Leveraging artificial intelligence for the assessment of severity of depressive symptoms


This project is supported by the Abdul Latif Jameel Clinic for Machine Learning in Health (J-Clinic), the National Institutes of Health, Massachusetts General Hospital, and the MIT Media Lab Consortium Member Companies.

Depressive disorders are ranked as the single largest contributor to non-fatal health loss (7.5% of all Years Lived with Disability), affecting an estimated 300 million people[1]. Evidence-based treatments are available and measurement-based care has been described as the gold standard. Monitoring of depressive symptoms is currently performed with self-administered and interview-based assessment methods conducted by clinicians in their offices. However, the shortage of mental health specialists and the limited resources available to primary care physicians who often manage patients with depression prevent close monitoring of symptoms, delaying optimal treatment, and potentially prolonging suffering. Passive recording of behavioral data (gathering information without an individual's direct input) has been identified as a potentially feasible method for long-term monitoring of depression. 

During the past decade, along with the development of wearable sensors, we have seen the progressive use of machine learning, which has allowed for the development of complex models. The combination of sensor technology and machine learning enables detailed measurement in real time of a wealth of behaviors predicting variation of depression. Our interdisciplinary team, including one of the leading labs on depression research at the Massachusetts General Hospital and the Affective Computing group at the MIT Media Lab, has conducted a study applying machine learning analytics to create a model combining wristband sensor data and phone-based passive measurements to assess depression. In our pilot study with chronically depressed patients monitored over eight weeks, we found that an algorithm based on biological and behavioral sensor data could estimate depression severity evaluated by a clinician with high accuracy, comparable to the inter-rater reliability.

In this project, we build from our pilot study and develop an objective, passive, sensor-based algorithm able to detect depression and early response as well as predict response. We will monitor  for 12 weeks 100 adults with Major Depression Disorder who just started treatment. The identification of reliable, objective, passive assessment of depression with biosensors will have significant ramifications for the monitoring of depression, early detection of response, and ultimately contribute to the advancement of precision medicine.

[1] World Health Organization. "Depression and other common mental disorders: global health estimates." (2017).


affective computing group