In a new study, Dr. John-Jose Nunez and his team trained artificial intelligence (AI) to predict whether cancer patients would go on to see a psychiatrist or counsellor, based solely on their initial oncology consultation documents.
And they were over 70 per cent accurate.
The most common mental health condition in cancer patients is called adjustment disorder: a period of low mood or anxiety after their initial diagnosis. For most people, it’s gone in a few days. But for 24 per cent, these symptoms can persist and develop into major depressive disorder.
“That can [make] it harder to proceed with their cancer treatment … such that mental illness in cancer patients can actually affect not only quality of life, but also survival rates,” said Nunez, a clinical research fellow and psychiatrist at the UBC Mood Disorders Centre and BC Cancer.
The main mental health resources offered by BC Cancer are counselling and psychiatry. Though the two are often used interchangeably, only psychiatrists are qualified to diagnose and prescribe medication for mental illnesses. Counsellors focus more on the social, cognitive and behavioural contributors to mental illness and offer various forms of therapy to help clients develop healthier lifestyles and thinking patterns.
While patients can refer themselves to counselling, there are barriers such as unawareness about available resources or doubts that mental health care would be beneficial.
For psychiatric treatment, patients must be referred by a physician, such as their oncologist or surgeon. But initial oncology consultations are lengthy, so it can be hard to fit in the time to explicitly ask about mental health.
Additionally, most patients exhibit symptoms of anxiety at their first appointment, so it can be difficult for physicians who are not experts in mental health to determine whether a patient requires mental health care.
Rather than taking upwards of 10 years of medical education, the AI can “see 36,000 patients within 24 hours … so much more efficient training,” said Nunez.
The training process began by providing the AI with around 36,000 patients’ initial oncology consultation documents. Then, the researchers told the model whether the patient went on to see a counsellor or psychiatrist. Equipped with this data, the AI was then given a new set of oncology consultation documents, asked to make its own predictions and corrected if needed.
Natural language processing (NLP) is the branch of AI that deals with how computers can process language like humans. NLP can be further divided into neural models and non-neural models.
Neural models analyze complex patterns and meanings in phrases and sentences, whereas non-neural models, such as the Bag of Words (BoW), analyze frequencies.
This study trained three different neural models and one non-neural model.
BoW, the simplest of the models, counts the number of times certain words, or “tokens,” are found in a document. Tokens that related directly to cancer, such as “myeloma” and “radiat” were the most important predicting factors for likelihood to see a psychiatrist.
This may be because the pathology or treatment for some cancers may make patients more susceptible to certain psychiatric disorders. For example, multiple myeloma, a cancer that causes a buildup of plasma cells in bone marrow, produces inflammatory proteins that can cause schizophrenia, a mental disorder that causes paranoia and hallucinations.
Important tokens in predicting counselling included demographic factors such as “retir” (retiree, retired) and “financi.”
Other common tokens in predicting counselling included “princ” and “georg,” which likely refer to Prince George — home to a BC Cancer site which serves northern BC.
The site is a long commute for many of its patients and many counsellors are also social workers who assist patients with their commute, which could explain the correlation between this rural town and likelihood of seeking counselling.
In order for models to work outside of BC or with illnesses other than cancer, they need to be tweaked.
“We can just give the model just a little bit of [new] data, let’s say from Alberta, and the model will just be able to adjust itself,” said Nunez.
In as little as 12 hours of fine-tuning, these models have the potential to be more widely applicable globally and for patients with a wide array of conditions.
Share this article
First online
