Generative AI and mental health

America emerged from three years of a pandemic with a mental health crisis. It is a crisis of care. There are effective interventions but, unlike Covid, few people who need care are receiving high quality, timely, effective treatment. The number of “deaths of despair”—a term that combines deaths from suicide, drug overdoses, and alcohol—has increased from 66, 392 in 2000 to 244,453 in 2021, with roughly a 30 percent increase during the pandemic. While over a million Americans died from Covid, most were people over age 65. Roughly 8,100 Americans under 30 died from Covid, while more than 140,000 died from deaths of despair. Morbidity has also increased, especially for youth. A recent CDC survey found that 57 percent of high school females describe persistent sadness and hopelessness, with 24 percent having a suicide plan.

The mental health care system has proven a poor match against these increases in morbidity and mortality. Too few providers are trained to deliver evidence-based treatments. More than half of people with anxiety, mood, and psychotic disorders do not seek professional care until they are in crisis. Care is reactive, delayed, and not integrated with medical treatment. And, in contrast to every other area of medicine, mental health care has been a data-free zone, without objective measurements of diagnosis or outcomes. It has not been accountable for improved outcomes at the individual or population level.

For the past four decades, psychiatrists have touted a biological revolution for people with mental illness, reframing mental disorders as brain disorders. With advances in neuroscience and genomics and a new generation of medications, there was both hope and hype that biology would bend the curve for morbidity and mortality. Alas, advanced neuroimaging has not yielded diagnostics, genomics has yet to identify causal factors, and the new medications have arguably been better for pharmaceutical companies than patients.

While a focus on the brain may yet deliver breakthroughs, a better near-term strategy is a focus on behavior, cognition, and mood. These brain “outputs” are the basis of how clinicians diagnose and track mental illness. Yet, our ability to capture these outputs objectively and generate responses to altered behavior, cognition, and mood has, until recently, not progressed beyond two people sitting in a room together, trying to understand each other.

During the pandemic, the locus of mental health care shifted rapidly from brick-and-mortar to telehealth. This shift increased access as people from rural and under-served communities could conveniently connect for care. Hundreds of thousands sought care online. The largest corporate providers of mental health care did not exist five years ago. But telehealth-1.0 did little to improve quality. While audio or video visits offered valuable data on voice, speech, face emotion, and even therapeutic rapport, none of this relevant data has been collected. Telehealth-2.0 will need to move mental health from a data-free zone toward measurement-based care.

Generative AI technology is the tool-driven revolution for an area of medicine that is primarily about behavior, cognition, and mood. Mental health diagnostics and treatments are based on language and observation, not surgical biopsies and blood biomarkers. Using the latest models from OpenAI, I have explored how GPT-4 could help transform care for people with mental illness. Four areas may offer near-term opportunities for developing a tool-driven revolution for mental health. Of course, using new interventions in health care requires demonstration of safety and efficacy, and in some cases, regulatory approval. These areas are suggested as domains for research.

Telehealth-2.0
In current practice, the wealth of data from patient interactions, in-person and online, is largely squandered. Imagine collecting objective scores of behavior, cognition, and mood with each interaction and using these scores for diagnostic coding and tracking progress across sessions. A digital scribe records the session and generative AI technology provides a summary with customized reports for the clinician, patient, and payer. These reports include objective measures as well as the potential for translations into culturally sensitive language.

Crisis Response
In July 2022, the U.S. launched 988, a new crisis response number that anyone in crisis could use to access immediate mental health support, rather than calling 911 for a police or fire department response. Unfortunately, the nation does not have a trained workforce to respond to a mental health crisis, so most calls are transferred to 911. AI has a dual role here: detecting the severity of a crisis and supporting 988 to respond in real time. Most crisis responders are not master clinicians; many are volunteers. By creating effective responses that are linguistically diverse, future versions of GPT-4 could revolutionize the 988 system. And by helping to train a new workforce, it could be the best way to scale up.

Clinical Decision Support
On the most recent version, GPT-4, clinical decision support has been the most surprisingly useful. Asked about a 27-year-old female with bipolar disorder who was concerned about weight gain on lithium prior to her wedding, GPT-4 summarized the literature on weight gain with lithium and noted alternate medications. It then advised that she “keep her eyes on the prize” since marriage to a supportive spouse could prove more important than the risk of a manic episode off lithium. This is only one example of ways in which the tech goes beyond a simple summary to provide a synthesis for guiding clinical decisions. If the clinician can learn the best prompts, the depth and range of support could upskill an ER nurse with a psychotic patient or a case manager in a community clinic or a volunteer on a crisis line. Of course, the same information would be available to patients and families.

Psychotherapy Chatbot
The concept of an autonomous therapist tracks back to the earliest days of AI. More than half a century later, this is the first health application that most people consider for generative AI. But GPT-4 does not agree. Prompted with, “Can AI replace human psychotherapists?” the response was, “While artificial intelligence (AI) has the potential to improve mental health care, it is unlikely that it will replace human psychotherapists entirely.” It went on to list these reasons: lack of emotional connection, difficulty with complex emotions, need for individualized treatment, and ethical concerns. But for treatments that are highly structured, like cognitive behavior therapy or VR exposure treatments for phobias, future versions of GPT-4 can deliver the same intervention, although research has shown that sustained engagement usually requires involvement of a human therapist. Even for less structured therapies, some data suggest that people will share more with a bot than a human therapist. It relieves concerns that they are being judged or need to please the human therapist. And for a generation of digital natives, the appeal of a human therapist may not be the same as it was for their parents and grandparents. The use of chatbots for supporting therapy is a work in progress, but should not be dismissed, even if GPT-4 seems unwilling to take on the role.

As mental disorders are assessed mainly through language and observation and treated with both medications and psychotherapy, AI advances are uniquely suited to address our mental health crisis. The latest version of GPT-4 and its successors combined with extant NLP tools can immediately facilitate the shift to Telehealth-2.0, build an improved crisis response, upskill the workforce with clinical decision support, and provide chatbot support for specific therapies. Of course, we may be in Act One of a five-act play. Recalling Freeman Dyson’s quote about a tool-driven revolution, can we even imagine the new things this tool can discover “that have to be explained?”