Artificial Intelligence continues to amaze with its thousands of applications. Every day there are articles of different markets telling us about uses that are serious or half-serious and others very serious. Just recently I came across another interesting and serious study, reported in an article published by the Proceedings of the National Academy of Sciences (PNAS) in the United States.
The assumption is interesting and serious: given that we use social media to communicate and, when we do so, we expose – even unconsciously – moods, representing them with words; some scholars decided to use analysis of Facebook posts as a first element of study for prediagnosing a tendency towards depression.
Early diagnosis on Facebook
In the United States, the country of the study, the problem of early diagnosis of depression is certainly relevant given that, as the study highlights, every year the percentage of the population living in a state of depression ranges between 7 and 26%, and only between 13 and 49% of them receive minimally adequate treatment.
This low diagnostic capacity suggested to the team that worked on the study the possibility of evaluating and finding alternative tools to identify patients at risk of depression, even before they present themselves at a healthcare facility.
The algorithm that reads sadness
We start from the axiom that the language we use is the litmus test of our unconscious: references to sadness, loneliness, hostility and an increase in self-referencing, could indicate an inclination towards a depressive state.
The team worked “backwards” on a sample of 683 consenting individuals, 114 of whom had already been diagnosed – in the standard way – with a state of depression, asking for and obtaining their consent to examine the archive of the Facebook posts of the 6 months prior to diagnosis.
In the first tests, the Machine Learning algorithm behaved like existing screening questionnaires – and currently used to identify depression – but with the clear advantage of being able to perform the analysis discretely and in the background.
Reading of the posts of patients that were given as a sample showed that an appropriately trained algorithm is able to read and detect linguistic behaviors typical of the first manifestations of the illness. In fact, the team built an algorithm that takes into account various elements of the posts, including content, length, frequency of publication and repetitiveness.
It then correlated the results of “diagnosis” of the algorithm with the presence or absence of depression of each patient in the reference cluster (diagnosis performed with classical methodology).
The result of the study is very interesting, because it demonstrates the ability of the system to identify the symptoms of a probable depression with an advance window of more than 3 months compared with the actual diagnosis (namely, when the patient perceives the problem and asks for help from a doctor).
Although this accuracy of prediction is still modest, it suggests nevertheless that, possibly in combination with other forms of non-invasive digital screening, the potential exists for developing indicators of mental disorders in advance of actual diagnosis, such as to help the patient before the situation worsens.