Frontiers in Digital Health

Background and Objective: Lung function reference equations are essential for interpreting spirometry and related modalities, yet existing tools are fragmented, modality- and language-specific (predominantly R), and lack integrated severity classifiers. We present PySpiro, an open-source Python library unifying 23 spirometry, 1 multiple breath washout, 2 diffusing capacity, 1 static lung volume, and 2 oscillometry reference equations under a single API, with 10 clinical severity classifiers. Methods: PySpiro was implemented in Python 3 using NumPy and pandas, with all equations coded directly from literature. The unified API exposes predicted medians, z-scores, limits of normal (LLN/ULN), and raw Lambda-Mu-Sigma (LMS) triplets per parameter, plus a vectorised compute() method for batch processing at the DataFrame level. Unit-tests covering every equation and classifier verify agreement with published values and confirm correct boundary, out-of-range, and missing-input handling. Results: PySpiro covers spirometry equations from 1971 to 2022 across Caucasian, Asian, Indian, and race-neutral global populations, plus non-spirometry modalities (multiple breath washout, diffusing capacity, static lung volumes, oscillometry). Classifiers span COPD, interstitial lung disease, pattern and severity grading, bronchodilator reversibility, and primary ciliary dyskinesia. Compared with rspiro (5 equations, 4 spirometry) and the only prior Python package (spiref, 2 equations, no z-scores or % predicted), PySpiro offers substantially broader coverage, multi-modality support, a richer output API, and 10 classifiers absent from both. Conclusions: PySpiro is freely available (PyPI: pyspiro; doi:10.5281/zenodo.15519193) and provides a comprehensive, validated, and extensible platform for lung function data analysis in research pipelines.

Background: Accurate patient understanding of lung cancer screening (LCS) results is critical for engagement and follow-up adherence, given the currently low screening uptake. While artificial intelligence (AI) systems show promise in facilitating patient communication, its effective use in clinical settings requires appropriate invocation of imaging tools and self-regulation by deferring certain questions to physicians. To this end, we developed and evaluated a planner-executor style multimodal agentic system to answer simulated patient questions about LCS CT results. Methods: This retrospective study utilized 116 LCS CT reports and images collected from a tertiary academic hospital. The system employed a multi-agent architecture (planner and executor) to mimic clinical reasoning. We developed a large language model (LLM)-based question generation framework to prepare a comprehensive set of 699 simulated patient questions, categorized as answerable and defer-to-doctor to assess self-regulation. Performance was evaluated on tool-calling, self-regulation accuracy (ability to correctly defer out-of-scope questions to a human provider) assessed by LLM-as-a-judge, and clinical quality assessed by a reader performance study with four physicians evaluating a subset of 100 responses. Results: The system demonstrated high overall tool-calling accuracy of 93.1% (646/694; 95% CI: 91.2%, 95.0%) and self-regulation accuracy of 92.6% (462/499; 90.2%, 94.8%). For answerable questions, the percentage of responses receiving perfect 5.0 scores across all readers included 78.9% (180/228) for clinical accuracy (inter-rater agreement: Gwet's AC2=0.92), and 73.2% (167/228) for patient understandability (0.86). For defer-to-doctor questions, the percentage of responses included 77.3% (133/172) for clinical accuracy (0.80) and 70.3% (121/172) for patient understandability (0.78). Conclusions: The planner-executor style multi-modal agentic system reliably answers patient-specific questions about LCS. Its high tool-calling and self-regulation performance demonstrates its potential as a safe and effective digital communication facilitator in LCS programs.

As digital self-triage becomes more widely used, quality assurance requires evaluation of both emergency recognition and overall agreement with clinical judgement, including how these outcomes differ across generations of large language models. The Benchmark Self-Triage (BeST) project intends to provide a framework to compare systems for digital self-triage against a medical 'gold standard' in Germany. In a first step developing the method and demonstrating its potential we developed a representative set of patient vignettes and compared two regulated medical-device triage systems, SmED and Infermedica, with five structured OpenAI configurations spanning multiple model generations: gpt-4.1, gpt-5.4, gpt-5.6-luna, gpt-5.6-sol, and gpt-6-astra. An LLM-based patient actor covered 480 physician-curated vignettes derived from German routine triage data. A pool of seven physicians provided preliminary reference classifications and consensus for 480 vignettes. Recommendations were mapped to three urgency categories 'emergency', 'acute', and 'not urgent' and evaluated using winsorised weights derived from routine triage frequencies. Weighted emergency sensitivity ranged from 51.3% (95% CI 35.5-66.3%) for gpt-4.1 to 99.9% (99.7-100.0%) for gpt-6-astra. Weighted exact three-category agreement (accuracy) ranged from 54.4% to 68.8%; gpt-4.1 achieved the highest agreement despite the lowest emergency sensitivity. No system performed best across all urgency categories. The later-generation OpenAI configurations showed a more precautionary triage profile resembling SmED, with higher sensitivity for emergency cases; gpt-6-astra showed the strongest shift. This cross-sectional comparison across model generations suggests a shift towards more precautionary triage behaviour, without establishing a longitudinal improvement in accuracy. These findings highlight the limitations of overall agreement for evaluating self-triage and support category-specific assessment. Broader expert consensus and evaluation, as well as further steps to reduce dependence on SmED-derived source data, are needed and planned before drawing conclusions about clinical safety.

At-home sampling can address obstacles to participation in transcriptomics research. We previously developed the homeRNA and homeRNAmax platforms; these custom-designed tubes containing RNA stabilizer interface with a blood tube compatible with upper-arm blood collection devices, allowing for blood collection and RNA stabilization to occur from a participant's home. In this work we introduce homeRNAdrop, an insert for a commercially available blood tube (BD Microtainer) that divides the tube into two compartments: the top compartment holds 200 L blood and the bottom compartment holds ~590 L liquid stabilizer (RNAlater). By combining blood collection and RNA stabilization into the same tube, homeRNAdrop streamlines sample preparation for participants and allows for more compact sample storage for the lab. In a remote pilot study, n=25 participants from across the United States tested the homeRNAdrop kit. The n=23 samples successfully received by the lab show that samples collected using homeRNAdrop are of sufficient RNA quality (all RINs > 6, a common cutoff for analysis) and yield (mean yield=1.44 g) for downstream gene expression analysis. Almost all participants who successfully collected blood using the Tasso Mini device found the homeRNAdrop device easy to use (n=23/24). Overall, this work adds a simplified, user-friendly platform to the growing suite of tools for remote blood RNA stabilization.

The explosive growth of ambient AI scribes in healthcare has focused almost exclusively on relieving clinician documentation burdens and enhancing clinician efficiency and reducing burnout. While most research has focused on the clinician experience, these same types of tools, such as a patient held AI Personal Health Assistant, can be put in the hands of patients to facilitate communication with healthcare providers, to provide greater autonomy and activation for patients in managing their own healthcare, and to facilitate the task of receiving and retaining complex new information. However, to ensure these tools serve as more than just administrative aids, we partnered with a large technology company to conceptually codesign how such a tool might be studied in real world settings in a way that reflected outcomes of greatest importance to patients and families. We used the Nominal Group Technique and engaged thirty three patients and caregivers and ten multidisciplinary clinicians in structured virtual sessions to identify the outcomes on which subsequent studies should hinge. Our findings reveal a significant misalignment between current industry trends and patient priorities. While the market emphasizes the scribe function for billing, stakeholders overwhelmingly prioritized the AIPHAs role in longitudinal care management and post visit coordination. The most critical outcome identified was the impact on communication and respect, suggesting that patients view AI as a mechanism to improve the human to human dynamic and ensure a shared understanding of the care plan. Our findings further highlighted that the primary hurdle for the use of patient held AI is not technical utility but clinician trust and the need for formal consent processes. Importantly, as this technology is new and integration into the clinical experience is still being explored, patients cannot assume they will have access to the transcript of their own encounter when that record is created and held by the health system. This gap highlights the potential value of a patient held tool, preserving discussions from their visit and follow up steps in clearer detail than they may receive in their clinical note or traditional visit summary. From a system perspective, this transition requires a strategic shift from measuring workflow focused efficiency to measuring patient side advocacy and engagement. By centering the evaluation of AI on these patient defined metrics and fundamentally designing subsequent AI investigations using a partnership model, health systems can move beyond simple automation toward a model of care delivery that empowers patients to navigate complex clinical journeys with greater autonomy and clarity.

Patient narratives can reveal aspects of health that standardized measures miss, yet collecting open-ended responses at population scale remains difficult. We evaluated whether speech-to-text (STT) could enable richer responses in online health surveys without the resource demands of interviews. Across four surveys involving people with post-COVID-19 condition, healthy adults, people who engage in sex work, and older adults, participants could answer open-ended questions by typing or using STT. We compared response length, lexical diversity, stopword proportion, and total unique content words, alongside STT uptake and participant feedback. Quantitative comparisons were made using a mixed-effects model including modality (STT vs typed) as fixed effect and a random intercept per participant. STT uptake varied markedly across populations (0.7-80.0%). Spoken responses were estimated to contain more characters than typed responses in two studies, by 245.0 and 891.8 characters on average (both p<0.001), and contained more unique content words, despite lower lexical diversity and higher stopword proportions. Participants generally valued STT for its convenience and spontaneity, but uptake was constrained by context, privacy, interface design and individual preferences. STT may enable population-scale collection of richer health narratives, provided implementation is tailored to the study population and research context.

INTRODUCTION: The digital clock drawing test (dCDT) is a cognitive screening tool capturing high-resolution pen-movement data. Existing approaches for classifying cognitive impairment from dCDTs rely on selected features or image-based methods. We propose representing dCDTs as structural networks to classify cognitive impairment and characterize spatial organization in clock drawings. METHODS: Weighted networks were constructed from 3,362 Framingham Heart Study participants, with and without cognitive impairment. Network representations were generated separately for command and copy conditions, with nodes representing pre-defined clock components (digits, hands, clock face) and edges capturing pairwise spatial relationships between components. Random forest models and graph convolutional networks were trained to predict cognitive impairment from the networks. Models were benchmarked against demographics-only and established summary feature baselines. RESULTS: Network-based models achieved AUC values comparable to the established summary feature baseline across command-only and copy-only conditions. Differences in spatial organization between the cognitively intact and impaired groups were greatest in the upper-right clock region, involving digits 1 and 2, the clock hands, and the surrounding clock face. Differences between the command and copy networks were larger among cognitively impaired participants. DISCUSSION: The proposed network representations provide an objective, scalable, and flexible framework for representing clock drawings for use in classification models. dCDT networks matched the predictive performance of established summary feature-based approaches and identified clock regions where spatial organization differed by cognitive status. After external validation, this approach has the potential to be incorporated within a clinical workflow to support automated and interpretable cognitive screening.

Depression in Alzheimer's disease (AD) is clinically challenging to characterise because affective and cognitive manifestations frequently overlap with the underlying neurodegenerative condition. This study investigates depression-related speech patterns in AD using an explainable multimodal framework that integrates linguistic and acoustic features from elicited picture-description speech. To address the absence of directly observed depression labels in the target AD cohort, we employ a source-informed cross-dataset transfer strategy. Unsupervised clustering identified two distinct speech sub-phenotypes within the AD cohort. By comparing these unsupervised representations to a depression-labelled source corpus, one cluster demonstrated mathematical proximity (measured via cosine similarity in a shared latent space) to the source depression centroid. We explicitly acknowledge that this proximity reflects geometric similarity in a shared latent space, which may encapsulate overlapping domain characteristics alongside affective variance. To validate the structural integrity and separability of these discovered clusters, a proxy supervised classification task was employed. The retained multimodal artificial neural network (ANN) integrated information from both modalities and achieved 92.73\% cluster separability accuracy. This demonstrates that the representation-derived two-cluster partition was highly learnable on the independent internal hold-out test set; it does not establish clinical validity, longitudinal stability, or diagnostic validity of the identified sub-phenotypes. For interpretability, Shapley Additive exPlanations (SHAP) analysis was utilised to extract feature attributions and organise influential acoustic and linguistic features into a multimodal taxonomy. Furthermore, a Multimodal SHAP Attribution Magnitude (MSAM) score was derived as an exploratory model-based attribution score rather than a clinically validated severity measure. Ultimately, this study provides an interpretable computational baseline for characterising depression-related speech variance in AD, establishing a structural foundation for future prospective validation using independently labelled clinical cohorts.

Awareness of comorbidities in people with dementia is vital context for dementia research and clinical management. Natural Language Processing (NLP) applied to free text in the electronic health record (EHR) may help ascertain comorbidity presence. Our aim was to assess the validity of the MedCAT Comorbidity NLP app in a cohort of people with dementia. We used text from the EHR of South London and Maudsley NHS Mental Health Trust via Clinical Records Interactive Search (CRIS), linked to structured data from local primary care (Lambeth DataNet) and hospital admission data in a cohort of 3,543 people with dementia. We evaluated MedCAT Comorbidity, examining internal validity via manual review of text and external validity by cross-validation against linked primary care and hospital data; and comparison of the predictive value of comorbidity measures for mortality. Internal precision was highest for atrial fibrillation (100%) and lowest for diabetes (78%). Cross-validation with linked data varied substantially across conditions, with precision from 96% to 59% and recall 77% to 4%. The median number of comorbidities identified was two using any single source, or three when combined. MedCAT Comorbidity demonstrated predictive ability for mortality (HR 1.11 per condition), although the model fit was better using comorbidity from hospital admission data. In people with dementia, the MedCAT Comorbidity app showed high precision for most conditions during internal validation. The variability of precision and recall on external validity assessment partly reflects incomplete documentation. Results from the MedCAT comorbidity app can assist researchers in understanding comorbidity of cohorts, particularly when primary care or hospital data is not available. Our findings highlight the need to develop quality assurance protocols that include internal and external validity checks before deploying general data extraction apps for specific clinical subpopulations.

Social media language and digital lifestyle behaviours are promoted as scalable signals of population mental health, but predictive models are often evaluated in ways that overstate accuracy. We analysed 20,000 tweets from 72 users, labelled by whether the account belonged to a user identified as depressed, and two synthetic datasets (100,000 and 30,000 records) linking screen time, sleep and social media use with self-reported stress. Under a conventional tweet-level split, a term-weighted logistic regression appeared accurate (area under the receiver operating characteristic curve 0.835), but when whole users were held out its performance fell to near chance (0.595; 95% confidence interval 0.505 to 0.686), because the model largely recognised users and their topics. Depression-associated tweets were more often negative (31.7% versus 24.2%), yet sentiment alone could not identify unseen users (0.539). The two behavioural datasets gave opposite answers: strong discrimination in one (0.926), driven by short sleep, and chance in the other (0.506), so synthetic results reflect how the data were generated. At a population prevalence of 10%, positive predictive values would be 0.10 to 0.30. Digital distress signals should be validated across individuals, on real linked data and at realistic prevalence before informing public health surveillance or prevention.

Objective: Retrospective evaluation of clinical pathway adherence using electronic medical record (eMR) data is limited by heterogeneous data structures and statistical measures that do not account for patient-specific context, care sequence, or timing. We designed a modular event-based architecture to dynamically evaluate the New South Wales Adult Sepsis Pathway over time. Method: Data from eMRs were collected from adult emergency department encounters across four hospitals between July 2022 and June 2024 and processed through three sequential layers: event-stream transformation, pathway-relevance filtering, and state-based pathway evaluation. Heterogeneous records were standardized and organized into a chronological timeline for each encounter. The framework then updated the pathway state as new clinical information became available and assessed whether recommended care was delivered within the relevant timeframe and sequence. Result: The architecture processed 451,124 encounters. This included 57,631,084 clinical events extracted and transformed into encounter-level event streams. After filtering for pathway relevance, 43,951,146 events remained (76.3%). A total of 129,907 (34.6%) encounters entered the sepsis pathway, 50,118 (38.6%) entered YELLOW state, indicating progression to suspected sepsis, and 3,349 (2.6%) encounters progressed to RED state from GREEN, indicating the initial shock. Of 50,118 encounters entering YELLOW state, 2,426 (4.8%) progressed to shock, 8,700 (17.4%) encounters were recognized with sepsis through a blood culture order, and 4,287 (8.6%) encounters were recognized with sepsis through an antibiotic order. This demonstrates the ability of the architecture to prioritize higher-acuity states during evaluation rather than going through an intermediate state. Conclusion: The architecture developed shows the feasibility of transforming fragmented routine eMR data into a reusable longitudinal representation that supports context-dependent pathway evaluation. Separating data preparation from pathway rules improves transparency, maintainability, and adaptability and provides a foundation for future interoperable, near-real-time pathway monitoring

Background: Kano State has one of the highest concentrations of zero-dose (ZD) children in Nigeria, challenging equitable immunization coverage. Digital engagement tools meant to strengthen timely data to address such gaps often underperform due to limited adaptability to user needs. Learning-by-doing (LxD), which brings together human-centered design and adaptive management, offers more context-responsive solutions. Intervention and Implementation Approach: MCH-Track, a digital platform for tracking and referral for childhood vaccinations was used by 56 female community mobilizers in two local government areas (LGAs) out of the 44 LGAs in Kano State, Nigeria. We conducted a program case study from: 1) August 2024-April 2025 (initial implementation); and 2) adapted implementation (May 2025-November 2025), where LxD was incorporated to address emergent challenges. LxD included structured shadowing, monthly interviews with users, after-action reviews, and change log to track adaptations alongside routine services. Performance was monitored using four indicators: feedback actioned, response time-to-user feedback, ease-of-use score, and mobilizer performance in tracking ZD children. Findings: After incorporating LxD, 17 issues were identified and led to 12 prioritized adaptations. All logged issues received responses, with median response time declining to approximately two weeks, while community mobilizers reported higher usability of the digital platform (e.g., mean ease-of-use score of 4.1 out of 5). Enhanced root-cause analysis enabled trend tracking and clearer escalation pathways for systemic issues like micro-stockouts at immunization points. Empathy tool records revealed mobilizer struggles with platform. This was immediately logged, prompting user experience redesign into a seamless, single-step process from registration-to-referral. During the LxD implementation period, an average of 33 ZD children per month were identified via MCH-Track, a substantial increase from the pre-LxD period (an average of 11 ZD children identified per month). Average referral-to-access rates also improved during the LxD period, rising from 44% of identified and referred ZD children accessing immunization services pre-LxD to 70% after incorporating LxD practices. Conclusion: Embedding LxD within MCH-Track strengthened co-creation, streamlined workflows, and improved community mobilizer performance, demonstrating the value of adaptive management in reaching ZD children.

Background: Patient Decision aids (PtDAs) are important tools that support shared decision-making (SDM) between clinicians and patients, enabling patients to be more informed and engaged in decisions regarding their care. The use of PtDAs can increase patient knowledge, reduce decisional regret, and engage the clinician and patient in meaningful dialogue. Shared decision-making relies on decision aids, but how patients receive the PtDA can vary. This qualitative study elucidated patient familiarity with PtDAs and the barriers, needs, and requirements of patients to access and confidently use DAs, including when received via a clinician-facing EHR-integrated Decision Aid Navigator System (DEANS). Methods: We conducted six focus groups (FG) with participants who were recruited from two academic medical centers. Focus groups included an introduction to what a PtDA is and an exploration of how patients or family members, i.e., people who had not made a medical decision for themselves using a PtDA, but who have helped a family member, had previously used PtDAs in their healthcare. The FGs focused on PtDA-users, affordances and barriers to using PtDAs, and PtDA-user needs and requirements of a proposed DEAN patient portal. We applied Rapid Qualitative Analysis (RQA) to summarize the data, using a consensus making approach to identify emerging themes. Results: We identified three overarching themes: 1) Patients had varied familiarity with PtDAs and often confused them with general health tools, but found value in SDM, 2) Patients perceived technology accessibility, PtDA credibility, and time as barriers potentially impeding PtDA use, and 3) Patients valued flexible, user-friendly, and accessible PtDAs that support patient-led decisions, particularly when introduced by clinicians and delivered via a patient portal. Conclusion: This study provides patient-derived insights into factors that influence PtDA uptake, highlighting common misconceptions about PtDAs and identifying key patient priorities such as accessibility, flexibility, clear content, and clinician-guided usage. Patient also expected clinicians to be familiar with PtDAs, especially when the PtDA covers a complex decision. These findings can inform the development of more patient-centered SDM infrastructure and strategies to support PtDA use in clinical practice, including clinician-supported delivery, accessibility, and integration into routine care to strengthen shared decision-making.

Abstract Background: Large language models (LLMs) are increasingly explored as adjunct tools in graduate medical education, but their reliability for answering discipline-specific certification examinations remains incompletely characterized. In dentistry in China, the national dental examination serves as a milestone for stomatology trainees. Prior studies predominantly rely on single metric, limiting reproducibility and granularity. Objective: To systematically evaluate the performance of five LLMs in answering the national standardized training examination for dental residents. Method: A total of 1,318 questions covering the six dentistry disciplines were selected. Five LLMs (DeepSeek-V3, DeepSeek-R1, Qwen3-32B, Claude 3.5, and GPT-4o) were evaluated via a single-response protocol on the Dify-based Chatflow framework. Performance combined human verification with automated scoring using the Retrieval-Augmented Generation Systems (RAGAS) framework across seven dimensions. Tasks involving question classification, key-point categorization, and instructional requirement summarization were designed to enable a horizontal comparison of models' cognitive capabilities, assessed using inter-model consistency scores. Results: DeepSeek-R1 achieved the highest correct choice rate (80.90%, p < 0.001) and answer correctness score (0.853, p < 0.001), significantly outperforming the others. All models exhibited limitations, including information insufficiency, redundancy, and hallucination-related issues (lowest 44.20% in DeepSeek-R1). In classification and summarization tasks, Claude 3.5 showed higher consistency scores (186/181) than Qwen3-32B (152/151). Conclusion: Evaluated LLMs demonstrated good performance on dental examination questions and in executing tasks. DeepSeek-R1 consistently outperformed the others. While LLMs showed potential as supportive tools in dental education, further refinement and validation are needed before formal adoption in graduate medical educational and clinical training settings.

Digital photography remains the gold standard for dermatological assessment, but reliance on standardized stationary image-capture systems limits trial scalability, whereas smartphones combined with computer vision artificial intelligence (AI) could enable hybrid and decentralized clinical trials (DCTs). This study assessed the facial feature grading of an AI-based system against expert panel consensus across three settings with decreasing acquisition control: Standard Studio Imaging (SSI), Clinic Smartphone Imaging (CSI), and Home Smartphone Imaging (HSI). Expert inter-rater reliability was good to excellent (median ICC 0.84-0.90), and the system showed high operator-independent test-retest repeatability across independently captured images (median ICC 0.97-0.98). Agreement with experts was maintained across all settings for macro-geometric features (ptosis {rho} = 0.74-0.81, nasolabial fold {rho} = 0.68-0.79) and for discrete pigmentation and inflammatory lesions (pigmented spot density {rho} = 0.62-0.76), declined for fine-texture features under uncontrolled acquisition (forehead wrinkles {rho} = 0.76 to 0.48), and remained limited for color-dependent features irrespective of setting. These findings support the AI-based system as an objective grading tool for a defined subset of facial endpoints in hybrid and DCTs in dermatology.

People from UK South Asian communities experience barriers to accessing dementia care. Evidence suggests that evidence-based psychosocial dementia interventions can be successfully adapted for different cultural contexts, but further evidence is needed on how adaptations are developed and refined in practice. We aimed to adapt GRACE-family (Goal diRected, Accessible and Evidence-based Care), an evidence-based, manualised post-diagnostic dementia intervention, comprising 4-8 sessions for family carers and where possible, people with dementia, for Bengali-speaking communities. In a theory-informed process, a co-adaptation group encompassing lived, academic and clinical/professional experience developed consensus regarding which intervention functions were core. They then explored how adaptation might, while retaining these, account for cultural and linguistic contexts, informed by focus groups with memory service and Bengali community organisation staff (n=14). We developed a prototype adaptation, then piloted it with four Bangladeshi families supporting a person with dementia. Post-intervention, participants were asked about goal attainment, intervention acceptability, and qualitatively interviewed. Pilot data and a focus group with Sylheti-speaking family carers informed further refinements (n=6). Across data sources, two themes informed adaptation: (1) stigma, trust and community networks shaped access to dementia support; (2) flexible delivery was needed to respond to diverse language, communication and cultural needs. 3/4 pilot families completed the prototype intervention, which was in English as initial consultation cautioned against developing a written manual in a predominantly spoken language. Participants reported improvement on 7/8 personalised goals set, and all considered the intervention acceptable. Pilot delivery identified challenges using English-language materials during interpreter-mediated sessions. Iterative refinements added standard Bengali materials and spoken Sylheti audiovisual resources to broaden delivery options and facilitator training for culturally responsive and interpreter-mediated delivery. Our co-adaptation approach enabled stakeholders to advise from different, including experiential perspectives. It may inform future approaches to cultural and linguistically adaptations of psychological interventions for minoritised communities.

Large language models (LLMs) are increasingly consulted for health advice, but they can produce fluent statements that are false or unsupported, known as hallucinations, and this risk is poorly characterised in low-resource languages such as Bangla. We measured hallucinations of locally deployable open-weight LLMs in Bangla diabetes question answering and tested whether retrieval-augmented generation (RAG) grounded in Bangladeshs National Guideline on Diabetes Mellitus reduces them. Forty patient questions spanning six clinical sub-domains were each answered twice, under zero-shot prompting, by Qwen3.5-9B, TigerLLM-9B, Qwen3.6-35B-A3B and Gemma-4-26B-A4B, and by BanglaDiaRAG, which adds hybrid dense (BGE-M3) and sparse (BM25) retrieval from a physician-reviewed Bangla guideline knowledge base to Qwen3.5-9B. Two physicians rated the 400 de-identified responses by consensus for medical accuracy, completeness, safety, readability and reliability on a 0-1 scale, and responses with accuracy below 0.70 were classified as hallucinated. Hallucination rates were 2.5% for Gemma-4-26B-A4B, 3.8% for TigerLLM-9B, 22.5% for Qwen3.6-35B-A3B and 77.5% for Qwen3.5-9B. Guideline grounding lowered the rate of Qwen3.5-9B to 11.3% (difference -66.3 percentage points; 95% CI -76.3 to -55.0) and raised every criterion by 0.24 to 0.38 points (Holm-adjusted p < 0.001). BanglaDiaRAG outscored Qwen3.6-35B-A3B on all five criteria and matched or exceeded TigerLLM-9B, but it remained below Gemma-4-26B-A4B in safety, readability and reliability, and its residual errors were specific, confident and potentially harmful. Guideline grounding can turn a weak general-purpose model into a much safer Bangla diabetes assistant, but the choice of base model mattered as much as retrieval and no configuration was free of hallucination, so clinical oversight remains essential before patient-facing use.

Background: Estimating nutrient intake from meal photographs with vision-language models (VLMs) is now practical, but patient photographs often cannot leave the institution. We developed PLANS (Personal-device Local Assessment of Nutrients System), in which a small open VLM runs entirely on a personal device. Methods: Open VLMs were compared on two image sets with reference values (NutriImage, 887 Japanese meals; SNAPMe, 1,134 US meals), and the selected model (Qwen3-VL-4B, 4-bit) was run on an iPad Air without network access. Energy (primary outcome), protein, fat, carbohydrate and salt were evaluated by median absolute percentage error (MedAPE) and Spearman rank correlation (rho). The effect of supplying the food weight, including errors in it, was examined. Results: Energy estimates on the iPad were identical to those of a desktop-computer run for 84-86% of images (other values, 75-87%). MedAPE on the iPad was 37.0% and 36.5% for energy (26-28% of images within +/-20%), 43-65% for the macronutrients and 68-101% for salt (rho 0.51-0.85). On the desktop computer, supplying the true weight reduced MedAPE of the 4B model on NutriImage from 37.0% to 32.5% for energy and from 100.0% to 66.8% for salt; energy MedAPE stayed below the base value for declared weights of 0.5-1.25 times the true weight or rounded to 50 g, but rose markedly when the container weight was included. Conclusions: An offline VLM on a tablet reproduced server-side energy accuracy and is a candidate tool for dietary counselling where photographs cannot leave the institution; given the wide limits of agreement, its usefulness in practice remains to be examined.

Background. Research data enclaves such as the All of Us Research Program do not allow participant-level data, or models fitted to them, to leave, which makes it hard to deliver case-finding models, which flag patients with possibly uncoded conditions for clinician review, to patients and clinicians. Objectives. We describe and technically evaluate a system that fits models outside an enclave from released summary counts and delivers them through standard health information technology interfaces. Methods. We exported dissemination-compliant summary counts from the All of Us Researcher Workbench and, outside the enclave, fitted versioned models for Long COVID, chronic fatigue, fibromyalgia, endometriosis, Lyme disease, lupus, migraine, depression, and anxiety. A model contract (a model file and a written specification), a Python reference implementation, and synthetic Fast Healthcare Interoperability Resources (FHIR) test patients define the expected outputs. One JavaScript scorer runs in a patient demonstration, a SMART on FHIR clinician application, and a CDS Hooks service. An atlas displays aggregate results. We evaluated the system on synthetic records with automated tests, cross-implementation output comparison, launches through the public SMART sandbox, and a counterexample review of the scoring and display rules. Results. The deployed Long COVID model had an area under the receiver operating characteristic curve of 0.67 in a random validation split. The eight self-report models reached 0.61 to 0.82 in identifying self-reporters without a code. The JavaScript scorer reproduced all 162 reference outputs exactly, and all 437 automated tests passed. A counterexample review found 5 classes of failure, including scoring ineligible patients and treating unavailable data as absent, all fixed and added as tests. Conclusions. Models fitted from released summary counts can be delivered through FHIR, SMART on FHIR, and CDS Hooks with reproducible scoring on synthetic records. Real-record interoperability, usability, and prospective clinical performance remain to be evaluated.

Background: Passively collected smartphone and wearable data, combined with genomic information, may enable scalable early identification of youth who are at risk for substance use disorder. However, heterogeneous data collected in naturalistic settings limit researcher control over data quality, requiring substantial preprocessing. Data cleaning and feature engineering decisions can meaningfully influence analytic outcomes yet are often minimally described or omitted in published work. Objectives: We provide rationales and recommendations for common preprocessing challenges in passive sensing research, including establishing inclusion criteria when "outliers" may reflect true behavioral variability, analyzing differences across smartphone operating systems, handling observations collected outside protocol windows, and engineering features for machine learning applications. Methods: We illustrate principled approaches to these challenges using smartphone and Fitbit data from the Adolescent Brain Cognitive Development Study (ABCD). The sample was drawn from the available half of the year-4 follow-up (data release 5.1; N = 4,754; ages 13-14). Results: Extreme outlying observations were concentrated among low-engagement participants with few observed days. Smartphone usage among Apple users appeared under-recorded relative to Android users when compared with self-report measures. Fitbit-derived behavior differed systematically between observations collected within versus outside protocol windows. We also identified potential errors in the study's default preprocessing pipeline. Conclusion: We recommend prioritizing person-level inclusion criteria that preserve meaningful behavioral variability, examining temporal distributions rather than assuming observations are exchangeable, and thoroughly investigating and reporting data quality issues. Although illustrated using the ABCD Study, these principles broadly apply to research using smartphones, wearables, and other digital data sources.