Qualitative coding 3,183 open-ended responses by hand would take 80+ hours. With an LLM-assisted workflow, Gino Carmona-Díaz and colleagues in Colombia did it in 10. ICC = .824. Full tutorial, prompts, and code are open access.
link.springer.com/article/10.3...
An AI-generated podcast featuring our work. So cool!
We also identified important limitations:
• Overreliance on a single benchmark dataset
• Incomplete reporting of performance metrics
• Limited clinical integration
Huge thanks to
@christianwebb.bsky.social, @nigeljaffe.bsky.social, Kristina Pidvirny, Anna Tierney, Mia Vaidean, and Poorvesh Dongre.
Key findings:
• Overall accuracy ≈ 80%, but balanced accuracy ≈ 70% when accounting for class imbalance
• Text type matters: structured clinical interviews perform best
• Simple linguistic features + traditional ML often perform comparably to more complex transformer models
Now out in npj Digital Medicine 🎉
www.nature.com/articles/s41...
Our systematic review and meta-analysis examines how well language-based models detect depression from text.
We reviewed 123 studies (40,000 + observations) using NLP and machine learning.
Podcast cover for JAMA Psychiatry Author Interviews, dated February 18, 19 minutes. Title: Predicting Adolescent Response to School-Based Mindfulness. A blue square displays 'Psyc' and initials 'JN'. A purple play button is at the bottom.
John Torous, MD, speaks with Christian A. Webb, PhD, of Harvard Medical School and McLean Hospital, about the limits of population-level prediction and the need for more potent and targeted interventions for #YouthMentalHealth.
🎧 Listen now:
ja.ma/3OSGiIw
Curious about AI and spirituality? Have a few minutes for an anonymous survey? Colleague and friend Roman Palitsky at Emory U is launching a project on this fascinating intersection: qualtrics.kcl.ac.uk/jfe/form/SV_...
Mental health services struggle to meet demand. Wagner et al tested whether a RoBERTa LLM can detect anxiety or depression in text, showing accuracy similar to human experts and paving the way for more accessible, scalable mental health evaluation.
www.nature.com/articles/s44...
Thanks Christian, appreciate it! 🙏 Couldn’t have done it without you and the team.
A huge thanks to @christianwebb.bsky.social, @nigeljaffe.bsky.social, Kristina Pidvirny, Anna Tierney, Mia Vaidean, and Poorvesh Dongre who went with me through all stages of meta-analysis grief from horror, despair and self-blame (what were we thinking?) all the way to hope and pride in this work.
Text-based detection could make early screening more scalable and accessible, but how well do these tools actually work?
Our goal was to bring clarity to the field synthesizing evidence and highlighting what works and what still needs work to build better tools for early screening and detection.
💫 Excited to share new preprint a systematic review & meta-analysis of 123 studies (40k+ ppl) on how well language-based models detect depression from text.
www.researchsquare.com/article/rs-8...
Thanks to @bbrfoundation.bsky.social for highlighting our lab's study (led by @hadarfisher.bsky.social & @nigeljaffe.bsky.social) using📱smartphone sensors + LLM-derived text ratings to track behavioral activation and symptom change in teens with anhedonia.
bbrfoundation.org/content/smar...
Huge thanks to @nigeljaffe.bsky.social, @christianwebb.bsky.social and our amazing team Habiballah Rahimi-Eichi, Erika Forbes, @diegopizzagalli.bsky.social and @drjbake.bsky.social
These patterns appeared at the individual level, showing the potential for personalized, real-time monitoring of therapy progress. In the future, such tools could help clinicians track clients’ progress outside the therapy room and deliver just-in-time interventions that enhance ongoing treatment.
• Smartphone mobility features predicted weekly improvements in anhedonia and depressive symptoms.
• GPT “activation” ratings correlated with both self-report of activation and mobility data (e.g., time away from home, number of places visited).
• Increases in GPT-rated activation were associated with higher daily positive and lower negative affect.
BA aims to increase activation, encouraging engagement in rewarding, goal-directed activities, which is assumed to reduce anhedonia and depressive symptoms.
Using GPT-4o-based ratings of daily EMA text responses and smartphone data, we found:
In this study, we tested whether passive smartphone data (GPS, accelerometer) and large language models (LLMs) like GPT could capture meaningful change among adolescents receiving behavioral activation (BA) therapy for depression and anhedonia.
Now out in NPP – Digital Psychiatry and Neuroscience! 📱🧠
Can LLMs and smartphone sensing help us track therapy progress in real time?
📱EMA is increasingly used in intervention studies to acquire a more fine-grained and ecologically valid assessment of change. But EMA is relatively burdensome. What's the added value? We tried to address this question in our new paper now out @jmirpub.bsky.social www.jmir.org/2025/1/e69297 1/n
My hope is that, with the right boundaries, knowledge, and caution, and with the incredible work of so many researchers in the field, AI can advance our long-standing efforts to close the urgent gap in mental health access.
I argue that we can’t stop this train, but as clinicians and researchers, we can and *should* shape its track. Rather than resist change, we should be more active in guiding it, wisely, ethically, and collaboratively (with the developers and the users).
In this letter, I respond to Ziv Ben-Zion’s important World View column, which recommended ways to stop, or at least slow down, the use of AI (such as ChatGPT) for emotional support.
👉Ben Zion’s world View column: www.nature.com/articles/d41...
My correspondence on emotionally responsive AI is now published in Nature (!)
I’m thankful for the opportunity to share this little drop of thought. 🌱
👉 Read the full correspondence here: www.nature.com/articles/d41...
Thanks so much for tagging me, Christian! This sounds like an absolutely amazing dataset. @eikofried.bsky.social I’m traveling now with little reception (and lots of kids’ noise 😅), but once I’m somewhere quiet I’ll send an email with my thoughts on de-identification and some collab suggestions.
How emotions unfold over time, their flexibility and adaptability, may be just as important as what people feel for understanding depression vulnerability.
This opens new doors for early identification and prevention.
📉 The results: Participants with more rigid emotional systems were significantly more likely to develop depressive symptoms later, even after controlling for risk factors, sex, emotional intensity, and variability.
This was specific to depression! Emotion rigidity didn't predict anxiety symptoms.
We followed adolescents without depression and had them report emotions 4x daily for a month. Then we tracked depressive symptoms for 2 years.
We used dynamic systems methods to build individual emotion networks and calculated emotional rigidity (how densely interconnected emotional states were).
