How well could AI get to know you?: The use of AI in the inference and prediction of mental health states.

Imagine a world where weather forecasting was never achieved…

Imagine people sweating through their jumpers during unprecedented heatwaves, people shivering in beer gardens on a particularly brisk spring evening, clad only in shorts and sandals. Thousands of perms across the country, ruined! All due to surprise thunderstorms!

Our ability to forecast weather allows us to plan all these things in advance (when it’s accurate!), avoiding those plans that may not work out & confirming those plans that sound just right.

Now imagine if we could forecast other fluctuations…what if we could predict mental health?

Relapse rates in mental health conditions

Mental health conditions are often chronic and relapsing in nature – 50% of people treated for one episode of depression relapse within the first six months of recovery. Similarly, people treated for anxiety have around a 40% chance of relapsing within 2 years of their first episode. With each episode the level of stress needed to trigger another episode lowers, and so as episodes increase, the likelihood of another one increases. Because episodes of poorer mental health can come and go, wouldn’t it be great to see them coming?

Many utilisations of Artificial Intelligence (AI) have been produced in the past decade to try and consistently monitor mental health to establish patterns that let us predict episodes of mental illness and act pre-emptively with interventions. Recent efforts have combined digital phenotyping and supervised machine learning to produce one product to do it all!

Digital phenotyping has been eloquently described as the ‘moment-by-moment quantification of the human phenotype in situ’. Basically, this just means taking continuous measurements of a person’s physiology throughout the day & finding ways to add a value to what we find. The perfect & most convenient way to do this is through passive sensors already existing in our smartphones! This method allows for completely non-intrusive yet continuous collection of the data we need to see how someone is doing.

Passive data

What kind of data are we collecting from people?

Linear acceleration (m/s) of the device collected through the smartphone’s accelerometer, incl. gravitational acceleration = Tracking physical activity.

Angular acceleration collected through the smartphone’s gyroscope = Tracking physical activity.

Device location expressed in latitude and longitude coordinates = Tracking physical activity.

Applications notifications = Tracking level of online interaction.

Application foreground changes = Tracking phone screen activity & online interaction.

Interactions with the screen (keyboard taps, clicks, long clicks, upward and downward scrolls) grouped by application = Tracking phone screen activity and online interaction.

Screen status changes (on, off, locked, unlocked) = Tracking phone screen activity

We relied on sensors measuring changes in location to guess how physically activity people were throughout the day, as lower levels of physical activity have been linked to increased severity of depression and anxiety (though not always! I hope we haven’t startled anyone reading from their bed, 5 episodes deep into a new Netflix series – please resume!). Sleep is another reliable predictor of mental health, with quality of sleep being proven to have an affect on quality of mental health. Given the sensors we had access to, we used changes in location and in screen status throughout the day to measure hours of sleep each night, as well as how consistent it was.

The Thrive: Well-Being app has recently began focusing on using digital phenotyping to predict mental health states. Currently, the company is running preliminary studies to help get into the practise of building models with the highest accuracy we can achieve, and right now research efforts are focused on learning how to infer before we can predict.

What do we mean by infer?

Thrive’s app operates based on user’s reports of their mental health. However, the user pool is predominantly working professionals, i.e., very busy people! So users often forget or don’t have the time to keep up to date with these measures. By collecting data passively obtained from sensors in smartphones and pairing this with scores for anxiety, depression, resilience and stress, as well as daily mood ratings, we aim to build an algorithm able to infer these missing self-assessment values and offer intervention at the right time.

What did we do?

To do this, we paired questionnaire scores achieved by participants with phenotypical data passively collected over the span of 8 weeks and built a Random Forest Classifier model to predict the scores a person would have achieved on a given day.

What did we find?

From the data we have collected so far, we can see that the model is better able to predict a participant’s GAD-7 score, regarding their levels of anxiety, than their PHQ-9 score regarding the severity of their depression.

What does this tell us?

Well firstly, it tells us that this method of monitoring mental health is viable and can be used to identify when a person may need help. Being able to monitor when treatment is needed using such an everyday device allows for continuous long-term treatment. However, the difference in the accuracy for predicting depression over anxiety may suggest that the physiological markers we focused for both disorders aren’t equally relevant. For example, people with depression can experience insomnia as well as hypersomnia, whereas people with anxiety are more likely to experience shorter sleep cycles with more disruption. Perhaps the sleep variable needs to be weighted differently in our model for each condition.

What’s next?

In future research, it might be important to consider how our methods of measuring certain markers may lead to inaccuracy. For instance, we relied on the location of a person’s phone to make assumptions about their activity – but of course not everyone travels everywhere with their phone on their person. Even the small difference between carrying a phone in your pocket or in a bag can give us different results.

References

Burcusa, S. L., Iacono, W. G. (2007). Risk for recurrence in Depression. Clinical Psychology Review, 27(8), 959–985. https://doi.org/10.1016/j.cpr.2007.02.005
Carek, P. J., Laibstain, S. E., & Carek, S. M. (2011). Exercise for the Treatment of Depression and Anxiety. The International Journal of Psychiatry in Medicine, 41(1), 15–28. https://doi.org/10.2190/PM.41.1.c
Moriarty, A.S., Paton, L.W., Snell, K.I.E. (2021) The development and validation of a prognostic model to PREDICT Relapse of depression in adult patients in primary care: protocol for the PREDICTR study. Diagn Progn Res 5, 12. https://doi.org/10.1186/s41512-021-00101-x
Nutt, D., Wilson, S., & Paterson, L. (2008). Sleep disorders as core symptoms of depression. Dialogues in clinical neuroscience, 10(3), 329–336. https://doi.org/10.31887/DCNS.2008.10.3/dnutt.
Sheikh, M., Qassem, M., & Kyriacou, P. A. (2021). Wearable, environmental, and smartphone-based passive sensing for Mental Health Monitoring. Frontiers in Digital Health, 3. https://doi.org/10.3389/fdgth.2021.662811.
Silva, L. R., Seguro, C. S., de Oliveira, C. G., Santos, P. O., de Oliveira, J. C., de Souza Filho, L. F., de Paula Júnior, C. A., Gentil, P., & Rebelo, A. C. (2020). Physical inactivity is associated with increased levels of anxiety, depression, and stress in Brazilians during the COVID-19 pandemic: A cross-sectional study. Frontiers in Psychiatry, 11. https://doi.org/10.3389/fpsyt.2020.565291.
Staner, L. (2003). Sleep and anxiety disorders. Dialogues in clinical neuroscience, 5(3), 249–258. https://doi.org/10.31887/DCNS.2003.5.3/lstaner.
Torous, J., Kiang, M. V., Lorme, J., & Onnela, J. P. (2016). New tools for new research in psychiatry: a scalable and customizable platform to empower data driven smartphone research. JMIR mental health, 3(2), e5165.
van Berkel, N., D’Alfonso, S., Kurnia Susanto, R. et al. AWARE-Light: a smartphone tool for experience sampling and digital phenotyping. Pers Ubiquit Comput (2022).
Yonkers, K., A., Bruce, S., E., Dyck, I., R., Keller, M., B. (2003). Chronicity, relapse, and illness-course of panic disorder, social phobia, and generalized anxiety disorder: findings in men and women from 8 years of follow-up. Depress Anxiety 2003; 17:173179.