PLOS Digital Health

Mobile devices offer a scalable opportunity to collect longitudinal data that facilitate advances in mental health treatment to address the burden of mental health conditions in young people. Sharing these data with the research community is critical to gaining maximal value from rich data of this nature. However, the highly personal nature of the data necessitates understanding the conditions under which young people are willing to share them. To answer this question, we developed the MindKind Study, a multinational, mixed methods study that solicits young peoples preferences for how their data are governed and quantifies potential participants willingness to join under different conditions. We employed a community-based participatory approach, involving young people as stakeholders and co-researchers. At sites in India, South Africa, and the UK, we enrolled 3575 participants ages 16-24 in the mobile app-mediated quantitative study and 143 participants in the public deliberation-based qualitative study. We found that while youth participants have strong preferences for data governance, these preferences did not translate into (un)willingness to join the smartphone-based study. Participants grappled with the risks and benefits of participation as well as their desire that the "right people" access their data. Throughout the study, we recognized young peoples commitment to finding solutions and co-producing research architectures to allow for more open sharing of mental health data to accelerate and derive maximal benefit from research.

BackgroundOne of the most difficult challenges in pediatric telemedicine is to accurately discriminate between the sick and not sick child, especially in resource-limited settings. Models that flag potentially sick cases for additional safety checks represent an opportunity for telemedicine to reach its potential. However, there are critical knowledge gaps on how to develop such models and integrate them into electronic clinical decision support (eCDS) tools. MethodsTo address this challenge, we developed a study design that utilized data from paired virtual and in-person exams at a telemedicine and medication delivery service (TMDS) in Haiti. Providers were allowed to mark respondent data as potentially unreliable. Artificial intelligence /machine learning (XGBoost) was applied to analyze paired data from participants across three consecutive implementation studies. Model derivation focused on identifying sick patients (not-mild) and those requiring escalation. An ensemble method, based on gradient boosted decision-trees, was used given the limited sample size. The area under the receiver operating characteristic curve (AUC) was the primary outcome measure. ResultsA total of 683 paired records were available for this secondary analysis from 2225 participants enrolled. The median age was 15 months and 47% were female. For prediction of a sick child, we found an AUC of 0.82 (95% CI 0.78-0.86) after 5-fold cross validation; calibration slope and intercept were 1.31 (95%CI:1.09-1.53) and 0.04 (95%CI:-0.14-0.23), respectively. For prediction of escalation, we found an AUC of 0.78 (95%CI:0.74-0.81); calibration slope and intercept were 0.63 (95%CI:0.52-0.74) and 0.05 (95%CI:0.52-0.74), respectively. Accounting for data marked as potentially unreliable had mixed effects. InterpretationThese methods and findings offer an innovative and important proof-of-concept to improve pediatric telemedicine. The models require external validation prior to eCDS integration and deployment. Once validated, the models are designed to provide a critical safety check for experienced providers and digitally convey expertise to new providers. FundingNational Institutes of Health (USA) grants to EJN (R21TW012332; DP5OD019893), internal funding at UF (Childrens Miracle Network), and private donations. RESEARCH IN CONTEXTO_ST_ABSEvidence before this studyC_ST_ABSWe conducted two Pubmed searches for reports published in all languages. The first search terms were (telemedicine) AND (artificial intelligence OR machine learning) AND (pediatrics OR paediatrics). The primary search criteria identified 153 publications and reviews were excluded leaving 101 papers all published after 1998. Enumerated results by named subspecialty were neurology (n=2), ophthalmology (n=13), otology (n=3), endocrinology (n=11), cardiology (n=6), pulmonology (n=3), gastroenterology(n=1), dermatology (n=1) and surgery (n=22). Ten of the publications focused on global health or low-middle income countries (LMIC) populations. The second search terms were ((Telemedicine) AND (delivery OR paramedicine) AND (pediatrics OR paediatrics)) AND (global health OR LMIC) which generated 99 publications and 76 papers remained after reviews were excluded, all published after 2015. After manual evaluation of the results from both searches, no publications were identified that fully met the scope of this paper. Examples of telemedicine research for concordance with paired exams does exist 1,2. Added value of this studyTo the best of our knowledge, this is the first study that investigated pediatric disease severity prediction using virtual and in-person exams in the context of telemedicine -- for either high or low resourced settings. Therefore, the added value of this study is an innovative and important proof-of-concept to improve telemedicine research and clinical practice beyond the scope of global health. Implications of all the available evidenceInside the field of global health, there is a need to develop evidence-based approaches to extend care early to pediatric patients who may be isolated by poverty, geography or unrest. This must be done safely and this paper offers an approach to develop and incorporate disease severity prediction models into eCDS tools. In addition, these tools may serve as a welcomed safety check for experienced providers and a method to digitally convey expertise to new providers as these services scale.

Bronchiolitis is an acute infection of the lower respiratory tract that predominantly affects children less than two years old. Although self-limiting, symptoms of bronchiolitis can be distressing for young children. Research has demonstrated that parents may not have the necessary information to be able to identify bronchiolitis symptoms, resulting in emergency department (ED) visits and hospitalizations. Parents have expressed that they feel unprepared, afraid, and that they lack information on their childs condition. Digital knowledge translation (KT) tools have the potential to convey complex health information to parents to support their healthcare decision-making needs. We worked with parents of children with bronchiolitis to develop and evaluate three digital tools on bronchiolitis (whiteboard animation video, infographic, and e-Book). Following prototype completion, usability testing was conducted using iPads in two Alberta ED waiting rooms. Parents were randomized to one out of the three tools. Overall, the tools were highly rated, suggesting that arts-based digital tools are useful in delivering complex health information to parents.

This study explores clinician understanding and perception at site lead level towards machine learning (ML) decision support tools for paediatric related emergency care across the UK and Ireland, essential in guiding safe and effective frontline implementation. A cross-sectional online survey was distributed via Paediatric Emergency Research United Kingdom and Ireland (PERUKI) to the lead for digital systems or PERUKI site lead, with one response sought per site. Survey development was in REDCap, and descriptive analysis (counts, percentages) was performed. The response rate was 86.7% (65/75), mostly from England (83.1%). While 80.0% understood Artificial Intelligence, fewer understood advanced concepts such as Deep Learning (32.3%). Most clinicians believed ML will support decision making (83.1%), would be willing to use (87.7%), and the future of decision making is a combination of human and ML (83.1%). Barriers included concerns about bias (61.5%), ML accuracy (56.9%), and inadequate information technology infrastructure (67.7%). Digital leads were more concerned about ML accuracy than non-digital (68.2% vs. 51.2%). Among potential applications, antimicrobial stewardship ranked highest (90.8%), and diagnosis of mental health conditions lowest (24.6%). Strong interest in ML tools for decision support in paediatric emergency care was evident, though concerns about bias, accuracy, and infrastructure must be addressed. Ongoing co-design with clinicians is critical in ensuring these tools are trusted, useful and suited to paediatric emergency care. Targeted education, digital leadership, and strategic investment in infrastructure and governance are essential for the successful adoption and integration of ML in clinical workflows. Author SummaryWithin emergency care we are seeing a rapid growth in the research, development and frontline use of machine learning based tools for decision support, yet very little is known about the intended users understanding, opinions, experience of this technology and supporting structures. With any new technology a greater understanding leads to better adoption and the clinicians who will use these tools should be directly involved in their design, implementation and evaluation to ensure that these tools are clinically relevant and usable in practice. Through our survey of clinical site leads (key drivers, influencing the adoption of ever advancing technology), the findings provide critical insight into what clinicians are most concerned about, their perceptions, understanding, what applications they view as most clinically relevant and their willingness to be involved in future research and development of these tools. The results also revealed many important themes such as infrastructure readiness, trust, explainability, clinical integration, targeted education and human-artificial intelligence collaboration. These findings will contribute to shaping the future of research, development, education, governance and policy within this rapidly growing area.

Asthma is the most common chronic condition in children with an estimated 15% of children and youth living with asthma in Canada. Acute asthma exacerbations, or asthma attacks, are the main reason for children to seek emergency care, contributing to financial burdens for families and healthcare systems. This burden highlights opportunities to reduce health system costs and improve patient and family education. We worked with parents of children with asthma to develop and evaluate two digital knowledge translation (KT) tools on asthma. These tools merge the best available research evidence with narratives of parent experiences, and use art and engaging media (video and interactive infographic) to optimize uptake and appeal. Following prototype completion, usability testing was conducted among 60 parents (30 parents per tool) in an urban Alberta emergency department waiting room. Parents viewed the tools on an iPad and answered questions via an electronic survey. Usability was assessed based on nine items with responses on a five-point Likert scale from 1=strongly disagree to 5=strongly agree. Overall, results were positive and the tools were highly rated across most usability items. Mean scores across usability items were 4.13 to 4.63 for the video and 4.10 to 4.43 for the infographic. The scores from the usability testing suggest arts-based digital tools are useful in sharing complex health information with parents about the care of a child with asthma and provide meaningful guidance on how to improve KT tools to better reflect the needs of parents of children with asthma.

Computer vision models for chest X-ray interpretation hold significant promise for global healthcare, but their clinical value depends on equitable development across diverse populations. We conducted a scientometric analysis to examine authorship patterns, geographic distribution, and dataset origins to assess potential disparities that could affect clinical applicability. We systematically reviewed literature on computer vision applications for chest X-rays published between 2017-2025 across multiple databases, including PubMed, Embase and SciELO databases. Using Dimensions API and manual extraction, we analyzed 928 eligible studies, examining first and senior author affiliations, institutional contributions, dataset provenance, and collaboration patterns across different income classifications based on World Bank categories. High-income countries dominated research leadership, representing 55.6% of first authors and 59.7% of senior authors; no first authors were affiliated with low-income countries. China (16.93%) and the United States (16.72%) led in first authorship positions. Most datasets (73.6%) originated from high-income settings, with the United States being the largest contributor (40.45%). Private datasets were most frequently used (20.52%). Cross-income collaborations were rare, with only 3.9% of publications involving partnerships between high-income and lower-middle-income countries. Findings reveal substantial disparities in who shapes computer vision research on chest X-rays and which populations are represented in training data. These imbalances risk developing AI systems that perform inconsistently across diverse healthcare settings, potentially exacerbating healthcare inequities. Addressing these disparities requires coordinated efforts to develop globally representative datasets, establish equitable international collaborations, and implement policies that promote inclusive research practices.

Abstract Background: Cardiovascular disease is the leading cause of mortality in Nigeria and across sub-Saharan Africa, with rising incidence attributable to urbanisation, sedentary lifestyles, and limited access to early detection tools. Concurrently, patient dropout from rehabilitation programs remains a critical operational challenge for Nigerian clinics, with many patients failing to return after their initial consultation. Methods: We developed CardioAI, an Explainable Artificial Intelligence system comprising two predictive modules. The cardiovascular risk module trained four machine learning models - Logistic Regression, Random Forest, Gradient Boosting (XGBoost), and a Multilayer Perceptron - on a combined UCI Heart Disease dataset of 1,025 patient records. A novel Lifestyle Risk Index was engineered from five modifiable clinical markers. SHAP (SHapley Additive exPlanations) was applied for per-prediction feature attribution. The patient retention module trained three classifiers on a synthetic dataset of 800 records, modelling 10 operational and behavioural dropout factors. An NLP and OCR pipeline using Tesseract v5.5 and spaCy was implemented for clinical document processing. Results: The cardiovascular risk module achieved an AUC-ROC of 0.999 (XGBoost), 0.998 (Random Forest), 0.994 (MLP), and 0.927 (Logistic Regression) on the held-out test set. Cross-validated AUC with constrained tree depth was 0.97, confirming generalisation. SHAP analysis identified the Lifestyle Risk Index, ST depression, resting blood pressure, exercise-induced angina, and cholesterol as the five most influential predictors. The retention module achieved AUC-ROC of 0.66 (Logistic Regression), demonstrating the difficulty of dropout prediction with synthetic data. Conclusions: CardioAI demonstrates that explainable machine learning can provide clinically actionable cardiovascular risk assessment and patient retention intelligence in a low-resource Nigerian healthcare context. The system is freely deployable, open-source, and designed for pilot validation in teaching hospitals across Lagos and Port Harcourt. Keywords: cardiovascular risk prediction, machine learning, explainable AI, SHAP, patient retention, clinical decision support, Nigeria, sub-Saharan Africa, XGBoost, random forest, digital health

This study aimed to optimise the balance between participant burden and algorithm performance for predicting high-risk moments in a smoking cessation just-in-time adaptive intervention (JITAI) by systematically varying ecological momentary assessment (EMA) prompt frequency, predictor count, and training data source. Thirty-seven participants completed 16 EMAs per day for the first 10 days of their smoking cessation attempt, reporting mood, context, behaviour, cravings, and smoking lapses. Random forest algorithms predicting lapses and cravings were evaluated in terms of F1-score and ROC-AUC via mixed effects models accounting for clustering within individuals Performance across out-of-sample individuals ranged from excellent to poor but was, on average, modest. Lapse prediction outperformed craving prediction, particularly for ROC-AUC (Median F1-score: Lapses 0.436 [IQR 0.180-0.625], Cravings 0.400 [IQR 0.048-0.649]; Median ROC-AUC: Lapses 0.659 [IQR 0.514-0.809], Cravings 0.628 [IQR: 0.510-0.729]). A substantial proportion of configurations fell below commonly used minimum performance thresholds, particularly for F1-score. Reducing EMA frequency had outcome- and metric-dependent effects. Lapse F1-scores improved with fewer prompts (16 EMAs: 0.254 [IQR 0.081-0.500], 3 EMAs: 0.588 [IQR 0.353-0.667]), while ROC-AUC showed a slight, inconsistent decline (16 EMAs: 0.661 [IQR 0.520-876], 4 EMAs: 0.613 [IQR 0.494-0.786], 3 EMAs: 0.704 [IQR 0.567-0.809]). For cravings, both metrics declined with fewer prompts (F1-score: 16 EMAs: 0.470 [IQR 0.141-0.745]; 3 EMAs: 0.333 [IQR 0.000-0.600]; ROC-AUC: 16 EMAs 0.700 [IQR 0.582-0.811], 3 EMAs 0.544 [IQR 0.421-0.676]). Feature reduction had negligible impact on lapse prediction (F1-score: all features 0.435, selected features 0.441; ROC-AUC: all 0.660, selected 0.657), but slightly reduced craving performance (F1-score: all 0.410 [IQR 0.117-0.646], selected 0.400 [IQR 0.000-0.650]; ROC-AUC: all 0.632, selected 0.622). Including participant-specific data improved lapse F1-scores (None 0.286 [IQR 0.000-0.571], 30pc 0.542 [IQR: 0.329-0.667]), but did not ROC-AUC (None 0.655 [IQR: 0.512-0.786], 30pc 0.694 [IQR 0.513-0.852]); and impaired craving ROC-AUC (None 0.650 [IQR: 0.544-0.734], 30pc 0.614 [IQR 0.493-0.730]; F1-score: None 0.424 [IQR 0.143-0.649], 30pc 0.400 [IQR 0.000-0.703]). Overall, EMA-based machine learning detected lapse risk but showed modest overall performance and substantial inter-individual variability. Using higher EMA density, larger predictor sets, and participant-specific training data did not consistently outperform over more parsimonious approaches. However, machine learning prediction alone is unlikely to be sufficient for real-world JITAI implementation, and may be best combined with complementary rules-based approaches.

ObjectiveLower urinary tract symptoms (LUTS), such as urinary urgency, frequency, and incontinence, affect the majority of the population, causing substantial morbidity, yet few receive effective care. Sizeable symptomatic overlap between LUTS categories leads to high rates of misdiagnosis. To improve diagnostic accuracy, we sought to employ machine learning approaches to LUTS categorization to generate diagnostic groupings based on patient-reported clinical data, creating a novel tool for diagnosis of patients with voiding complaints. MethodsQuestionnaire responses in a Development Dataset of 514 female subjects were used for model development, identifying 4 major clusters and 9 specific phenotypes of LUTS using agglomerative hierarchical clustering. Each cluster and phenotype was assigned a clinical identity consistent with recognized causes of voiding dysfunction by the consensus of two urologic specialists. Then, a random forest classifier was trained to assign unseen patients into these phenotypes. That model was then applied to a Validation Dataset of 571 additional subjects to confirm the diagnostic algorithm. ResultsThis data-driven, hierarchical clustering approach captured overlapping symptoms inherent in typical patients, recognizing common uncomplicated diagnoses (i.e., overactive bladder) as well as several underrecognized diagnostic categories (i.e., myofascial pelvic pain). A diagnostic algorithm derived by supervised machine learning to assign unseen subjects into these phenotypes demonstrated good reproducibillty of the phenotypes and their symptomatic patterns in an independent Validation Dataset. ConclusionsWe describe the generation of a machine learning algorithm relying only on validated, patient-reported symptoms for diagnostic classification. Given a growing physician shortage and increasing challenges for patients accessing specialist care, this type of digital technology holds great potential to improve the recognition and diagnosis of functional urologic conditions.

Urinary tract infections (UTI) are a common source of acute illness for infants and children. Approximately 7-8% of girls and 2% of boys will experience a UTI before they are 8 years old. UTIs may be difficult to identify and treat as symptoms in children are different from expected adult symptoms. A previously conducted systematic review identified four common information needs expressed by parents. More specifically, the research identified that parents had difficulty recognizing signs and symptoms of UTIs, felt disappointed by health care providers responses, needed timely and relevant information, and feared the unknown due to lack of UTI knowledge. This demonstrates that more effective knowledge translation tools are needed to satisfy parent information needs. The purpose of this research was to work with parents to develop and test the usability of an interactive infographic and video about UTIs in children. Prototypes were evaluated by parents through usability testing in two Alberta emergency department waiting rooms. Results were positive and overall, the tools were highly rated across all usability items, suggesting that arts-based digital tools are useful mediums for sharing health information with parents.

Artificial intelligence (AI) is increasingly positioned as a transformative tool in education and health. Yet empirical evidence on AI readiness in low-and middle-income countries, particularly among youth, remains scarce. This study examined patterns of adoption, equity determinants, and ethical awareness among Ghanaian youth to inform responsible AI integration in education and health systems. A cross-sectional survey was conducted among 200 youth aged 18-35 years in Ghana. Descriptive statistics, chi-square tests, and logistic-regression analyses were used to assess AI adoption, equity patterns, and predictors of readiness. Most participants reported current (89%) or prior (65%) use of AI tools. Accessibility was a significant positive predictor of adoption ({beta} = 0.142, p = 0.001), whereas limited internet connectivity ({beta} = -0.088, p = 0.049) and perceived exclusion or inequity ({beta} = -0.109, p = 0.026) were significant negative predictors. Gender and age differences indicated persistent digital inequities. Ethical concerns were widespread: 51% were somewhat concerned and 39% very concerned about data privacy, algorithmic bias, and transparency. Ghanaian youth exhibit high AI readiness, but it is distributed in structurally uneven and ethically contested contexts. Readiness is best understood as a dynamic interaction between technical access, social inclusion, and trust. Translating readiness into equitable implementation will require investments in digital infrastructure, ethical governance, and participatory design. This study provides one of the first quantitative assessments of AI readiness among African youth and offers an evidence base for developing trustworthy, inclusive AI applications, such as healthcare and educational chatbots, that are grounded in local realities. Author SummaryArtificial intelligence (AI) is often presented as a solution to challenges in healthcare and education. However, there remains limited evidence on peoples readiness to use AI in low-and middle-income countries and on the ways in which equity and ethics shape that readiness. We surveyed 200 youth in Ghana to understand their use of AI tools, perceptions of fairness and ethical concerns. Most participants were already using AI, yet adoption was uneven. Access to reliable internet and devices strongly increased use, while perceptions of exclusion and limited connectivity reduced it. Many youths expressed concern about data privacy, bias, and transparency in AI systems. These findings show that Ghanaian youth are eager but cautious adopters who value fairness and accountability. Building equitable and trustworthy AI systems in education and health will require improving access, addressing social inequalities, and involving youth directly in the design and governance of new technologies.

Several triage systems have been developed, but little is known about their performance in low-resource settings. Evaluating and comparing novel triage systems to existing triage scales provides essential information about their added value, reliability, safety, and effectiveness before adoption. This prospective observational study included children aged < 15 years who presented to the emergency departments of two public hospitals in Kenya between February and December 2021. We compared the performance of Emergency Triage Assessment and Treatment (ETAT) guidelines and Smart Triage (ST) models (ST-only model, ST model with independent triggers, and recalibrated ST model with independent triggers) in categorizing children into emergency, priority, and non-urgent triage categories. We visualized changes in classification of participants using Sankey diagrams. 5618 children were enrolled, and the majority (3113, 55.4%) were aged between one and five years of age. Overall admission and mortality rates were 7% and 0.9%, respectively. ETAT classified less children, 513 (9.2%), into the emergency category compared to 790 (14.1%), 1163 (20.8%) and 1161 (20.7%) by the ST-only model, ST model with independent triggers and recalibrated model with independent triggers, respectively. ETAT also classified more children, 3089 (55.1%), into the non-urgent triage category compared to 2442 (43.6%), 2097 (37.4%) and 2617 (46.7%) for the respective ST models. ETAT classified 191/395 (48.4%) of admitted patients as emergency compared to more than half by all the ST models. ETAT and the ST-only model classified 25/49 (51%) children who died as emergencies, while the ST models with independent triggers classified 39/49 (79.6%) children as emergencies. Smart Triage shows potential for identifying critically ill children in low-resource settings, particularly when combined with independent triggers. Additionally, it performs comparably to ETAT. Evaluation of Smart Triage in other contexts and comparison to other triage systems is required. Author summaryPrioritizing children according to the level of severity of illness in the outpatient department is crucial to ensure very sick children are identified and receive life-saving treatment while those with less severe symptoms can safely wait in the queue. Appropriate triage prevents avoidable paediatric mortality. As new triage systems are developed, it is essential to evaluate their performance before being used by healthcare professionals to manage patients. In this study, we compared a newly developed triage algorithm, Smart Triage, to the World Health Organizations Emergency Triage Assessment and Treatment (ETAT) guidelines. Here, we highlight how participants were categorised into emergency, priority, and non-urgent categories by both triage systems. We also assessed changes in triage categorization by comparing the Smart Triage model only (with and without site specific recalibration) and the model with independent emergency and priority triggers aligned with ETAT. Our study shows that Smart Triage had comparable performance to ETAT, and it can be used to triage children in resource-limited settings. Smart Triage can be integrated into a digital device allowing frontline healthcare workers to rapidly triage children presenting to the outpatient department and recognize very sick children faster, so that they can be treated in a timely manner.

Digital health interventions (DHIs) refer to discrete technological functionalities designed to achieve specific objectives in addressing health system challenges through digital health applications. These interventions are tools for strengthening health systems, particularly in low- and middle-income countries. This study consolidates findings from Ethiopia, Ghana, and Zimbabwe, examining three distinct digital health applications with varying intervention capabilities that contribute to health system strengthening within their respective primary healthcare contexts. The interventions analyzed include Ethiopias District Health Information System 2 (DHIS2), Ghanas dual system namely the District Health Information Management System (DHIMS) and the Local Health Information Management System (LHIMS and Zimbabwes Impilo electronic health record (E-HR) system. In Ethiopia, DHIS2 supports DHIs focused on data aggregation, reporting, and performance monitoring at the public health level. At the health system level, DHIS2 has enhanced accountability and data quality, leading to improved decision-making and resource distribution. In Ghana, DHIMS functions as a public health-level DHI, facilitating data-driven performance monitoring, while LHIMS operates at the patient level, supporting patient tracking and management, improving patient workflows and resource tracking. However, a lack of interoperability between these two systems has led to data duplication challenges. Zimbabwes Impilo E-HR, a patient-level DHI, has streamlined clinical workflows, improved information sharing, and enhanced decision-making at the point of care. Despite these successes, challenges persist, including infrastructure limitations, high staff turnover, and insufficient technical capacity among users. Interoperability issues, particularly in Ghana and Ethiopia, hinder seamless data exchange, while sustainability concerns such as funding gaps and inadequate government support limit the full potential of these systems. The studies highlight the need for targeted training, improved infrastructure, and enhanced integration of health information systems to maximize the benefits of DHIs. Authors SummaryHealth systems are built on six fundamental building blocks, yet they face numerous challenges, particularly in the African context. Digital health interventions (DHIs) have been shown to strengthen one or more of these building blocks by improving healthcare efficiency, data management, and service delivery. These interventions are especially valuable in low- and middle-income countries, where they help address systemic issues such as inefficient data management, limited access to care, and health workforce efficiencies. Our paper presents findings from three independent studies examining digital health applications implemented in Ethiopia, Ghana, and Zimbabwe, each playing a unique role in strengthening their respective health systems. We explore how DHIs have contributed to key health system functions, highlighting successes in data utilization, patient management, and decision-making. At the same time, we report on the challenges faced by implementers, including infrastructure limitations, interoperability issues, and sustainability concerns. Based on our analysis and experience of the environment, we offer insights and practical recommendations for overcoming these challenges, ensuring that DHIs can maximize their impact on healthcare delivery. By addressing these barriers, DHIs have the potential to enhance health system resilience and efficiency, ultimately improving health outcomes in resource-limited settings.

BackgroundPneumonia is a leading cause of death among children under 5 years in low- and-middle-income-countries (LMICs), causing an estimated 700,000 deaths annually. This burden is compounded by limited diagnostic imaging expertise. Artificial intelligence (AI) has potential to improve pneumonia diagnosis from chest radiographs (CXRs) through enhanced accuracy and faster diagnostic time. However, most AI models lack validation on prospective clinical data from LMICs, limiting their real-world applicability. This study aims to develop and validate an AI model for childhood pneumonia detection using Nigerian CXR data. MethodsIn a multi-center cross-sectional study in Ibadan, Nigeria, CXRs were prospectively collected from University College Hospital (a tertiary hospital) and Rainbow-Scans (a private diagnostic center) radiology departments via cluster sampling (November 2023-August 2024). An AI model was developed on open-source paediatric CXR dataset from the USA, to classify the local prospective CXRs as either normal or pneumonia. Two blinded radiologists provided consensus classification as the reference standard. The models accuracy, precision, recall, F1-score, and area-under-the-curve (AUC) were evaluated. ResultsThe AI model was developed on 5,232 open-source paediatric CXRs, divided into training (1,349 normal, 3,883 pneumonia) and internal test (234 normal, 390 pneumonia) sets, and externally tested on 190 radiologist-labeled Nigerian CXRs (93 normal, 97 pneumonia). The model achieved 86% accuracy, 0.83 precision, 0.98 recall, 0.79 F1-score, and 0.93 AUC on the internal test, and 58% accuracy, 0.62 precision, 0.48 recall, 0.68 F1-score, and 0.65 AUC on the external test. ConclusionThis study illustrates AIs potential for childhood pneumonia diagnosis but reveals challenges when applied across diverse healthcare environments, as revealed by discrepancies between internal and external evaluations. This performance gap likely stems from differences in imaging protocols/equipment between LMICs and high-income settings. Hence, public health priority should be developing robust, locally relevant datasets in Africa to facilitate sustainable and independent AI development within African healthcare. Author SummaryPneumonia is a leading cause of death in children under five, especially in low-resource settings like Nigeria, where access to diagnostic tools and expertise is limited. Our study explores how artificial intelligence (AI) can help address this gap by detecting pneumonia from chest X-rays. We trained an AI model using a large dataset of childrens X-rays from the United States and tested it on images collected in Nigeria. While the AI model performed well on the U.S. data, its accuracy dropped significantly when tested on the Nigerian X-rays. This reveals how differences in imaging techniques and equipment between countries can affect the performance of such models. It highlights the need for AI systems to be adapted to local contexts to ensure they are reliable and effective in real-world settings. Our findings underline the importance of creating high-quality, locally relevant datasets in Africa to support the development of AI tools that address the unique challenges of the region. By investing in such efforts, we can improve access to life-saving technologies, particularly for vulnerable populations in resource-limited healthcare systems.

AbstractO_ST_ABSBackgroundC_ST_ABSTuberculosis (TB) remains a leading global cause of preventable death, with 10.8 million cases and 1.3 million deaths reported in 2023. Current methods for TB screening include symptom-based screening, and chest X-ray (CXR) with computer-aided detection (CAD-CXR). Each method has limitations related to cost, accessibility, and screening efficacy. As a result, an estimated 2.6 million TB cases were missed in 2023. AI-based TB screening using lung sounds captured by a digital stethoscope offers a potential solution to these challenges, enhancing access, efficacy, and cost-efficiency. MethodsA dataset comprising 49,770 anonymized chest auscultation recordings from 1,659 participants (cases and controls) were collected by trained nurses in South Africas Western Cape province from June 2021 to November 2022 using AI Diagnostics prototype digital stethoscope. Consenting participants suspected to have TB that reported a recent sputum TB Xpert Ultra test were recruited from 34 primary care clinics. After stratification and data preparation, a final dataset of 1,169 participants was partitioned into an 80% training and 20% hold-out test set. A pre-trained transformer- based architecture was fine-tuned using K-fold cross-validation. The ensemble models ability to predict pulmonary TB was evaluated on the hold-out test set, with sputum Xpert Ultra as the reference standard. ResultsThe AI model achieved a mean Area under the Receiver Operating Curve (AUC-ROC) of 0.79 (95% CI: 0.73-0.85). At a sensitivity of 89.9% (95% CI: 82.4%-94.4%), the ensemble model has a specificity of 50.4% (95% CI: 42.0%-58.7%) for predicting pulmonary TB using lung sounds. ConclusionAI-based digital chest auscultation for TB, with a sensitivity of 89.9% and specificity of 50.4% in this study, shows early promise as an alternative or adjunct to current TB screening methods. In addition, the methods portability and low cost have the potential to significantly improve TB screening access. Future independent studies in diverse, unselected populations with high TB prevalence are required to validate model generalizability. Key messages What is already known on this topicTuberculosis (TB) is a major global health issue, especially in low-resource areas. Existing screening methods like symptom checks, chest X-rays, and CAD tools are often costly, hard to access, or not sensitive enough. AI has shown promise in detecting other lung conditions using sound, but its use for TB screening has not been well studied. What this study addsThis is the first large study showing early promise that AI can detect TB from lung sounds using a digital stethoscope. This technology could be further developed as a low-cost and portable screening tool which aligns well with the World Health Organizations End TB Strategy. How this study might affect research, practice, or policyThis study would encourage further research into AI-based auscultation in different populations and settings, helping build more the evidence base supporting the use of AI in disease screening. Further research would also support the development of more accurate and generalizable models. In clinical practice AI-based digital stethoscopes could be used for early TB screening, allowing faster diagnosis and treatment. This would be especially important in asymptomatic TB cases where symptom-based screening would miss all cases. From a policy perspective, the results of this study would support further research which may support the inclusion of this technology in national and global TB screening guidelines and WHO endorsement. This study was commercially funded by the technology provider, AI Diagnostics Pty (Ltd).

Cardiac auscultation is an accessible diagnostic screening tool that can help to identify patients with heart murmurs for follow-up diagnostic screening and treatment for abnormal cardiac function. However, experts are needed to interpret the heart sounds, limiting the accessibility of auscultation for cardiac care in resource-constrained environments. Therefore, the George B. Moody PhysioNet Challenge 2022 invited teams to develop algorithmic approaches for detecting heart murmurs and abnormal cardiac function from phonocardiogram (PCG) recordings of the heart sounds. For the Challenge, we sourced 5272 PCG recordings from 1568 pediatric patients in rural Brazil, and we invited teams to implement diagnostic screening algorithms for detecting heart murmurs and abnormal cardiac function from the recordings. We required the participants to submit the complete code for training and running their algorithms, improving the transparency, reproducibility, and utility of their work. We also devised an evaluation metric that considered the costs of screening, diagnosis, treatment, and diagnostic errors, allowing us to investigate the benefits of algorithmic diagnostic screening and facilitate the development of more clinically relevant algorithms. We received 779 algorithms from 87 teams during the course of the Challenge, resulting in 53 working codebases for detecting heart murmurs and abnormal cardiac function from PCGs. These algorithms represent a diversity of approaches from both academia and industry. The use of heart sound recordings for identifying heart murmurs and abnormal cardiac function allowed us to explore the potential of algorithmic approaches for providing accessible pre-screening in resource-constrained environments. The submission of working, open-source algorithms and the use of novel evaluation metrics supported the reproducibility, generalizability, and clinical relevance of the research from the Challenge. Author summaryCardiac auscultation is an accessible diagnostic screening tool for identifying heart murmurs. However, experts are needed to interpret heart sounds, limiting the accessibility of auscultation in cardiac care. The George B. Moody PhysioNet Challenge 2022 invited teams to develop algorithms for detecting heart murmurs and abnormal cardiac function from phonocardiogram (PCG) recordings of heart sounds. For the Challenge, we sourced 5272 PCG recordings from 1568 pediatric patients in rural Brazil. We required the participants to submit the complete code for training and running their algorithms, improving the transparency, reproducibility, and utility of their work. We also devised an evaluation metric that considered the costs of screening, diagnosis, treatment, and diagnostic errors, allowing us to investigate the benefits of algorithmic diagnostic screening and facilitate the development of more clinically relevant algorithms. We received 779 algorithms from 87 teams during the Challenge, resulting in 53 working codebases and publications that represented a diversity of approaches to detecting heart murmurs and identifying clinical outcomes from heart sound recordings.

Background & ObjectiveDigital Service Providers have come up with certain Digital Well-Being Features as a step towards tackling harmful effects of screen overuse on physical and mental health. However, the awareness and use of the same remains scant. Our objective was to assess the knowledge, attitudes and practices regarding Digital Well-being features in the adult population of Maharashtra, India and the associations and correlations of the practice of using these features with screen time and degree of screen addiction. MethodsA cross-sectional online questionnaire-based study was conducted among 335 participants who were selected using quota sampling and were administered a Smartphone Addiction Scale and a self-designed questionnaire. ResultsKnowledge attitudes were good and total of 65.4% participants were digital wellbeing feature users. Correlation of digital hygiene score and digital wellbeing score was found neither with addiction nor with the average screen time. Interpretation & ConclusionKnowledge, attitude and practices regarding Digital Well-Being features were adequate among the urban population of Maharashtra. However, their use was not found to be associated with reduced screen time or a low screen addiction score. With further development and standardization, these features can be a useful tool for prevention of screen overuse and addiction.

IntroductionThe adoption of advanced reasoning models, such as ChatGPT O1 and DeepSeek-R1, represents a pivotal step forward in clinical decision support, particularly in pediatrics. ChatGPT O1 employs "chain-of-thought reasoning" (CoT) to enhance structured problem-solving, while DeepSeek-R1 introduces self-reflection capabilities through reinforcement learning. This study aimed to evaluate the diagnostic accuracy and clinical utility of these models in pediatric scenarios using the MedQA dataset. Materials and MethodsA total of 500 multiple-choice pediatric questions from the MedQA dataset were presented to ChatGPT O1 and DeepSeek-R1. Each question included four or more options, with one correct answer. The models were evaluated under uniform conditions, with performance metrics including accuracy, Cohens Kappa, and chi-square tests applied to assess agreement and statistical significance. Responses were analyzed to determine the models effectiveness in addressing clinical questions. ResultsChatGPT O1 achieved a diagnostic accuracy of 92.8%, significantly outperforming DeepSeek-R1, which scored 87.0% (p < 0.00001). The CoT reasoning technique used by ChatGPT O1 allowed for more structured and reliable responses, reducing the risk of errors. Conversely, DeepSeek-R1, while slightly less accurate, demonstrated superior accessibility and adaptability due to its open-source nature and emerging self-reflection capabilities. Cohens Kappa (K=0.20) indicated low agreement between the models, reflecting their distinct reasoning strategies. ConclusionsThis study highlights the strengths of ChatGPT O1 in providing accurate and coherent clinical reasoning, making it highly suitable for critical pediatric scenarios. DeepSeek-R1, with its flexibility and accessibility, remains a valuable tool in resource-limited settings. Combining these models in an ensemble system could leverage their complementary strengths, optimizing decision support in diverse clinical contexts. Further research is warranted to explore their integration into multidisciplinary care teams and their application in real-world clinical settings.

ProblemRapid and unplanned urbanization in LMICs exacerbates health inequities, particularly among the urban poor living in slums. In Madhya Pradesh (MP), India, the state government introduced Sanjeevani clinics to provide free, quality healthcare to these populations. This study highlights lessons from digitizing these clinics and making them AI-ready with an open-source AI framework. ApproachA multisectoral collaboration between the state government, implementation, and academic partners led to their successful digitization and AI-readiness of Sanjeevani clinics. An innovative, three-tablet task-shifting model and Smart Clinic Application reduced the data entry burden. ChikitsaChakra, an open-source, probabilistic decision support framework with a voice interface was developed to glean insights from data. Local settingMadhya Pradesh has a population of over 84 million, with approximately 24 million (29%) living in urban settings. Of these, 6.7 million (28%) resided in urban weak economic clusters. Prior to the introduction of Sanjeevani clinics, the population was heavily reliant on private healthcare, incurring significant out-of-pocket expenditures (OOPE). Relevant changesDigitization created an evidence base for over 2.6 million primary care consultations, saving over $33 million in cumulative OOPE while also revealing patterns in antibiotic prescribing. Lessons learnedThe Sanjeevani clinics" success hinged on multisectoral collaboration among government bodies, nonprofits, academia, and private partners, crucial for digitalization and AI-readiness. The innovative three-tablet model and application improved data quality, reducing healthcare workers" burdens and enhancing monitoring and evaluation. ChikitsaChakra provided essential insights.

The costs of participating in training programs that rely on video conferencing vary by mechanics of use and the specific platform. We proposed practical solutions to limiting costs in low resource settings with the use of video conferencing calls. Scenarios in which facilitators have their video on and expect learners to participate with continuous video result in the greatest data burden, while use of intermittent video by both facilitator and learners can significantly lower data use, and thus costs. The choice of a platform also impacts teleprogramming, with creative options for use of lower cost platforms to reduce participant and training organization costs. These might include sharing educational content or video via chat groups and limiting conference to audio alone. In the context of COVID-19 where virtual meetings have become prevalent, it is critical that data burden is considered by program directors and funders. Looking forward, hybrid training that includes virtual and in-person training will likely become the norm in global health settings, but achieving this model will still require thoughtful consideration of data costs. Further, our findings are relevant to many other fields and advocate for evaluation of costs and data burden along with the growing use of teleprogramming in these settings. Author SummaryThe COVID-19 pandemic restricted travel and in-person gatherings. These restrictions also impacted access to important in-person training programs for healthcare workers, especially in areas where financial resources are limited. However, one positive impact of the pandemic has been improved access and experience with video conferencing tools (like Zoom) for many healthcare training programs. These video conferencing tools provide a way to complete essential training when in-person options may be limited. However, video-based training can have significant costs including internet data costs, mobile device costs, and healthcare worker professional time. Our study examined the data costs associated with video conferencing using several video conferencing applications across 5,554 mobile data plans in 228 countries. We found that costs are highest when trainers have their video on and learners participate with video on throughout the training. Intermittent video use by both trainers and learners can significantly lower costs. We also found that using lower-cost video-conferencing tools may help reduce costs. Additionally, there are training methods that can reduce costs including sharing educational content or video via chat groups and limiting conferences to audio only. Virtual training is a powerful and common tool in healthcare settings, but it is essential to consider costs, especially in areas with limited resources.