I’ve written a few times about both commercial applications and research projects that aim to provide earlier detection of various medical conditions, including Alzheimer’s and Parkinson’s, simply from listening to our voice. The range of applications continues to grow, with a team from the University of Alberta showcasing a system that can perform the same trick for depression.
The work, which was documented in a recently published paper, used standard benchmark sets of audio recordings to develop a method that utilizes a range of machine-learning algorithms to provide a more accurate diagnosis than traditional approaches.
The study built upon previous research that highlighted the key role the timbre of our voice plays in helping to identify our mood. The team hopes that they will ultimately be able to develop a commercial application to help users to self-diagnose.
“A realistic scenario is to have people use an app that will collect voice samples as they speak naturally. The app, running on the user’s phone, will recognize and track indicators of mood, such as depression, over time. Much like you have a step counter on your phone, you could have a depression indicator based on your voice as you use the phone.”
With depression affecting around 15% of Canadians at some point in their lives, the team believes that such a tool could be vital in helping people to better reflect on their moods over time and gain a better insight into their own mental wellbeing.
“This work, developing more accurate detection in standard benchmark data sets, is the first step,” the authors conclude.
It joins a number of other interesting projects working along similar lines. 18 months or so ago I wrote about some interesting new research that was able to accurately predict depression purely from listening to our voice. Whilst it’s perhaps fair to say that such work has yet to make it to market, a recent study from the University of Vermont highlights the work still being done in the field.
It reveals the use of machine learning to spot signs of depression and anxiety in one of the 20% of young children who are believed to suffer from some form of anxiety or depression. Diagnosing the condition in children that young is difficult as they’re not able to reliably articulate the feelings they’re having, which can lead to many children remaining undiagnosed, and therefore untreated.
As with so many conditions, the earlier you can diagnose mental health issues in children the better, but this is especially so in young children as their brains are still developing and so untreated conditions can easily develop into things such as substance abuse later in life. The traditional method of diagnosis involves a semi-structured interview with a clinician, but the researchers believe AI can do a better job, both in terms of accuracy and speed.
The researchers developed a modified version of the Trier-Social Stress Task that’s designed to invoke feelings of stress and/or anxiety in a subject. This involved asking a group of volunteer children to improvise a short story, with each story judged based upon how interesting it was. The judge was instructed to give either neutral or negative speech, whilst holding a stern expression throughout the speech. Alongside the test, each child was also given a structured clinical interview to test for various issues.
A machine learning algorithm was then used to analyze various statistical features of the audio recordings of each story to relate them to the diagnosis of that child. The researchers found that this algorithm was successful at diagnosing the children, with the middle point in the story the most valuable part in terms of predictive ability. The researchers explain,
“The algorithm was able to identify children with a diagnosis of an internalizing disorder with 80% accuracy, and in most cases that compared really well to the accuracy of the parent checklist.”
The team plans to further develop the system into a screening tool that can be used in a clinical environment, potentially even via an app. It could even be combined with motion analysis or other diagnostic tools to better identify children at risk of depression and anxiety, with the team confident that they can provide the diagnosis before even their parents suspect anything is wrong.