PLOS digital health最新文献

Despite rising use of machine learning (ML) methods to detect depression within social media data, few are developed with and for adolescents. This is unfortunate, because adolescents may be more likely than adults to experience somatic than emotional symptoms and may be less likely to express emotions on social media. Accordingly, ML methods that focus on emotional symptoms may undercount adolescents at risk for depression. As a step toward developing an adolescent-centered ML method, we co-developed an interview guide with Latino adolescents to understand 1) social media norms for expressing somatic and emotional symptoms; and 2) identify potential signals of each. For the latter, we adopted a novel approach of asking interviewees to take on the "human classifier" role and tell us what they look for within social media data. Using framework analysis on 43 interviews with Latino adolescents, we find evidence suggesting norms prescribe more strongly against conveying emotional symptoms than somatic symptoms on social media. Additionally, rather than literal statements conveying they are experiencing depression, adolescents appear to use audiovisual cues to signal emotional symptoms and posting behavior (time of post, posting less) for somatic symptoms. Accordingly, norms may hinder opportunities for leveraging social media data to detect depression among adolescents, particularly when using ML methods that search for literal statements of depression or signals of emotional symptoms. Because peers tend to recognize depression in an adolescent earlier than medical experts, these findings suggest the need to develop and validate ML methods that incorporate a set of signals for somatic symptoms, particularly audiovisual cues and posting behavior. We discuss the benefits of "centering at the margins," which is focusing on a population that is understudied within this domain, and the need for ML methods developed with adolescent input.

Short message service text reminders (SMS nudges) aimed to help vulnerable patient populations remember to fill their prescriptions are becoming more common but accurately measuring their effects on improving prescription fill and readmission rates remains challenging. Patients who presented to the emergency department (ED) with a primary diagnosis of congestive heart failure (CHF) were included in the study. We conducted a retrospective cohort study of CHF patients who did and did not interact with SMS nudges, then matched patients who were prescribed medications at any point in their hospital visit with records of subsequent prescription fills. Patients that interacted with SMS nudges had 19% higher odds of filling prescriptions overall (1.19 OR (95% CI: 1.15 - 1.24), p < 0.001) and 6% lower odds of being readmitted to the hospital (0.94 (95% CI: 0.9 - 0.99), p = 0.009) than patients who did not interact with SMS nudges. Interactive SMS nudges via a novel tool may improve prescription fill rates across multiple groups of CHF patients, and contribute to a reduction in readmissions.

[This corrects the article DOI: 10.1371/journal.pdig.0001026.].

Digital food marketing to youth is concerning given its widespread reach, engagement strategies and influence on lifelong food behaviours. Nonethless, little is known about youth's engagement (i.e., liking/sharing/following food companies on social media, having food company/restaurant/delivery service apps downloaded) with food companies via digital media, particularly in Canada. This study examined whether youth's digital engagement with food companies differed by sociodemographic characteristics in Canada. An observational cross-sectional online survey was conducted in 2023 among 1162 Canadian children (aged 10-12 years) and adolescents (13-17 years). Participants self-reported their sociodemographic information and engagement with food companies via digital media. Descriptive analyses and logistic regression models examined differences in engagement by gender, age group, race/ethnicity and income adequacy. Among all participants, 20.9% reported having liked, shared, or followed food/restaurant companies on social media, 23.1% had food/restaurant company apps on their smartphones, and 16.6% had apps for food delivery services. White participants and youth from medium income adequacy households had lower odds of having liked/shared/followed food companies on social media than racial/ethnic minority group participants (OR: 0.59; 95% CI: 0.43, 0.80) and those from low income adequacy households (OR: 0.57; 95% CI: 0.41, 0.80), respectively. Children and White participants had lower odds of reporting food company apps on their smartphones than adolescents (OR: 0.54; 95% CI: 0.41, 0.72) and racial/ethnic minority group participants (OR: 0.48; 95% CI: 0.35, 0.64), respectively. Children and White participants also had lower odds of reporting food delivery service apps on their smartphones than adolescents (OR: 0.52; 95% CI: 0.38, 0.72) and racial/ethnic minority group participants (OR: 0.32; 95% CI: 0.23, 0.44), respectively. No significant differences were observed between genders. Overall, many Canadian youth are engaging with food companies via digital media. Government-led food marketing regulations that extend to social media and food company and delivery service apps are warranted.

The integration of artificial intelligence (AI) into health data research promises to transform precision medicine, especially by managing complex and chronic conditions like hypertension through decision support. Yet health AI also furthers surveillance, with serious ethical and social impact. Nevertheless, surveillance in health, in particular data-AI research and innovation, is understudied. This paper provides a conceptual analysis of health data-AI surveillance using the Hypermarker research project as a case study. We trace the evolution of surveillance within medicine, public health, data-driven research, and the proliferation of digital health technologies, before examining how the development of AI technologies amplifies and transforms these existing practices. We analyse health data-AI surveillance's implications of pervasiveness and unobtrusiveness, hypercollection and function creep, hypervisibility and profiling, informational power, and the formation of a surveillant assemblage, followed by an assessment of the safeguards and measures implemented by the Hypermarker project. Our analysis exposes several key challenges for responsible surveillance practices in health data-AI research: strengthening trustworthiness through fairness and equity, ensuring accountability through transparency, and fostering public control and oversight. To this end, we recommend advancing responsible governance by implementing arrangements such as community advisory panels, independent review boards and oversight bodies, data-AI justice frameworks and dialogues, transparency dashboards and public AI portals, stewardship committees, accountability assemblies, and open oversight cycles.

[This corrects the article DOI: 10.1371/journal.pdig.0000988.].

Autism spectrum disorder (ASD) is a neurological and developmental disorder that manifests in social and behavioral deficits. The onset of symptoms may begin in early childhood, but diagnosis is often subjective, and scores can vary between specialists. Several studies suggest that diffusion tensor imaging (DTI)-derived indicators of anisotropy in water diffusion at microstructural level could be biomarkers for this disorder. Emerging advances in neuroimaging and machine learning can provide a fast and objective alternative for its early diagnosis. We propose and evaluate a machine-learning (ML)-powered computer-aided diagnosis (CAD) system for the detection of ASD from DTI. For the development and validation of the system we used the ABIDE II database (n = 150). The system involves processing the raw DTI to obtain fractional anisotropy (FA), mean diffusivity (MD), radial diffusivity (RD) and axial diffusivity (AD) in 25 ASD-relevant regions of interest defined in the JHU ICBM-DTI-81 White-Matter Labeled Atlas to train a ML binary classifier. We evaluated the use of support vector machine (SVM) with various kernels and random forest (RF) optimized for computational efficiency. The best configuration, which used RF, had a sensitivity of 100%, accuracy of 95.65%, precision of 91.67%, and a specificity of 91.67%. An external test yielded 94.73% sensitivity, 97.37% accuracy, and 100% in precision and specificity. Results in this small sample show the generalization power of the best model, and the utility of carefully leveraging imaging information with clinical knowledge on relevant white matter regions commonly affected by ASD to design a CAD system for ASD.

We performed a prospective, longitudinal investigation to determine whether magnetic resonance imaging (MRI)-based radiomic features from thigh intramuscular fat (IMF) can predict future hamstring strain injury (HSI). Further, we sought to determine if muscle imbalance or injury profile along with radiomics could increase prediction accuracy. This study analyzed IDEAL MRI scans of 93 professional American football players (9 injured, 84 uninjured). Radiomic features relating to textural patterns of IMF were extracted from bilateral hamstring and quadriceps muscles. Feature selection identified non-correlated features that were more strongly associated with future HSI. The K-nearest neighbor classifier was employed to assess the performance of the following models: radiomics of hamstring IMF [Formula: see text] and quadriceps IMF [Formula: see text] muscle imbalance features (Mb) and injury profile features (Mi), as also integrated models for Mr, Mb and [Formula: see text], and integrated Mr and Mb (Mr+b) where [Formula: see text] [Formula: see text] (area under the curve (AUC)=0.79; 95%CI:0.78-0.79) significantly outperformed [Formula: see text] (AUC = 0.69; 95% CI: 0.68-0.70), [Formula: see text] (AUC = 0.74; 95% CI: 0.73-0.75), [Formula: see text] (AUC = 0.68; 95% CI: 0.67-0.69), Mi (AUC = 0.68; 95% CI: 0.68-0.69) as well as Mb (AUC = 0.64; 95% CI: 0.63-0.65). The results indicate that future HSI can be predicted when incorporating radiomics features from hamstrings IMF with muscle imbalance and injury profile data. These novel findings merit further validation in a larger population, one that includes populations of injured and uninjured participants, a limitation acknowledged in current study. This approach could inform future strategies to identify factors to mitigate the risk of HSI not just in elite male athletes but also in athletes of both sexes and any level of participation.

Artificial intelligence (AI) and large language models (LLMs), when combined with human expertise in collaborative intelligence (CI), can enhance medical decision-making, reduce bias in guideline development, and support precision care. New obesity management medications (OMMs) such as GLP-1 receptor agonists and dual incretin mimetics complement metabolic bariatric surgery but currently lack clear integration strategies. To address this gap, IFSO released consensus guidelines in 2024. This study evaluates their robustness by comparing expert recommendations with LLM outputs, highlighting the role of AI in assessment and strengthening clinical consensus. Thirty-one IFSO consensus statements were tested across eleven advanced LLMs on June 1, 2025. Models received standardized prompts that required binary "AGREE" or "DISAGREE" outputs, supported by brief, evidence-based rationales. Individual responses were aggregated to form an overall "LLM consensus," and mean percentage agreement was calculated against the original IFSO expert grades-Fleiss' κappa quantified inter-model reliability beyond chance. Incorporating the AI responses led to shifts in the consensus grade for 2 of the 31 statements. One statement originally rated A + was downgraded to A after some LLMs' outputs indicated disagreement, citing nuanced evidence on pre- and post-MBS OMM use and comparative effectiveness. One statement on combining OMMs with endoscopic therapies was upgraded from C to B due to unanimous support from the LLM. The remaining 29 statements maintained their original grades, demonstrating strong overall alignment between LLM outputs and expert consensus. Overall concordance between LLMs and experts was 93%, with substantial inter-model agreement(κ = 0.81 [95% CI 0.74-0.87]). Integrating AI, especially LLMs, into collaborative intelligence frameworks strengthens clinical consensus when evidence is limited. This study shows that concordance between LLMs outputs and expert consensus should not be taken as evidence of objectivity; rather, it may simply reflect overlap between the published evidence base and the model's training data or retrieval sources.

Electronic Health Records (EHRs) have improved access to patient information but substantially increased clinicians' documentation workload. Large Language Models (LLMs) offer a potential means to reduce this burden, yet real-world deployments in live hospital systems remain limited. We implemented a secure, GDPR-compliant, on-premises LLM assistant integrated into the Epic EHR at a European university hospital. The system uses Qwen3-235B with Retrieval Augmented Generation to deliver context-aware answers drawing on structured patient data, internal and regional clinical documents, and medical literature. A one-month pilot with 28 physicians across nine specialties demonstrated high engagement, with 64% of participants using the assistant daily and generating 482 multi-turn conversations. The most common tasks were summarization, information retrieval, and note drafting, which together accounted for over 70% of interactions. Following the pilot, the system was deployed hospital-wide and adopted by 1,028 users who generated 14,910 conversations over five months, with more than half of clinicians using it at least weekly. Usage remained concentrated on information access and documentation support, indicating stable incorporation into everyday clinical workflows. Feedback volume decreased compared with the pilot, suggesting that routine use diminishes voluntary reporting and underscoring the need for complementary automated monitoring strategies. These findings demonstrate that large-scale integration of LLMs into clinical environments is technically feasible and can achieve sustained use when embedded directly within EHR workflows and governed by strong privacy safeguards. The observed patterns of engagement show that such systems can deliver consistent value in information retrieval and documentation, providing a replicable model for responsible clinical AI deployment.