As a yoga instructor in a clinic for people with eating disorders, I don’t see many straight cisgender male clients. But when I do, many have one thing in common: military service.
Conditions: Autism; Intellectual Disabilities With Other Behavioral Symptoms; Intellectual Disability
Interventions: Behavioral: Regulating Together group therapy
Sponsors: Children’s Mercy Hospital Kansas City; National Institutes of Health (NIH); National Center for Advancing Translational Sciences (NCATS); Frontiers Clinical & Translational Science Institute
<strong>Background:</strong> With the increasing burden of chronic diseases, self-management education (SME) is crucial. Traditional SME based on face-to-face delivery by clinicians is resource-intensive, and general digital tools such as web-based platforms often provide limited interactivity for patient learning. Although chatbots based on large language models (LLMs) show promise in interactivity, their real-world effectiveness lacks empirical evidence. <strong>Objective:</strong> This study aimed to explore the feasibility and preliminary effectiveness of an LLM-based chatbot specifically designed for osteoporosis SME. <strong>Methods:</strong> A formative randomized controlled trial was conducted in a tertiary hospital from February 2024 to March 2025. Adults aged ≥18 years with osteoporosis were recruited and randomly assigned (1:1) to either the intervention (OPBot) group or a control group receiving traditional health education. The chatbot provided interactive educational content and question-and-answer support, while the control group received face-to-face education and written materials. Osteoporosis knowledge was assessed using the Osteoporosis Knowledge Assessment Tool at baseline and discharge. Nurses’ time spent on health education was self-recorded during each intervention session and aggregated across sessions. Adherence to disease management was assessed at 1, 3, and 6 months after discharge via telephone using Likert-scale questionnaires. The reliability of OPBot responses was evaluated by 2 clinician assessors using a 5-point Likert scale, with interrater agreement calculated using Cohen <i>κ</i>. Group comparisons were conducted using 2-tailed independent <i>t</i> tests, Mann-Whitney <i>U</i> tests, and chi-square tests, and adherence outcomes were analyzed using mixed-effects models. <strong>Results:</strong> A total of 100 participants were randomized; 12% (12/100) were excluded due to loss to follow-up, refusal of the second knowledge assessment, or death, leaving 88% (88/100) participants for analysis (n=45, 51.1% in the OPBot group and n=43, 48.9% in the control group). The OPBot group showed significantly higher postintervention knowledge scores than the control group (median 80.0, IQR 70.0-89.0 vs median 75.0, IQR 65.5-80.0; <i>P</i>=.01). Nurses in the OPBot group spent lesser time on SME than those in the control group (median 5.0, IQR 2.0-17.0 vs median 23.0, IQR 20.0-25.0 minutes; <i>P</i><.001). For adherence outcomes, a significant group×time interaction was observed for calcium supplement intake (odds ratio 1.49, 95% CI 1.08-2.06; <i>P</i>=.02), indicating differing adherence trajectories over time. The OPBot group also showed higher odds of consuming calcium-rich foods across time points (odds ratio 2.87, 95% CI 1.04-7.89; nominal <i>P</i>=.04), although this association did not remain significant after Holm correction. No significant effects were observed for sun exposure (<i>P</i>=.56), exercise (<i>P</i>=.79), or total adherence scores (<i>P</i>=.33). In the question-and-answer module, most OPBot responses were rated as highly reliable 89.4% (76/85), with high interrater agreement (Cohen <i>κ</i>=0.83). <strong>Conclusions:</strong> LLM-based chatbots specifically designed for osteoporosis SME may improve patient knowledge, supporting adherence behaviors, and reducing healthcare workload. However, further large-scale studies are needed to confirm these findings.
Background: Self-harm is the strongest risk factor for suicide and an important outcome for mental health care. Although prevalent in clinical populations, it is often imprecisely captured in routinely collected clinical data, where it is often recorded and stored as unstructured free text. Contemporary language models, such as GPT (OpenAI) and Gemini (Google), can analyze free-text clinical notes, but such models may violate data governance of processing sensitive patient data. Objective: This study aimed to evaluate whether a privacy-preserving language model running entirely within an institution’s secure computing infrastructure (here, the UK National Health Service [NHS]) could accurately identify the presence and timing of self-harm using electronic health records from secondary mental health care. Methods: Clinical notes were drawn from Oxford Health NHS Foundation Trust using a multistage workflow: (1) a random sample of 1000 patients with a psychiatric diagnosis, defined according to the (; codes F00–F99); (2) candidate-note identification using a Gemma3-4b language model to flag notes containing self-harm content; and (3) from those candidates, 1352 randomly sampled notes were selected for expert annotation, resulting in gold-standard corpus enriched for self-harm content. Clinical notes were annotated for the presence of self-harm and its timing (≤90 days, >90 days, or unknown). A privacy-preserving locally served 27-billion-parameter Gemma 3 language model (“Gemma3-27b”) was used as the core model. Prompts were systematically developed and refined using a labeled development set to identify self-harm and generate a structured output per clinical record. Gemma3-27b performance was compared against a strong baseline multilabel text classification model based on robustly optimized BERT pretraining approach (RoBERTa), a transformer-based language model architecture. Model performance was evaluated using precision, recall, and the -score (harmonic mean of precision and recall), with 95% CIs estimated from 1000 bootstrap samples with replacement. Results: Gemma3-27b outperformed the RoBERTa classifier across all categories, achieving Precision=0.92, Recall=0.92 (sensitivity), and -score=0.92 for notes containing self-harm, and Precision=0.97, Recall=0.97 (specificity), and -score=0.97 for notes without self-harm. For the 51 notes labeled as recent self-harm in the held-out test set, Gemma3-27b achieved Precision=0.84, Recall=0.75, and -score=0.79. The global weighted -score of Gemma3-27b across all categories was 0.88, compared to 0.85 for RoBERTa. Conclusions: With systematic prompt development on a labeled development set, but no gradient-based fine-tuning, the current Gemma3-27b language model matched or exceeded a fine-tuned RoBERTa classifier for ascertaining self-harm events and their timing. Aggregate gains were modest, while improvements were largest in the most challenging, lower-frequency timing categories. On a simplified binary recent-versus-other task, RoBERTa performed marginally better, indicating that supervised classifiers remain highly effective when the task is simplified and sufficient labeled data exist. This work demonstrates the technical feasibility of privacy-preserving self-harm detection within a secure NHS research environment.
The global health care sector is under increasing strain.
Decades of chronic underinvestment and constraints in recruitment have coincided with a surge in demand for services for aging populations. Gaps in provision are already taking a toll, with fragmented access to care and high rates of stress and burnout among staff. And it’s getting worse. The World Health Organization has warned that current shortfalls will increase to 11 million workers by 2030.
In their urgent hunt for a solution, many health-care providers are now pinning their hopes on agentic AI, with more than two-thirds (68%) having already adopted AI agents into their workforce, according to KPMG.
The technology is being deployed to automate complex back-office processes, collaborate with medical teams, and even triage patients, all in a bid to reduce the cognitive load on clinicians and improve quality of care for patients as the supply of human health-care workers dwindles.
A different type of digitalization
Until now, the benefits of digitalization within health care have been limited.
Many staff have blamed slow or outdated technology for adding to the administrative burden rather than alleviating it. For example, U.S. patient data was migrated to electronic health records (EHRs) in the early 2000s, but this data remains fragmented and reliant on manual inputs.
New telehealth services and digital care tools, like remote monitors, have had similar shortcomings, says Ashis Barad, MD, chief digital and technology officer at Hospital for Special Surgery (HSS), an academic medical center in New York that focuses on musculoskeletal health. Both technologies have helped improve access to health care by removing geographical barriers, he says, but they’ve failed to replicate the quality of in-person care or win trust from patients.
Agentic AI is different from these existing technologies, he insists.
Rather than relying on manual inputs or defaulting to human workers for any case that sits slightly outside a rigid framework, AI agents can handle nuanced, complex scenarios. They can make autonomous decisions, retrieve information from expert clinical sources, and iterate over time, freeing clinicians to focus on higher-level patient care. As Dr. Barad puts it: “Agentic AI takes your workflow and collapses it, augments it, supercharges it, and makes it more performant.”
At HSS, AI agents have already been deployed in multiple areas. They handle complex backend processes, such as insurance claims that previously took several weeks to complete and involved both HSS staff and a third-party contractor to handle the volume. Now, says Dr. Barad, AI agents complete 1,100 claims per month. They’ve reduced the appeals stage from 45 minutes to five and improved the success rate of those appeals from 65% to 100% in the nine months since implementation. HSS now handles all claims in-house.
Building on that success, HSS is now deploying AI agents in non-clinical patient-facing settings with an AI scheduling and triage service, as part of a collaboration with enterprise agentic AI developer Ema Unlimited. The service is accessible 24/7 via web, text, or phone. It uses conversational AI to ask patients clarifying questions about their condition and then books appointments with the most appropriate clinician, factoring in location, insurance coverage, and physician availability. “It completes the whole loop,” says Dr. Barad. The AI agent is trained on “all of our context, all of our rules, and all of our knowledge base,” he adds, providing patients with streamlined access to highly specialist knowledge from world-leading surgeons.
Given the high-stakes decisions delegated to AI agents, the triage service has built-in safeguards—sensitive, complex, or uncertain scenarios are escalated to human specialists. Every decision made by the AI agent is auditable and human staff can step in at any point. Patient data is kept secure and the system is trained on all HSS protocols, policies, and care pathways. By keeping humans in the loop, Ema says its technology strikes the balance between efficient automation, patient-first safety, and human-informed decision making.
As the technology becomes more prolific, it will be incumbent on providers to ensure they have these sorts of guardrails embedded into systems, says Dr. Barad. At HSS all decisions around the technology are filtered through an AI subcommittee that Dr. Barad co-chairs alongside a senior nursing executive. AI agents that may touch on patient care will be scrutinized with far more rigor than, say, backend processes, he explains.
AI agents prompt systems-level change
For example, Dr. Barad has plans to create a dedicated AI lab at the HSS main campus in New York City—a move that aims to democratize access to the technology across the organization. It will be open to all staff looking to understand or build AI agents, he explains, with informative classes and one-on-one training. “We’re getting agentic AI into everybody’s hands,” he says. This echoes research by Deloitte, which found that leading agentic AI adopters in health care were far more likely to have opted for multiagent solutions, redesigning end-to-end workflows rather than sticking to narrow solutions or individual use cases.
The key, it appears, is to integrate AI agents across the entire enterprise, treating them as a general-purpose technology. As Dr. Barad puts it: “It’s wrong to think of agentic AI in use cases… It’s a general-purpose technology, analogous to electricity.”
In practice, this means health-care providers need to set the right foundation to achieve value with agentic AI. This includes creating a unified data strategy, one that integrates fragmented data sources across an organization to create a single, comprehensive source of truth. In health care, data is often split across multiple departments and providers, each with their own legacy IT system.
In systems that rely on fragmented data sources, metrics often lack standardized definitions too. For example, Dr. Barad says that each hospital he’s worked in has had a slightly different definition for “time to start surgery,” a metric commonly used to gauge operating room efficiency. This level of fragmentation impedes AI agents from retrieving information from different sources or applications and assimilating the tacit knowledge that differentiates them from other technologies.
By creating greater interoperability of data at HSS, patient-facing AI agents can draw from a patient’s clinical care history and existing recommendations from their clinician, combine this information with current symptoms, and decide whether a situation requires escalation before notifying the correct specialist and informing the patient.
Building better outcomes
For Dr. Barad, the potential for AI agents to overhaul health care and alleviate the current pressures on resources, access, and patient care is huge.
He envisions a future in which 90% of non-clinical health-care tasks could be administered by AI agents, freeing clinicians up for what he calls white-glove work, meaning the most complex, specialized, and sensitive cases.
Most health-care providers seem equally optimistic. According to research by KPMG, 84% of providers are already comfortable handing decision making about specific processes over to AI agents.
“We’re spending so much time on keyboards and computers right now that we’re actually not doing what we should be doing,” says Dr. Barad. “This is going to rehumanize health care.”
This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.
