From the Editor
Thought logs. Homework. Exposure. Psychotherapy is evidenced for the treatment of many mental disorders – but access is problematic. Can technology help? Is AI a game changer?
In a new NEJM AI paper, Dr. Michael V. Heinz (of Dartmouth College) and his co-authors attempt to answer these questions. In an RCT, they compared 210 participants receiving a chatbot intervention (Therabot) with a control group, analyzing symptoms of mood, anxiety, and disordered eating. “As the first RCT of its kind, our study supports the feasibility, acceptability, and effectiveness of a fine-tuned, fully GenAI–powered chatbot for treating mental health symptoms.” We consider the paper and its implications, and tap the expertise of Dr. John Torous (of Harvard University).
An AI bot for better mental health?
In this week’s other selection, we look at a new review from The American Journal of Psychiatry. Drs. Nicholas H. Neufeld and Daniel M. Blumberger (both of the University of Toronto) provide an update on neuromodulation strategies for schizophrenia, drawing on nearly 140 citations and reviewing different technologies “that span electrical, magnetic, and ultrasound forms of stimulation.” They note: “The evolution of interventions holds the promise of fewer adverse effects and a noninvasive approach, increasing the scale at which these interventions may be offered in hospital and community settings.”
DG
Selection 1: “Randomized Trial of a Generative AI Chatbot for Mental Health Treatment”
Michael V. Heinz, Daniel M. Mackin, Brianna M. Trudeau, et al.
NEJM AI, 27 March 2025
The prevalence and burden of mental health disorders have increased significantly over the past three decades… Digital therapeutics (DTx) – automated, evidence-based software for the treatment or diagnosis of medical conditions – offer a solution to bridge this gap. While DTx aim to improve the accessibility and scalability of evidence-based mental health interventions, these approaches have been plagued by attrition and low rates of engagement… Artificial intelligence (AI) represents a promising direction for improving DTx. Chatbots hold particular promise given their capacity to imitate human conversation and dialogue, long known to be integral parts of psychotherapeutic treatments – that is, the talking cure… Although such chatbots (e.g., Woebot) have shown benefits in clinical trials, and in some cases a capacity to promote a therapeutic alliance, they are inherently limited by their reliance on an explicitly programmed decision trees and restricted inputs.
Recent advances in computing and machine learning now allow for sophisticated systems capable of learning, adapting, and understanding context in natural language, removing the necessity for explicit programming. Pushing these bounds even further in the language domain, the advent of generative AI (Gen-AI), recently popularized by ChatGPT, has enabled the automated production of novel and highly personalized responses to human input. To date, conversational agents using Gen-AI have fallen under general purpose, wellness, or companion applications, rather than software intended for the diagnosis and treatment of mental health disorders… Paired with existing frameworks for DTx, Gen-AI chatbots have unprecedented potential to address existing problems with engagement while powering the development of new and personalized interventions.
So begins a paper by Heinz et al.
Here’s what they did:
- They conducted “a national, randomized controlled trial of adults with clinically significant symptoms of major depressive disorder (MDD), generalized anxiety disorder (GAD), or at clinically high risk for feeding and eating disorders (CHR-FED).”
- Participants were randomly assigned to a 4-week chatbot intervention or waitlist control.
- The intervention: Therabot “utilizes a generative large language model (LLM) fine-tuned on expert-curated mental health dialogues. The dialogues were developed by our research team…”
- Primary outcome: “symptom changes from baseline to postintervention (4 weeks) and to follow-up (8 weeks).” They used the Patient Health Questionnaire 9 (PHQ-9), the Generalized Anxiety Disorder Questionnaire (GAD-Q-IV), and the Weight Concerns Scale (WCS). Secondary outcomes included user engagement and therapeutic alliance.
Here’s what they found:
- In total, 1 707 people were screened. Participants included 210 adults, randomly assigned to intervention and waitlist control (WLC).
- Depression symptoms. Therabot users showed significantly greater reductions in symptoms of MDD (mean changes: −6.13 vs. −2.63 at 4 weeks; −7.93 vs. −4.22 at 8 weeks) relative to controls at postintervention and follow-up.
- Anxiety symptoms. Significantly greater reductions in symptoms of GAD (mean changes: −2.32 vs. −0.13 at 4 weeks; −3.18 vs. −1.11 at 8 weeks).
- CHR-FED. Significantly greater reductions in symptoms of CHR-FED (mean changes: −9.83 vs. −1.66 at 4 weeks; −10.23 vs. −3.70 at 8 weeks).
- Utilization. “Therabot was well utilized (average use >6 hours).” See below for a heat map of messages per day.
- Therapeutic alliance. Participants rated the therapeutic alliance as comparable to that of human therapists.
A few thoughts:
1. This is a timely and relevant study, published in a good journal.
2. The main finding in a sentence: “Critically, as compared with the WLC, Therabot users showed a greater reduction in depression, anxiety, and CHR-FED symptoms at postintervention (4 weeks) and at follow-up (8 weeks).”
3. How to understand the finding? I asked Dr. John Torous (of Harvard University) for his thoughts.
“The field of chatbot research is moving quickly, and each week, we see exciting new studies around AI and large language models. This week was no exception with a fascinating study on Therabot that aimed to offer help for people with depression, anxiety, or at risk for eating disorder. Unlike most AI chatbot studies, this study was a randomized controlled trial, often considered the highest level of evidence for evaluating the effectiveness of interventions due to their ability to minimize bias and establish causal relationships. But an intriguing question emerged – what is the ‘C’ in RCT for this study? In other words, what was the control group? For this study, the answer was nothing, and those in the control group served as a waitlist control. The old adage, ‘something is better than nothing’ and the clinical study wisdom that expectations and placebo effects can drive outcomes suggest that any digital intervention will beat a waitlist control. That was the case in this study of Therabot. Adding active digital controls, even ones as simple as a countdown timer, mood tracker, and the game Tetris, have, in different studies, proven as effective as different digital mental health technology interventions. This is not to say Tetris is a treatment for depression but rather to highlight the importance of considering the control group in digital studies. Picking the right control group is tricky, and our team has shared a framework to make that decision easier. And in 2025, when there are thousands of free apps, websites, and now even AI to chat with, comparing a therapy chatbot to, a waitlist control makes it impossible to judge effectiveness. Further, given recent findings that untuned out-of-the-box, chatbots may have better accuracy than finely tuned chatbots, the necessity of comparing any finely tuned chatbot to a more basic chatbot is even more critical. We can all agree that chatbots are showing they can be exceptional, but that does not grant them an exception from demonstrating that. And we can all agree the authors of this paper did a fine job with this study and were very clear in their aims and goals. The post-publication publicity the paper garnered reminds us of the great interest and expectations for chatbots – which is all the more reason to control for that in studies.”
– John Torous
The full NEJM AI paper can be found here:
https://ai.nejm.org/doi/full/10.1056/AIoa2400802
Selection 2: “An Update on the Use of Neuromodulation Strategies in the Treatment of Schizophrenia”
Nicholas H. Neufeld and Daniel M. Blumberger
The American Journal of Psychiatry, April 2025
Schizophrenia is a heterogeneous and chronic psychiatric disorder that manifests clinically with psychotic symptoms in positive (e.g., hallucinations) and negative (e.g., amotivation) domains and cognitive dysfunction. As neuroimaging approaches have progressed from smaller case-control studies to larger consortium-based analyses, our knowledge of the brain networks impacted by this disorder has increased. Schizophrenia afflicts an estimated 1% of the global population and is challenging to treat. Even with trials of antipsychotic medication at therapeutic dosages and adequate duration, up to 30% of patients do not respond to first-line antipsychotics… In this context of treatment resistance, various neuromodulation strategies stimulating the cortex and subcortex have been proposed as options…
So begins a paper by Drs. Neufeld and Blumberger.
They focus on several interventions; here we highlight four.
Electroconvulsive Therapy
“ECT was initially introduced as a treatment for patients with schizophrenia in 1938 as a replacement for chemically induced seizures. Despite the initial use of ECT for schizophrenia, the APA Task Force on ECT lists schizophrenia as an ‘other diagnostic indication,’ largely due to a limited number of studies. More specifically, ECT is indicated for augmenting antipsychotic treatment in patients with schizophrenia when a patient has responded to ECT, not responded to psychopharmacology, or when symptoms are severe. Consistent with these indications, response rates to ECT in this population approach those for patients with depression, and there is emerging evidence that ECT is an equal or perhaps better option for treatment-resistant schizophrenia…
“The most recent Cochrane review concluded that ECT has a positive effect on medium-term clinical response for people with treatment-resistant schizophrenia. However, the review was more tepid with respect to longer-term outcomes.”
Magnetic Seizure Therapy
“Magnetic seizure therapy (MST) is an emerging form of convulsive therapy that utilizes high-frequency, high-intensity transcranial magnetic stimulation to induce a therapeutic seizure via electromagnetic induction. Like ECT, MST requires anesthesia; unlike ECT, however, MST delivers a more focal stimulus to the brain that avoid regions involved in memory, thus minimizing or avoiding cognitive adverse effects altogether. Promising preliminary studies in depression have been corroborated by a recently published clinical trial comparing MST and ECT in patients experiencing a major depressive episode that found comparable remission rates and much fewer cognitive adverse effects.
“Similar results have been found for patients with schizophrenia. Open-label studies have demonstrated high response and remission rates and negligible cognitive adverse effects. In a study comparing MST to bitemporal ECT, similar response rates were observed, with less cognitive impairment in the MST group. Interestingly, the neural correlates of this preserved cognition have begun to be examined. In a neuroimaging study nested within this trial, MST was found to have no effect on hippocampal volume, and ECT increased bilateral hippocampal volume. As the most recent Cochrane review notes, there is a dearth of studies on MST for schizophrenia, and well-designed clinical trials are needed to understand the efficacy and tolerability of MST in this population.”
Deep Brain Stimulation
“Deep brain stimulation (DBS) is a form of precision neuromodulation that employs a neurosurgical approach. Initially developed for treating movement disorders, the first application of DBS for psychiatric disorders occurred in 2005, when DBS was used to stimulate the subgenual anterior cingulate cortex (sgACC) in patients with treatment-resistant depression. A DBS trial for schizophrenia was proposed in part based on the success of targeting the sgACC in depression and the nucleus accumbens (NAcc) in obsessive-compulsive disorder (OCD). In terms of target selection, and beyond its success in treating depression, inhibitory DBS of the sgACC was rationalized on the basis of dysfunction in the default mode network… In an initial trial of DBS targeting these regions in seven patients with schizophrenia, two of the four patients with sgACC electrode placement and two of the three patients with NAcc electrode placement met criteria for symptomatic improvement. A crossover phase in the same study demonstrated worsening when the stimulation was discontinued. Improvements in positive symptoms have been found to endure for both sgACC and NAcc stimulation after 3 years…”
Transcranial Magnetic Stimulation
“Transcranial magnetic stimulation (TMS) was developed in 1985 to focally and noninvasively modulate brain activity. Repetitive TMS (rTMS) delivers multiple pulses within a brief time frame, causing enduring excitability and network changes, with high-frequency (>5 Hz) stimulation associated with increased excitability and low-frequency (1 Hz) stimulation with decreased cortical excitability… In schizophrenia, although TMS studies of positive and negative symptoms as well as cognition have all been conducted, the progression from standard frequencies to TBS to accelerated protocols has not yet been realized. In terms of positive symptoms, auditory hallucinations have garnered the most research, and initial studies used low-frequency (1 Hz) rTMS to modulate hyperactivity in the temporal lobes that was thought to give rise to auditory hallucinations. Evidence has mounted implicating the temporoparietal cortex as a core region of hyperactivity associated with hallucinations, and inhibitory rTMS to this region has continued to be developed despite nonreplication of earlier study findings. In a procedure known as ‘priming,’ a brief period of stimulation at a different frequency is delivered to enhance the neurophysiological effect of the stimulation of interest. A study that explored stimulation of the primary auditory cortex compared 1-Hz stimulation, 6-Hz priming of 1-Hz stimulation, and sham stimulation, but did not find clinical response to be better than sham stimulation in either the primed or nonprimed stimulation conditions…
“The prefrontal cortex (PFC) has been proposed as a neural substrate for negative symptoms based on observations of reduced metabolism in the PFC. Excitatory protocols targeting the PFC have been employed with moderate efficacy. Patient factors (age, shorter illness duration) and TMS factors (higher pulse frequencies, greater stimulus intensity, and longer treatment durations) have been associated with greater efficacy. Excitatory (10 Hz) rTMS delivered to other sites, such as the dorsomedial PFC and cerebellar vermis, have also shown promise for the treatment of negative symptoms…
“Overall, and in contrast to TMS for depression, TMS for schizophrenia remains a more challenging endeavor, with promise and pitfalls.”
A few thoughts:
1. This is a practical, well-written update, published in a major journal.
2. It’s also impressive, with 137 references.
3. Here are three take-aways:
- Yes, think about neuromodulation for those with schizophrenia. As the authors remind us, “While the majority of neuromodulation studies have focused on patients with mood disorders, predominantly depression, there is an unmet need for patients with schizophrenia, who are in dire need of novel therapeutic options.”
- ECT may be old and (perhaps) out of fashion for those with treatment-resistant schizophrenia – but there is evidence, including a Cochrane review suggesting positive effect.
- The future is bright: advancing neuroimaging and data analysis may allow for better targeting of brain networks and nodes. “Continued advances will hopefully move neuromodulation from a tool typically considered in treatment-resistant illness to earlier adoption to act synergistically with antipsychotic medication and other treatment strategies and promote functional recovery.”
4. Neuromodulation has been considered in past Readings, including a podcast interview with Dr. Blumberger. “As far as medical treatments go, ECT is the safest medical procedure in all of medicine.” You can find it here:
The full AJP paper can be found here:
https://psychiatryonline.org/doi/10.1176/appi.ajp.20250068
Reading of the Week. Every week I pick articles and papers from the world of Psychiatry.
Recent Comments