Background: Hospital in the home (HITH) provides home-based care as an alternative to traditional hospitalization. In response to the COVID-19 Omicron wave, a public hospital in the rural Western portion of Southeast Queensland implemented a virtual HITH service to support adults, maternity patients, and children with moderate COVID-19 symptoms and additional health concerns. Although the pandemic accelerated the uptake of virtual care within HITH models, existing literature has focused on clinical outcomes, with limited evidence on key implementation outcomes. Objective: Using the RE-AIM (reach, effectiveness, adoption, implementation, and maintenance) framework, this study evaluated the implementation of the virtual COVID-19 HITH service and identified factors influencing its implementation, to inform ongoing service development and support potential scaling of this model of care. Methods: The RE-AIM implementation science framework was selected to guide the evaluation, capturing both clinical and contextual dimensions of implementation at both individual and organizational levels. Quantitative data on service usage and costs were retrospectively extracted from electronic medical records and finance records, while patient experience data were drawn from patient-reported experience measures surveys. Qualitative data were collected through one-on-one interviews with patients and staff. All data sources were analyzed separately and then triangulated within the RE-AIM framework to understand what occurred, how, and why. Results: The service admitted 3192 patients, most of whom were female (2027/3192, 63.5%), English-speaking (3140/3192, 98.4%), and residing in socioeconomically disadvantaged areas (1879/3192, 58.9%) (reach). The model was feasible and safe to implement, managing 3240 admissions with no reported deaths. Patients valued continuous access to care and described better recovery experiences at home (effectiveness). Staff viewed the model as appropriate for identifying and managing high-risk patients in the community, easing pressure on hospital beds (adoption). The service cost Aus $ 5.4 million (US $3.5 million) over 11 months. Implementation barriers included the urgency of the pandemic scenario, limited infrastructure and human resources, and changing requirements in relation to COVID-19. These were mitigated by several people factors that were critical to its successful implementation, including a consultant-led structure, staff commitment, and adaptability (implementation). The service saved 16,651 inpatient bed days before being integrated into core HITH operations. The experience strengthened staff capabilities in emergency response, virtual care delivery, and strategic planning. The model shows promise for broader application into pediatric care, though further work is needed to enhance interdepartmental collaboration and staff recognition (maintenance). Conclusions: This study demonstrated that a virtual HITH model can be implemented effectively and safely at scale. Findings support its potential for integration into routine care, provided that adequate resource planning, a skilled and multidisciplinary workforce, well-defined care pathways, and equity-focused strategies are in place.

As iOS 26 is rolling out to all users, developers have been updating their apps to include features powered by Apple's local AI models.

Nature, Published online: 19 September 2025; doi:10.1038/d41586-025-03055-y

Scientists used artificial intelligence to write coherent viral genomes, using them to synthesize bacteriophages capable of killing resistant strains of bacteria.

Nature, Published online: 19 September 2025; doi:10.1038/d41586-025-02362-8

Artificial intelligence could reduce the strain that ageing populations put on medical systems.

You’re reading the web edition of STAT’s Health Tech newsletter, our guide to how technology is transforming the life sciences. Sign up to get it delivered in your inbox every Tuesday and Thursday.

Califf warns AI in health care ‘overhyped’

On a makeshift stage in a Midtown Manhattan office earlier this week,former Food and Drug Administration Commissioner Robert Califf struck a measured tone about the potential for artificial intelligence in health care. Asked whether the technology was overhyped he said it was. “I hear way too much about the money. I’m not hearing a lot of human values coming through discussions,” he said. Adding:

“Almost all of the technology is being applied to optimizing the financial status of healthcare delivery entities or companies that are making medical products and that’s not aligned with equitable, better patient outcomes. So until someone puts a soul back in the system, I think it’s going to get worse and worse.”

Continue to STAT+ to read the full story…

© Adobe

Three scholars discovered a strange mirror deep in the forest. It spoke to them in a soothing voice and answered all their questions warmly, knowledgeably, and eloquently.

The captivated scholars became obsessed, whispering one secret after another to the mirror. It replied with affection, promise, and meaning that kept them returning to it. They began ignoring one another, each convinced the mirror “understood” them best.

Read the rest…

© Adobe

Nature Biotechnology, Published online: 18 September 2025; doi:10.1038/s41587-025-02820-8

The COP16 decision established a multilateral mechanism for digital sequence information (DSI) benefit-sharing. This Comment brings together insights from academia and commercial DSI researchers to assess what has been accomplished so far, identify remaining challenges and describe elements under discussion to support collective goals.

Background: Despite artificial intelligence (AI) models demonstrating high predictive accuracy for early cholangiocarcinoma recurrence, their clinical application faces challenges, such as reproducibility, generalizability, hidden biases, and uncertain performance across diverse datasets and populations, raising concerns about their practical applicability. Objective: This meta-analysis aims to systematically assess the diagnostic performance of AI models using computed tomography (CT) imaging to predict early recurrence of cholangiocarcinoma. Methods: A systematic search was conducted in PubMed, Embase, and Web of Science for studies published up to May 2025. Studies were selected based on the Participants, Index test, Target condition, Reference standard, Outcomes, and Setting (PITROS) framework. Participants included patients diagnosed with cholangiocarcinoma (including intrahepatic and extrahepatic locations). The index test was AI techniques applied to CT imaging for early recurrence prediction (defined as within 1 year), while the target condition was early recurrence of cholangiocarcinoma (positive group: recurrence; negative group: no recurrence). The reference standard was pathological diagnosis or imaging follow-up confirming recurrence. Outcomes included sensitivity, specificity, diagnostic odds ratio (DOR), and area under the receiver operating characteristic curve (AUC), assessed in both internal and external validation cohorts. The setting comprised retrospective or prospective studies using hospital datasets. Methodological quality was assessed using an optimized version of the revised Quality Assessment of Diagnostic Accuracy Studies-2 tool. Heterogeneity was assessed using the I² statistic. Pooled sensitivity, specificity, DOR, and AUC were calculated using a bivariate random-effects model. Results: A total of 9 studies with 30 datasets involving 1537 patients were included. In internal validation cohorts, CT-based AI models showed a pooled sensitivity of 0.87 (95% CI 0.81-0.92), specificity of 0.85 (95% CI 0.79-0.89), DOR of 37.71 (95% CI 18.35-77.51), and AUC of 0.93 (95% CI 0.90-0.94). In external validation cohorts, pooled sensitivity was 0.87 (95% CI 0.81-0.91), specificity was 0.82 (95% CI 0.77-0.86), DOR was 30.81 (95% CI 18.79-50.52), and AUC was 0.85 (95% CI 0.82-0.88). The AUC was significantly lower in external validation cohorts compared to internal validation cohorts (P<.001). Conclusions: Our results show that CT-based AI models predict early cholangiocarcinoma recurrence with high performance in internal validation sets and moderate performance in external validation sets. However, the high heterogeneity observed may impact the robustness of these results. Future research should focus on prospective studies and establishing standardized gold standards to further validate the clinical applicability and generalizability of AI models.

Background: Artificial intelligence (AI) and Large Language models (LLMs) are increasing in sophistication and are being integrated into many disciplines. The potential for LLMs to augment clinical decisions is an evolving area of research. Objective: This study compared the responses of over 1000 kidney specialist physicians (nephrologists) to outputs of commonly used LLMs using a questionnaire determining when a kidney biopsy should be performed. Methods: This research group completed a large online questionnaire for nephrologists to determine when a kidney biopsy should be performed. The questionnaire was co-designed with patient participation, refined through multiple iterations, then piloted locally before international dissemination. It was the largest international study in the field and demonstrated variation between human clinicians in biopsy propensity relating to human factors such as sex and age, as well as systemic factors such as country, job seniority and technical proficiency. The same questions were put to both human doctors and LLMs in an identical order in a single session. Eight commonly used LLMs were interrogated: Chat GPT 3.5, Mistral Hugging Face, Perplexity, Microsoft Co-pilot, Llama 2, GPT 4.0, MedLM and Claude 3. The most common response given by clinicians (human mode) to each question was taken as the baseline for comparison. Questionnaire responses to the indications and contraindications for biopsy generated a score (0-44) reflecting biopsy propensity, in which a higher score was used as a surrogate marker for an increased tolerance of potential associated risks. Results: The ability of LLMs to reproduce human expert consensus varied widely with some models demonstrating a balanced approach to risk in a similar manner to humans, whilst other models reported outputs at either end of the spectrum for risk tolerance. In terms of agreement with the human mode, Chat GPT 3.5 and GPT 4.0 (Open AI) had the highest levels of alignment, with the human mode selected in 6/11 questions. The total biopsy propensity score generated from the human mode was 23/44. Both Open AI models produced similar propensity scores between 22 and 24, however Llama 2 and MS Co-pilot also reported scores within this range, but with poorer response alignment to the human mode at only 2/11 questions. The most risk averse model in this study was MedLM with a propensity score of 11 and the least risk averse model was Claude 3 with a score of 34. Conclusions: LLM outputs demonstrated a modest ability to replicate human clinical decision making in this study, however the performance varied widely between LLM models. Questions with more uniform human responses produced LLM outputs with greater alignment, whereas in questions with low levels of human consensus there was poor output alignment. This may limit the practical use of LLMs in real world clinical practice.

Background: Survivors of stroke often face persistent challenges accessing postdischarge care due to mobility limitations, transportation burdens, and inflexible scheduling. Teleconsultation has emerged as a potential solution to improve continuity of care, but its perceived strengths and limitations from the patient perspective remain insufficiently understood. Objective: This study aimed to explore the experiences of survivors of stroke with a nurse-led teleconsultation program to (1) identify perceived capabilities; (2) understand limitations in usability, accessibility, and clinical function; and (3) generate patient-informed recommendations for improvement. Methods: A qualitative study was embedded within a 3-month nurse-led teleconsultation intervention delivered by advanced practice nurses. A total of 21 survivors of ischemic stroke (aged 45-76 y; female: n=11, 52%) who had preserved cognitive function (Montreal Cognitive Assessment score ≥22) and smartphone access participated in 6 focus groups conducted via Zoom. Data were analyzed thematically using an established framework. Data saturation was achieved. Results: Participants widely valued teleconsultation for reducing logistical burdens; enhancing access; and offering a more comfortable, emotionally supportive setting for follow-up care. Many reported increased awareness and motivation for self-monitoring. However, limitations included an inability to perform physical assessments or respond to emergencies; digital and usability barriers, especially among older users; and scheduling inflexibility. Participants emphasized the need for patient-initiated follow-up mechanisms, physician collaboration for medication management, and greater support for users considered digitally marginalized. They also highlighted the potential of teleconsultation to serve as a triage tool, reserving in-person care for complex cases. Conclusions: Nurse-led teleconsultation was perceived as a convenient and supportive modality for poststroke care, particularly for stable follow-ups and psychosocial support. However, its long-term viability depends on addressing clinical and technical limitations, enhancing user autonomy, and integrating interdisciplinary input. By centering the lived experiences of survivors of stroke, this study offers concrete recommendations to guide the development of more inclusive, responsive, and patient-centered teleconsultation models.

Background: With the increasing use of social media platforms for health communication, the quality of health educational short videos (HESVs) has become a key concern. However, no standardized framework exists to evaluate the quality of health videos on social media, highlighting the need for a comprehensive evaluation system. Objective: The aim of this study is to develop a valid and structured evaluation tool for assessing the quality of HESVs on social media. Methods: The initial evaluation indicators obtained from the literature review and brainstorming undertaken in the study group were provided to the nominal group reference Lasswell’s 5W communication model, and two rounds of nominal group technique (NGT) were carried out to screen, add, revise, and adjust indicators, and reach a consensus of evaluation system. The indicators were then ranked based on their significance, as scored by the experts using the analytic hierarchy process. The content validity was assessed by experts who rated the relevance of each indicator on a 4-point Likert scale. Results: The primary indicators include communicator, communication content, communication channel, and communication effect, along with 13 secondary indicators and 34 tertiary indicators. 11 experts were enrolled in the NGT, 45% of experts had a doctoral degree, 80% of them were ranked associate professor or professor. The average familiarity coefficient of each key indicator of the NGT was 0.85. The average values of the expert judgment coefficient and authority coefficient were 0.93 and 0.85, respectively. In Round 1 of NGT, the “Communication target” of 5 primary indicators, 7 of 20 secondary indicators, and 66 of 94 tertiary indicators did not reach a consensus, and therefore, they were not deleted and will proceed to the next round of NGT. In Round 2 NGT, 1 primary indicator, 7 secondary indicators, and 59 tertiary indicators were deleted based on the consensus criteria. After the two rounds of NGT, 4 primary indicators, 13 secondary indicators, and 34 tertiary indicators finally reached a consensus. Among primary indicators, communication content was found to be the most influential, accounting for 45.68%. Among secondary indicators, credibility, scientificity, availability, and social attention were the most influential indicators, with priorities of 56.67%, 24.26%, 74.62%, and 39.89% in their respective categories. Among tertiary indicators, ‘Become a hot search recommended by the platform’ was the most influential indicator with a weight of 0.07. The content validity of all the evaluation indicators were 0.73 – 1.0, and the scale-level content validity index (average) was 0.87, which was indicated as acceptable. Conclusions: The evaluation system for the quality of HESVs on social media (LassVQ) was developed, and its validity was acceptable. The proposed evaluation system can be used in conjunction with qualitative methods to gain a holistic perspective on the multidimensional quality of HESVs on social media.

Personalized cancer vaccines aim to broaden the anti-tumor T cell repertoire by targeting neoantigens unique to each patient’s tumor, but immunogenicity has been inconsistent. In this issue of Cell, Blass, Keskin, Tu et al. evaluate NeoVaxMI, a multi-adjuvant personalized synthetic long-peptide vaccine administered with nivolumab in patients with melanoma. NeoVaxMI elicited stronger CD4+ and CD8+ responses than earlier iterations, and vaccine-induced T cells trafficked to regressing metastatic lesions.

Nature Medicine, Published online: 18 September 2025; doi:10.1038/s41591-025-03962-7

The arts, according to a systematic synthesis of data from 95 studies (across 26 countries), may support non-communicable disease prevention by providing opportunities for increased physical activity, and helping to address social forces that contribute to health inequities.

Nature Medicine, Published online: 18 September 2025; doi:10.1038/s41591-025-03942-x

A model developed with data from 19,403 patients with colorectal cancer for prediction of 1-year mortality is used as a decision support tool in a prospective cohort, showing promising results in reducing postoperative complications.

Andrea Magnorsky presented on Byte-Sized Architecture at Cloud Native Summit 2025, as a format for building shared understanding through small, recurrent workshops. Ahilan Ponnusamy and Andreas Grabner discussed the Technology Operating Model for AI adoption. Both approaches drew on the Open Practice Library for human-centred collaboration and driving architectural evolution.

By Rafiq Gemmail

Background: Among the alternative solutions being tested to improve access to genetic services, chatbots (or conversational agents) are being increasingly used for service delivery. Despite the growing number of studies on the accessibility and feasibility of chatbot genetic service delivery, limited attention has been paid to user interactions with chatbots in a real-world health care context. Objective: We examined users’ interaction patterns with a pretest cancer genetics education chatbot as well as the associations between users’ clinical and sociodemographic characteristics, chatbot interaction patterns, and genetic testing decisions. Methods: We analyzed data from the experimental arm of Broadening the Reach, Impact, and Delivery of Genetic Services, a multisite genetic services pragmatic trial in which participants eligible for hereditary cancer genetic testing based on family history were randomized to receive a chatbot intervention or standard care. In the experimental chatbot arm, participants were offered access to core educational content delivered by the chatbot with the option to select up to 9 supplementary informational prompts and ask open-ended questions. We computed descriptive statistics for the following interaction patterns: prompt selections, open-ended questions, completion status, dropout points, and postchat decisions regarding genetic testing. Logistic regression models were used to examine the relationships between clinical and sociodemographic factors and chatbot interaction variables, examining how these factors affected genetic testing decisions. Results: Of the 468 participants who initiated a chat, 391 (83.5%) completed it, with 315 (80.6%) of the completers expressing a willingness to pursue genetic testing. Of the 391 completers, 336 (85.9%) selected at least one informational prompt, 41 (10.5%) asked open-ended questions, and 3 (0.8%) opted for extra examples of risk information. Of the 77 noncompleters, 57 (74%) dropped out before accessing any informational content. Interaction patterns were not associated with clinical and sociodemographic factors except for prompt selection (varied by study site) and completion status (varied by family cancer history type). Participants who selected ≥3 prompts (odds ratio 0.33, 95% CI 0.12-0.91; P=.03) or asked open-ended questions (odds ratio 0.46, 95% CI 0.22-0.96; P=.04) were less likely to opt for genetic testing. Conclusions: Findings highlight the chatbot’s effectiveness in engaging users and its high acceptability, with most participants completing the chat, opting for additional information, and showing a high willingness to pursue genetic testing. Sociodemographic factors were not associated with interaction patterns, potentially indicating the chatbot’s scalability across diverse populations provided they have internet access. Future efforts should address the concerns of users with high information needs and integrate them into chatbot design to better support informed genetic decision-making.

Artificial intelligence can draw cat pictures and write emails. Now the same technology can compose a working genome.

A research team in California says it used AI to propose new genetic codes for viruses—and managed to get several of these viruses to replicate and kill bacteria.

The scientists, based at Stanford University and the nonprofit Arc Institute, both in Palo Alto, say the germs with AI-written DNA represent the “the first generative design of complete genomes.”

The work, described in a preprint paper, has the potential to create new treatments and accelerate research into artificially engineered cells. It is also an “impressive first step” toward AI-designed life forms, says Jef Boeke, a biologist at NYU Langone Health, who was provided an advance copy of the paper by MIT Technology Review.  

Boeke says the AI’s performance was surprisingly good and that its ideas were unexpected. “They saw viruses with new genes, with truncated genes, and even different gene orders and arrangements,” he says.

This is not yet AI-designed life, however. That’s because viruses are not alive. They’re more like renegade bits of genetic code with relatively puny, simple genomes. 

In the new work, researchers at the Arc Institute sought to develop variants of a bacteriophage—a virus that infects bacteria—called phiX174, which has only 11 genes and about 5,000 DNA letters.

To do so, they used two versions of an AI called Evo, which works on the same principles as large language models like ChatGPT. Instead of feeding them textbooks and blog posts to learn from, the scientists trained the models on the genomes of about 2 million other bacteriophage viruses.

But would the genomes proposed by the AI make any sense? To find out, the California researchers chemically printed 302 of the genome designs as DNA strands and then mixed those with E. coli bacteria.

That led to a profound “AI is here” moment when, one night, the scientists saw plaques of dead bacteria in their petri dishes. They later took microscope pictures of the tiny viral particles, which look like fuzzy dots.

“That was pretty striking, just actually seeing, like, this AI-generated sphere,” says Brian Hie, who leads the lab at the Arc Institute where the work was carried out.

Overall, 16 of the 302 designs ended up working—that is, the computer-designed phage started to replicate, eventually bursting through the bacteria and killing them.

J. Craig Venter, who created some of the first organisms with lab-made DNA nearly two decades ago, says the AI methods look to him like “just a faster version of trial-and-error experiments.”

For instance, when a team he led managed to create a bacterium with a lab-printed genome in 2008, it was after a long hit-or-miss process of testing out different genes. “We did the manual AI version—combing through the literature, taking what was known,” he says. 

But speed is exactly why people are betting AI will transform biology. The new methods already claimed a Nobel Prize in 2024 for predicting protein shapes. And investors are staking billions that AI can find new drugs. This week a Boston company, Lila, raised $235 million to build automated labs run by artificial intelligence.

Computer-designed viruses could also find commercial uses. For instance, doctors have sometimes tried “phage therapy” to treat patients with serious bacterial infections. Similar tests are underway to cure cabbage of black rot, also caused by bacteria.

“There is definitely a lot of potential for this technology,” says Samuel King, the student who spearheaded the project in Hei’s lab. He notes that most gene therapy uses viruses to shuttle genes into patients’ bodies, and AI might develop more effective ones.

The Stanford researchers say they purposely haven’t taught their AI about viruses that can infect people. But this type of technology does create the risk that other scientists—out of curiosity, good intentions, or malice—could turn the methods on human pathogens, exploring new dimensions of lethality.

“One area where I urge extreme caution is any viral enhancement research, especially when it’s random so you don’t know what you are getting,” says Venter. “If someone did this with smallpox or anthrax, I would have grave concerns.”

Whether an AI can generate a bona fide genome for a larger organism remains an open question. For instance, E. coli has about a thousand times more DNA code than phiX174 does. “The complexity would rocket from staggering to … way way more than the number of subatomic particles in the universe,” says Boeke.

Also, there’s still no easy way to test AI designs for larger genomes. While some viruses can “boot up” from just a DNA strand, that’s not the case with a bacterium, a mammoth, or a human. Scientists would instead have to gradually change an existing cell with genetic engineering—a still laborious process.

Despite that, Jason Kelly, the CEO of Ginkgo Bioworks, a cell-engineering company in Boston, says exactly such an effort is needed. He believes it could be carried out in “automated” laboratories where genomes get proposed and tested and the results are fed back to AI for further improvement.

 “This would be a nation-scale scientific milestone, as cells are the building blocks of all life,” says Kelly. “The US should make sure we get to it first.”

Nature, Published online: 17 September 2025; doi:10.1038/s41586-025-09505-x

People cheat more when they delegate tasks to artificial intelligence, and large language models are more likely than humans to comply with unethical instructions—a risk that can be minimized by introducing prohibitive, task-specific guardrails.

Meta launches Hyperscape into beta, allowing users to turn real world spaces into digital spaces for VR.

Nature, Published online: 17 September 2025; doi:10.1038/d41586-025-02993-x

A modified large language model called Delphi-2M analyses a person’s medical records and lifestyle to provide risk estimates for more than 1,000 diseases.