Psychiatry is both an art and a science, and artificial intelligence can provide tools to allow it to better understand what drives us.
True to this duality, I’ve always loved working with people as much as technology, especially the interaction between the two. I approach my work with engineering and record-keeping precision, but I also pay close attention to the “messy” psychology of creation, as well as the more medical biology of our “nature”. In this sense, I usually tell patients that we intend to work on the “hardware and software” to better understand ourselves and optimize how we feel and act.
Human behavior is complex, noisy and sometimes even erratic. The arrival and popularization of usable consumer-facing artificial intelligence in the form of GPT-3 marks a step towards understanding this complexity. With the ability to analyze and extract patterns from vast amounts of data, AI can provide nuanced and specific insights into what makes each of us happy, healthy and productive. It’s not here yet, but it will soon be a breakthrough for healthcare professionals, patients, and everyone who tracks mood, steps, heart rate, and sleep.
a visual art
Humans are highly visual creatures and a disproportionate amount of our brains are devoted to vision. Coincidentally, GPT-3 showed us its amazing ability in art and linguistic analysis. A picture is worth a thousand words, and on a visual level, no one can argue that the artwork produced by DALL-E, an AI image generator, is magnificent. Sure, there’s the occasional hallucination, extra finger, or creepy facial expression, but those glitches pale in comparison to the overall impressive “creativity” we see when we tell DALL-E to draw any scene in Van Gogh’s style.
We see visual patterns much more easily than text. That’s why we like graphics and illustrations. As visual creatures, we can stop to appreciate the level of achievement and expertise in AI art. It is easier to recognize a great work of art at first sight than an eloquent text. If AI can do what it can for language in art, it’s pretty impressive.
AI art can be stylistically inspiring, appearing both creative and falsely emotional. And the art looks great. Add more steps and processing and it’s almost perfect (gulp). Most people will sooner understand the meaning or style of a work of art than a written poem or a mathematical theorem. Our eyes are also much quicker at finding flaws and imperfections and breaks in patterns. I know GPT-3’s art and chat is an illusion of creativity and experience, but it’s still pretty impressive in the visual realm. His writing and textual analysis are not far behind.
a nebulous science
Psychiatry and, more globally, human humors represent a “fuzzy field” in particular. I often compared it with forecasting the weather – with seasons, daily variations and many variables that influence the result. We are still not good at predicting the weather more than seven days in advance. Almost the same; no biometric device or app has been able to tell me I’m having a rough day.
It’s hard to measure mood. Unlike other fields of medicine, there are few objective tests in psychiatry. There are some labs to check and diagnostic tests (like ECGs) to order. Otherwise, it’s questionnaires and interviews. Variations in sleep, diet, daily activity, socialization, and demands at home and work can make it very difficult to predict mood or the outcome of a given intervention (such as starting an exercise program or a new supplement or medication).
For anyone who has tracked their moods, it’s clear that the data is so variable, so dependent on so many things, that it can be nearly impossible to tell what’s working and what’s not. This may be why so few of us embrace new interventions – from taking supplements to regular exercise and an extra hour of sleep. The results can be so variable and inconsistent, so marred by a bad day that it’s really hard to see what’s working and what’s not. The signal is too noisy. Patterns are very variable and dependent on many things. This is why psychiatry is said to be both an art and a science; It’s subjective and complicated.
It’s all about pattern recognition
Early in my career, one of my closest mentors told me, “Psychiatry is all about pattern recognition.” Was he right. More generally, all of medicine is about pattern recognition. All human interactions, from making coffee to finishing a project at work, involve the same thing. Furthermore, human wisdom may well be the height of experiencing and recognizing patterns embedded in feelings.
Essential Readings in Psychiatry
Fascinatingly and frighteningly, we are at a crossroads in human history where computers are beginning to catch up with that “intelligence” that has always defined us as a wise man or “wise man”. AI is becoming increasingly adept at pattern recognition – especially with language, and it can now write screenplays, essays, poetry and explain general relativity to a third grader.
Current AI (such as ChatGPT) is similar to a sophisticated parrot – one that has been “fed” with 45 terabytes of data from the Internet, starting with GPT-3. Can sound quite authoritative and well-versed, sometimes without really knowing what he’s talking about. Numerous articles have been written about how changing the question or prompt results in wildly varied responses. Experts have found this to be dead wrong on several detailed issues. In fact, not all birds can fly. However, even without deeper wisdom (which is the promise of artificial general intelligence), pattern recognition is a powerful tool that can tremendously benefit the nebulous science of mental health.
Big Data Pearls
We generate more information than we can understand. From biometric tracking devices, heart rate monitors, step counters, and sleep and mood trackers, we’re swamped. Medical records are full of detailed notes about symptoms, treatments and timing of interventions, often with years of data. Even in a well-designed randomized control trial with many controlled factors, the advantage of one intervention over another can be quite subtle. In daily life, especially in mental health, the benefit of one intervention over another is easily lost. Did those supplements you took for a month change anything?
For years, I’ve wondered what patterns exist in my data that I’m not aware of. The same goes for my patients: Is it better to exercise in the morning or in the afternoon? Sleep seven hours or aim for eight? Did the ashwagandha supplements I took for a month improve anything? Did meditating daily for 10 minutes have any effect on my anxiety? My sleep? Or, in my work, does a given patient tend to do better on one antidepressant than another? Does increased sleep or socialization play a greater role in overall mood?
Lots of data. Lost associations. Currently, research-level statistics would be needed to reveal these nuances. The data is there. It is not yet practical to aggregate, analyze and extract causes and effects.
I have always loved biometrics, both for myself and my patients, and I approach my charts as a coder. I make consistent notes on regular data points at each visit, mood, anxiety, sleep, impulse control, and cognitive/learning ability (MASEIC for short). My hope has always been that one day AI-safe analytics could be applied to my large patient dataset. The goal would be to find patterns and associations that I missed. To tell me more about what works and what doesn’t, or what works best.
Until GPT-4 starts listening, or we start measuring heart rate, skin conductivity, speech rate, and pitch during visits, I’ll continue to take good notes for later analysis. In the case of AI, more data is no problem.
For the confusing field of psychiatry, pattern recognition will be a gold mine. Artificial Intelligence will serve as a powerful tool in a field where so much information is subjective, variable and multifactorial. But are we going to hear the results or suggestions?
Awareness can be key, as we all notice our speed when passing a “Your speed is…” sign. The hope is that with better goal feedback and clearer causality, we will be more motivated towards healthy behaviors. For practitioners, this can provide additional support for choosing one intervention over another and provide objective feedback on progress and change.
The AI may need to show you that you are objectively happier if you spend time with friends, go to bed an hour earlier, exercise more, and stop eating after sunset. These interventions will be more actionable and “fixed” when an AI analysis actually confirms they are working and shows the data to support them.
Like a blood pressure or ECG reading, AI can finally provide some valuable metrics for the confusing art of psychiatry and mental health, which so often lacks objective metrics and tests.