading a new study aimed at finding an accurate way of detecting mood episodes in people with BD, researchers from the Brigham and Women’s Hospital (BWH) in Boston turned to a now-ubiquitous health-monitoring device, the Fitbit.
“Most people are walking around with personal digital devices like smartphones and smartwatches that capture day-to-day data that could inform psychiatric treatment,” said Jessica Lipschitz, PhD, from BWH’s Department of Psychiatry and the study’s lead author. “Our goal was to use that data to identify when study participants diagnosed with bipolar disorder were experiencing mood episodes.”
Studies have shown that most people with BD, which used to be called manic depressive disorder or manic depression, experience a change in symptom severity and mood ‘polarity’ at least three times a year. This includes going from feeling very happy, irritable, with a marked increase in activity level (mania), to feeling sad, indifferent, or hopeless with very low activity levels (depression). Hypomania is like mania but less severe; it doesn’t cause the impairment in social or work functioning that manic episodes do.
There are two types of BD: bipolar I disorder and bipolar II disorder. BP-I is defined by manic episodes that last at least seven days (most of the day, almost every day) or mania so severe that hospitalization is required. Separate depressive episodes usually occur as well and typically last at least two weeks. Some people with BP-I experience what’s referred to as ‘rapid cycling,’ where they have more than four episodes of mania or depression in one year. BP-II is characterized by a pattern of depression and hypomania.
For the present study, the researchers recruited 54 adults diagnosed with BP-I or BP-II and asked them to wear a Fitbit continuously for nine months. The Fitbit Inspire was chosen for its ability to collect data on activity, heart rate, and sleep. The participants were also asked to self-report depression and mania symptoms every two weeks over the same nine-month period.
The data, which included 17 variables such as step count, very active minutes, sedentary minutes, heart rate and resting heart rate, total sleep time, sleep efficiency score, deep sleep duration, REM sleep duration, and bedtime, was used to train a predictive machine learning algorithm. The algorithm was able to figure out the importance of each variable in predicting clinically significant symptoms of depression and mania.
The algorithm accurately predicted 89.1% of clinically significant hypomanic or manic symptoms (with a sensitivity of 80.0% and a specificity of 90.1%) and 80.1% of clinically significant depressive symptoms (sensitivity of 71.2%, specificity of 85.6%). Sensitivity refers to a test’s ability to correctly identify patients with a condition; specificity is its ability to correctly identify people without that condition.
The five variables that contributed most to predictions of depression were duration of awakenings, total sleep time, median bedtime, resting heart rate, and percentage of sleep spent in deep sleep. For predictions of mania or hypomania, the top five variables were heart rate, sleep efficiency, percentage of sleep spent in REM sleep, number of very active minutes, and median bedtime.
“Our findings are particularly noteworthy because all input was passively collected, none of the metrics utilized were invasive in terms of privacy, we used mainstream consumer devices, and our methods did not demand high levels of Fitbit compliance,” the researchers said. “Other researchers have achieved more accurate mood predictions with more invasive data collection protocols that use data like geolocation and voice features, which may raise privacy concerns, and textile wearables, which may feel restrictive to patients.”
The findings have the potential to transform models of care in BD and improve treatment precision.
“In the future, our hope is that machine learning algorithms like ours could help patients’ treatment teams respond fast to new or unremitting episodes in order to limit negative impact,” said Lipschitz.
The study was published in the journal Acta Psychiatrica Scandinavica, and is available as an early view PDF.
Source: BWH via EurekAlert!