Forum

AI offers huge benefits for mental health field

Editorials featured in the Forum section are solely the opinions of their individual authors.

In August of last year a team of researchers from New York, Argentina, and Brazil published a proof of concept paper concluding that machine learning techniques that analyzed people’s speech patterns helped predict the onset of psychosis. This idea has caught on with many other psychologists, including NeuroLex Diagnostics CEO Jim Schwoebel, who is using the idea to develop a tool to help diagnose schizophrenia.

A data-based approach is a necessary step in the future of psychology and psychiatry and can help the field respond to its patients’ needs more efficiently and efficiently.
All the way back in 2006, researchers from West Virginia University published a paper concluding that psychiatric diagnoses were not reliable, meaning that the same symptoms could turn up different diagnoses with different patients and doctors. They attributed the unreliability to the lack of standardization among psychiatrists and proposed a method of diagnostics. However, psychiatric care continues to be a highly individualized process with different doctors having different methods. If diagnoses mean different things to different doctors, then a patient never gets help because their symptoms mean one thing to their psychiatrist and another thing to a psychopharmacologist, making communication and collaboration between their care providers impossible. This disconnect prevents patients from receiving treatment, and makes the medical intervention a waste of time and money. These problems are compounded by the high price, both in time and money, of psychiatric care, and the fact that many doctors are overloaded with to many patients stuck in the endless cycle of failed diagnosis.

A lack of consistency is a problem machine learning is particularly useful for. Humans are excellent at discerning patterns where there are none through their own creativity and biases, so it’s easy for psychiatrists to get trapped in their own intuitions despite the evidence. Machines don’t have personal inclinations and basis, and won’t weight data in any particular direction. Machine learning could help organize the huge amount of information being thrown at a psychiatrist with each patient and would make headway into standardizing the process of diagnosis so treatment can be standardized and distributed well. Even if the final result is a machine learning algorithm just narrowing the field to several disorders, the algorithm could save hours of time for psychiatrists who might be overwhelmed with as many as fifty patients relying on them for care or weeks of the wrong treatment for a person that the psychiatrist’s intuition was wrong about.

Like any algorithmically based approach to anything, though, the idea of machine learning in psychiatric diagnostics is ruffling a few feathers.
The Atlantic feature on Schwoebel mentions two potential issues they have with his program. The first is that the artificial intelligence may not be able to pick up on cultural or personal differences. The second is a version of the first, saying that the findings of the algorithm will be biased towards particular demographic groups based on who is available for training the algorithm.

These problems could also be exacerbated because the stigma on mental health issues is stronger in some cultures than others. However, this is a reason that it might be difficult for a machine learning approach to be perfect, not a reason to abandon it. The algorithm has to be trained on as many people and as many segments of the population as it can. The process of picking up on a cultural speech pattern or accent is just using data — a person’s accent or vernacular compared to their overall speech pattern — to reach a conclusion about whether this person is or is not likely to have a certain mental disorder. That is exactly what data-based processes do.

While psychiatry as a whole needs to be modernized and take advantage of many new approaches beyond just this one, the reflex to apply machine learning and other data-based techniques is the right one. The early returns say speech patterns are useful in some cases and if proof-of-concept papers come out supporting many symptoms detected by this technology, it can be a really powerful tool even if it is outstripped by more specific tests. It requires continued interaction between psychiatrists and patients, so that information is recorded and remains cheap and easy to get. If it turns out to be a versatile tool that is useful in diagnosing many mental disorders, it can help narrow down the more specific tests someone might need or improve the accuracy of those tests by giving a doctor something more specific to look for.

People with mental health problems float through a vacuous network of aimless question and answer sessions that lead to guesses as to what might be causing their distress. These machine learning techniques can add structure and goals to that process and help get results for people who really need them.