The AI that can Spot Markers of Depression in Instagram Photos
A team of researchers from Harvard and the University of Vermont have developed an AI program which can identify clinical depression in Instagram users at an accuracy rate of 70 percent- just by looking at their posts.
The new machine learning program was reportedly applied to monitor and analyse the content on 166 separate Instagram profiles, with a combined 43,950 different images. According to the study, published by Andrew G. Reece and Christopher M. Danforth, the AI program analyses content to detect specific potential indicators of depression.
Alongside an analysis of metadata, the system studies aspects such as colour, and uses facial recognition to imitate personal assessment. By looking at earlier studies, the team were able to deduce that depressed individuals often opted for darker, greyer colour-scales.
Other factors included an examination of posting rates and a look at the sort of engagement generally received. According to previous investigations, depressed individuals often received a low rate of likes next to a higher proportion of comments.
The study says this:
“Using only photographic details, such as colour and brightness, our statistical model was able to predict which study participants suffered from depression, and performed better than the rate at which unassisted general practitioners typically perform during in-person patient assessment.”
[An analysis of earlier studies found general practitioners correctly diagnosed depression at a rate of 42 percent.]
As a comparison study, the group also had a selection of participants (unaware that the study was linked to identifying depression) rate each photograph, in order to better gauge how the average person is able to spots depression markers in Instagram photos. The same machine learning system was also applied to photos only posted before each individual was diagnosed, which performed at a successful diagnosis rate of over 50%.
Although the system is still in its infancy, with one or two shortcomings, it is nevertheless a solid foundation for subsequent models.