The proposal is conceptual, not a tested system, but it offers a technically plausible route to shortening the interval between a deviation and its detection. The authors, Markus Knott and Markus Krebs, describe the shadow as a unidirectional, purpose-bound and deliberately incomplete projection of institutional conformity. It would observe but not act; its findings would inform clinicians and managers rather than make or alter clinical decisions.

This terminology matters because “digital twin” has become an umbrella phrase. One influential engineering classification distinguishes a digital model with no automated exchange, a digital shadow with one-way data flow, and a digital twin with bidirectional flow. The authors adopt that stricter distinction, even though ISO/IEC 30173:2023 uses a broader concept based on data connections and synchronization.

Their narrower definition foregrounds accountability: in a clinical organization, whether a representation can act back on the institution is not merely technical. The organizational focus is not new, and the paper does not treat it as unexplored territory. Research on organizational digital twins has used operational and contextual data to represent how resources are deployed and value delivered, while process-mining work has modeled organizational processes with object-centric Petri nets that track interacting entities such as patients, specimens and board decisions.

Yet researchers have challenged the fidelity promised by the twin metaphor. Organizational routines contain performative actions that leave event records, but also an ostensive layer of shared, partly tacit understanding. A synchronized model would also have to represent agency, conflict and emergence.

These are treated less as missing features than as properties that make a complete organizational twin difficult to justify. Healthcare research includes several well-populated digital-twin strands involving patient physiology, medical devices and hospital operations. Hospital-operations twins model patient flow, bed allocation and capacity, while the broader healthcare literature includes patient-level physiological twins and device-level twins.

These systems generally address throughput, resource allocation and safety rather than whether an institution currently satisfies the requirements that define a certified cancer center. The proposed shadow instead uses an externally maintained certification standard as its reference model, spanning structure, process, outcome and governance. Versioning is essential because a change in a finding might result from a revised requirement rather than altered practice.

Every finding would therefore need to identify the requirement version used, the data used and the time at which the assessment occurred. The architecture has four layers with deliberately different update speeds. The structure layer draws on personnel qualifications, continuing-education records, equipment and infrastructure registers, cooperation agreements and network-partner status, generally updating over months as events occur.

The process layer includes tumor-board case volumes, pretherapeutic presentation rates, recommendation completeness and implementation, pathway timestamps and trial-screening records, with potential daily refreshes. The outcome layer uses cancer-registry submissions, survival and recurrence data, complications, morbidity and patient-reported outcomes, whose latency ranges from months to years. The fourth, reflexive governance layer records audit findings, corrective actions, critical-incident reports, complaints and internal-audit minutes.

Continuous therefore means latency-aware, not instantaneous; each alert must disclose the age of its underlying data. A worked example shows how a requirement could become an automated case assessment followed by human review. Indicator 2a of the German Colorectal Cancer Center Programme requires elective rectal-cancer patients and elective patients with stage IV colon cancer to be presented to an interdisciplinary tumor board before therapy, with a target of at least 95%.

A proposed rule would compare the tumor-board recommendation with the first treatment timestamp for cases in the defined denominator, then calculate the period rate and flag it if it fell below 95%. Individual cases would be classified as conformant, non-conformant or unassessable; the 95% threshold would not be applied to each case separately. The system would retain case and source-record identifiers, rule and standard versions, and evaluation time.

A quality officer would then confirm the finding, document a justified exception, or open corrective action. The software could assemble evidence and flag gaps; deciding whether a gap reflects incomplete documentation, a breached requirement or a legitimate clinical circumstance would remain human work. The authors distinguish four stages of maturity along two separate axes: technical coupling and analytical function.

ODS-0 is the familiar manual digital model that reports last year’s indicator values. ODS-1 adds automated one-way flow to provide current descriptive values and is described as technically within reach, although not yet demonstrated in a certified center. ODS-2, the central proposal, compares continuous event data with machine-readable certification requirements and reports when and in which dimension conformity may have been lost.

ODS-3 would add predictive analytics, but forecasting would not make the architecture more bidirectional; only automated actuation would cross the shadow–twin boundary, which the authors deliberately leave out of scope. None of these proposed levels has been implemented or validated in a certified center. A possible near-term test is narrower than a general-purpose digital twin.

Certification bodies would first need to express indicator definitions in machine-readable form. A single-center demonstrator could then apply process-mining conformance checking to tumor-board data. A multicenter study would run the shadow alongside conventional audits, measuring concordance, detection time, false-positive rates and staff time.

The authors regard workload as a co-primary outcome because a more frequent quality system that adds documentation would defeat its purpose. In an earlier intervention at one certified center, redesigning documentation for 12 tumor boards reduced coordinators’ preparation time from 66.0 to 46.5 hours per week and preparation workload for radiologists and pathologists by 22.9% and 52.7%, respectively. That result concerned one data-heavy process rather than organizational conformity, but it illustrates the relief that could make automation legitimate.

Governance is not a secondary concern. If the shadow is intended to observe institutional rather than individual conformity and does not enter individual clinical decisions, the authors classify it as a management instrument outside the EU Medical Device Regulation’s device definition. That boundary could shift if the system were used to support individual clinical decisions.

Any AI component, including language models extracting signals from free-text tumor-board minutes, would also require implementation-specific assessment under the EU AI Act. The authors consider an institutional conformity system unlikely to be high-risk under Annex III, while warning against uses that identify individual clinicians or drift toward clinical decision support. Data-protection analysis must likewise consider the changed purpose and aggregation of data already processed for treatment, documentation, quality assurance and registry reporting; the European Health Data Space provides mechanisms for secondary use, not a blanket permission.

The hardest risk is that continuous measurement can create the appearance of quality without improving it. Measures can become targets in ways that distort behavior, and evidence from the English health service suggests real-time transparency may intensify reactive gaming and allow the substitution of remote surveillance for site visits. Continuous observation should be most useful when detecting unintended drift, such as a required discipline becoming unreliable or case volume falling below an annual minimum, and when findings support interactive managerial discussion rather than punishment.

A reflexive layer may expose suspicious patterns, including alerts dismissed without review or accumulating exceptions, but only human auditors can interpret them. The proposal’s most important test is not how comprehensively a digital shadow sees a cancer center, but whether clinicians and quality managers perform less manual evidence-gathering after it exists. It would not reproduce every tacit feature of multidisciplinary care, replace auditors or guarantee better clinical outcomes.

Nor does the paper demonstrate that a center has already reduced non-conformity. Its achievement is a disciplined architecture for using information that certified centers already generate, combining technical integration with versioned conformity evaluation and retaining human accountability for interpretation and action. If validated, that approach could narrow the blind spot created by annual snapshots and shorten the time between a center departing from its own standard and anyone noticing.

The result would be a monitoring signal, not a self-acting digital twin—and that boundary is presented as a feature of responsible design rather than a failure of ambition.

Subject of Research: A proposed digital-shadow architecture for continuous conformity monitoring in certified cancer centers

Article Title: The organizational digital shadow of the certified cancer center—from yearly snapshots to continuous signals

Article References: Knott, M., & Krebs, M. (2026). The organizational digital shadow of the certified cancer center—from yearly snapshots to continuous signals. Frontiers in Digital Health, 8, Article 1955074. https://doi.org/10.3389/fdgth.2026.1955074

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1955074

Keywords: digital shadow, certified cancer centers, continuous auditing, quality assurance, health informatics, process mining, organizational conformity, oncology, digital twins, governance, organizational, digital

The study, published in BMC Health Services Research, used a qualitative descriptive design and individual semi-structured interviews conducted online between February and April 2025. The participants, whose mean age was 59.9 years, represented a demographic that is often overlooked in digital health research: middle-aged and older adults in a middle-income country, many of whom had limited access to structured follow-up care after finishing treatment. The research team analyzed the interview transcripts using inductive thematic analysis, an approach that lets patterns emerge from the data itself rather than testing predefined categories, and reported the work in accordance with the COREQ guideline, a 32-item checklist designed to make qualitative research transparent and reproducible. Ethics approval came from both The Chinese University of Hong Kong and Nam Dinh University of Nursing in Vietnam, and all participants gave documented informed consent.

Three major themes emerged from the analysis. The first captured what the researchers call multidimensional survivorship experiences: physical changes that persisted long after treatment ended, emotional turmoil that survivors frequently kept hidden, financial challenges associated with cancer care, a sense of navigating the journey in isolation, and a search for spiritual support. This clustering is significant because it underscores that cancer survivorship is not simply a medical state defined by the absence of disease. It is a lived experience in which bodily changes, psychological distress, economic pressure and existential questions interact and compound one another. A survivor who cannot afford transport to a follow-up visit is also a survivor who may miss early detection of recurrence, and the anxiety of not knowing what symptoms matter can amplify both the emotional and financial burdens at once.

The second theme mapped the survivors’ supportive care needs, and it reads almost like a specification document for a digital health intervention. Participants wanted information from reliable sources, which is a pointed request in an era when cancer misinformation spreads readily on social media. They wanted help recognizing the symptoms of cancer recurrence, so that normal post-treatment sensations could be distinguished from warning signs that warrant a clinic visit. They wanted reminders for post-treatment follow-up, a low-technology feature with potentially high-impact consequences for early detection. And they wanted health promotion strategies, concrete guidance on diet, activity and lifestyle choices that could reduce the risk of the cancer returning or a second cancer developing. Each of these needs maps directly onto a function that a well-designed smartphone application can deliver at scale and at low marginal cost.

The third theme addressed how such a program should actually be designed, and here the survivors offered technical and cultural guidance that developers rarely hear from this population. Participants expressed preferences for specific features of mHealth design, but they also flagged age-related digital support as essential. With a mean age approaching sixty, many survivors are not digital natives, and an application that assumes fluency with menus, notifications and account creation risks excluding precisely the people it aims to serve. The survivors indicated they would need assistance and simplified interfaces to use such tools confidently. They also asked that cultural elements be incorporated into the program, reflecting beliefs, practices and communication styles familiar to Vietnamese patients, and they identified practical barriers to using mHealth, from unfamiliarity with smartphones to concerns about privacy and the sheer novelty of managing health through a screen.

The researchers situated their findings within the Cancer Survivorship Care Framework, a conceptual model that organizes survivorship care around the domains of surveillance for recurrence, prevention of new cancers, intervention for delayed physical and psychosocial effects, and coordination of care among specialists and primary providers. The alignment matters because frameworks only become useful when they can be translated into interventions that patients will actually adopt. The study’s results provide exactly that bridge: each theme corresponds to a framework domain, and each expressed need corresponds to a candidate feature for an mHealth program. Information from reliable sources supports prevention and health literacy; symptom-recognition guidance supports surveillance; follow-up reminders support coordination of care; and attention to emotional and spiritual needs supports intervention for psychosocial effects.

Why does this matter on a global scale? Colorectal cancer is among the most commonly diagnosed cancers worldwide, and as screening and treatment improve, the population of survivors continues to grow. Health systems in high-income countries are already straining to provide survivorship care, and in countries like Vietnam the challenge is magnified by uneven access to specialists, long travel distances and resource constraints at the primary care level. Mobile health offers a technological shortcut, but only if the programs are built around what survivors actually need rather than what developers assume they want. Studies like this one, grounded in the lived experience of patients, are the raw material for that co-design process. The authors explicitly frame their findings as informing the future co-design and development of a culturally appropriate mHealth survivorship program involving colorectal cancer survivors and relevant healthcare stakeholders.

The emphasis on co-design is more than a methodological flourish. Digital health has a well-documented attrition problem: applications are downloaded with enthusiasm and abandoned within weeks, often because they were designed without meaningful input from end users. For older cancer survivors, the stakes of abandonment are higher than engagement metrics suggest. A survivor who stops opening an application may also stop tracking symptoms, miss a follow-up reminder, and delay reporting a recurrence signal that could have been caught early. By interviewing survivors before a single line of code is written, the research team is following a model that treats patients as partners in engineering, not merely recipients of a finished product. The preferences on digital support for older users and cultural tailoring are the kind of specifications that emerge only when the people who will use the system are asked directly.

The study also carries a quieter but equally important message about isolation. The theme of navigating the journey in isolation and searching for spiritual support suggests that many survivors leave the hospital with their medical needs addressed but their social and existential needs unmet. An mHealth program cannot replace human connection, but it can reduce isolation by providing a continuous, reliable channel of information and reassurance between clinical visits, and by signaling to survivors that their ongoing needs are seen and taken seriously. Whether delivered as tailored educational content, structured check-ins or follow-up reminders, that continuity of care is precisely the gap that survivors in this study described.

For now, the findings are descriptive rather than interventional; no application has been built or tested, and the study captured the perspectives of thirty survivors in a single national context. The authors note that the research received no specific grant funding, and the published version is open access under a Creative Commons Attribution 4.0 license, allowing clinicians, designers and patient advocates anywhere to build on it. The next step, informed by this work, will be developing and evaluating an mHealth survivorship program designed with colorectal cancer survivors rather than for them. If it succeeds, the result could turn the most isolating phase of cancer care, the quiet months and years after treatment ends, into one where a trusted companion lives on the survivor’s own phone, ready with answers, reminders and reassurance whenever the uncertainty returns.

Subject of Research: Survivorship experiences and mobile health support preferences among colorectal cancer survivors in Vietnam

Article Title: Cancer survivorship experiences and preferences for mHealth program among colorectal cancer survivors: a qualitative study

Article References: Vi, D. P. N., Khanh, N. T., Duyen, D. T. H., Tuyet, N. H. A., Quyen, T. T. M., Tse, M., & Wong, C. L. (2026). Cancer survivorship experiences and preferences for mHealth program among colorectal cancer survivors: a qualitative study. BMC Health Services Research. https://doi.org/10.1186/s12913-026-15648-y

Image Credits: AI Generated

DOI: 10.1186/s12913-026-15648-y

Keywords: colorectal cancer, cancer survivorship, mHealth, mobile health, supportive care, qualitative research, Vietnam, digital health, health informatics, thematic analysis, patient-centered design, Cancer

The study, published in Frontiers in Digital Health, tackles a well-known weakness of vision-language models, the class of artificial intelligence systems designed to connect what a machine sees with what it can say. When these models generate radiology reports, they rely entirely on the representations they learned during training. If the relevant contextual information is not captured, the model may produce statements that are incorrect, unsupported, or outright fabricated, a phenomenon researchers call hallucination. Even state-of-the-art systems evaluated by trained radiologists have been shown to generate reports containing missing or false references, a serious problem in a domain where precision of diagnostic terminology matters.

The new framework, developed by Anandhu Saji, Fousia M. Shamsudeen, Chinnu Jacob, Cristina Popirlan, Claudiu-Ionut Popirlan, and D. Jude Hemanth, builds on an established vision-language model called R2Gen, which was designed specifically for medical report generation. The workflow operates in three stages. First, R2Gen processes the input chest X-ray through a ResNet-101 convolutional encoder that converts the image into a dense visual feature vector, and its language decoder produces a preliminary report. Second, a retrieval stage searches an external repository of radiology reports for cases similar to the one at hand. Third, a refinement stage powered by the Zephyr 7B large language model combines the preliminary report with the retrieved textual context and produces the final, polished report.

The retrieval mechanism is the technical heart of the system. The researchers used FAISS, an efficient library for similarity search developed for exactly this kind of vector-based lookup. For every image-report pair in the repository, the visual feature vector extracted by the R2Gen encoder is stored as the searchable vector in the FAISS index, alongside a fixed-dimensional semantic embedding of the corresponding report generated by the sentence transformer model all-MiniLM-L6-v2. At inference time, the query image yields both an image embedding and, via its preliminary report, a text embedding. A joint scoring function based on cosine similarity combines visual and semantic similarity, and the top-k most similar cases are retrieved. Their associated report information is then handed to the refinement model as textual context.

One of the most practical aspects of the design is what it does not require: retraining. The retrieval and refinement stages operate entirely at inference time, meaning the pretrained R2Gen component is left untouched. There is no additional fine-tuning of the underlying report generation model, and the retrieval repository can be extended or updated independently simply by rebuilding the retrieval index when needed. In a field where most recent advances have come from computationally expensive fine-tuning of large language models, this modular, plug-in approach offers a flexible alternative. The authors describe the two specialized modules, a Retrieval Agent and a Report Generation Agent, as functionally specialized components within a predefined pipeline rather than an autonomous multi-agent reasoning system.

To test the framework, the team used two benchmark datasets. IU X-Ray contains 7,470 chest X-ray images and 3,955 radiology reports, each structured into indication, comparison, findings, and impression sections. CheXpert Plus is far larger, with 223,000 image-report pairs from 64,700 patients and annotations for 14 chest conditions. In both cases the data were split 80:20, with the larger portion used to build the FAISS vector database and the remaining 20 percent reserved for evaluation. The retrieval databases were constructed independently for each dataset, patient-level separation was maintained, and the ground truth reports of the evaluation cases were used only for scoring, never as input to the generation process.

The results were evaluated with standard natural language generation metrics, each capturing a different aspect of text quality. BLEU measures the precision of n-gram matches, ROUGE-L focuses on recall-oriented overlap of the longest common sequences, METEOR accounts for stemming and synonym matching, and token-overlap precision, recall, and F1-score quantify how closely generated tokens match the reference. The gains over the R2Gen baseline were substantial. On CheXpert Plus, the baseline achieved a BLEU-4 score of 0.081, while the proposed framework reached 0.2085, with ROUGE-L rising from 0.246 to 0.3622 and METEOR from 0.318 to 0.3917. On IU X-Ray, BLEU-4 climbed from 0.165 to 0.3939, ROUGE-L from 0.371 to 0.5070, and METEOR from 0.187 to 0.5382.

The comparison extended to previously published state-of-the-art methods, including R2GenGPT, R2GenCSR, MambaXray-VL-L, METTransformer, Token-Mixer, and BootstrappingLLM. On CheXpert Plus, the proposed method’s BLEU-4 of 0.2085 and F1-score of 0.4546 exceeded the 0.112 and 0.335 reported for MambaXray-VL-L, the strongest prior result in that comparison. On IU X-Ray, the framework’s BLEU-4 of 0.3939, ROUGE-L of 0.5070, and METEOR of 0.5382 surpassed the 0.206, 0.401, and 0.215 reported for R2GenCSR, while also achieving precision, recall, and F1 values of 0.5997, 0.6976, and 0.6338. Qualitative analysis reinforced the numbers: the generated reports captured key diagnostic elements such as clear lung fields, absence of effusions or consolidations, a normal cardiac silhouette, and intact bony structures.

The work sits within a rapidly growing body of research on retrieval-augmented generation, or RAG, in radiology. A recent scoping review of 45 RAG-based studies found that retrieval consistently improves report generation and question answering performance, but also flagged hallucinations, inconsistent evaluation practices, and limited clinical validation as persistent challenges. Other recent approaches have explored semantically guided retrieval, factual evidence fusion, and multi-agent conversational systems. The present study distinguishes itself less by inventing a new retrieval algorithm and more by its modular organization: it shows that simply pairing an established report generator with inference-time retrieval and a separate language model for refinement can deliver measurable gains without touching the trained model.

The authors are careful about the scope of their claims. Automatic metrics like BLEU and ROUGE measure textual similarity, not clinical correctness, and improvements in these scores should not be read as direct evidence of diagnostic accuracy. The retrieval stage is also only as good as the repository behind it, so the characteristics of the stored reports inevitably shape the context supplied during refinement. Future work, they suggest, should incorporate dedicated clinical factuality measures, extend evaluation to additional datasets and imaging modalities, and explore more clinically informed retrieval strategies. Still, the message is compelling: sometimes the best way to make an AI a better writer is not to train it harder, but to give it a reference library and let it check its work against the accumulated experience of the reports that came before.

Subject of Research: Retrieval-augmented chest X-ray report generation using an agent-based framework combining R2Gen, FAISS similarity search, and Zephyr 7B report refinement

Article Title: Enhanced vision language framework for medical report generation using agent-based retrieval

Article References: Saji, A., Shamsudeen, F. M., Jacob, C., Popirlan, C., Popirlan, C.-I., & Hemanth, D. J. (2026). Enhanced vision language framework for medical report generation using agent-based retrieval. Frontiers in Digital Health, 8, Article 1989756. https://doi.org/10.3389/fdgth.2026.1989756

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1989756

Keywords: chest X-ray, radiology report generation, retrieval-augmented generation, vision-language models, R2Gen, FAISS, Zephyr 7B, medical imaging, artificial intelligence, hallucination, natural language processing, health informatics

The answer, published in BMC Health Services Research, is a qualified yes, with important caveats about what happens after a positive test. Benjamin Sloan, Annie N. Cowell, Christina Schultz, Amy Sitapati, Marie-Reine Lauer, Ryan Moran and colleagues implemented an EHR-based algorithm to identify patients at risk of latent tuberculosis infection, or LTBI, using discrete data elements that trigger an Epic Health Maintenance topic flag. The Health Maintenance topic is a structured, patient-specific EHR feature used to document and track completion of recommended preventive care. In other words, rather than relying on clinicians to remember which patients need TB risk assessments, the system itself flags the recommendation, embedding it in the same preventive-care machinery that reminds providers about vaccines and cancer screening.

The technical architecture matters because LTBI screening is fundamentally a data problem. Risk factors for LTBI, such as birth or travel in high-prevalence countries, known exposure to active TB, immunosuppression, and homelessness or incarceration, are typically recorded in the chart as structured fields: demographic entries, problem-list items, social history elements. The UC San Diego team built staff- and clinician-centered workflows on top of these discrete data elements so that risk assessment determination could be completed during routine encounters. The algorithm does not diagnose LTBI; it identifies candidates for risk assessment and, where risk factors are confirmed, for Interferon-Gamma Release Assay testing, the modern blood-based diagnostic that has largely replaced the older tuberculin skin test in primary care.

To evaluate whether the algorithm actually changed behavior, the researchers conducted a retrospective quasi-experimental study tracking monthly QuantiFERON, or QFT, orders before and after the Health Maintenance topic went live in June 2025. The statistical approach was interrupted time series segmented regression over 47 months, from January 2022 through November 2025, a method that models the baseline trend, the immediate level change at intervention, and the post-intervention slope separately. Because seasonal patterns can distort such comparisons, the team also ran a seasonally matched sensitivity analysis comparing June through November 2025 against the same months in 2022 through 2024.

The results were striking. The interrupted time series analysis demonstrated a significant immediate increase of 547.42 additional QFT orders at implementation, with a p-value below 0.001. That is, in the month the algorithm switched on, testing volume jumped by roughly 550 orders, an effect far too large to attribute to chance. The post-intervention slope subsequently declined by 103.17 orders per month, p = 0.006, suggesting the initial surge partially receded as the backlog of flagged patients was worked through. The seasonally matched comparison confirmed significantly higher mean monthly QFT orders in June through November 2025, at 2055.17, versus 1795.33 in the same months of 2022 through 2024, a mean difference of 259.83 orders with a 95 percent confidence interval of 1.19 to 518.48 and p = 0.049. The effect size, Cohen’s d of 1.35, indicates a large difference by conventional statistical standards.

Over five months of post-implementation monitoring, more than 13,000 patients were identified by the algorithm, and QFT positivity rates ran at 10 percent. That positivity rate is notable: it suggests the algorithm is not merely casting a wide net but flagging patients whose risk profiles genuinely correlate with infection. In a low-prevalence setting, indiscriminate testing would yield mostly negative results; a 10 percent yield among flagged patients indicates meaningful enrichment, the kind of signal that makes population screening economically and clinically defensible.

But the study’s most sobering findings lie downstream of the blood draw. The researchers conducted a structured chart review of QFT-positive patients to evaluate care cascade outcomes, including chest radiography completion and LTBI treatment initiation. Among 103 QFT-positive patients, 75, or 72.8 percent, had chest radiography ordered. Of those, only 50, or 66.7 percent, had a documented imaging result available. And among patients with normal imaging, the standard prerequisite for ruling out active disease before starting preventive therapy, only 23.3 percent initiated LTBI treatment. The attrition is severe at every step: roughly one in four positive patients never got a chest image ordered, a third of those ordered never produced a documented result, and fewer than one in four eligible patients actually started the antibiotics that would eliminate their infection risk.

This cascade gap is where the study’s implications become sharpest for TB elimination strategy. Latent tuberculosis infection is asymptomatic, so treatment offers no immediate benefit the patient can feel; it requires months of medication, sometimes with side effects, for a benefit that is entirely probabilistic and deferred. Public health frameworks have long recognized that screening without treatment completion does little to reduce future TB cases, since the bacterium remains in the population. The UC San Diego findings highlight that the EHR algorithm succeeded at the front end, converting a mandate into testing volume, but that diagnostic follow-up and treatment initiation remain the persistent barriers. The authors point to low primary-care encounter ordering rates and incomplete follow-up of positive results as key gaps, framing them as opportunities for workflow optimization to advance AB 2132 compliance and regional TB elimination goals.

The technical lesson for health informatics is that EHR-based screening algorithms can be built quickly and can produce measurable population-level effects, but they function best as part of a closed loop rather than a one-way flag. A Health Maintenance topic that identifies 13,000 at-risk patients and drives a 10 percent positivity rate is a triumph of data engineering; a system in which only 23 percent of eligible positive patients start treatment is a triumph of identification without completion. Closing that loop will likely require the same discrete-data discipline applied downstream: automated tracking of chest radiography orders and results, automated outreach to patients with normal imaging, and treatment initiation prompts embedded in the same preventive-care machinery that drove the testing surge. The UC San Diego team has demonstrated that the front door of TB elimination can be opened with code. The harder work, as their own cascade data show, is what happens once patients walk through it.

Subject of Research: Health Services Research

Article Title: EHR-based algorithm for latent tuberculosis infection screening in primary care: early implementation under California AB 2132

Article References: Sloan, B., Cowell, A. N., Schultz, C., Sitapati, A., Lauer, M.-R., & Moran, R. (2026). EHR-based algorithm for latent tuberculosis infection screening in primary care: early implementation under California AB 2132. BMC Health Services Research. https://doi.org/10.1186/s12913-026-15658-w

Image Credits: AI Generated

DOI: 10.1186/s12913-026-15658-w

Keywords: latent tuberculosis infection, EHR algorithm, AB 2132, primary care, San Diego, QuantiFERON, interrupted time series, care cascade, health informatics, public health policy, EHR-based, algorithm

The research team, drawing on pharmacists recruited through professional social networks and pharmacy association groups between March and June 2025, designed the investigation in two phases. The first phase was a cross-sectional survey intended to map who was using AI and why. The second was a quasi-experimental test of whether a single hour of targeted education could soften the attitudes that keep capable professionals away from the technology. Together, the two phases offer one of the most detailed portraits yet of how the digital divide plays out inside a specific health profession in a developing but technologically ambitious country.

The numbers from the first phase are striking. Of the 901 pharmacists surveyed, 75.1 percent reported familiarity with AI, yet only 41.7 percent of the entire cohort said they had actively used AI tools within the previous twelve months. Among those who were familiar with the technology, 55.5 percent had used it. And what they used it for matters: the most common application, cited by 42.3 percent of users, was verifying medication information, while a mere 8.3 percent used AI to analyze patient data or recommend dosages. The researchers suggest this pattern reflects the easy availability of large language models alongside persistent privacy concerns about patient data, a trend consistent with surveys of primary care providers elsewhere.

When the team ran the sociodemographic variables through a multivariable logistic regression, three factors emerged as independent predictors of AI use. Pharmacists who graduated after 2020 had nearly twice the odds of using AI compared with earlier graduates, an adjusted odds ratio of 1.943. Those holding a postgraduate degree had more than twice the odds of diploma holders, at 2.402. And the availability of digital infrastructure in the workplace, meaning reliable internet access, scalable storage, and application hosting, roughly doubled the odds of adoption as well, with an adjusted odds ratio of 2.099. Hospital pharmacists, who benefit from institutional information technology systems, were also significantly more likely to use AI than colleagues in privately owned or chain pharmacies.

The most intriguing finding came from an interaction analysis. The researchers hypothesized that the quality of the digital environment might moderate the apparent generational gap in AI adoption, and the statistics bore this out. Among pharmacists who graduated before 2020 and lacked workplace digital infrastructure, only 16.0 percent used AI. Yet post-2020 graduates maintained a 67.3 percent adoption rate even in low-infrastructure settings, likely by relying on personal mobile devices. When proper digital infrastructure was available, the pre-2020 cohort’s adoption rate jumped to 52.8 percent, shrinking the generational gap from 51.3 percentage points to just 7.1. The authors describe this as a technological friction effect: older pharmacists are not resisting the technology so much as being locked out of environments where it can run.

This framing challenges a common narrative that attributes slow AI uptake in developing countries to generational resistance. The Jordanian data suggest that basic connectivity is an essential leveler, echoing calls from the European Association of Hospital Pharmacists to invest in digital infrastructure, electronic prescribing, and interoperability as prerequisites for AI integration. Comparable studies in Eastern Romania, Switzerland, Greece, and Bulgaria have documented similar structural and educational barriers, indicating that the problem is not confined to the Middle East but mirrors the digitalization struggles of European pharmacy itself.

Income and workload also mattered. Pharmacists earning more than 600 Jordanian dinars per month had higher odds of AI use than the lowest earners, although the wide confidence interval on that estimate suggests some statistical uncertainty. Seeing more than fifty patients a day was independently associated with adoption, perhaps because high patient volumes create pressure to find efficiency tools. Gender, by contrast, dropped out of the picture once other variables were controlled: although the unadjusted chi-square analysis showed slightly higher use among men, the regression model found no meaningful difference between male and female pharmacists.

The second phase of the study addressed a different question: can attitudes be changed cheaply and quickly? Of the original 901 participants, 301 volunteered for a 60-minute educational webinar delivered by an external expert in health informatics and computational pharmacology. The session combined a thirty-minute lecture on AI basics and job-displacement concerns, a twenty-minute live demonstration of AI checking drug interactions, and a ten-minute ethics discussion on data privacy and algorithmic bias. Sessions were repeated bimonthly from April to July 2025 to accommodate different shift patterns, and attitudes were measured immediately before and after using paired categorical analysis with McNemar’s test.

The results were dramatic, at least in the short term. The share of pharmacists who believed AI will replace the need for pharmacists plummeted from 70.0 percent to 24.6 percent, a drop of 45.4 percentage points. Belief that AI involves high operational costs fell from 88.4 percent to 41.5 percent. Concerns about patient data confidentiality dropped by nearly 40 points, while agreement that AI helps provide optimal pharmaceutical care more than tripled, rising from 20.9 percent to 75.4 percent, and trust in AI’s reliability for drug-related information nearly quadrupled, from 15.3 percent to 60.9 percent. Notably, one belief barely moved: agreement that AI requires human oversight rose only from 19.6 percent to 22.9 percent, a change the authors interpret not as resistance but as evidence that the training successfully framed AI as a decision-support aid rather than an autonomous authority, a stance aligned with the Pharmaceutical Group of the European Union’s position that AI should assist, never replace, professional expertise.

The authors are careful about the limits of these findings. The cross-sectional design cannot establish causation, the intervention lacked a control group, and the posttest was administered immediately after the webinar, so only short-term attitudinal shifts are documented, not lasting changes in practice. The volunteers for Phase II may also have been more motivated or technologically adept than their peers, and the online recruitment method may underrepresent pharmacists in remote areas with poor internet access, precisely the population the infrastructure findings highlight. Still, the study’s practical prescriptions are clear: pharmacies need minimum standards for digital infrastructure, continuing education should incorporate AI fundamentals, and mentoring schemes that pair recent graduates’ digital fluency with senior pharmacists’ clinical experience could help close the gap from both directions. Whether a one-hour webinar produces pharmacists who actually use AI years later remains an open question, but the Jordanian data make a strong case that the first barrier to remove is not fear. It is the missing wiring.

Subject of Research: Socioeconomic determinants of artificial intelligence adoption among Jordanian pharmacists and the effect of an educational intervention on their attitudes

Article Title: The digital divide in pharmacy: Socioeconomic determinants of AI adoption and the impact of educational interventions in Jordan

Article References: Al-Tarawneh, A., Abu Dayyih, W., & Abdel-Qader, D. H. (2026). The digital divide in pharmacy: Socioeconomic determinants of AI adoption and the impact of educational interventions in Jordan. Pharmacia, 73, Article e204786. https://doi.org/10.3897/pharmacia.73.e204786

Image Credits: AI Generated

DOI: 10.3897/pharmacia.73.e204786

Keywords: artificial intelligence, pharmacy, Jordan, digital divide, pharmacists, educational intervention, digital infrastructure, health informatics, cross-sectional survey, technology adoption, job displacement fears, continuing education

The research, published in Discover Artificial Intelligence, comes from Bo Xie and Ruoke Zhang, who analyzed their own experience running the Nursing AI Studio, a year-round innovation incubator created to help frontline health professionals turn real-world problems into governable AI tools. The team conducted a retrospective comparative case study of four AI-powered web applications developed through the studio between spring 2025 and spring 2026: a clinical evaluation tool for nursing education, a tailored support tool for dementia caregivers, a patient simulation tool for nursing students, and a statistical assistance tool for large-scale biomedical and social-behavioral datasets. The first three served internal university users and were built with the help of a single graduate research assistant in computer science; the fourth was an NIH-funded project involving researchers at two universities.

The authors defined deployment burden as the organizational, technical, and coordination work required to move a locally developed functional AI-powered web application into legitimate institutional use. They operationalized it across five dimensions: temporal burden measured in business days, pathway-discovery burden reflected in gaps and uncertainty about the next responsible office, dependency burden from steps that could not proceed until prior approvals were complete, enterprise integration burden covering authentication and security review, and collaboration-context burden arising from external collaborators or changing enterprise requirements. Data came from structured deployment logs and dated operational records including email confirmations, approval notifications, and internal deployment notes.

The results reveal a striking pattern. Across the three internal tools, early infrastructure steps were consistently fast and stable: obtaining cloud storage access took one business day, subdomain setup took three, and enterprise AI licenses and Information Security Office review were completed in one to two business days. Enterprise integration, by contrast, was the longest and most variable stage: six business days for the first tool, seventeen for the second, twenty-five for the third, and more than three months for the cross-institutional project. These numbers suggest that early provisioning was not the primary bottleneck. Instead, the pathway slowed most when the application needed deeper enterprise handling around authentication, institutional access, or both.

Perhaps the most revealing finding concerns what the authors call pathway-discovery burden. For the first tool, approximately two weeks separated AI-license approval on October 17, 2025, from initiation of the cloud storage request on October 31, 2025, and another roughly two weeks separated cloud storage setup from initiation of the subdomain request on November 17, 2025. These lapses were not periods of team inaction but time spent trying to discover what the next institutional step was, whom to contact, and whether any prerequisite existed. Later experience suggested that acquiring an enterprise AI license did not actually need to precede cloud storage setup, meaning at least part of the delay reflected an initially misread sequence rather than an unavoidable dependency.

Some burdens proved reusable. Once an enterprise AI license and a workable internal path were established, later internal tools did not require a new AI license request. Cloud storage access and subdomain setup became standardized enough to function as reusable assets. However, repetition did not eliminate all obstacles. The third tool showed that prior experience still did not prevent substantial delays when downstream credential and integration issues arose. The first tool’s integration was also expedited through a deliberate administrative intervention linked to an existing memorandum of understanding with the Chief Information Officer’s office, meaning it should not be treated as a neutral baseline.

The fourth tool introduced a qualitatively different problem. Although it remained a health-related research tool, this NIH-funded project involved collaborators from two universities, so the challenge was no longer only how to move through one institution’s internal pathway. It became a cross-institutional access problem involving AI and computing resources, control of deployment infrastructure, and uncertainty about how external personnel could participate in building a tool that would ultimately live under the lead institution’s domain. The case suggests that a pathway workable for internal personnel can become fragile when collaboration crosses institutional boundaries.

The pathway also proved dynamic. During the period in which the manuscript was being prepared, an additional Information Security Office compliance requirement emerged for this workstream, underscoring that the deployment pathway continues to change even after teams believe they have learned it. Deployment burden is therefore not only a steep first-team learning curve but also an ongoing maintenance burden of keeping up with a moving enterprise and compliance environment. The authors argue that deployment knowledge cannot be treated as a one-time checklist that each team learns and keeps forever.

The practical implication is not that institutions should lower standards or bypass governance. Rather, they should stop forcing each AI builder team to rediscover the same pathway independently. The authors propose adapting a studio-like model across campus for clinically relevant AI projects in the health and social sciences: a small, shared service could help teams scope projects early, identify needed collaborators and expertise, maintain reusable templates and accounts, interpret sequencing, and escalate complex integration issues. They also recommend a staged pathway that pre-provisions what can be reused, allows controlled web deployment and pilot use when appropriate, and escalates to deeper enterprise integrations only when their value justifies the added burden.

For prospective researchers and frontline health professionals, the lesson is that internally developed tools are not equivalent to creating a custom GPT or configuring a commercial chatbot for personal use. Meaningful institutional deployment requires early decisions about collaborators, access, hosting, authentication, review, maintenance, and escalation points. The study has limitations: the evidence base was brief and retrospective, the sample was small, the cases arose in one institutional environment, and the authors were participant-observers. The tools were clinically relevant but not bedside clinical decision-support systems, so the pathway may differ for higher-risk applications. Still, the findings offer a practice-informed map that similar academic health settings can examine in their own infrastructure, making visible a hidden form of implementation waste that is easy to miss if one looks only at formal review times.

Subject of Research: Deployment pathways and organizational burden for internally developed AI tools in clinical education and health research

Article Title: Analyzing and redesigning deployment pathways for internally developed artificial intelligence tools in clinical education and health research

Article References: Xie, B., & Zhang, R. (2026). Analyzing and redesigning deployment pathways for internally developed artificial intelligence tools in clinical education and health research. Discover Artificial Intelligence, 6(1), Article 1240. https://doi.org/10.1007/s44163-026-02220-0

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02220-0

Keywords: artificial intelligence, health informatics, clinical education, enterprise integration, deployment burden, Nursing AI Studio, information security review, cross-institutional collaboration, health AI governance, implementation science, academic health research, e-health

The research, conducted by Cheng-Huan Yu and Mohammad Masum, introduces a framework called WellFusion that converts continuous wearable measurements into hybrid numerical–semantic representations. Instead of forcing a language model to process bare numbers alone or stripped-down category labels alone, WellFusion presents both simultaneously. A day’s activity might appear in a prompt as “Steps = 11,896 (Active)” or “Sleep = 387 (Fair)” — the raw value preserved alongside a human-readable descriptor. The authors contrast this with two alternatives: WellCat, which discards the numbers and keeps only categorical labels such as Sedentary, Active, or Highly Active, and WellNum, which keeps the raw values and drops the descriptors entirely.

The motivation stems from a fundamental mismatch in how foundation models process information. Large language models are trained on text and operate most naturally on language-compatible representations, yet wearable sensors emit continuous numerical observations with heterogeneous scales, units, and distributions. Categorical descriptors compress these measurements into a vocabulary the model can interpret semantically, but compression inevitably discards within-category differences — two people labeled “Active” may have step counts that differ by thousands, potentially reflecting meaningfully different physiological states. Preserving both channels, the authors hypothesized, might give the model the best of both worlds: fine-grained magnitude information plus contextual interpretation.

The study reformulates wellness prediction as a multi-target classification problem. Rather than predicting a single outcome, the model generates one structured output containing ordered labels — Below Normal, Normal, or Above Normal — for four self-reported wellness states simultaneously: stress, fatigue, sleep quality, and readiness. These labels are derived from the statistical distribution of training-participant scores, meaning “Normal” reflects position relative to the pooled training data rather than any clinical reference range. The authors are careful to note this is an operational modeling choice, not a validated clinical boundary.

The experimental foundation is the publicly available PMData dataset, comprising longitudinal records from 16 participants wearing Fitbit Versa 2 smartwatches in free-living conditions. After participant-level alignment and quality control, the final dataset contains 1,836 daily observations pairing passive wearable measurements with self-reported wellness assessments. Crucially, the study restricts model inputs to passively collected signals — steps, calories, sleep duration, and resting heart rate — deliberately excluding active self-reports to better reflect realistic deployment scenarios. Participants were partitioned exclusively into 10 for training, 2 for validation, and 4 held out for testing, with no participant appearing in more than one group.

The headline results come from a five-day evaluation using OpenAI’s GPT-4o-mini with full fine-tuning. Under the WellFusion representation, the model achieved target-level accuracies of 0.86 for fatigue, 0.83 for stress, 0.72 for sleep quality, and 0.69 for readiness, averaging 0.775 across the four targets. The categorical-only WellCat representation yielded consistently lower figures — 0.83, 0.74, 0.66, and 0.60 respectively, averaging 0.708. Every entry summarizes five repeated runs, and the hybrid representation outperformed the categorical-only version on all four targets in this descriptive comparison.

However, the picture becomes considerably more nuanced — and more honest — in the study’s additional three-day analyses. When the authors compared all three representations under a matched Mistral-7B configuration with LoRA adaptation, the numerical-only WellNum representation actually achieved the highest accuracy (0.596) and exact-profile match (0.129), outperforming WellFusion (0.380 accuracy) and WellCat (0.260). Yet WellFusion led on macro-F1, balanced accuracy, and quadratic weighted kappa, the ordinal-agreement metric. The representation ranking, in other words, depends on which metric and configuration you examine — a finding the authors present with unusual candor rather than burying.

Perhaps most striking is the performance of conventional baselines. A simple majority-class dummy classifier achieved the highest raw accuracy (0.677) among all listed three-day configurations, beating every trained model including the fine-tuned language models. XGBoost posted the best macro-F1 and balanced accuracy among conventional baselines. The authors explicitly state these results “do not support a general accuracy advantage for foundation models” — a refreshing admission in a field often eager to declare LLM supremacy. None of the tested three-day accuracy contrasts reached statistical significance after Holm adjustment, partly because with only four held-out participants, the minimum attainable p-value for the exact sign-permutation test was 0.125, making significance at the conventional 0.05 threshold mathematically unattainable.

The study also probes the fragility of its own categorical mappings. In a sensitivity analysis, shifting each numerical cut point upward by 10 percent caused category reassignment for 79.4 percent of resting heart rate values, 30.9 percent of calorie measurements, 28.9 percent of sleep durations, and 19.5 percent of step counts. This instability underscores that the descriptors — labels like “Normal” or “Excellent” — are operational tokens defining the model’s input vocabulary, not validated clinical states. The authors caution, for instance, that labeling sleep above 510 minutes as “Excellent” must not be read as a claim that longer sleep is always healthier.

On the adaptation side, the study compared Low-Rank Adaptation (LoRA), which trains small adapter parameters while freezing pretrained weights, against full-parameter fine-tuning. On the matched Mistral backbone, full fine-tuning produced higher accuracy and macro-F1 for both representations — WellFusion reached 0.560 accuracy with full fine-tuning versus 0.380 with LoRA — though the authors note that differing optimization schedules prevent attributing the gap to adaptation strategy alone. Zero-shot performance, without any fine-tuning, lagged substantially behind, with Mistral achieving only 0.22 to 0.26 accuracy across representations.

The authors are refreshingly explicit about limitations. The cohort is small — 16 participants, one fixed partition, no participant-level cross-validation. Wellness outcomes are self-reported rather than clinically measured. Complete-case filtering may introduce selection bias. The fatigue-directed SMOTE oversampling procedure inherits non-fatigue labels for synthetic instances without validating that assumption. Sleep duration, a predictor, may account for much of sleep-quality performance, but no feature-removal ablation was performed to disentangle this. The same-day design means results do not demonstrate prediction before measurements were available, and no claims are made about real-time deployment readiness, clinical utility, or computational efficiency.

What emerges is less a victory lap than a careful cartography of an emerging research frontier. The study demonstrates that representation design — how numbers become language — is a first-order modeling decision for LLM-based health analytics, not an afterthought. Hybrid representations that preserve both magnitude and meaning showed descriptive advantages in the five-day GPT-4o-mini comparison and on class-balanced metrics in three-day analyses, but numerical-only inputs won on raw accuracy in the matched LoRA setting. The authors frame their contribution as characterizing representation trade-offs within a fixed evaluation design, explicitly declining to claim statistical superiority, physiological mechanism, or clinical usefulness. In a domain where hype routinely outruns evidence, that restraint may be the study’s most valuable finding — a template for how foundation-model health research should communicate uncertainty while still advancing the field. Future work, they suggest, should prioritize participant-level cross-validation, joint-versus-independent prediction comparisons, sleep-duration removal ablations, and matched no-SMOTE training to establish which representation strategies genuinely generalize beyond a single small cohort.

Subject of Research: Using hybrid numerical and categorical representations of wearable sensor data for large language model-based multi-target wellness classification

Article Title: WellFusion: Hybrid numerical–semantic representations for multi-target wellness classification from wearable measurements

Article References: WellFusion: Hybrid numerical–semantic representations for multi-target wellness classification from wearable measurements. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: wearable devices, large language models, wellness prediction, WellFusion, PMData dataset, fine-tuning, LoRA, health informatics, machine learning, self-reported wellness, GPT-4o-mini, representation learning

The research, published in BMC Medical Education, was led by Hesamaddin Kamalzadeh and colleagues, who set out to test whether the classic Technology Acceptance Model, or TAM, still tells the full story in the age of generative AI. TAM, developed decades ago to explain why people adopt new technologies, rests on two pillars: perceived ease of use, meaning how effortless a tool feels to operate, and perceived usefulness, meaning how much a person believes the tool improves their performance. These two perceptions shape a user’s attitude toward the technology, which in turn drives their behavioral intention to keep using it. The Iranian team extended this framework with two constructs they considered essential for the AI era: AI literacy, which measures how well a person understands what artificial intelligence can and cannot do, and perceived security and privacy, which captures how safe users feel sharing information with a chatbot.

To gather their evidence, the researchers distributed a structured online survey to 292 medical students at Hormozgan University between May and June 2026. Eligibility was deliberately restricted to students who had already used a generative AI chatbot for academic purposes, ensuring that responses reflected genuine experience rather than curiosity about an unfamiliar tool. Of the 292 students invited, 267 completed the survey, and after screening, 252 participants with prior academic use of GenAI chatbots formed the final analytical sample. The cohort had a mean age of 22.64 years with a standard deviation of 2.28, and 147 participants, or 58.3 percent, were female. All participants provided electronic informed consent through a mandatory confirmation item before taking part, and the study received ethics approval from the university’s committee under code IR.HUMS.REC.1404.273.

The analytical engine behind the study was partial least squares structural equation modeling, or PLS-SEM, a statistical technique widely used in behavioral research to test networks of hypothesized relationships between latent constructs that cannot be directly observed. Rather than measuring AI literacy or perceived usefulness with a single question, the researchers used multi-item questionnaire scales and then verified that these scales behaved as coherent, reliable measurements. The measurement model passed standard checks for reliability and validity, giving the team confidence that the constructs were being captured accurately before any structural relationships were tested.

The structural results revealed a striking asymmetry in how the two new constructs behaved. AI literacy showed a strong positive association with perceived ease of use, with a path coefficient of 0.479 and a p-value below 0.001, indicating that students who understood AI better found chatbots significantly easier to operate. Perceived security and privacy, by contrast, was positively associated with perceived usefulness, with a smaller but statistically significant coefficient of 0.157 and a p-value of 0.005. In other words, feeling safe with a chatbot did not make it feel easier to use, but it did make students believe the tool was more valuable for their academic work. This dissociation is one of the study’s most technically interesting findings, because it suggests that literacy and trust operate through entirely separate psychological channels rather than merging into a single attitude.

Downstream in the model, the classic TAM sequence held firm. Perceived ease of use and perceived usefulness both fed positively into students’ attitudes toward using generative AI, and attitude in turn showed the strongest single association with behavioral intention, with a path coefficient of 0.524 and a p-value below 0.001. This means that the emotional and evaluative stance a student holds toward chatbots, their attitude toward using GenAI, was a more powerful predictor of their intention to keep using these tools than either ease of use or usefulness alone. The model as a whole explained 38.7 percent of the variance in behavioral intention, with R-squared values ranging from 0.229 to 0.387 across the endogenous constructs, a level of explanatory power that is respectable for behavioral research in complex educational settings.

Why do these numbers matter beyond the statistics classroom? Medical education is a domain where the stakes of AI adoption are unusually high. A chatbot that helps a student draft a study plan is one thing; a chatbot that shapes clinical reasoning is quite another. If acceptance is driven partly by AI literacy, then students with a sophisticated understanding of how large language models work, including their tendency to hallucinate plausible but incorrect information, may be better positioned to use these tools critically rather than credulously. Conversely, students with low literacy might either reject the technology outright or embrace it uncritically, and the study’s finding that literacy boosts perceived ease of use suggests that training programs could lower the psychological barrier to adoption while simultaneously improving the quality of use.

The privacy finding carries its own practical weight. Medical students routinely handle sensitive clinical scenarios, and even in academic exercises, the material they type into a chatbot can touch on patient-like details. The study’s demonstration that perceived security and privacy strengthens perceived usefulness implies that institutional policies are not merely bureaucratic safeguards but active drivers of adoption. If a university clearly communicates how student data is handled, what is logged, and what protections exist, students may come to see the chatbot as more genuinely useful, creating a virtuous cycle in which trust and engagement reinforce each other. The authors conclude that strengthening AI literacy and establishing clear privacy and security policies may support the informed and responsible integration of generative AI chatbots into medical education.

It is worth noting the study’s design boundaries. As a cross-sectional survey conducted at a single institution, it captures associations at one point in time rather than proving causation, and its findings from one Iranian medical school may not generalize universally. The sample was also restricted to students with prior academic use of chatbots, so the model says nothing about what convinces a skeptic to try the technology in the first place. Still, the effect sizes and the coherent structure of the results give the findings real weight, and the extended TAM framework offers a template that other institutions can adapt to their own student populations.

What emerges from this research is a portrait of a generation that has already moved past the novelty phase of generative AI. These are not students dabbling with a trendy tool; they are future physicians whose daily workflows are being quietly reshaped by systems that can converse, summarize, and reason in natural language. The study suggests that the institutions that will integrate these tools most successfully are those that invest in two things simultaneously: teaching students how AI actually works, and guaranteeing, in policy as well as practice, that their data is protected. Attitude, the study shows, is the fulcrum on which adoption turns, and attitude is built from both competence and confidence. As generative AI continues its march through medicine, the psychological blueprint uncovered in Bandar Abbas may prove to be a surprisingly universal map.

Subject of Research: Psychological and technological factors influencing medical students' acceptance of generative AI chatbots in medical education

Article Title: Factors associated with generative AI chatbot acceptance among medical students with prior academic use: a cross-sectional study based on an extended technology acceptance model

Article References: Kamalzadeh, H., Boushehri, E., Hayavi-Haghighi, M. H., Salari, M. S., & Abedini, S. (2026). Factors associated with generative AI chatbot acceptance among medical students with prior academic use: a cross-sectional study based on an extended technology acceptance model. BMC Medical Education. https://doi.org/10.1186/s12909-026-10647-3

Image Credits: AI Generated

DOI: 10.1186/s12909-026-10647-3

Keywords: generative AI, chatbots, medical education, Technology Acceptance Model, AI literacy, data privacy, medical students, PLS-SEM, behavioral intention, health informatics, cross-sectional study, Iran

The research, conducted by Mei Chen and Shibin Wang, tackles a blind spot in how we evaluate medical AI. Most assessments of large language models in health communication focus on hallucinations—the invention of facts that were never in the source text. That focus is understandable, since fabricated claims are dramatic and easy to spot once you know to look. Chen and Wang argue, however, that the opposite error deserves equal attention: the quiet, systematic omission of knowledge during text simplification. A summary that leaves out a drug’s key risk factor or a disease’s underlying mechanism does not announce its own incompleteness. It simply reads as complete.

To measure this hidden loss, the researchers developed a method they call knowledge triplet comparison, or KTC. The approach borrows a structure from knowledge graphs, the databases that represent factual information as connected triples of entities and relations. In the medical domain, a triplet might link a disease to its cause, a chemical to a disease it triggers, or a condition to one of its complications. By extracting such triplets from a source text and from the model’s adapted version, and then comparing the two sets, KTC turns the fuzzy question of whether meaning survived the rewrite into a countable, quantifiable measurement. Knowledge that appears in the original but not the adaptation counts as omitted; knowledge that appears only in the adaptation counts as added; and triplets that are altered between the two versions register as distortions of their own.

The team put this method to work on the Chemical-Disease Relations dataset, a benchmark corpus in which sentences from biomedical literature are annotated with relationships between chemicals and diseases. Six large language models were asked to adapt the material across eight different scenarios, each simulating a real-world use case: simplifying text for readers at a middle school reading level, shortening it, lengthening it, and other transformations of style, length, and audience. For each combination of model and scenario, the researchers measured how much of the original knowledge survived and how the surviving knowledge was distributed across categories of medical content.

The headline finding is stark. Across every adaptation scenario and every model tested, the systems retained only about half of the medical knowledge present in the source text. In other words, a patient who relies on an AI-generated plain-language version of a medical abstract is, on average, receiving roughly fifty percent of the factual content of the original. At the same time, the models inserted new material that was not in the source, adding roughly ten percent more knowledge than the original contained. This added fraction is the familiar territory of hallucination risk, and it is the smaller of the two problems the study documents.

The losses were not random, either. When the researchers categorized the omitted and distorted knowledge, a clear pattern emerged. Knowledge about etiology and pathological mechanisms—the explanations of why and how a disease develops—was among the most frequently lost categories, as was knowledge about complications and risk factors. These are not trivial details. Mechanistic understanding helps patients grasp why a treatment works or why a behavior matters, and risk-factor information often underpins prevention and early detection. A simplified text that systematically sheds exactly this kind of content may remain grammatically fluent and factually accurate in every sentence it keeps, while still leaving the reader dangerously under-informed.

Perhaps the most sobering result concerns the attempted fixes. The researchers tested whether targeted prompt engineering—explicitly instructing the models to preserve key knowledge, to avoid omissions, or to stay faithful to the source—could reduce the problem. It could not. Neither the quantity of knowledge lost nor its categorical distribution changed significantly in response to the corrective prompts. The distortion, in other words, is not a quirk of careless prompting but a persistent and regular property of how these models perform simplification. The authors suggest this regularity is itself valuable: if omissions follow predictable patterns, future research can characterize those patterns precisely and build detection and correction systems around them, rather than treating each summary as an unpredictable black box.

The implications reach well beyond the laboratory. Patient-friendly adaptation of medical literature is one of the most socially valuable applications proposed for language models, promising to close the gap between specialist publications and the general public. Hospitals, health websites, and patient advocacy organizations are already experimenting with AI tools that translate clinical jargon into accessible language. The study’s findings do not mean such tools are useless, but they do mean that fluency and readability are poor proxies for fidelity. A summary can score well on every readability metric while discarding half of what the source actually said, and current evaluation pipelines—built largely around hallucination detection and style assessment—would never flag it.

Chen and Wang frame their contribution as groundwork. By quantifying knowledge distortion systematically and showing that it follows stable, category-level patterns, the study lays the foundation for research into why models lose particular kinds of knowledge during simplification and how that loss might be prevented. The KTC method itself offers a template: rather than asking whether a summary sounds right, developers and regulators can ask, triplet by triplet, what survived the transformation. Until such safeguards mature, the study’s message to anyone using AI to make medicine readable is a cautionary one—the greatest danger of a simplified medical text may not be what it gets wrong, but everything it silently leaves out.

Subject of Research: Knowledge distortion in large language models during patient-friendly adaptation of medical literature

Article Title: Prevalent knowledge omissions: An unseen risk of LLMs in patient-friendly adaptation of medical literature

Article References: Chen, M., & Wang, S. (2026). Prevalent knowledge omissions: An unseen risk of LLMs in patient-friendly adaptation of medical literature. PLOS Digital Health, 5(10), e0001739. https://doi.org/10.1371/journal.pdig.0001739

Image Credits: AI Generated

DOI: 10.1371/journal.pdig.0001739

Keywords: large language models, medical literature, knowledge distortion, patient communication, hallucination, health informatics, text simplification, knowledge triplets, prompt engineering, PLOS Digital Health, AI safety, biomedical NLP

The framing of the paper is deliberately sober about the global picture. In 2015, UN member states adopted 17 Sustainable Development Goals to be achieved by 2030, with SDG 3 pledging to ensure healthy lives and promote well-being for all at all ages through 12 targets. Across all 17 goals, 33 targets and 57 indicators are classified as health-related. Yet the United Nations’ own 2023 assessment found that less than 10 percent of SDG 3 targets were on track at the midpoint. The global maternal mortality ratio remains more than three times higher than the 2030 target of 70 deaths per 100,000 live births, and nearly three-quarters of countries requiring accelerated progress on under-five mortality are in Sub-Saharan Africa. A systematic review of SDG implementation cited in the study found a striking absence of guidance on how to engage stakeholders, while pointing to academic institutions as underused engines of knowledge generation and policy guidance. AMPATH, the authors argue, is precisely the kind of stakeholder engagement model the global community has been missing.

The partnership’s origin story is rooted in the HIV epidemic. Beginning in 2001, AMPATH launched efforts to expand access to antiretroviral drugs at Moi Teaching and Referral Hospital in Eldoret and in surrounding communities. Backed by philanthropy, the United States Agency for International Development, and the Centers for Disease Control and Prevention, the initiative grew into one of Africa’s largest and most comprehensive HIV care systems, with a network of clinics that has cared for more than 250,000 patients living with HIV or AIDS. What distinguishes AMPATH from many donor-driven programs, the authors contend, is its tripartite academic mission of education, research, and care. Lessons learned from building an HIV treatment infrastructure were subsequently applied to non-communicable diseases, including mental and cardiovascular health, maternal and child health, and cancer, aligning the partnership’s portfolio with SDG target 3.4.

Some of the most technically consequential work has come through the Global Network for Women’s and Children’s Health Research, a collaboration through which AMPATH investigators have contributed to a series of landmark trials targeting SDG goals 3.1 and 3.2 on maternal and child mortality. The Azithromycin Prevention in Labor Use Study, A-PLUS, was a multi-country, double-blind, placebo-controlled randomized trial in six low- and middle-income countries including Kenya. It enrolled 29,278 pregnant women planning vaginal births and randomized them to receive either a single 2-gram oral dose of azithromycin or placebo during labor. The result: intrapartum prophylactic azithromycin reduced observed maternal infection or death through six weeks postpartum by approximately 35 percent. For a single, cheap antibiotic dose administered at delivery, that effect size has few parallels in maternal health.

Equally striking is the AFRINEST trial, which tested whether community health workers could safely manage seriously ill newborns at home in three Sub-Saharan African countries when hospital referral was not feasible. The trial demonstrated that three simplified outpatient regimens of affordable antibiotics, given by injection and by mouth, were as effective as the gold-standard reference treatment in preventing treatment failure among infants with probable sepsis. Follow-up studies confirmed the model’s feasibility elsewhere in the region. The findings fed directly into the WHO guideline on managing possible serious bacterial infection in young infants when referral is not feasible, and have since been incorporated into national maternal and child health programs in the Democratic Republic of Congo, Kenya, and Nigeria. It is a textbook example of the pathway the case study champions: local evidence, generated in the settings where the burden is highest, translated into global guidelines and national policy.

The ASPIRIN trial pushed a similarly elegant intervention into the global mainstream. Across seven sites in six low- and middle-income countries, 11,976 nulliparous women were randomized to daily low-dose aspirin or placebo in early pregnancy. The aspirin group showed significant reductions in preterm delivery and fetal loss. Subsequent analyses established that the regimen is effective even in women without additional risk factors, is safe, is cost-effective, and delays the onset of hypertensive disorders of pregnancy. The data have been used to advocate for low-dose aspirin therapy as a standard of care in low-resource settings and are now reflected in WHO guidance. Meanwhile, the Antenatal Corticosteroids Trial, a cluster-randomized study across Sub-Saharan Africa, Asia, and Latin America, delivered a more nuanced lesson: antenatal corticosteroids for women at high risk of preterm birth reduce neonatal mortality only when delivered within tertiary care facilities where preterm infants and mothers in premature labor can receive specialized care, a caution that reshaped implementation strategies worldwide.

On the neonatal front, AMPATH investigators collaborated with the American Academy of Pediatrics, WHO, and other partners on Helping Babies Breathe, a training program launched in 2010 to teach basic neonatal resuscitation to health workers in low- and middle-income countries. The collaboration extended to curriculum design, evaluation, and the development of digital education and quality improvement tools, including randomized trials of virtual reality and video-based refresher training. The scale of dissemination is remarkable: the program has been rolled out in more than 80 countries, over one million health care providers have been trained, and meta-analytic evidence links the training to reduced stillbirth and perinatal mortality. Because birth asphyxia remains a leading cause of neonatal death, the public health significance of embedding resuscitation skills in frontline providers is difficult to overstate.

HIV work followed a parallel but distinct pathway, one built on data. Through the International Epidemiology Databases to Evaluate AIDS, or IeDEA, a consortium established in 2006 with seven regional data centers, AMPATH investigators lead the East Africa region. The network pools and analyzes clinical data from 1.7 million patients receiving HIV care in 46 countries, using common data exchange standards. In 2014, IeDEA formalized a partnership with the WHO to accelerate the translation of research into policy, a mechanism that bypasses the often lengthy lag of journal publication. Observational data from the network provided evidence that informed the WHO’s pivotal recommendation to offer HIV treatment to all people living with the virus, and has been critical to monitoring progress toward the 90-90-90 goals: 90 percent of people with HIV knowing their status, 90 percent of those diagnosed on treatment, and 90 percent of those treated virally suppressed. Since the partnership began, IeDEA has completed 12 multiregional analyses and supported three global HIV progress reports. AMPATH researchers also supplied WHO working groups with pediatric data showing that self- and caregiver-reported adherence measures can underestimate non-adherence compared with automated monitoring systems and plasma drug levels, findings that shaped WHO recommendations for pediatric HIV care.

The fourth pillar of the case study is perhaps its most quietly revolutionary: informatics. As AMPATH’s HIV clinics multiplied across western Kenya, the partnership built a custom electronic health record with the Regenstrief Institute, the AMPATH Medical Record System, which became essential to achieving viral suppression and prevention targets and was later extended to chronic disease clinics. With funding and support from the WHO, the Rockefeller Foundation, and Partners in Health, that system was transformed into OpenMRS, an open-source electronic health record platform designed for low-resource settings. Today OpenMRS operates at more than 5,500 sites in over 60 countries, including facilities in the United States, Canada, and Europe, supporting care for infectious diseases and non-communicable conditions alike. Reliable electronic records, the authors argue, are not merely administrative conveniences; they are the measurement infrastructure without which progress on the health-related SDGs cannot be tracked, and without which countries lagging behind cannot be identified in time to act.

Taken together, the AMPATH case study makes a claim with implications far beyond western Kenya: that academic institutions, when woven into long-term partnerships with health ministries and multilateral agencies, can fill critical data gaps, test innovations under real-world constraints, and generate the evidence base on which global guidelines depend. Whether the intervention is a two-gram dose of azithromycin during labor, a simplified antibiotic regimen delivered by a village health worker, or a shared database spanning four dozen countries, the mechanism is the same: local evidence, gathered rigorously and shared openly, translated into standards that change practice everywhere. With the 2030 deadline looming and most health targets off track, the authors suggest that replicating this model, rather than inventing new frameworks from scratch, may be the most credible route left for turning the SDG promise into reality.

Subject of Research: Academic global health partnerships advancing the UN health-related Sustainable Development Goals through the AMPATH collaboration in Kenya

Article Title: Progress Toward the Health-Related Sustainable Development Goals: A Case Study of the Academic Model Providing Access to Healthcare (AMPATH)

Article References: Kelly, C. M., Bucher, S. L., Esamai, F., Wools-Kaloustian, K., Vreeman, R., Braitstein, P. K. A., Gardner, A., Carter, E. J., Kimaiyo, S., & Tierney, W. M. (2026). Progress Toward the Health-Related Sustainable Development Goals: A Case Study of the Academic Model Providing Access to Healthcare (AMPATH). Journal of General Internal Medicine. https://doi.org/10.1007/s11606-026-10732-9

Image Credits: AI Generated

DOI: 10.1007/s11606-026-10732-9

Keywords: AMPATH, Sustainable Development Goals, global health, HIV care, maternal mortality, neonatal health, WHO, OpenMRS, electronic health records, Kenya, health informatics, clinical trials