JMIR Medical Informatics

Machine Learning–Based Multidimensional Oximetry for Obstructive Sleep Apnea Screening: Development and External Validation

2026-05-08T13:30:03-04:00

Background: Obstructive sleep apnea (OSA) affects nearly one billion people globally and poses a substantial public health threat. Effective and accessible methods for OSA risk identification are urgently needed. Objective: This study aims to develop and externally validate a machine learning model derived from multi-parameter pulse oximetry (SpO₂) for OSA screening, and to evaluate its performance, interpretability, and robustness across sex and age subgroups. Methods: Of 4156 screened participants, 2195 underwent polysomnography (internal cohort) and 446 received home sleep apnea testing (external cohort). Eight SpO₂-derived parameters, including oxygen desaturation index (ODI), hypoxic burden (HB), and ST90 (percentage of sleep time with SpO₂ < 90%), were used to construct models. Six machine learning algorithms were trained, with F₁-score as the primary metric and area under the curve as the secondary metric. Model interpretability was assessed using Shapley additive explanations and intrinsic feature importance scores. Results: Nonlinear parameter-risk relationships were observed between oximetry indices and OSA probability. The 4-parameter ODI-HB-MinSpO₂-ST90 model achieved optimal performance (F₁-score = 0.9516, area under the curve = 0.9879), surpassing all single-parameter models. Shapley additive explanations analysis identified ODI, HB, and MinSpO₂ as key predictors. The ODI-HB-MinSpO₂-MeanSpO₂ configuration demonstrated superior performance in female and younger subgroups, whereas the ODI-HB-MinSpO₂-ST90 model remained optimal for male and older participants. Categorical boosting outperformed other algorithms across multiple metrics and remained robust in both subgroup and external validation analyses. Conclusions: The multi-parameter oximetry model based on the categorical boosting algorithm provides a simple and accurate tool for OSA screening. Sex- and age-stratified strategies can further enhance its clinical applicability. Trial Registration:

Clinical Context Variables Collectively Rival Model Choice in Embedding-Based Retrieval: Multi-Corpus Benchmark Study

2026-05-07T17:30:26-04:00

Background: Retrieval-augmented generation (RAG) systems increasingly support clinical decision-making by grounding large language model outputs in verifiable evidence. The retrieval component is foundational: if the correct document is not retrieved, downstream generation cannot recover it. Despite this, embedding model selection for clinical RAG remains guided by general-domain benchmarks with limited clinical coverage. Given the heterogeneity of clinical documentation across institutions, specialties, and electronic health record systems, it is unclear whether general-domain model rankings generalize to clinical retrieval tasks. Objective: This study evaluated whether clinical context variables, corpus type (encompassing differences in document length, medical specialty, and structural characteristics), and query format have effects on retrieval performance comparable to or exceeding those of embedding model choice. Methods: Ten primary embedding models plus two ablation variants and a BM25 lexical baseline (13 retrieval configurations total) were benchmarked on three clinical corpora (MTSamples medical transcriptions, n=500; PMC-Patients case reports, n=500; Mistral-7B-generated synthetic clinical notes, n=500). Twelve embedding configurations were evaluated across 3 corpora × 2 query formats (keyword vs natural language) × 4 chunking strategies, yielding 294 experimental conditions. Primary metrics included MRR@10, P@1, Recall@10/20/50/100, and NDCG@10, with bootstrap confidence intervals. Relative factor contributions were quantified using factorial ANOVA with η² effect sizes, including all two-way interactions. Results: In a factorial ANOVA across 288 balanced embedding conditions, embedding model choice explained 40.8% of variance in MRR@10 (η²=0.408), corpus type 24.6%, and query format 19.2%. Chunking strategy explained minimal variance (η²=0.002). The model × query format interaction (η²=0.029, P<.001) indicated differential query sensitivity across models. A model × corpus interaction (η²=0.040, P<.001) indicated that model rankings shifted meaningfully across corpora. Combined context variables (corpus + query format + context interactions) explained 49.0% of total variance, compared with 47.6% for model-related effects. Model rankings were moderately unstable under keyword queries (Kendall τ=0.59, 95% CI [0.21, 0.89]) but highly stable under natural language queries (τ=0.82–0.87). BM25 achieved near-perfect retrieval on PMC-Patients in this known-item setting (MRR@10=0.999). Domain-specific models (BioBERT, ClinicalBERT) performed worse than general-purpose embeddings despite biomedical pretraining, with mean pairwise cosine similarity exceeding 0.90, indicating that all embeddings clustered in a narrow cone. A validation experiment using reduced-lexical-dependence queries—generated from GPT-4o-extracted metadata rather than document text—supported rank stability across query derivations (Kendall τ=0.59–0.90, mean 0.76, all P≤.004) and showed that BM25 remained strong on structured case reports (MRR@10=0.980). Conclusions: Clinical context variables explained as much variance in retrieval performance as embedding model choice, and model × corpus interactions showed that rankings are not portable across documentation types. Validation with reduced-lexical-dependence queries supported rank stability across query derivations. These results argue against reliance on general-domain leaderboards for clinical RAG deployment and support mandatory local validation as a methodological requirement.

Cough Audio Recognition for Early Detection of Respiratory Diseases: Algorithm Development and Validation Study

2026-05-07T14:15:17-04:00

Background: Coughing is a common clinical symptom and a protective respiratory reflex closely associated with various respiratory system diseases. The acoustic characteristics of cough sounds are influenced by underlying pathological factors, with distinct acoustic signatures corresponding to different etiologies. Through rigorous analysis of these sounds, rapid identification and preliminary diagnosis of related conditions may be achieved. This approach holds great potential for broad application in mobile health and ubiquitous health platforms. Objective: This study aimed to explore the application of acoustic analysis of cough sounds in the diagnosis of respiratory diseases to enhance the diagnostic efficiency of health care professionals. Methods: In this study, we conducted extensive data collection, including voluntary cough audio recordings from patients diagnosed with respiratory diseases (eg, chronic obstructive pulmonary disease, lung cancer, COVID-19, and pneumonia) and from healthy participants. A total of 2610 audio samples were collected. We incorporated a channel attention mechanism (CAM) into the final convolutional block of each residual block in the ResNet18 neural network, thereby constructing the CAM-ResNet18 neural network model. The recorded cough audio samples were converted into spectrograms to form the input dataset for model training. The CAM-ResNet18 model was trained on the training set of this dataset, with iterative parameter adjustments until convergence was achieved. Finally, spectrograms from the test set were fed into the pretrained model for accurate classification of the cough-related conditions. Results: Experimental results on the collected audio dataset demonstrate that the proposed CAM-ResNet18 model achieves an accuracy of 83.9% and an average -score of 82.52% in classifying 5 types of cough sounds. In comparison, the traditional ResNet18 model achieves an accuracy of 78.16% and an average -score of 78.29%, indicating a clear performance improvement with the integration of the CAM. Conclusions: The experimental results validate the effectiveness of the proposed method, highlighting its significant potential for application in clinical diagnosis.

Resource Use Patterns in US Telehealth Services: Machine Learning and Clustering Analysis Across 4 Specialties

2026-05-07T13:00:05-04:00

Background: The expansion of telehealth services, particularly during the COVID-19 pandemic, has transformed health care delivery in the United States. Telehealth promises greater access and resource efficiency by reducing wait times and appointment lengths, especially in specialties like psychiatry, behavioral health, bariatrics, and sleep medicine. However, disparities exist in adoption based on demographics, geography, and socioeconomic status, raising concerns about equitable access and optimal resource use. Objective: This study aims to evaluate how telehealth impacts health care resource use across 4 specialties by examining 2 key metrics: patient-to-provider ratios and appointment durations. It seeks to understand how factors such as patient demographics, facility characteristics, and social determinants influence telehealth adoption and efficiency using a national dataset spanning from 2018 to 2023. Methods: We analyzed a deidentified dataset from Epic Cosmos, covering outpatient visits across 48 US states (2018-2023). After data preprocessing and feature engineering, we applied 3 machine learning (ML) models (random forest, extreme gradient boosting, and deep neural networks) to predict resource use. Using the model performing the best, feature importance was assessed using Shapley Additive Explanations values. We then used k-means clustering to group facilities into clusters per specialty. Comparative analyses were conducted to evaluate differences in use among clusters, during and after the pandemic. Results: Telehealth use peaked in 2020 and has remained above prepandemic levels since then. In 2018-2023, telehealth adoption reached 36.9% (4,543,021/12,311,710) in psychiatry, 23.9% (5,321,099/22,264,013) in behavioral health, 21.2% (924,333/4,360,061) in bariatrics, and 16.8% (851,803/5,070,256) in sleep medicine. Telehealth visits were consistently shorter than office visits (mean reduction 12.24 minutes; SD 3.33 minutes; P=.18), while patient-to-provider ratios varied significantly across specialties. Among ML models, extreme gradient boosting regression achieved the best performance (patient-to-provider ratios: R²=0.96-0.99; appointment durations: R²=0.61-0.69). Shapley Additive Explanations analysis identified visit type, telehealth use, facility size, rurality, and Social Vulnerability Index household vulnerability as the strongest predictors. Comparative analyses showed significant differences across clusters (all P<.05). Conclusions: Telehealth has become a sustainable component of health care, enhancing access and efficiency across both rural and urban areas. However, its impact varies across specialties and regions, highlighting the need for targeted strategies such as staffing support for vulnerable populations, infrastructure investments in rural facilities, and reimbursement models that reflect telehealth’s resource use. This study provides robust evidence from ML and clustering analyses, demonstrating how telehealth shapes resource use and offering actionable insights for equitable and sustainable integration.

Correction: Evaluation of Large Language Models for Radiologists' Support in Multidisciplinary Breast Cancer Teams: Comparative Study

2026-05-07T10:30:03-04:00

Digital Health Communication for Deaf Individuals: Scoping Review of Technologies, Strategies, and Outcomes

2026-05-06T15:30:17-04:00

Background: Hearing loss affects approximately 432 million adults globally, with Deaf individuals representing a distinct linguistic and cultural minority that faces significant barriers to accessing health information. These challenges contribute to health disparities by limiting preventive education and timely health interventions. Objective: This scoping review examines the effectiveness of digital communication technologies in promoting health literacy, awareness, and health-related skills among Deaf adults and children. Methods: A comprehensive literature search was conducted across 5 major databases: MEDLINE, Embase, Scopus, Web of Science, and PubMed, focusing on peer-reviewed studies published in English within the past 10 years. Seventeen studies were included, encompassing a variety of research designs, including randomized controlled trials, cross-sectional surveys, and mixed methods approaches. Data extraction focused on intervention type, outcomes, and target populations. Results: Findings indicate that video-based interventions are the most prevalent and effective, leveraging sign language, subtitles, and animations to enhance accessibility and comprehension. These digital tools improved health awareness, knowledge acquisition, and the practical application of health-related skills across both adult and child populations. Interventions ranged from stroke preparedness and cancer education to breast self-examination and cardiopulmonary resuscitation training. Social media platforms, SMS text messaging campaigns, and eHealth programs were also identified as effective in promoting preventive health behaviors. Despite these promising outcomes, several challenges remain, including limited digital literacy, inconsistent access to technology, and a lack of culturally and linguistically appropriate content. Additionally, most studies were geographically concentrated in the United States, with a limited number of high-quality randomized trials. Conclusions: This review highlights the transformative potential of accessible digital technologies to reduce health disparities and promote health equity among Deaf individuals. Future research should prioritize inclusive, culturally sensitive, and user-centered designs and explore emerging platforms to maximize engagement and improve health outcomes. Trial Registration: OSF Registry; https://osf.io/rezgm/overview

AI and Internet of Things for Chronic Obstructive Pulmonary Disease Remote Monitoring: Systematic Review of Exacerbation Prediction and Key Physiological Variables

2026-05-06T15:15:17-04:00

Background: Chronic obstructive pulmonary disease (COPD) is a leading cause of morbidity and mortality worldwide, with frequent exacerbations of COPD (ECOPD) significantly impacting patient health and health care systems. Predicting ECOPD early would increase patients’ quality of life and decrease the economic burden. The advancement of wearable technologies and Internet of Things (IoT) sensors has enabled continuous remote monitoring (RM), offering new opportunities for early ECOPD prediction. However, effectively leveraging wearable data requires robust artificial intelligence (AI) frameworks capable of processing heterogeneous physiological and environmental information. Objective: This systematic review aims to provide a comprehensive overview of both hardware and software solutions for predicting ECOPD using RM. From the reviewed literature, we first focus on key physiological and environmental variables essential for COPD monitoring that can be extracted from wearables and IoT sensors. Second, we describe the wearable and IoT devices currently deployed in COPD management. Finally, we review machine learning, including deep learning models, used for ECOPD prediction, discussing limitations for real-world implementation. By bridging AI-driven data processing with real-world sensor applications, this review aims to outline the current landscape, existing challenges, and future directions for developing effective RM solutions for ECOPD predictions. Methods: A comprehensive search was conducted following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines to identify studies using AI or machine learning techniques for predicting ECOPD in in-home contexts. Results: This review identified 26 studies that met the inclusion criteria. Twenty studies aimed at predicting or detecting exacerbations at the onset. The variables tracked most frequently were heart rate (n=9), peripheral oxygen saturation (n=9), and symptoms (n=8). Daily or weekly sampling was most common (n=14). Most studies (n=13) applied machine learning models—primarily random forest (n=5), CatBoost (n=2), decision trees (n=2), and support vector machines (n=2). Deep learning was used in 3 papers, while the remaining applied rule-based logics and probabilistic models. Wearables and IoT were used in only 6 out of 20 studies. Six papers analyzed changes in vital parameters during prodromal phases, defined as the period shortly before the onset of an exacerbation. Three studies collected data continuously, 2 daily, and 1 compared once-daily versus overnight monitoring; 4 of these 6 used wearable devices. Conclusions: Overall, current evidence highlights the potential of continuous monitoring of physiological and environmental variables for early ECOPD prediction, offering advantages over questionnaires or once-daily measurements. While wearables and IoT devices show promise, their use remains limited. Many studies rely on balanced datasets that do not mirror real-world exacerbation patterns and lack external validation across diverse populations. Future research should emphasize large-scale validation, integration of multimodal data, and translation of AI models into clinically feasible tools to enable timely intervention and improve COPD management. Trial Registration: PROSPERO CRD420251051302; https://www.crd.york.ac.uk/PROSPERO/view/CRD420251051302

General Practitioners and Pharmacists’ Perspectives on Electronic Prescribing for Multidose Drug Dispensing: Mixed Methods Study

2026-05-06T14:15:14-04:00

Background: Medication safety remains a significant challenge in health care, particularly for patients managing complex treatment regimens. In Norway, the introduction of electronic prescribing (e-prescribing) for multidose drug dispensing (eMDD) aims to improve medication adherence and minimize errors by seamlessly integrating with the national e-prescription infrastructure. Objective: This study aimed to investigate the challenges faced by general practitioners (GPs) and pharmacists in using eMDD in Norway. Additionally, it sought to gather their recommendations for system improvements to guide future development and nationwide implementation. Methods: A parallel mixed methods design was used, integrating both quantitative and qualitative data. A structured online survey was distributed to 54 pharmacies and 190 GP surgeries across Norway. The survey included a combination of multiple-choice and open-ended questions. Qualitative responses were analyzed thematically using NVivo, while quantitative data were processed using the built-in analytical tools in Nettskjema. Results: A total of 70 health care professionals participated in the study, revealing 7 key themes: training, system and technology, communication and interaction, division of responsibilities, medication safety, time and resource use, and implementation challenges. GPs reported inadequate training and an overwhelming volume of communication, while pharmacists identified issues with system integration and unclear role definitions. Both groups emphasized the need for improved system usability, stronger interprofessional collaboration, and a more defined governance structure. Conclusions: While the eMDD system has the potential to improve medication safety and optimize workflows, its success depends on addressing technical inefficiencies, improving user training, and clarifying role responsibilities. Actively involving end users in system development and policy planning is critical for achieving effective national implementation and ensuring integration with broader eHealth initiatives, such as the Patient’s Medication List.

Correction: Complete Blood Count–Derived Inflammation Indices to Predict 3-Year All-Cause Mortality in Patients With Diabetes and Acute Myocardial Infarction in Critical Care: Retrospective Cohort Study With Single-Center External Validation

2026-05-05T15:30:31-04:00

The Ethics of AI Scribes as Epistemic Agents

2026-04-30T17:00:25-04:00

Artificial intelligence (AI) scribes using ambient documentation technology that capture clinician-patient dialogue and auto-generate visit notes promise to alleviate documentation burden and reduce clinician burnout. In discussing empirical evidence, highlighting research gaps, and emphasizing technology-related ethical issues beyond established AI and data ethics, we show how this promise comes along with epistemic and relational risks. We proceed in 5 steps: first, we conceptually distinguish ambient documentation from broader ambient intelligence, frame it as a “tech-fix” for documentation-related burnout, and establish the notion of AI scribes as epistemic agents rather than mere transcription tools; second, we summarize empirical evidence on AI scribes, especially with regard to their impact on physicians, highlighting risks such as cognitive deskilling, clinical deprofessionalization, and shifts in epistemic accountability; third, we analyze effects on the patient-physician relationship, focusing on relational and interpretive dimensions, including changes in communication patterns and the omission of narrative nuance; fourth, we highlight risks to patient agency and epistemic justice; and fifth, we propose a design framework for ethical deployment beyond techno-solutionism. We argue that the usefulness of AI scribes should not be justified by short-term effects, but must be assessed in the context of clinical reasoning to improve not only the working conditions of physicians, but also the quality of patient care. The paper proposes a research and design agenda to counter simple “tech-fixes” for systemic problems, envisioning AI scribes that safeguard clinical reasoning and honor patient narratives while delivering relief from documentation burdens.