Psychiatry Realizes the Value and Bias of Big Data and AI
Language and biology meet the unconscious in a world of questionable databases designed to improve the quality of care
Cure her of that! Canst thou not minister to a mind diseased, pluck from the memory a rooted sorrow, raze out the written troubles of the brain, and with some sweet oblivious antidote cleanse the stuffed bosom of that perilous stuff which weighs upon her heart — Macbeth, William Shakespeare
The current state of big data in psychiatry has evolved from the epidemiologic foundation reported in 1984 and laid well before. The foundation was built on the development of defined diagnosis, diagnostic assessment tools that could be used in large samples…
Psychiatry is an evolving medical profession that is in the process of investigating the utility of complex data sets to solve complex human mental health disorders. One of the main stumbling blocks, which is quickly apparent, is that the data is biased by the fallible humans who originally entered it into any records.
Unlike other disciplines in medicine that depend on physical measures such as blood pressure, serologic analysis, or imaging, psychiatry is dependent on communication. Language and how it is used or interpreted is Alice’s rabbit hole. Biological psychiatry may have changed the field, but it is still heavily dependent on language.
Depending on the accuracy of the original records can be an algorithmic nightmare. Accept the data as it is provided, knowing that there’s no way to correct it, and you build bias into the database. How can you screen this bias out to ensure a cleaner endproduct?
One Clinical Trials Example of Misinformation
I was once charged with traveling to 17 research sites in the United States in my role as the clinical monitor to ensure the accuracy of administration and scoring of a test for cognition in Alzheimer’s patients.
In that role, I saw psychologists, psychiatrists, and advanced-practice nurses who were to provide data from their scoring of our test materials.
One portion of the test required that the patient identify objects presented to them on a table. A specific item, a Lone Ranger-type mask, has only one correct answer, a mask. Any other response would be scored as incorrect and an indication of impaired recognition of the target item.
The test had been designed and validated by psychologists working at a major medical center on the East Coast of the US. This fact led to problems for any patient who was tested in the Southern States. In Midwestern states, a similar situation arose with another item.
In the South, door knockers were often shaped in a mask-like configuration, and they are called not a “mask” but a “door face.” Southerners who gave this regionally-correct response were scored against in the test. They lost points.
A similar problem arose with one or two other items in the test where regional expressions were given rather than the East Coast-equivalent of it. Bias personified.
How many patients in clinical trials of various medications receive incorrect scoring because of regional bias or language differences?
Clinical Interviews of Patients
Psychiatric patients come from all walks of life, different educational levels, and cultures. Some of them have accents that are misinterpreted by clinicians who then make inaccurate notes. These notes will become part of a clinical record prime for database inclusion.
I sat with a psychiatrist who grew up outside the United States. She was interviewing a Latino patient regarding a clinical matter. During the interview, the man used a street term for another patient, and the psychiatrist thought he said something else.
He hadn’t, and I knew it because I was familiar with the word and how inflammatory it could be. What did she write in her notes? He was belligerent and argumentative. He wasn’t.
He was making a case for why he had been provoked into some action, but she didn’t see it that way — classic culture clash with resulting consequences for him. I had to explain to her what the term meant and the reason it was so offensive.
Patients in psychiatric hospitals who are “out of control” and lash out are medicated and put in a quiet room. These rooms are often referred to as a “padded cell” although they don’t always have padding. It was a mistake in this instance. His chart then had a notation about this behavior and the intervention. All of it goes into a clinical database.
The Amazing Power of AI and Big Data
A massive trove of clinical data is available worldwide, but much of it, from the previous decades, is in the form of handwritten notes. How can these notes be retrieved in a data-receptive form to be included in a database?
In fact, progress has been made in AI’s deciphering and utilizing handwriting. One of the advances is in the MNIST database. But this still leaves unresolved the initial interpretation by the clinician of the patient’s psychiatric disorder.
Until deep learning has the ability to recognize the hundreds or thousands of clinicians’ scribbles, much of the data, biased or not, is lost to us.
But now clinicians are utilizing software to encode their session or testing notes, and this creates a more straightforward path for analysis. Symptoms and diagnoses are one area where AI can genuinely prove its mettle.
The algorithm can quickly retrieve, analyze, and make determinations about factors unseen by the clinician. All of this is done at lightning speed for instant incorporation or a warning regarding the conclusions the human calculated. Hal’s ever-watchful eye is there to aid, not override in the best-case scenario.
Churning data from thousands of sources may uncover relationships unseen and suggest a new direction for research or treatment. This is one of the gems of AI in psychiatry.
Why try four different medications for a specific mental illness when AI could pick out the most promising initially? Now, it is a trial-and-error method despite advances in the genetic predisposition for psychiatric illness.
A Rosy Future for Psychiatry and AI?
The efforts to contain the COVID-19 virus include wearing masks, and here is a similarity to psychiatric databases. We cannot have clean data, free from bias if we don’t have clinicians who are sufficiently skilled in interviewing and interpreting both tests as well as language/cultural aspects of patients. The computer expression of “garbage in, garbage out” is significantly salient in this regard.
I’ve seen psychiatrists, untrained in psychological testing, administer tests with faults that any doctoral supervisor would have corrected. But the psychiatrists don’t have this type of supervision, and they shouldn’t be using tests for which they have questionable competence. Even the inter-rater reliability of some psychiatric tests is of concern.
In one instance, the psychiatrist was giving body cues regarding the answers he expected from the patient. He didn’t realize he was invalidating the testing. What were the consequences for that patient? I don’t know.
The future use of AI in all of medicine and psychiatry, specifically, can be a promising one if the inherent bias is addressed before the databases begin to affect people’s lives in untoward ways.