Background: Systems based on large language models (LLMs), multimodal LLMs, and vision-language foundation models are increasingly being evaluated for medical report generation in imaging and related clinical workflows. Existing reviews have summarized technical architectures, radiology applications, readability, and benchmark performance, but clinical readiness remains uncertain because safety, human oversight, and workflow outcomes are sparsely and inconsistently reported. Objective: The aim of this study is to assess the effectiveness (expert acceptance and blinded preference), safety (clinically significant, omission, and commission errors), and workflow burden (reporting time, corrections, edit distance, and editing burden) of LLM-based medical report generation. Methods: We searched PubMed/MEDLINE, Embase, Web of Science Core Collection, Scopus, and the Cochrane Library for studies published from January 1, 2016, through May 15, 2026. Eligible studies evaluated LLMs, multimodal LLMs, or vision-language foundation models for image-to-report generation, impression generation from findings, report drafting, or structured reporting in imaging workflows. Two reviewers performed screening, extraction, risk-of-bias assessment, and Grading of Recommendations Assessment, Development, and Evaluation–informed narrative certainty assessment. Outcomes were clinically significant error rate, omission error rate, commission error rate, reporting time, edit burden, expert acceptance, and blinded expert preference. Meta-analysis was not performed because no comparable outcome had at least 2 studies with compatible task structure and analyzable data. Results: A total of 101 studies were included. Chest x-ray was the largest modality group (36 studies), followed by computed tomography, magnetic resonance imaging (MRI), ultrasound, endoscopy, pathology, ophthalmic, electrocardiographic, dental, and mixed-modality contexts. No study was judged at low risk of bias; 15 were moderate, 72 high, and 14 serious. Safety and workflow evidence remained heterogeneous and largely nonpoolable. In a chest x-ray study, AI report acceptance was similar to that of radiologist reports (6047/8580, 70.5% vs 6288/8580, 73.3%), but false-negative findings were slightly higher (1584/8580, 18.5% vs 1527/8580, 17.8%). In a clinician-collaboration chest x-ray study, AI reports were equivalent or preferred in 233 of 300 (77.7%) and 170 of 303 (56.1%) cases across 2 datasets; yet, clinically significant errors persisted. In a brain MRI study, AI assistance reduced reading time from 61 to 53 seconds, whereas impression drafting increased editing time and edit distance. Conclusions: This review shifts the synthesis from plausible report generation to clinically interpretable effectiveness, safety, and workflow effects. Expert acceptance and preference suggested assistive value in selected supervised settings, but these signals were limited by inconsistent reporting of clinically significant errors, omissions, commissions, and failed generations. Workflow effects were mixed, with some studies reporting shorter reading time or drafting support, and others reporting greater editing time or edit distance. The evidence remains too heterogeneous, biased, and sparse on case-level end points to support a pooled meta-analysis or autonomous clinical-readiness claims. Adoption should remain locally validated, clinician-supervised, and accompanied by standardized reporting of acceptance, preference, omissions, commissions, failed generations, reporting time, corrections, and editing burden. Trial Registration: PROSPERO CRD420261302844; https://www.crd.york.ac.uk/PROSPERO/view/CRD420261302844

Background: Postoperative outcome assessment is often based on discrete follow-up visits, limiting characterization of individual recovery trajectories, and the timely identification of adverse events (AEs). Longitudinal smartphone-based monitoring may overcome these limitations by enabling frequent, resource-efficient collection of patient-reported outcomes and complications throughout recovery. Such data may provide a more patient-centered understanding of the postoperative course and complement conventional clinical surveillance. Objective: This study aimed to evaluate the feasibility of smartphone-based longitudinal monitoring of quality of life, subjective well-being, and AEs after elective neurosurgery and compare postoperative recovery trajectories and agreement between patient- and clinician-reported AEs. Methods: This interim analysis of a prospective cohort study included adult patients undergoing elective lumbar decompression, lumbar fusion, supratentorial craniotomy, or infratentorial craniotomy at a Swiss tertiary referral center between June 2023 and January 2025. Participants used a smartphone app to longitudinally report subjective well-being (Subjective Well-Being Index; 0‐10), quality of life (EQ-5D-5L), and AEs for up to 1 year postoperatively. Complications were self-reported using the Therapy-Disability-Neurology (TDN) classification and retrospectively adjudicated by physicians. Descriptive analyses assessed data density, engagement, and concordance between patient- and clinician-reported events. Mixed-effects models were used to evaluate factors associated with postoperative well-being. Results: Of the 100 enrolled patients (median age 64.0, IQR 52.95‐71.6 years; n=45, 45% women), 86 (86%) provided postoperative data. During a median follow-up of 3.2 (IQR 0.2‐11.3) months, participants submitted 4354 longitudinal well-being entries. Patients reported 22 unique AEs, whereas physicians identified 44 AEs, with overlap for 9 (20.5%) events. Most physician-reported AEs were mild (30/44, 68.2%; TDN grade 1‐2), and no grade 4 or 5 events occurred. Patient-reported AEs primarily reflected symptomatic and functional impairments, whereas physician-reported events more often included clinically detected or subclinical findings. In mixed-effects models, time since surgery was associated with improved well-being, and no other factors were statistically significant. Conclusions: Smartphone-based postoperative monitoring was feasible in this elective neurosurgical cohort and generated dense longitudinal patient-reported data beyond routine follow-up. Patient and clinician AE reporting captured partly distinct aspects of postoperative recovery, suggesting that smartphone-based self-reporting may complement rather than replace clinical surveillance. Trial Registration: ClinicalTrials.gov NCT06352710; https://clinicaltrials.gov/study/NCT06352710

Background: Overdose and suicide due to nonprescribed fentanyl use have increased significantly, yet health care systems lack reliable methods to identify patients who use nonprescribed fentanyl. codes are inconsistent and do not specify nonprescribed fentanyl use. Objective: This study aimed to develop natural language processing approaches to identifying nonprescribed fentanyl use in electronic health record (EHR) documentation. Methods: This retrospective study included Veterans Health Administration patients seen between April 5, 2023, and December 23, 2024. A term list was developed to identify fentanyl-related mentions in clinical text, and 250-character snippets surrounding identified mentions were extracted. Veterans (n=3878) were randomly sampled from 5 predefined groups based on the presence of 1 of 4 terms (“fent,” “blues,” “M30s,” and “tranq”) in their EHR documentation. Physician annotators classified snippets into “nonprescribed fentanyl use,” “prescribed fentanyl use,” or “other,” with interannotator agreement evaluated using the mean pairwise Cohen κ. Cross-validation folds were constructed at the patient level between training and test sets. Penalized logistic regression, Bio-ClinicalBERT, Llama 3-8B, and Mistral-7B were trained on labeled data and compared. Model performance was evaluated using precision, recall, and -scores for each class, with a focus on the nonprescribed fentanyl use class as the primary label of clinical interest using bootstrapped 95% CIs. A fairness analysis and Shapley additive explanations analysis were performed using Bio-ClinicalBERT. External validation was performed using Bio-ClinicalBERT on an independent sample of 200 snippets, each representing a unique patient from January 2025 to June 2026, with precision reported as the primary validation metric. Results: Of 7389 snippets, 9.6% (n=709) were classified as “nonprescribed fentanyl use,” 40.3% (n=2981) were classified as “prescribed fentanyl use,” and 50% (n=3699) were classified as “other.” Interannotator agreement was high (κ=0.822). Llama 3-8B achieved the highest -score for nonprescribed fentanyl use (0.87, 95% CI 0.83-0.92), followed by Mistral-7B (0.80, 95% CI 0.75-0.84), Bio-ClinicalBERT (0.80, 95% CI 0.74-0.85), and penalized logistic regression (0.74, 95% CI 0.73-0.75). Performance was consistent across demographic subgroups, with lower performance for the nonprescribed fentanyl use class observed in female and Hispanic subgroups. Shapley additive explanations analysis revealed clinically meaningful discriminating terms for each class, although subword tokens required contextual interpretation. External validation of Bio-ClinicalBERT demonstrated a precision of 0.79 for nonprescribed fentanyl use. Conclusions: Natural language processing can identify nonprescribed fentanyl use in EHR documentation, although model performance for this class was lower than overall model performance, reflecting the clinical complexity of identifying nonprescribed use and the variable ways in which clinicians document this problem. This approach may support risk prediction and targeting of interventions to patients exposed to nonprescribed fentanyl.

Background: Psychiatric distress is common among patients with cancer; yet, systematic interview-based screening remains difficult to scale in routine clinical care. Large language models (LLMs) have shown promise as scalable tools for mental health assessment, but most existing evidence is derived from clinician-authored records, translated text, or proxy data. The performance characteristics, error patterns, and explanatory behaviors of contemporary LLMs when applied to authentic, non-English psychiatric interviews remain insufficiently characterized. Objective: This exploratory study evaluated how contemporary LLMs reproduced psycho-oncologists’ item-level symptom ratings from real-world Korean psycho-oncology interviews, focusing on concordance, directional bias, and clinician-adjudicated error characteristics. Methods: Between April 2024 and May 2025, 101 adults receiving oncologic care in South Korea underwent semistructured interviews. Board-certified psycho-oncologists provided real-time, time-stamped ratings of 39 items. Generative pretrained transformer 4o (GPT-4o), Claude 3.5 Sonnet, and Gemini 2.5 Flash generated item scores and brief rationales using identical Korean zero-shot rubrics. Concordance between clinician ratings was evaluated using ordinal and binary screening metrics. A paired Wilcoxon test compared patient-level total symptom burden. Binary mismatches were clinically adjudicated as ambiguous or definite overestimation or underestimation with an 8-etiology taxonomy. Model-generated rationales were further meta-evaluated using GPT-5.4, Claude Sonnet 4.6, and Gemini Pro 3.1 across 4 dimensions: citation (use of quoted supporting statements), structure (logical organization of the rationale), mapping (consistency between the rationale and the assigned item rating), and expansion (degree of interpretive elaboration beyond the explicit transcript content). The associations between meta-evaluation results and absolute error were examined using cross-classified mixed-effects models. Results: Out of 101 participants, 88 (87.1%) were predominantly female and had breast cancer as the primary cancer type (n=70, 69.3%). Across 3931 item-level ratings, all models showed good agreement with clinicians (intraclass correlation coefficient: 0.816‐0.872), with GPT-4o showing the highest agreement. Claude 3.5 and Gemini 2.5 yielded significantly higher patient-level symptom burden (adjusted

Background: Colorectal cancer remains a leading cause of death despite being largely preventable through polypectomy. AI systems designed to enhance polyp detection during colonoscopy have shown promise, but the extent to which they improve detection of different-sized polyps remains unclear. Objective: This study compared the size-stratified efficacy of AI-assisted colonoscopy vs standard colonoscopy using the Hartung-Knapp-Sidik-Jonkman (HKSJ) method, and generated exploratory rankings while acknowledging all cross-platform comparisons are indirect. Methods: This systematic review and network meta-analysis (NMA) searched PubMed, Embase, Cochrane CENTRAL, and Web of Science from inception to July 25, 2026, supplemented by citation searching. We included randomized controlled trials (RCTs) comparing AI-assisted vs standard colonoscopy in adults (≥18 years of age), reporting mean polyp detection counts stratified by size (≤5 mm, 6-9 mm, and ≥10 mm). Two reviewers screened studies, extracted data, and assessed risk of bias using the Cochrane Risk of Bias 2.0. We conducted frequentist NMA using the HKSJ method with restricted maximum likelihood estimation, calculated 95% prediction intervals (PIs), and assessed heterogeneity using I² and τ². Certainty of evidence was rated using the GRADE (Grading of Recommendations Assessment, Development, and Evaluation) framework. Results: A total of 13 RCTs (4156 participants) compared 8 AI systems to standard colonoscopy, forming a network without direct AI comparisons. For diminutive polyps (≤5 mm), AI showed a modest advantage (standardized mean difference [SMD] 0.21, 95% CI 0.07 to 0.35, 95% PI –1.12 to 1.54), but substantial heterogeneity (I²=86.6%) and wide PI crossing the null indicated high uncertainty. EndoScreener showed the most consistent evidence (SMD 0.36, 95% CI 0.18-0.54). For small and large polyps, effects were minimal (SMD 0.02, 95% CI –0.02 to 0.06, 95% PI –0.03 to 0.07; SMD 0.01, 95% CI 0.00-0.02, 95% PI –0.01 to 0.03). GRADE certainty was very low for diminutive polyps and low for small and large polyps. Sensitivity analysis excluding Tianjin YuJin did not materially change findings. Conclusions: AI may modestly enhance diminutive polyp detection, but effects on small and large polyps are minimal, with no platform superiority. Given very low to low certainty, findings are hypothesis-generating. This exploratory NMA provides size-stratified comparisons that can inform future head-to-head trial design. Unlike prior reviews aggregating all polyp sizes, we show the overall AI benefit is driven by diminutive polyp detection, providing a framework for targeted deployment—prioritizing AI for diminutive polyp screening, with limited value for larger lesions. Head-to-head trials are urgently needed. Trial Registration: PROSPERO International Prospective Register of Systematic Reviews CRD420251266932; https://www.crd.york.ac.uk/PROSPERO/view/CRD420251266932

Background: Large language models (LLMs) are increasingly being deployed in health care, but their use and deployment in many real-world hospital environments pose significant challenges and concerns. In particular, state-of-the-art commercial models store or process data externally, which is often in conflict with ensuring patient data protection. At the same time, using LLMs locally is limited by the lack of available computing infrastructure. Small open-source LLMs that do not require substantial computing resources could offer a practical way to resolve these tensions, but their medical utility in real-world local contexts, especially in non-English languages, has not been sufficiently evaluated. Objective: This study aimed to evaluate the feasibility of small, locally deployable open-source LLMs for clinically relevant tasks in a resource-constrained hospital setting and to propose a reproducible framework for institution-specific evaluation before deployment. Methods: We evaluated 6 open-source LLMs ranging from 8B to 24B parameters (from the Mistral, Phi4, Falcon3, Llama3.1, and Meditron3 families) in a zero-shot setting across 7 tasks covering 4 clinical use cases: information extraction, medical text translation, text generation, and clinical decision support. We used deidentified French clinical data from a Swiss tertiary hospital, including discharge letters, clinical notes, and structured electronic health records. Performance was assessed using task-specific metrics, such as precision, recall, F₁-score, embedding-based semantic similarity, recall-oriented understudy for gisting evaluation (ROUGE) score, readability indices, and human review by clinicians. Results: Model performance varied substantially between tasks. In the simplest retrieval task (needle-in-the-haystack), several models performed strongly, with Llama3.1 achieving an F₁-score of 99.81% and Mistral-small achieving 99.71%. In contrast, performance was poor in more complex tasks. For detecting protected health information, the best-performing LLMs achieved only modest overall macro–F₁-scores (0.33-0.34), substantially below a fine-tuned Robustly Optimized BERT Pretraining Approach (RoBERTa) baseline (0.94). In the task of extracting immune-related adverse events from discharge notes, the highest overall macro–F₁-score was 0.35 with Phi4. For medical text translation, Phi4 ranked highest in embedding-based evaluation, whereas Meditron3-Phi4 performed the worst, with clinician reviews identifying hallucinations in 55% of its outputs. In the task of summarizing discharge letters, quality was low across all models, with the best penalized ROUGE score reaching only 0.169 with Llama3.1. In the tasks of generating patient-friendly discharge note summaries and clinical decision support, clinician ratings generally ranged from dissatisfied to neutral, and no model achieved consistently satisfactory performance. Conclusions: Small open-source LLMs appear feasible for simple retrieval-oriented tasks in local hospital deployments but are currently inadequate for more complex applications, such as clinical decision support, deidentification, extraction of adverse events, and medical summarization. These findings highlight the importance of locally grounded evaluation tailored to specific use cases and the need for robust institutional evaluation frameworks to ensure safe and reliable deployment. Trial Registration:

Background: New HIV cases among Malaysian men who have sex with men continue to rise, with young men who have sex with men (YMSM) accounting for 44% of new infections and experiencing high rates of comorbidities. Mobile health (mHealth) apps offer a promising approach to addressing these challenges by providing discreet access to health information, screening tools, and linkage to services. Given the near-universal smartphone ownership among Malaysian YMSM and high levels of mobile gaming engagement, gamified mHealth apps may be particularly effective in sustaining engagement and promoting HIV prevention behaviors and comorbidity management. However, realizing their full potential requires identifying the features and design principles that are most important to Malaysian YMSM and that can support sustained engagement and improve health outcomes in this vulnerable population. Objective: This study used an interdisciplinary expert panel approach to identify the key features, design principles, and gamification elements to be incorporated into MY-Hero, a gamified mHealth app designed to support HIV prevention, comorbidity management, and overall well-being among Malaysian YMSM. Methods: Guided by the integrated behavioral model (IBM), we conducted an interdisciplinary expert panel comprising experts in health, technology, and design sciences, as well as health care professionals with expertise in the Malaysian YMSM population, along with local lesbian, gay, bisexual, transgender, queer, and others (LGBTQ+) community leaders. Through a structured series of discussions and thematic analysis, we identified culturally appropriate design principles and key features for MY-Hero that align with the health needs and sociocultural context of Malaysian YMSM. Results: Three expert panel sessions were conducted with 9 experts. Several key themes emerged: (1) the importance of establishing clinical affiliation and facilitating linkage to HIV testing, pre-exposure prophylaxis (PrEP), and related services, including harm reduction services; (2) the need to enhance user interface (UI) and user experience (UX) by optimizing usability, interactivity, and engagement to maintain user interest; (3) the incorporation of customizable health content to tailor interventions based on individual characteristics and preferences; (4) the use of gamification mechanisms, such as reward systems and progression tracking to promote adoption and sustained engagement with health services; and (5) the importance of privacy and data security as critical design considerations for ensuring user safety, confidentiality, and trust. Conclusions: The findings highlight the importance of user-centered design, contextualized health content, and gamification mechanisms in mHealth tools for Malaysian YMSM. These insights provide practical guidance for developing culturally appropriate gamified mHealth interventions for vulnerable populations by informing strategies to enhance user engagement, reduce HIV prevention fatigue, and support the well-being of YMSM in stigmatized contexts.

Background: Nonpharmacologic interventions are important for treating knee pain due to osteoarthritis or after total knee arthroplasty (TKA), and extended reality (XR) technology may enhance treatments for these indications. Objective: This systematic review aimed to evaluate XR interventions for pain due to knee osteoarthritis (KOA) or for recovery after TKA. Methods: Databases were searched through May 2023 and updated in December 2025. Eligible trials evaluated XR interventions to treat KOA pain or after TKA. We classified interventions by depth of immersion and clinical mechanism. We used the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) criteria to determine the certainty of evidence for prioritized outcomes. Meta-analyses were performed when ≥3 studies evaluated similar comparisons, outcomes, and time points. Results: Eligible trials addressed KOA (k=12) or recovery after TKA (k=9). Sample sizes ranged from 36 to 306 participants, and most studies had a follow-up of ≤3 months. Nineteen studies assessed pain-related functioning and pain intensity, and 5 assessed adverse events (AEs). For KOA, 10 studies examined interactive digital rehabilitation (IDR), and 2 examined virtual reality (VR)–digitally augmented exercise (DAE). IDR for KOA may result in better pain-related functioning (low certainty of evidence [COE]; pooled standardized mean difference [SMD] −0.59, 95% CI −1.11 to −0.06; prediction interval [PI] −1.72 to 0.55; k=5) and lower pain intensity at 6‐8 weeks (low COE; pooled SMD −0.46, 95% CI −0.92 to 0.00; PI −1.39 to 0.47; k=4). VR-DAE for KOA (k=2) produced inconsistent results (very low COE). For post-TKA studies, 5 examined IDR, 2 examined VR-DAE, 1 examined VR-distraction, and 1 examined VR-psychoeducation. Post-TKA IDR may result in better pain-related functioning (low [k=4] and moderate COE [k=1]) but little to no difference in pain intensity (low-moderate COE; pooled SMD at 3‐4 months −0.12, 95% CI −0.75 to 0.52; PI –1.63 to 1.27; k=3). VR-psychoeducation probably results in lower pain at 4 weeks (moderate COE; k=1), and VR-distraction may result in 6 months (low COE; k=1), whereas VR-DAE produced mixed findings (k=2; very low COE). IDR was not associated with AEs, and VR may not be associated with AEs for KOA (high and low COE), though AE reporting was uncommon (k=5) and evidence was very uncertain for post-TKA. Conclusions: IDR may augment treatment for KOA and post-TKA recovery, and VR may benefit post-TKA rehabilitation. This review is the first to stratify by level of immersion, clinical mechanism, and follow-up duration and to systematically evaluate AEs. IDR may be ready for integration into KOA care, while use after TKA needs more evidence. Randomized controlled trials with implementation outcomes could determine how XR interventions can be used for KOA, whereas trials evaluating efficacy and AEs are needed before their use for post-TKA. Trial Registration: PROSPERO CRD42023439903; https://www.crd.york.ac.uk/PROSPERO/view/CRD42023439903

This commentary extends recent discussion of the solidarity gap associated with patient-facing AI by examining gaps in governance and risk allocation, health care professionals’ responsibilities in practice, and opportunities for professional stewardship and advocacy. We argue that equitable implementation requires shared accountability and meaningful health care professional participation in the design, evaluation, reimbursement, governance, and oversight of patient-facing AI before ambiguity results in patient harm.