Summary: Yes, a Fitbit has the potential to improve bipolar disorder treatment by giving providers real-time insight into physiological markers for manic and depressive phases and symptoms.
Key Points:
- Wireless fitness devices can act as monitors for specific physiological metrics related to BD
- Providers can check in on patients on a regular schedule
- Data can help tailor treatment to match symptoms
- Research is needed to determine best methods and practices, but early precision and accuracy results are promising
Tracking Mental Health Symptoms to Fine-Tune Treatment
Ongoing research into optimal treatment protocols for mental health disorders is an essential component of public health. In this article, we’ll review a recent study on a new concept: using data from wearable fitness tracking devices, such as a Fitbit, to improve bipolar disorder treatment.
In 2024, a group of researchers published the study “Digital Phenotyping in Bipolar Disorder: Using Longitudinal Fitbit Data and Personalized Machine Learning to Predict Mood Symptomatology.” As we mention above, the researchers hypothesized data from wearable fitness trackers could help identify current, real-time symptomology in patients with bipolar disorder.
Here’s how they describe the study
“[We] evaluated whether a novel, personalized machine learning approach, trained entirely on passive Fitbit data, with limited data filtering, could accurately detect mood symptomatology in [patients with bipolar disorder]”
The machine learning (ML) approach they tested is called Binary Mixed Model (BiMM). While an in-depth discussion of machine learning and what makes BiMM a novel ML approach is beyond the scope of this article, here’s a quick explainer on how ML works:
- Algorithm collects data
- Data processed, prepared for analysis
- Programmers choose a model to generate outcomes
- In this case, BiMM
- Programmers fine-tune model parameters
- Model analyzes data
- Model generates a prediction, conclusion, or inference.
In this study, researchers used a previously trained model – the BiMM we mention above – to analyze Fitbit data and determine – or predict – the current state of symptomology for patients with bipolar disorder.
Let’s take a closer look.
Identifying Current Symptoms With Fitbit Data + Machine Learning
The two categories of symptomology researchers examined were depression and hypomania, which are the two primary categories of symptoms associated with bipolar disorder (BD). Let’s clarify that, with a basic definition of bipolar disorder:
“In bipolar disorder, episodes of depression alternate with episodes of mania, or a less severe form of mania, hypomania. Mania is characterized by excessive feelings of elation and confidence or physical activity out of proportion to the situation, while depression is characterized by persistent feelings of sadness and hopelessness.”
To learn more about bipolar disorder(s), please navigate to our bipolar disorder treatment page:
Bipolar Disorder Treatment at Crownview
To learn more about depression, please navigate to our depression treatment page:
Inpatient Depression Treatment Center
Understanding the content on those pages will help understand the symptoms the researchers examined, which brings us back to the study.
As we mention above, the symptoms researchers examined were depression and hypomania. The sample set included 54 adult patients with bipolar disorder. Patients wore Fitbits for six months and completed self-report/self-assessment questionnaires every two weeks. Here’s the data researchers collected from the fitness trackers:
- Step count
- Sedentary minutes
- Very active minutes
- Heart rate
- Resting heart rate
- Total sleep time
- Sleep efficiency
- Deep sleep: count (periods of deep sleep), duration (time in deep sleep), percent (percent of time from falling asleep to getting up next day)
- REM sleep: count, duration, percent (see deep sleep)
- Awakenings during sleep: count, duration, percent (see deep sleep/REM sleep)
- Median bedtime
After collecting and examining data, researchers identified the metrics that contributed the most significantly to predicting/identifying symptomology.
For depression, the following five variables contributed most:
- Duration of awakenings
- Total sleep time
- Median bedtime
- Resting heart rate
- Percentage of sleep spent in deep sleep
For mania/hypomania, the following five variables contributed most:
- Heart rate
- Sleep efficiency
- Percentage of sleep spent in rem sleep
- Number of very active minutes
- Median bedtime
Let’s see what they found.
Fitbit Data for Bipolar Disorder Treatment: Can it Help?
No spoilers: yes, it probably can help. More than probably. If this vein of research continues yielding similarly promising results, it’s very likely that in the near future, machine learning algorithms will be able to help clinicians monitor symptomology in patients in real-time or near-real-time, using data from devices like wearable fitness monitors, e.g. Fitbits.
In this study, the chosen ML algorithm, BiMM, accurately identified/predicted bipolar symptoms at the following rates:
- Depression: 80.1% accuracy
- Sensitivity: 71.2%
- Specificity: 85.6%
- Hypomania: 89.1% accuracy
- Sensitivity: 80.0%
- Specificity: 90.1%
In this context, sensitivity means true accuracy and specificity means how well a model eliminates false positives.
Regarding depression, the information above means that 71.2 percent of the time, the ML prediction matched self-report and clinician data on depression, and 85.6 percent of the time, it eliminated patients without depressive symptoms from the result, generating an overall accuracy score of 80.1 percent.
Regarding hypomania, the information above means that 80.0 percent of the time, the ML prediction matched self-report and clinician data on hypomania, and 90.1 percent of the time, it eliminated patients without symptoms of hypomania from the results, generating an overall accuracy score of 89.1 percent.
Is That Accurate Enough?
While we’re not statisticians, it appears as if these tests are approaching a level of accuracy that may be beneficial for patients with bipolar disorder.
To put this in context, it’s important to understand that almost no medical tests are one hundred percent accurate, and almost none achieve perfect sensitivity and specificity. For example, information from this article shows the following:
- Digital mammography tests have 97% sensitivity and 64.5% specificity, meaning they accurately diagnose breast cancer in nearly all people who have it (and take the test), but they may yield a significant number of false positives.
- A standard test for diabetes – the HbA1c test – shows 21% sensitivity and 94% specificity, meaning it’s bad at identifying diabetes, but good at ruling out the presence of diabetes among patients tested.
Let’s consider something else: this study is not about diagnosis. It’s about using a Fitbit or similar device – and this is oversimplifying – to allow a clinician or provider to check in on their patient, analyze the data, and understand where – in terms of the disorder – their patient is at that time.
Therefore, using this approach, if a provider notes the patient is in a depressive phase, and they’re between appointments, they can reach out to their patient and offer support. Likewise, if they check the data and note their patient is in a manic or hypomanic phase, and between appointments, they can reach out and offer appropriate support.
While this method is not ready quite yet, it is promising. We’ll close with an observation by lead researcher Dr. Jessica Lipschitz, interviewed in the online journal Science Daily:
“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.”
In other words, the hope is that this method will allow treatment providers to respond more quickly to immediate patient needs, and improve outcomes: if and when that happens, it may be an important step forward and improve bipolar disorder treatment for years to come.
Gianna Melendez
Jodie Dahl, CpHT