Frontiers in digital health最新文献

Background: Screening methods are essential for detection of numerous pathologies. Chest x-ray radiography (CXR) is the most widely used screening modality. During the screening, radiologists primarily examine normal radiographs, which results in a substantial workload and an increased risk of errors. There is an increasing necessity to automate radiological screening in order to facilitate the autonomous sorting of normal studies.

Objective: We aimed to evaluate the capabilities of artificial intelligence (AI) techniques for the autonomous CXRs triage and to assess their potential for integration into routine clinical workflow.

Methods: A rapid evidence assessment methodology was employed to conduct this review. Literature searches were performed using relevant keywords across PubMed, arXiv, medRxiv, Elibrary, and Google Scholar covering the period from 2019 to 2025. Inclusion criteria comprised large-scale studies addressing multiple pathologies and providing abstracts in English. Meta-analysis was conducted using confusion matrices derived from reported diagnostic performance metrics in the selected studies. Methodological quality and the overall quality of evidence were assessed using a combination of QUADAS-2, QUADAS-CAD, and GRADE frameworks.

Results: Out of 327 records, 11 studies met the inclusion criteria. Among these, three studies analyzed datasets reflecting the real-world prevalence of pathologies. Three studies included very large cohorts exceeding 500,000 CXRs, whereas the remaining studies used considerably smaller samples. The proportion of autonomously triaged CXRs ranged from 15.0% to 99.8%, with a weighted average of 42.3% across all publications. Notably, in a study conducted under real-world clinical conditions on continuous data flow, this proportion was 54.8%. Sensitivity was 97.8% (95% CI: 94.8%-99.1%), and specificity was 94.8% (95% CI: 53.0%-99.7%). Fifty-five percent of the studies were classified as having a low risk of bias. Primarily, elevated risk of bias and heterogeneity of results were attributed to variability in sample selection criteria and reference standard evaluation.

Conclusions: Modern AI systems for autonomous triage of CXRs are ready to be implemented in clinical practice. AI-driven screening can reduce radiologists' workload, decrease sorting errors and lower the costs associated with screening programs. However, implementation is often hindered by regulatory and legislative barriers. Consequently, comprehensive clinical trials conducted under real-world conditions remain scarce.

By leveraging everyday technologies such as mobile apps, wearables, and AI-enabled tools, digital health interventions (DHIs) offer new pathways to integrate self-management and intervention programs into the fabric of daily life, while bridging gaps in care through continuous, context-aware support. Yet many tools underperform clinically because digital engagement ("screen time") is conflated with impact, while behavioral science is retrofitted, if applied at all. We propose the ENGAGE Framework: a cyclical, six-step model of precision engagement that integrates user needs, behavioral science, and adaptive personalization to transform initial curiosity into sustained real-world habits. By leveraging available data, users can be segmented according to their need (Step 1: Enroll & Segment), targeted with the most relevant and engaging message to increase micro-engagement (Step 2: Nudge & Hook), and persuaded to engage in real-world health behavior change (Step 3: Guide Behavior). From this macro-engagement step, additional core behavioral science principles are used to reinforce the real-world behaviors long enough to positively impact health outcomes (Step 4: Anchor Habits), while measuring progress (Step 5: Generate Evidence) to inform adaptive and optimized engagement strategies (Step 6: Expand & Evolve with AI) for tailored interventions and communications based on user characteristics, context, and clinical data for both new and existing users. Each step of the ENGAGE Framework maps to evidence-based techniques, implementation tactics (e.g., integration pathways and operational deployment strategies), and metrics that help translate superficial engagement into long-lasting behavior change and measurable clinical outcomes. We synthesize relevant engagement literature, identify gaps and challenges (e.g., measurement heterogeneity, lack of focus on macro-engagement, product development challenges, ecosystem barriers), and offer a practical checklist for innovators. By focusing on who needs what support, when and why, ENGAGE aims to help DHI developers and researchers design interventions that are effective, equitable, and empirically testable.

Digital biomarkers are revolutionizing medicine in ways that were unimaginable a few years ago. Consequently, precision medicine approaches now realistically can promise personalization, i.e., the right treatments for the right patients at the right time, including earlier, targeted interventions which lead to a major paradigm shift in how medicine is practiced from reactive to preventive action. Although the scientific evidence is clear on the power of digital biomarkers, there is an unmet need for translating these findings into actionable insights in clinical practice. In this paper, we focus on Alzheimer's disease and related dementias (ADRD), and how digital biomarkers could empower clinical decision making in its preclinical stages. We argue that a new all-encompassing score is needed, akin to a BrainHealth Index linked to the established and validated risk stratifications frameworks and is directed at the prevention of ADRD. Specifically, we propose the new concept "Digital Neuro Fingerprint (DNF)", built with simultaneous collection of multimodal digital biomarkers (speech, gait, eye movements etc.) from smartphone based augmented reality or virtual reality while an individual is immersed in activities of daily living. Fusing the captured multimodal digital biomarkers, data is automatically analyzed with custom combinations of machine- and deep-learning approaches and enhanced with explainable artificial intelligence (XAI) and uncertainty quantifications. We argue that DNF is useful for capturing ADRD progression and should supersede the biomarkers that are invasive and expensive to obtain, offering a sensitive and highly specific score that measures meaningful aspects of health for the patients in high-frequency intervals.

Assistive technologies can enhance safety, independence, and quality of life for people with blindness and low vision. Despite their benefits, abandonment of these technologies remains widespread, and recent research on this issue is limited. In this Perspective article, we draw on both professional experiences and relevant scientific literature to examine adoption and abandonment in the context of new artificial intelligence-powered applications. We highlight risks arising from misaligned design, inconsistent industry support, and inadequate user training. We synthesize existing knowledge on factors that influence abandonment and propose three priorities to realign assistive technology development: participatory and transdisciplinary research, integrated technology ecosystems, and socially supported engagement. Taken collectively, these priorities ensure that emerging assistive technologies better align with the needs of people with blindness and low vision, promoting lasting adoption rather than abandonment.

Background: Audiovisual tools are increasingly used in healthcare to improve patient education and engagement. However, few studies, particularly in dermatology, have evaluated their effectiveness in enhancing patient understanding during the informed consent process. This study aimed to compare the effectiveness of a motion-graphic educational video with conventional verbal consent for patients undergoing platelet-rich plasma (PRP) therapy for androgenetic alopecia (AGA).

Methods: In this randomized controlled trial, participants aged 18-55 years with AGA were recruited at the Dermatology Clinic, Walailak University Hospital, between December 2024 and March 2025. Participants were randomized to receive informed consent through either an educational video (Group A) or a conventional verbal explanation (Group B). Pre- and post-intervention knowledge and anxiety levels were assessed, and satisfaction was evaluated in Group A.

Results: Thirty-four participants completed the study (73.5% male; median age: 39.5 years, IQR: 23.0). Median baseline knowledge and anxiety scores were 0.0 (IQR: 2.0) and 6.0 (IQR: 3.0), respectively. Post-intervention knowledge scores increased significantly in both groups (Group A: 9.0, IQR: 1.0; Group B: 7.0, IQR: 2.0; p < 0.001), with a greater knowledge gain in Group A (8.0, IQR: 3.0) compared to Group B (6.0, IQR: 2.0; p = 0.009). Anxiety scores remained unchanged in both groups. Group A reported a high usefulness score for the video (median, 10.0; IQR, 1.0). No significant correlations were found between demographic factors and baseline knowledge or anxiety.

Conclusions: The motion-graphic educational video improved patient knowledge compared with conventional verbal explanations, without reducing anxiety. Participants reported high satisfaction, supporting the use of audiovisual media as an effective adjunct to the informed consent process.

Clinical trial registration: https://www.thaiclinicaltrials.org/show/TCTR20241222004, identifier TCTR20241222004.

Background: Knee osteoarthritis (KOA) is a chronic joint disorder that significantly affects the quality of life in the older adult. It is primarily characterized by notable knee pain following activity, which typically alleviates with rest. With the rapid growth of the internet, people increasingly rely on social media to obtain health-related information. Short-form videos, as an emerging format, play an important role in information dissemination. TikTok is currently the world's most downloaded application platform primarily dedicated to short-form video content. Against this backdrop, we observed a substantial number of KOA-related videos on TikTok, the quality and reliability of which have not yet been systematically evaluated.

Objective: To assess the quality and reliability of KOA-related videos available on the domestic TikTok platform.

Methods: A total of 100 KOA-related videos were retrieved and screened from TikTok. Basic metadata were extracted, and video content and format were categorized through coding. The source of each video was also documented. Two independent raters evaluated video quality using the DISCERN instrument, the Journal of the American Medical Association (JAMA) benchmark criteria, and the Global Quality Score (GQS).

Results: Of 100 analyzed videos, 96 were posted by medical staff and 4 by science communicators. Eighty videos were audio-based (41% outpatient daily, 39% general science popularization), with others using graph-text formats. The video content is divided into 7 groups: disease prevention, diagnosis, symptoms, description, life-style and therapy, among which the video related to disease description is the most. The average DISCERN, JAMA and GQS scores of videos were 36.29, 1.24 and 2.45, respectively, and the overall quality was low. Further analysis shows that there are significant differences in video quality between science communicators and medical staff. The number of likes, comments, collections, and shares are strongly positively correlated with each other, and they are weakly positively correlated with the number of upload days and DISCERN scores.

Conclusion: KOA-related content on TikTok demonstrates concerning quality limitations, with significant variation across source types. Given TikTok's expanding influence in health communication, urgent improvements and standardized quality control measures are needed.

Introduction: Hereditary angioedema (HAE) is a rare disease imposing a significant quality of life burden. Affect monitoring via Ecological Momentary Assessment (EMA) could offer personalized psychological support by collecting repeated, ecological data in real-life, overcoming the limitations of traditional methods. This study assessed the feasibility and acceptability of an EMA protocol for affect monitoring in HAE patients vs. healthy controls (CTR).

Methods: HAE patients and CTR were recruited for a 16-week EMA study. Participants received weekly EMA surveys assessing affect via REDCap™. Feasibility was evaluated through recruitment, response, and completion rates. Acceptability was assessed via a post-study questionnaire through a visual analogue scale ranging from 1 to 100.

Results: Twenty-eight Caucasian subjects were contacted, 12 HAE [median age: 50 (22) years, 5 males] and 14 CTR [age: 30 (32) years, 6 males] agreed to participate, resulting in a recruitment rate of 93%. Response and completion rates were ≥92% and ≥96% respectively in both groups. Completion time was brief and did not differ between groups [HAE: 1' 28″ (29″) vs. CTR: 1' 15' (15″), P = 0.274]. The protocol was considered acceptable by both groups [HAE: rate 83.5 (18.8) vs. CTR: 72.0 (13.0), p = 0.27] with HAE rating the experience as helpful [79 (39.8)] and thought-provoking [67 (33)].

Conclusion: EMA is a highly feasible and acceptable method for affect monitoring in HAE. The presence of a rare disease does not appear to be a barrier to its application, supporting its use in this clinical setting.

Background: Valvular heart disease and heart failure are major global health burdens, yet access to comprehensive echocardiography is often limited, particularly in resource-constrained settings. Artificial intelligence (AI) may enable rapid, point-of-care cardiac assessment using simplified imaging protocols.

Objectives: To evaluate whether a deep learning model can accurately detect significant valvular and ventricular dysfunction using only a single two-dimensional apical four-chamber echocardiographic view, including images acquired by non-cardiologists with handheld ultrasound devices.

Methods: We retrospectively analyzed 120,127 echocardiographic studies from a tertiary medical center to train and validate a deep learning model for identifying moderate-or-greater mitral or tricuspid regurgitation, right ventricular dysfunction, and reduced left ventricular ejection fraction (≤40%). A prospective cohort of 209 patients underwent handheld point-of-care cardiac ultrasound performed by non-cardiologist physicians, with same-hospitalization comprehensive echocardiography as the reference standard.

Results: In retrospective testing, model areas under the curve (AUCs) were 0.883 for mitral regurgitation, 0.913 for tricuspid regurgitation, 0.940 for right ventricular dysfunction, and 0.982 for reduced ejection fraction. In the prospective cohort, AUCs were 0.72, 0.87, 0.95, and 0.97 for the same respective targets.

Conclusions: A single-view deep learning model demonstrated strong diagnostic accuracy for detecting significant valvular and ventricular dysfunction across both standard and handheld ultrasound acquisitions. This approach may facilitate rapid, scalable cardiac function screening by non-cardiologists in diverse clinical environments.

Clinical trial registration: identifier NCT05455541.

Introduction: The classification of lesion types in Digital Breast Tomosynthesis (DBT) images is crucial for the early diagnosis of breast cancer. However, the task remains challenging due to the complexity of breast tissue and the subtle nature of lesions. To alleviate radiologists' workload, computer-aided diagnosis (CAD) systems have been developed. The breast lesion regions vary in size and complexity, which leads to performance degradation.

Methods: To tackle this problem, we propose a novel DBT Dual-Net architecture comprising two complementary neural network branches that extract both low-level and high-level features. By fusing different-level feature representations, the model can better capture subtle structure. Furthermore, we introduced a pseudo-color enhancement procedure to improve the visibility of lesions on DBT. Moreover, most existing DBT classification studies rely on two-dimensional (2D) slice-level analysis, neglecting the rich three-dimensional (3D) spatial context within DBT volumes. To address this limitation, we used majority voting for image-level classification from predictions across slices.

Results: We evaluated our method on a public DBT dataset and compared its performance with several existing classification approaches. The results showed that our method outperforms baseline models.

Discussion: The use of pseudo-color enhancement, extracting high and low-level features and inter-slice majority voting proposed method is effective for lesion classification in DBT. The code is available at https://github.com/xiaoerlaigeid/DBT-Dual-Net.