Wearable sensors such as smartwatches that collect physical and physiological data may be powerful tools in the effort to better understand brain and behavioral illnesses and their genetic drivers, a new Yale study finds.
Using smartwatch data collected from more than 5,000 adolescents, Yale researchers were able to train AI models to predict whether individuals had different psychiatric illnesses and uncover illness-associated genes. The findings, published Dec. 19 in the journal Cell, suggest wearable sensors may enable a much more nuanced understanding — and treatment — of psychiatric illness.
“In traditional psychiatry, a doctor will assess your symptoms and you’ll either be diagnosed with an illness or you won’t,” said co-senior author Mark Gerstein, the Albert L. Williams Professor of Biomedical Informatics and professor of molecular biophysics and biochemistry, of computer science, and of statistics and data science. “But in this study, we focused on processing the wearable data in a meaningful way that could both be leveraged to predict illnesses more comprehensively, and to better connect with underlying genetic factors.”
Detecting illness in such a quantitative way is difficult. But wearable sensors, which collect data continuously over time, may be the answer.
These findings suggest smartwatch data can give us insights into how physical and behavioral temporal patterns relate to different psychiatric illnesses.
For the new study, researchers used data from the Adolescent Brain Cognitive Development Study, the largest long-term assessment of brain development and child health in the United States. The data used in the study — collected from smartwatches worn by adolescents between the ages of 9 and 14 — included measurements of heart rate, calorie expenditure, activity intensity, steps taken, sleep level, and sleep intensity.
“Smartwatch data, when processed correctly can be used as a ‘digital phenotype,’” said Jason Liu, a postdoctoral associate in Gerstein’s lab and co-lead author of the study. Phenotype refers to observable traits, such as eye color or height. The researchers propose using the term “digital phenotype” to describe the traits that can be measured and tracked with digital tools like smartwatches.
“One advantage of doing that means we can use the digital phenotype almost as a diagnostic tool or a biomarker, and also bridge the gap between disease and genetics,” said Liu.
To that end, the researchers also developed a way to take the massive amount of smartwatch data, align it so that the various time points matched up across individuals, and convert the raw data into something that an AI model could be trained on.
“This was a new problem to solve in the research world and it was very technically challenging,” said Gerstein.
The researchers trained machine learning models to predict whether an individual had attention-deficit/hyperactivity disorder (ADHD) or an anxiety disorder based on either the dynamic smartwatch data or a snapshot of the data that summarized what was collected over time. They found that digital phenotypes significantly improved model accuracy and the best models leveraged the dynamic data (rather than the snapshot), suggesting the additional temporal details were useful in characterizing illness.
Heart rate was the most important measure for predicting ADHD, the team found, while sleep quality and stage (the different cycles a body passes through during sleep) were more important for identifying anxiety.
“These findings suggest smartwatch data can give us insights into how physical and behavioral temporal patterns relate to different psychiatric illnesses,” said Gerstein.
Moreover, the data could also help differentiate between different subtypes of the disease.
“For instance, within ADHD there are different forms,” said Beatrice Borsari, a postdoctoral associate in Gerstein’s lab and co-lead author of the study. “Maybe we can expand this work to help distinguish between inattentive and hyperactive forms, which typically respond to different pharmacological treatments.”
Having shown that the digital phenotype could be used to predict psychiatric illness, the researchers then investigated whether it could also help identify underlying genetic factors. When they examined whether genetic mutations affected the smartwatch-collected data differently in healthy individuals than it did in those with ADHD, they identified 37 genes related to ADHD. But when they ran a similar analysis to determine whether particular genes were associated with an ADHD diagnosis they found none. This, they said, highlights the added value of using the continuous smartwatch data.
Together, the findings link psychiatric illness, digital phenotypes, and genotypes and show how wearable sensors may yield a deeper understanding of psychiatric disease.
“Such a method holds significant promise for addressing longstanding challenges in psychiatry and may ultimately reshape how we understand the genetics and symptom structure of psychiatric disorders,” said Walter Roberts, assistant professor of psychiatry at Yale School of Medicine and co-senior author on the study.
And while the study focused on ADHD and anxiety, the researchers expect the approach may be more broadly applicable. For instance, it may be useful for understanding neurological disease or neurodegeneration. Further, they hope their findings can serve as inspiration to move beyond traditional clinical diagnoses and adopt more quantitative behavioral measurements that can have greater utility in identifying genetic biomarkers.
