Mobile health (mHealth) applications are increasingly integral to healthcare delivery, yet traditional randomized trials face practical challenges in evaluating these dynamic tools. Real-world data (RWD), collected during routine app use, offers a complementary pathway to real-world evidence (RWE) that may reflect how mHealth applications are used in everyday settings. We conducted a scoping review to map how naturally emerging RWD are currently used in peer-reviewed studies to evaluate patient-facing mHealth applications. We systematically searched PubMed, Scopus, and Web of Science (January 2007-November 2024) and extracted data on application type, RWD characteristics and study design aspects. Study-level, design-centred evidence levels for RWD-based effectiveness evaluations were assigned using a combined Oxford Centre for Evidence-Based Medicine and FDA RWE framework. Seventy-two studies evaluating 61 unique mHealth applications were included. Most studies focused on mental health or metabolic conditions and relied predominantly on data actively reported by users, often via in-app surveys, with comparatively limited use of device-generated data or integration with system-generated data such as clinical or claims data. Single-group pre-post approaches were most frequently observed, while only a minority employed comparative observational, quasi-experimental, or randomized designs. These findings illustrate current patterns in the use of RWD for mHealth evaluation and highlight both opportunities and constraints in longitudinal and comparative assessments of mHealth applications in real-world contexts.

Atrial fibrillation (AF) is the most common sustained arrhythmia and a leading cause of ischemic stroke. Existing risk scores, such as CHA₂DS₂-VASc, offer limited predictive accuracy and fail to capture complex clinical patterns. To improve generalizability and clinical utility, we developed and externally validated clinically interpretable machine learning models using only age, comorbidities, and medication use to predict 1-year stroke risk in patients with newly diagnosed AF. Both logistic regression (LR) and Platt-calibrated extreme gradient boosting (XGB) models achieved high discrimination in internal (AUCs = 0.915 and 0.914) and external validation cohorts (AUCs = 0.877-0.886), significantly outperforming CHA₂DS₂-VASc (AUCs = 0.614-0.621; p < 0.001). Calibration curves and decision curve analysis confirmed strong clinical utility. Long-term follow-up demonstrated superior risk stratification and treatment responsiveness in LR-defined high-risk groups. These models provide accurate, individualized stroke risk estimates to guide direct oral anticoagulant (DOAC) initiation in real-world hospital settings.

Digital health technologies (DHTs) such as wearables, smartphones, and connected devices have immense potential for supporting self-care in chronic disease management, yet engagement remains highly variable and is often measured by usage metrics alone. Guided by the AIM-ACT framework, this mixed-methods study examined how psychosocial and contextual factors shape DHT engagement among 146 adults with heart failure who completed a 6-month digital intervention involving multiple devices (ClinicalTrials.gov ID NCT05056129). Using k-medoids clustering of survey, ecological momentary assessment, and device-log data, we identified three distinct psychosocial phenotypes-Challenged Survivors, Activated Learners, Engaged Self-Regulators-reflecting differences in motivation, psychosocial resources, and interaction with DHTs. Qualitative interviews with 26 participants contextualized these phenotypes and revealed mechanisms linking psychosocial traits to DHT engagement quality. Findings underscore the value of psychosocial phenotyping for understanding heterogeneity in DHT engagement to inform the design of adaptive and equitable digital interventions that remain effective as they scale across diverse populations.

Concept mapping links free-text descriptions to standardized concepts in biomedical terminologies and ontologies. Existing approaches fall into two main categories: rule-based systems, which are interpretable but struggle with ambiguity and scalability, and learning-based methods, which leverage contextual signals but perform poorly on long-tail concepts due to limited training data and offer limited explainability. To address these limitations, we propose Medical Concept Mapping (MCM), a novel agentic workflow that uses language models to rephrase ambiguous mentions into explicit, standardized descriptions prior to concept linking. This reformulation substantially improves mapping accuracy for underrepresented and abbreviated concepts. Across multiple benchmarks, MedMentions, ST21pv, and MCN, MCM outperforms state-of-the-art baselines, including KrissBERT, SciSpaCy, and USAGI, achieving Recall@1 scores of 63.3, 60.0, and 67.9, respectively. On zero-shot abbreviated mentions, MCM maintains strong performance, exceeding baseline Recall@1 by up to 24.8 points. Human evaluation confirms the quality of LLM-generated expansions: 79.4% were rated as reasonable and useful, and GPT-OSS received the highest approval (85.1%). These results demonstrate that MCM enables more accurate, interpretable, and robust long-tail concept normalization for biomedical applications.

Magnetic Resonance Imaging is a critical imaging modality in clinical diagnosis and research, yet its complexity and heterogeneity hinder scalable, generalizable machine learning. Although foundation models have revolutionized language and vision tasks, their application to MRI remains constrained by data scarcity and narrow anatomical focus. We present Decipher-MR, a 3D MRI-specific vision-language foundation model trained on 200,000 MRI series from over 22,000 studies spanning diverse anatomical regions, sequences, and pathologies. Decipher-MR integrates self-supervised vision learning with report-guided text supervision to build robust representations for broad applications. To enable efficient use, Decipher-MR supports a modular design that enables tuning of lightweight, task-specific decoders attached to a frozen pretrained encoder. Following this setting, we evaluate Decipher-MR across disease classification, demographic prediction, anatomical localization, and cross-modal retrieval, demonstrating consistent improvements over existing foundation models and task-specific approaches. These results support Decipher-MR as a promising and reusable foundation for MRI-based AI, within the scope of the tasks and datasets evaluated.

Physicians frequently confront time-sensitive decisions under uncertain conditions, necessitating reliable tools for forecasting clinical outcomes. Although clinical predictive models have the potential to assist in these critical decisions, their widespread adoption is hindered by complexities in data handling, model development, and integration into clinical workflows. This study introduces a novel framework (Hopkins LLM) leveraging structured electronic health records (EHRs) data to develop and deploy clinical large language models (LLMs) that act as multi-task-capable predictive engines to support clinically constrained decision-support tasks with minimal barriers to implementation. Employing the advanced LLaMA architecture, consisting of 7 billion parameters, our model was pre-trained on a comprehensive corpus and subsequently fine-tuned and tested on a dataset of 42,160 patients within Johns Hopkins Health System, addressing a spectrum of clinical and operational prediction tasks. We validated our model across three diverse external health systems and four key prediction tasks involving 1,329 patients, including 30-day all-cause readmissions, 90-day all-cause mortality, 30-day intensive care unit (ICU) admissions, and treatment recommendations. The proposed Hopkins-LLM framework achieved a mean area under the receiver operating characteristic curve (ROC-AUC) of 0.84 [0.82, 0.88], yielding a significant 0.28 advancement over zero-shot baseline LLMs (p<0.05). These findings underscore the promise of LLMs as unified, user-friendly clinical prediction systems, adept at reasoning across diverse data sources to enhance decision-making at the point of care.

High-definition transcranial alternating current stimulation (HD-tACS) is an emerging neuromodulatory intervention for major depressive disorder (MDD). Here, we evaluate the efficacy and safety of a portable, low-intensity (1.6 mA), 10 Hz HD-tACS protocol targeting the bilateral dorsolateral prefrontal cortex (DLPFC) in adults with MDD. This randomized clinical trial was conducted including 120 adult MDD participants. Participants underwent 20 sessions (5 sessions/week for 4 weeks) of either active 10 Hz HD-tACS (1.6 mA) or sham stimulation a ratio of 1:1 and were followed for an additional 4 weeks. The primary outcome was the percentage reduction in 17-item Hamilton Depression Rating Scale scores at week 4. The active group showed significantly greater improvement in depression than the sham group (48% ± 14% vs 24% ± 13%, p < 0.001), sustained through the 4-week follow-up period. Additionally, significant improvements in the active group were also found in anxiety, sleep efficiency, quality of life components of physical, vitality, social functioning, and mental health. Adverse events including two cases of new-onset hypomania occurred, but no serious adverse events were observed. Thus, this trial supported the efficacy and good tolerability of 10 Hz HD-tACS for patients with MDD. Further research requires more focus on the underlying neural mechanisms and the optimized stimulation protocols. Trial Registration: China Clinical Trial Registry Identifier: ChiCTR2500103570. ClinicalTrials.gov Identifier: NCT06891326.

Scoliosis is the most common developmental spinal deformity, but its genetic underpinnings remain only partially understood. To identify scoliosis-related loci, we utilized dual energy X-ray absorptiometry (DXA) scans from 57,588 individuals in the UK Biobank (UKB), and quantified spinal curvature using deep learning-based vertebral segmentation and landmarking to measure cumulative horizontal displacement. On a subset of 150 individuals, our automated image-derived curvature measurements showed a correlation of 0.83 with clinical Cobb angle assessments, supporting their validity as a proxy for scoliosis severity. To connect spinal curvature to genetics, we conducted a genome-wide association study (GWAS). Our quantitative imaging phenotype identified 2 novel loci associated with scoliosis in a European population. These loci are in SEM1/SHFM1 and on an lncRNA on chr 3 located between EDEM1 and GRM7. Genetic correlation analysis revealed significant overlap between our image-based GWAS and ICD-10-based GWAS in the UKB and the Biobank of Japan. We show that our quantitative GWAS identifies more genome-wide significant loci than a case-control scoliosis dataset with ten times the sample size. Our results illustrate the potential of quantitative imaging phenotypes to uncover genetic associations that are challenging to capture with medical records alone and identify new loci for functional follow-up.

The rising prevalence of vision-threatening retinal diseases poses a significant burden on the global healthcare systems. Though deep learning (DL) techniques offer promising avenues for improving diagnostic efficiency, data scarcity and imbalance issues persist in training robust diagnostic models, particularly for rare eye diseases. Here, we introduce EyeDiff, a generative foundation model capable of synthesizing lesion-preserving ophthalmic images from textual descriptions. Both objective metrics and expert human evaluations confirmed EyeDiff's ability to generate high-fidelity images across multiple imaging modalities, accurately reflecting textual descriptions of diverse retinal diseases and lesion types. By augmenting minority classes across 11 globally sourced datasets, EyeDiff consistently boosted the diagnostic accuracy for both common and rare eye diseases across different foundation model types, including modality-specific, multimodal and vision-language foundation models trained solely on real data. These results underscore EyeDiff's potential as a general-purpose text-to-image foundation model, offering a scalable and flexible approach to generate balanced, disease-relevant data for advancing retinal disease diagnosis.

Lengthy acquisition time remains a key bottleneck for the widespread use of MRI in clinics. While accelerated MRI can reduce scan duration, it often introduces increased noise, compromising image quality and diagnostic reliability. In this study, we present a unified deep learning-based denoising model for multi-organ accelerated MRI, designed to operate directly on reconstructed images from commercial MRI systems. Our model was trained on a prospectively collected, large-scale real-world dataset comprising 148,930 noisy-clean image pairs from six clinical centers and four major MRI vendors, spanning six organs and 96 MRI protocols. On a test set of 20,143 real-world image pairs, our model consistently outperforms state-of-the-art denoising methods. Importantly, downstream evaluation using tissue segmentation demonstrates a 7.05% improvement in Dice score across multiple organs compared to noisy images. The model further generalizes effectively to 46,870 external clinical images from four independent cohorts, highlighting its robustness across various scanners and acquisition protocols. To assess clinical utility, two experienced radiologists conducted blinded evaluations across multiple organs, focusing on overall image quality, diagnostic confidence, and disease diagnosis. The denoised images retained high visual fidelity and yielded diagnostic performance equivalent to clean images even with acceleration factor of 3× compared to clinical scanning setup, such that many acquisitions can be completed within one minute. This unified MRI denoising model holds great potential for various clinical applications.