2f:I[703058,["/_next/static/immutable/chunks/3bhq_5eh1d760.js","/_next/static/immutable/chunks/1855-ogyi8eh4.js","/_next/static/immutable/chunks/2pfsjpn6k6zlg.js","/_next/static/immutable/chunks/26c7v3_3uwnwu.js","/_next/static/immutable/chunks/25a3ozcrvi1ff.js","/_next/static/immutable/chunks/2r4wx50nvjh_4.js","/_next/static/immutable/chunks/23fsw06hsi5b1.js","/_next/static/immutable/chunks/06_yt270h41rg.js","/_next/static/immutable/chunks/3ezsiz-7shkna.js","/_next/static/immutable/chunks/20x2y96ldu5wx.js","/_next/static/immutable/chunks/2lxntcm3pd1_h.js","/_next/static/immutable/chunks/00dgco8irq3jw.js","/_next/static/immutable/chunks/00nwe-kq0rboc.js","/_next/static/immutable/chunks/2051ffx3xzio9.js"],"PaperListPageClient"] 30:T55a,To develop and validate a deep learning-based AI system for the dynamic, real-time differentiation of benign and malignant gastric ulcers during endoscopy, with the goal of enhancing diagnostic precision. This was a multicenter, retrospective study collecting endoscopic images and videos from four tertiary hospitals in China. An improved YOLOv8 model, incorporating an illumination attention module, was developed for real-time instance segmentation and classification. The dataset comprised 9,820 benign ulcer images, 1,727 malignant ulcer images, and 15,791 normal mucosa images, split into training, testing, and validation sets at an 8:1:1 ratio. Performance was evaluated based on precision, recall, specificity, and processing latency. On the validation set, the AI model achieved an overall precision, recall, and specificity of 0.91, 0.91, and 0.95, respectively. For malignant ulcer recognition specifically, the precision, recall, and specificity were 0.90, 0.91, and 0.99. The model demonstrated strong real-time performance with a latency of 8.84 ms per frame and a processing speed of 113 frames per second. The developed AI model enables accurate, real-time discrimination between benign and malignant gastric ulcers during endoscopy. It holds potential to augment clinical decision-making, standardize diagnostic quality, and optimize biopsy strategies.31:T7cf,Abstract Aim To develop and validate a deep learning-based AI system for the dynamic, real-time differentiation of benign and malignant gastric ulcers during endoscopy, with the goal of enhancing diagnostic precision. Methods This was a multicenter, retrospective study collecting endoscopic images and videos from four tertiary hospitals in China. An improved YOLOv8 model, incorporating an illumination attention module, was developed for real-time instance segmentation and classification. The dataset comprised 9,820 benign ulcer images, 1,727 malignant ulcer images, and 15,791 normal mucosa images, split into training, testing, and validation sets at an 8:1:1 ratio. Performance was evaluated based on precision, recall, specificity, and processing latency. Results On the validation set, the AI model achieved an overall precision, recall, and specificity of 0.91, 0.91, and 0.95, respectively. For malignant ulcer recognition specifically, the precision, recall, and specificity were 0.90, 0.91, and 0.99. The model demonstrated strong real-time performance with a latency of 8.84 ms per frame and a processing speed of 113 frames per second. Conclusion The developed AI model enables accurate, real-time discrimination between benign and malignant gastric ulcers during endoscopy. It holds potential to augment clinical decision-making, standardize diagnostic quality, and optimize biopsy strategies. 25:[[["$","script","0",{"type":"application/ld+json","dangerouslySetInnerHTML":{"__html":"{\"@context\":\"https://schema.org\",\"@type\":\"CollectionPage\",\"name\":\"Artificial Intelligence / Deep Learning: modality Research & Clinical Takeaways\",\"description\":\"Explore peer-reviewed papers on Artificial Intelligence / Deep Learning (modality). Clinician-written summaries and actionable practice insights. Updated daily.\",\"url\":\"https://www.reachrx.ai/hub/tag/modality/Artificial%20Intelligence%20%2F%20Deep%20Learning\",\"breadcrumb\":{\"@type\":\"BreadcrumbList\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https://www.reachrx.ai\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Research Hub\",\"item\":\"https://www.reachrx.ai/hub\"},{\"@type\":\"ListItem\",\"position\":3,\"name\":\"Modalities\",\"item\":\"https://www.reachrx.ai/hub/tag/modality\"},{\"@type\":\"ListItem\",\"position\":4,\"name\":\"Artificial Intelligence / Deep Learning\",\"item\":\"https://www.reachrx.ai/hub/tag/modality/Artificial%20Intelligence%20%2F%20Deep%20Learning\"}]}}"}}]],["$","$L2f",null,{"title":"Artificial Intelligence / Deep Learning: Research & Clinical Takeaways","description":"Review evidence by treatment modality: Artificial Intelligence / Deep Learning — concise summaries and practical next steps.","papers":[{"doi":"10.1001/jamacardio.2026.3829","pmid":"42661496","title":"Artificial Intelligence–Enabled Acquisition and Interpretation for Screening Aortic Stenosis","journal_title":"JAMA Cardiology","journal_issn":"2380-6583","published_at":"2026-08-28T00:00:00.000Z","url":"https://doi.org/10.1001/jamacardio.2026.3829","source_score":31,"is_retracted":false,"is_open_access":false,"raw_metadata":{"pubmed":true,"crossref":true,"unpaywall":true},"created_at":"2026-09-01T06:20:58.074Z","updated_at":"2026-09-01T06:20:58.074Z","tags":[{"paper_doi":"10.1001/jamacardio.2026.3829","tag_type":"age_group","value":"adult","value_numeric":null},{"paper_doi":"10.1001/jamacardio.2026.3829","tag_type":"category","value":"Cardiac & Cardiovascular Systems","value_numeric":null},{"paper_doi":"10.1001/jamacardio.2026.3829","tag_type":"condition","value":"Aortic Stenosis","value_numeric":null},{"paper_doi":"10.1001/jamacardio.2026.3829","tag_type":"impact_flag","value":"practice_changing","value_numeric":null},{"paper_doi":"10.1001/jamacardio.2026.3829","tag_type":"level_of_evidence","value":"2","value_numeric":null},{"paper_doi":"10.1001/jamacardio.2026.3829","tag_type":"modality","value":"Artificial Intelligence / Deep Learning","value_numeric":null},{"paper_doi":"10.1001/jamacardio.2026.3829","tag_type":"modality","value":"Focused Cardiac Ultrasound","value_numeric":null},{"paper_doi":"10.1001/jamacardio.2026.3829","tag_type":"sample_size","value":"n=1302","value_numeric":1302},{"paper_doi":"10.1001/jamacardio.2026.3829","tag_type":"sex","value":"mixed","value_numeric":null},{"paper_doi":"10.1001/jamacardio.2026.3829","tag_type":"study_design","value":"Diagnostic Accuracy Study","value_numeric":null},{"paper_doi":"10.1001/jamacardio.2026.3829","tag_type":"warning","value":"PPV of 49.4% for AI alone highlights a potential for false positives in lower-prevalence settings — prevalence in this screened population is not explicitly stated, which limits direct extrapolation.","value_numeric":null},{"paper_doi":"10.1001/jamacardio.2026.3829","tag_type":"warning","value":"The sample_size tag reflects the prospective novice-operator cohort (n=1302) as the primary prospective evaluation group; total participants across all development and validation cohorts exceed 11,000. Clinicians should interpret the numeric tag in this context.","value_numeric":null},{"paper_doi":"10.1001/jamacardio.2026.3829","tag_type":"warning","value":"The study design sits between a prospective cohort study and a diagnostic accuracy study; labeled as 'Diagnostic Accuracy Study' per the paper's own description.","value_numeric":null}],"summaries":{"what_studied":"A deep learning algorithm was developed and validated for detecting moderate or greater aortic stenosis (AS) using AI-guided focused cardiac ultrasound (FoCUS), including prospective evaluation by novice operators across multiple Mayo Clinic sites (Midwest, Arizona, Florida) from 2005–2025.","key_result":"The model achieved AUROC of 0.99 across all validation cohorts. In prospective novice-operator use (n=1302), 96.6% of exams were analyzable, with sensitivity 93% and specificity 96%; adding expert review of AI-positive and uninterpretable exams raised PPV from 49.4% to 91.1% (sensitivity 85.4%).","limitations":"Single health system (Mayo Clinic), which may limit generalizability; novice operators still required expert review of flagged/uninterpretable cases to achieve acceptable PPV; PPV of 49.4% with AI alone reflects a relatively low prevalence of moderate+ AS in the screened population.","practice_takeaway":"AI-guided FoCUS operated by novice users can reliably screen for moderate or greater AS — a hybrid model pairing AI triage with targeted expert review of positive and uninterpretable exams may expand access in resource-limited settings. Clinicians should note that AI-alone PPV (~49%) requires expert confirmation before acting on positive results."}},{"doi":"10.1016/j.cgh.2026.07.021","pmid":"42567461","title":"Use of Artificial Intelligence for Duodenal Biopsy for Celiac Disease: Developing a Prediction Model for Histopathology Diagnosis","journal_title":"Clinical Gastroenterology and Hepatology","journal_issn":"1542-3565","published_at":"2026-08-07T00:00:00.000Z","url":"https://doi.org/10.1016/j.cgh.2026.07.021","source_score":7,"is_retracted":false,"is_open_access":true,"raw_metadata":{"pubmed":true,"crossref":true,"unpaywall":true},"created_at":"2026-08-14T13:24:54.067Z","updated_at":"2026-08-14T13:24:54.067Z","tags":[{"paper_doi":"10.1016/j.cgh.2026.07.021","tag_type":"age_group","value":"adult","value_numeric":null},{"paper_doi":"10.1016/j.cgh.2026.07.021","tag_type":"category","value":"Gastroenterology & Hepatology","value_numeric":null},{"paper_doi":"10.1016/j.cgh.2026.07.021","tag_type":"condition","value":"Celiac Disease","value_numeric":null},{"paper_doi":"10.1016/j.cgh.2026.07.021","tag_type":"condition","value":"Villous Atrophy","value_numeric":null},{"paper_doi":"10.1016/j.cgh.2026.07.021","tag_type":"impact_flag","value":"limited_evidence","value_numeric":null},{"paper_doi":"10.1016/j.cgh.2026.07.021","tag_type":"level_of_evidence","value":"3","value_numeric":null},{"paper_doi":"10.1016/j.cgh.2026.07.021","tag_type":"modality","value":"Artificial Intelligence / Deep Learning","value_numeric":null},{"paper_doi":"10.1016/j.cgh.2026.07.021","tag_type":"sample_size","value":"n=2406","value_numeric":2406},{"paper_doi":"10.1016/j.cgh.2026.07.021","tag_type":"study_design","value":"Retrospective Cohort Study","value_numeric":null},{"paper_doi":"10.1016/j.cgh.2026.07.021","tag_type":"warning","value":"Sex breakdown of the study population is not reported. The abstract does not clarify whether this is a single-center or multi-center dataset, which affects generalizability. The study design is labeled retrospective cohort based on the 2008–2021 data collection period, but the paper does not explicitly state this label.","value_numeric":null}],"summaries":{"what_studied":"A deep learning model (clustering-constrained attention multiple-instance learning) was evaluated on 2,406 whole-slide duodenal biopsy images from >2,100 adults to diagnose celiac disease across 5 histopathological categories, using data from 2008–2021.","key_result":"The model distinguished celiac disease from normal histology with AUC 0.992 and 97% accuracy. Across all 5 categories (normal, celiac disease, IELs only, IELs + positive tTG-IgA, seronegative villous atrophy), average AUC was 0.872 with 74% accuracy; celiac vs. seronegative villous atrophy yielded AUC 0.810 and 76% accuracy.","limitations":"Single-center or retrospective dataset design limits generalizability; overall 5-class accuracy of 74% leaves meaningful diagnostic uncertainty; patients on a gluten-free diet were excluded, narrowing applicability to untreated populations.","practice_takeaway":"This AI model shows strong potential as a screening tool to flag normal vs. celiac biopsies, but its 74% overall multi-class accuracy means it is not yet ready to replace pathologist review. Clinicians should view it as a triage aid, not a standalone diagnostic."}},{"doi":"10.1053/j.gastro.2026.07.009","pmid":"42521098","title":"AI-Assisted Risk Stratification in Stage II Colorectal Cancer: Multi-Institutional Validation of Semantically-Enhanced Deep Learning","journal_title":"Gastroenterology","journal_issn":"0016-5085","published_at":"2026-07-28T00:00:00.000Z","url":"https://doi.org/10.1053/j.gastro.2026.07.009","source_score":72,"is_retracted":false,"is_open_access":true,"raw_metadata":{"pubmed":true,"crossref":true,"unpaywall":true},"created_at":"2026-07-31T13:37:02.281Z","updated_at":"2026-07-31T13:37:02.281Z","tags":[{"paper_doi":"10.1053/j.gastro.2026.07.009","tag_type":"age_group","value":"adult","value_numeric":null},{"paper_doi":"10.1053/j.gastro.2026.07.009","tag_type":"category","value":"Gastroenterology & Hepatology","value_numeric":null},{"paper_doi":"10.1053/j.gastro.2026.07.009","tag_type":"condition","value":"Colorectal Cancer","value_numeric":null},{"paper_doi":"10.1053/j.gastro.2026.07.009","tag_type":"condition","value":"Stage II Colorectal Cancer","value_numeric":null},{"paper_doi":"10.1053/j.gastro.2026.07.009","tag_type":"impact_flag","value":"practice_changing","value_numeric":null},{"paper_doi":"10.1053/j.gastro.2026.07.009","tag_type":"level_of_evidence","value":"3","value_numeric":null},{"paper_doi":"10.1053/j.gastro.2026.07.009","tag_type":"modality","value":"Artificial Intelligence / Deep Learning","value_numeric":null},{"paper_doi":"10.1053/j.gastro.2026.07.009","tag_type":"modality","value":"Computational Pathology","value_numeric":null},{"paper_doi":"10.1053/j.gastro.2026.07.009","tag_type":"sample_size","value":"n=1608","value_numeric":1608},{"paper_doi":"10.1053/j.gastro.2026.07.009","tag_type":"sex","value":"mixed","value_numeric":null},{"paper_doi":"10.1053/j.gastro.2026.07.009","tag_type":"study_design","value":"Retrospective Multi-Institutional Validation Study","value_numeric":null},{"paper_doi":"10.1053/j.gastro.2026.07.009","tag_type":"warning","value":"Study design labelled as 'Retrospective Multi-Institutional Validation Study' — the paper includes both internal and external validation cohorts, which is a strength, but the retrospective nature limits causal inference.","value_numeric":null},{"paper_doi":"10.1053/j.gastro.2026.07.009","tag_type":"warning","value":"The combined multivariate cohort (n=1,220) is smaller than the total WSI cohort (n=1,608), suggesting some patients were excluded from this analysis; reasons are not detailed in the abstract.","value_numeric":null},{"paper_doi":"10.1053/j.gastro.2026.07.009","tag_type":"warning","value":"The 'practice_changing' impact flag is applied cautiously — the study is retrospective and prospective validation has not yet been reported; adoption in clinical workflows would be premature without further evidence.","value_numeric":null}],"summaries":{"what_studied":"SÉMIL (Semantically-Enhanced Multiple Instance Learning), an AI model integrating vision-language foundation models with attention-based MIL, was developed and validated on 1,608 H&E-stained whole slide images (WSIs) from three stage II colorectal cancer cohorts (Austin n=697, MCO n=478, DYNAMIC n=433) for automated invasive front assessment and prognostic risk stratification.","key_result":"SÉMIL outperformed non-semantic MIL on external validation (AUC 0.713–0.821 vs. 0.686–0.803). Survival prediction HRs were 4.73 (Austin, p=0.0012), 2.84 (MCO, p=0.0032), and 2.10 (DYNAMIC, p=0.0396). Among NCCN high-risk stage II patients specifically, HRs ranged 2.96–3.50 (all p<0.05). In multivariate analysis of the combined cohort (n=1,220), SÉMIL remained independently prognostic (HR=1.98, p=0.005) after adjusting for T stage, MMR status, and lymph node examination adequacy.","limitations":"- The study is retrospective across three cohorts, which may introduce selection bias.\n- Manual pathologist assessment was used as a comparator but concordance analysis was not a pre-specified primary endpoint.\n- Generalizability beyond the three included institutions and their specific H&E staining protocols is not yet established.","practice_takeaway":"For stage II colorectal cancer patients flagged as high-risk by NCCN criteria — where adjuvant chemotherapy decisions are most uncertain — SÉMIL may provide an additional, reproducible prognostic layer from routine H&E slides to help guide treatment. Clinicians should note this tool still requires prospective validation before routine adoption."}},{"doi":"10.1111/ddg.70347","pmid":"42494241","title":"AI‐assisted diagnosis of nail unit melanoma and melanonychia using a clinical deep learning model","journal_title":"JDDG: Journal der Deutschen Dermatologischen Gesellschaft","journal_issn":"1610-0379","published_at":"2026-07-24T00:00:00.000Z","url":"https://doi.org/10.1111/ddg.70347","source_score":1,"is_retracted":false,"is_open_access":null,"raw_metadata":{"pubmed":true,"crossref":true,"unpaywall":false},"created_at":"2026-07-28T14:02:32.694Z","updated_at":"2026-07-28T14:02:32.694Z","tags":[{"paper_doi":"10.1111/ddg.70347","tag_type":"age_group","value":"adult","value_numeric":null},{"paper_doi":"10.1111/ddg.70347","tag_type":"category","value":"Dermatology","value_numeric":null},{"paper_doi":"10.1111/ddg.70347","tag_type":"condition","value":"Melanonychia","value_numeric":null},{"paper_doi":"10.1111/ddg.70347","tag_type":"condition","value":"Nail Unit Melanoma","value_numeric":null},{"paper_doi":"10.1111/ddg.70347","tag_type":"impact_flag","value":"limited_evidence","value_numeric":null},{"paper_doi":"10.1111/ddg.70347","tag_type":"level_of_evidence","value":"3","value_numeric":null},{"paper_doi":"10.1111/ddg.70347","tag_type":"modality","value":"Artificial Intelligence / Deep Learning","value_numeric":null},{"paper_doi":"10.1111/ddg.70347","tag_type":"sample_size","value":"n=294","value_numeric":294},{"paper_doi":"10.1111/ddg.70347","tag_type":"study_design","value":"Retrospective Diagnostic Accuracy Study","value_numeric":null},{"paper_doi":"10.1111/ddg.70347","tag_type":"warning","value":"External validation was mentioned as a goal but not yet performed; real-world performance in diverse populations remains unknown.","value_numeric":null},{"paper_doi":"10.1111/ddg.70347","tag_type":"warning","value":"Sex of study participants is not reported in the abstract or results; sex tag omitted.","value_numeric":null},{"paper_doi":"10.1111/ddg.70347","tag_type":"warning","value":"The study design is retrospective with a single-institution cohort; level of evidence set at 3, though some may classify diagnostic accuracy studies differently.","value_numeric":null}],"summaries":{"what_studied":"A CNN-based AI model was developed and validated to distinguish nail unit melanoma (NUM) from benign melanonychia using clinical images from 172 melanonychia and 122 NUM patients; human raters (dermatologists, residents, non-medical participants) were assessed with and without CNN assistance on 40 representative images.","key_result":"DenseNet-121 achieved the highest AUROC (0.954); ResNet-50 achieved the highest sensitivity (87.2%). CNN assistance improved human diagnostic accuracy from 70.0% to 80.8%, with dermatology residents showing the largest gain.","limitations":"Retrospective single-institution design with a relatively small cohort (n=294); pediatric cases excluded, limiting generalizability; no multi-ethnic or multi-institutional external validation performed.","practice_takeaway":"A CNN tool can meaningfully boost clinician accuracy in distinguishing nail unit melanoma from benign melanonychia, especially for less experienced raters. It should be used as a supportive screening aid, not a replacement for clinical judgment or biopsy when warranted."}},{"doi":"10.1007/s00417-026-07273-6","pmid":"42171726","title":"AI-driven multimodal retinal imaging for early detection and risk stratification of vascular and neurodegenerative diseases","journal_title":"Graefe's Archive for Clinical and Experimental Ophthalmology","journal_issn":"0721-832X","published_at":"2026-05-22T00:00:00.000Z","url":"https://doi.org/10.1007/s00417-026-07273-6","source_score":7,"is_retracted":false,"is_open_access":true,"raw_metadata":{"pubmed":true,"crossref":true,"unpaywall":true},"created_at":"2026-05-31T13:35:31.901Z","updated_at":"2026-05-31T13:35:31.901Z","tags":[{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"age_group","value":"mixed","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"category","value":"Ophthalmology","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"condition","value":"Alzheimer's Disease","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"condition","value":"Chronic Kidney Disease","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"condition","value":"Coronary Artery Disease","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"condition","value":"Diabetic Retinopathy","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"condition","value":"Hypertension","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"condition","value":"Mild Cognitive Impairment","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"condition","value":"Multiple Sclerosis","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"condition","value":"Parkinson's Disease","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"condition","value":"Stroke","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"impact_flag","value":"limited_evidence","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"level_of_evidence","value":"5","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"modality","value":"Artificial Intelligence / Deep Learning","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"modality","value":"Fundus Photography","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"modality","value":"Multimodal Retinal Imaging","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"modality","value":"OCT Angiography","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"modality","value":"Optical Coherence Tomography","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"sample_size","value":"n=197","value_numeric":197},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"sex","value":"mixed","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"study_design","value":"Narrative Review","value_numeric":null},{"paper_doi":"10.1007/s00417-026-07273-6","tag_type":"warning","value":"This is a narrative review without formal systematic meta-analysis or risk-of-bias scoring; conclusions reflect expert synthesis rather than pooled quantitative evidence. The n=197 sample size tag reflects included studies, not patients. Technology readiness level (TRL) estimates are the authors' own framework and not externally validated. Most cited performance metrics (e.g., AUROC values) come from individual heterogeneous studies rather than this review's own analysis.","value_numeric":null}],"summaries":{"what_studied":"This narrative review evaluated how AI-integrated multimodal retinal imaging (fundus photography, OCT, OCTA, and emerging metabolic modalities) can enable early detection and risk stratification of systemic vascular diseases (hypertension, stroke, coronary artery disease, diabetic complications) and neurodegenerative disorders (Alzheimer's, Parkinson's, multiple sclerosis, cognitive decline), covering deep learning, self-supervised learning, and multimodal fusion approaches.","key_result":"Evidence across reviewed studies shows retinal AI can predict cardiovascular risk factors, estimate blood pressure, detect subclinical cerebrovascular disease, grade diabetic retinopathy, and identify neurodegenerative changes—with technology readiness levels (TRL) ranging from ~5–7 for DR screening to ~2–4 for most neurodegenerative applications; multicenter AD detection models achieved internal AUROC ~0.93 and external AUROCs ~0.73–0.91.","limitations":"This is a narrative rather than formal systematic review, with no quantitative risk-of-bias assessment. Most reviewed studies are retrospective, single-center, or case-control designs without prospective clinical outcome validation. Significant ophthalmic confounders (glaucoma, myopia, media opacity, imaging artifacts) can distort AI-derived retinal biomarkers and are inconsistently controlled across studies.","practice_takeaway":"Retinal AI is most ready for clinical use as a triage and surveillance tool for diabetic retinopathy screening (TRL 5–7), while applications for stroke, Alzheimer's, and Parkinson's remain pre-deployment (TRL 2–5) and should not yet replace standard diagnostic workups. Clinicians should treat AI retinal risk scores as decision-support outputs requiring human review, uncertainty quantification, and confirmatory testing rather than standalone diagnoses."}},{"doi":"10.1016/j.jacc.2026.03.084","pmid":"42126358","title":"ECG Signatures and Long-Term Ischemic Stroke Risk","journal_title":"JACC","journal_issn":"0735-1097","published_at":"2026-05-13T00:00:00.000Z","url":"https://doi.org/10.1016/j.jacc.2026.03.084","source_score":28,"is_retracted":false,"is_open_access":false,"raw_metadata":{"pubmed":true,"crossref":true,"unpaywall":true},"created_at":"2026-05-16T13:22:28.281Z","updated_at":"2026-05-16T13:22:28.281Z","tags":[{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"age_group","value":"adult","value_numeric":null},{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"category","value":"Cardiac & Cardiovascular Systems","value_numeric":null},{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"condition","value":"Atrial Cardiopathy","value_numeric":null},{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"condition","value":"Cardioembolic Stroke","value_numeric":null},{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"condition","value":"Ischemic Stroke","value_numeric":null},{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"impact_flag","value":"practice_changing","value_numeric":null},{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"level_of_evidence","value":"3","value_numeric":null},{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"modality","value":"Artificial Intelligence / Deep Learning","value_numeric":null},{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"modality","value":"Electrocardiography","value_numeric":null},{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"sample_size","value":"n=204033","value_numeric":204033},{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"sex","value":"mixed","value_numeric":null},{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"study_design","value":"Retrospective Cohort Study","value_numeric":null},{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"warning","value":"All cohorts are from Boston-area academic medical centers; external generalizability to community or non-US populations is uncertain.","value_numeric":null},{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"warning","value":"Study design is labeled as 'Retrospective Cohort' based on longitudinal care data use; the paper does not explicitly state this term—there is minor ambiguity.","value_numeric":null},{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"warning","value":"The 'practice_changing' impact flag reflects the novelty and performance parity with FSRP, but prospective validation of clinical utility has not yet been demonstrated.","value_numeric":null},{"paper_doi":"10.1016/j.jacc.2026.03.084","tag_type":"warning","value":"The total sample size combines the development cohort (101,496) and three test cohorts; the n=204,033 tag is an approximation of the full study population. The paper reports test cohort sizes separately.","value_numeric":null}],"summaries":{"what_studied":"A deep learning model (ECG2Stroke) trained on 12-lead ECGs was evaluated for predicting 10-year incident ischemic stroke risk across three independent hospital cohorts (MGH, BWH, BIDMC; total >200,000 patients), and compared against the revised Framingham Stroke Risk Profile (FSRP).","key_result":"ECG2Stroke achieved moderate discrimination across all sites (10-year AUCs: 0.795 at MGH, 0.774 at BWH, 0.772 at BIDMC) with low calibration error, performing similarly to FSRP. It was strongly associated with cardioembolic stroke (cause-specific HR 2.17 per 1-SD; 95% CI 1.64–2.87) but not noncardioembolic stroke; saliency maps pointed to the P-wave as the key signal.","limitations":"- Training and validation cohorts are all from a single academic health system network in Boston, limiting generalizability to diverse populations.\n- The model's associations are observational; causality and impact on clinical outcomes if acted upon are unproven.\n- Stroke subtype classification (cardioembolic vs. noncardioembolic) was only available in a subset, potentially introducing ascertainment bias.","practice_takeaway":"A routine 12-lead ECG, analyzed by AI, can stratify 10-year ischemic stroke risk as well as the Framingham Stroke Risk Profile—without requiring lab values or clinical history. Clinicians may consider AI-ECG screening as an efficient, low-burden tool to flag patients who warrant further workup for cardioembolic stroke prevention."}},{"doi":"10.1186/s12876-026-04848-9","pmid":"42082915","title":"Development and validation of a real-time AI model for differentiating benign and malignant gastric ulcers : a multicenter retrospective study","journal_title":"BMC Gastroenterology","journal_issn":"1471-230X","published_at":"2026-05-05T00:00:00.000Z","url":"https://doi.org/10.1186/s12876-026-04848-9","source_score":7,"is_retracted":false,"is_open_access":true,"raw_metadata":{"ooir":{"day":"2026-05-10","doi":"10.1186/s12876-026-04848-9","days":"7","issn":"1471-230X","rank":"8","field":"Clinical Medicine","score":"7","category":"Gastroenterology & Hepatology","categories":["Gastroenterology & Hepatology"],"_categories":["Gastroenterology & Hepatology"]},"pubmed":{"doi":"10.1186/s12876-026-04848-9","title":"Development and validation of a real-time AI model for differentiating benign and malignant gastric ulcers : a multicenter retrospective study.","journal":"BMC gastroenterology","methods":"","results":"On the validation set, the AI model achieved an overall precision, recall, and specificity of 0.91, 0.91, and 0.95, respectively. For malignant ulcer recognition specifically, the precision, recall, and specificity were 0.90, 0.91, and 0.99. The model demonstrated strong real-time performance with a latency of 8.84 ms per frame and a processing speed of 113 frames per second.","abstract":"$30","keywords":["Artificial intelligence","Benign and malignant differentiation","Deep learning","Endoscopic diagnosis","Gastric ulcer"],"pubmed_id":"42082915","conclusions":"The developed AI model enables accurate, real-time discrimination between benign and malignant gastric ulcers during endoscopy. It holds potential to augment clinical decision-making, standardize diagnostic quality, and optimize biopsy strategies.","publication_date":"2026-05-05T00:00:00"},"crossref":{"DOI":"10.1186/s12876-026-04848-9","URL":"https://doi.org/10.1186/s12876-026-04848-9","ISSN":["1471-230X"],"link":[{"URL":"https://link.springer.com/article/10.1186/s12876-026-04848-9","content-type":"text/html","content-version":"vor","intended-application":"text-mining"}],"page":null,"type":"journal-article","issue":null,"score":1,"title":["Development and validation of a real-time AI model for differentiating benign and malignant gastric ulcers : a multicenter retrospective study"],"author":[{"given":"Yibo","family":"Tan","sequence":"first","affiliation":[]},{"given":"Yongjun","family":"Wu","sequence":"additional","affiliation":[]},{"given":"Mei","family":"Yang","sequence":"additional","affiliation":[]},{"given":"Yan","family":"Li","sequence":"additional","affiliation":[]},{"given":"Xiaofei","family":"Bi","sequence":"additional","affiliation":[]},{"given":"Song","family":"He","sequence":"additional","affiliation":[]},{"given":"Zhihang","family":"Zhou","sequence":"additional","affiliation":[]},{"given":"Junyu","family":"Lu","sequence":"additional","affiliation":[]}],"issued":{"date-parts":[[2026,5,4]]},"member":"297","prefix":"10.1186","source":"Crossref","volume":null,"created":{"date-time":"2026-05-04T04:53:56Z","timestamp":1777870436000,"date-parts":[[2026,5,4]]},"indexed":{"date-time":"2026-05-05T02:58:12Z","timestamp":1777949892396,"date-parts":[[2026,5,5]]},"license":[{"URL":"https://creativecommons.org/licenses/by/4.0","start":{"date-time":"2026-05-04T00:00:00Z","timestamp":1777852800000,"date-parts":[[2026,5,4]]},"delay-in-days":0,"content-version":"tdm"},{"URL":"https://creativecommons.org/licenses/by/4.0","start":{"date-time":"2026-05-04T00:00:00Z","timestamp":1777852800000,"date-parts":[[2026,5,4]]},"delay-in-days":0,"content-version":"am"}],"subject":[],"abstract":"$31","language":"en","relation":{},"resource":{"primary":{"URL":"https://link.springer.com/10.1186/s12876-026-04848-9"}},"subtitle":[],"assertion":[{"name":"received","group":{"name":"ArticleHistory","label":"Article History"},"label":"Received","order":1,"value":"3 February 2026"},{"name":"accepted","group":{"name":"ArticleHistory","label":"Article History"},"label":"Accepted","order":2,"value":"15 April 2026"},{"name":"first_online","group":{"name":"ArticleHistory","label":"Article History"},"label":"First Online","order":3,"value":"4 May 2026"},{"name":"Ethics","group":{"name":"EthicsHeading","label":"Declarations"},"order":1},{"name":"Ethics","group":{"name":"EthicsHeading","label":"Ethics approval and consent to participate"},"order":2,"value":"The study was conducted in accordance with the Declaration of Helsinki (as revised in 2024). The study protocol, including the request for waiver of informed consent, was initially reviewed and approved by the Institutional Review Board (The Ethics Committee of the Second Affiliated Hospital of Chongqing Medical University; Approval No. 2024 Ethical Review No.129). The requirement for informed consent was waived by the Institutional Review Boards due to the retrospective nature of the study and the use of fully anonymized data."},{"name":"Ethics","group":{"name":"EthicsHeading","label":"Consent for publication"},"order":3,"value":"Not applicable."},{"name":"Ethics","group":{"name":"EthicsHeading","label":"Competing interests"},"order":4,"value":"The authors declare no competing interests."}],"deposited":{"date-time":"2026-05-04T04:53:57Z","timestamp":1777870437000,"date-parts":[[2026,5,4]]},"issn-type":[{"type":"electronic","value":"1471-230X"}],"published":{"date-parts":[[2026,5,4]]},"publisher":"Springer Science and Business Media LLC","reference":null,"short-title":[],"update-policy":"https://doi.org/10.1007/springer_crossmark_policy","alternative-id":["4848"],"content-domain":{"domain":["link.springer.com"],"crossmark-restriction":false},"original-title":[],"container-title":["BMC Gastroenterology"],"reference-count":0,"published-online":{"date-time":null,"timestamp":null,"date-parts":[[2026,5,4]]},"references-count":0,"short-container-title":["BMC Gastroenterol"],"is-referenced-by-count":0},"unpaywall":{"doi":"10.1186/s12876-026-04848-9","is_oa":true,"title":"Development and validation of a real-time AI model for differentiating benign and malignant gastric ulcers : a multicenter retrospective study","updated":"2026-05-05T11:59:32+00:00","published_date":"2026-05-04","best_oa_location":{"url":"https://doi.org/10.1186/s12876-026-04848-9","pmh_id":null,"url_for_pdf":null,"url_for_landing_page":"https://doi.org/10.1186/s12876-026-04848-9"}}},"created_at":"2026-05-14T13:23:01.287Z","updated_at":"2026-05-14T15:13:26.918Z","tags":[{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"age_group","value":"mixed","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"category","value":"Gastroenterology & Hepatology","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"condition","value":"Gastric Cancer","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"condition","value":"Gastric Ulcer","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"impact_flag","value":"limited_evidence","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"level_of_evidence","value":"3","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"modality","value":"Artificial Intelligence / Deep Learning","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"modality","value":"Endoscopy","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"sample_size","value":"n=27338","value_numeric":27338},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"study_design","value":"Diagnostic Accuracy Study","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"study_design","value":"Retrospective Multicenter Validation Study","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"warning","value":"Design label approximated; this is a retrospective, image-based diagnostic model development/validation study.","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"warning","value":"Metrics are from the study’s validation set; no prospective clinical testing was reported.","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"warning","value":"No direct comparison to endoscopist diagnostic accuracy is provided, making clinical added value difficult to assess.","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"warning","value":"Sample size reflects number of images; the number of videos/patients was not reported.","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"warning","value":"Sample size reflects total images, not unique patients — patient-level N is not reported, which may inflate performance metrics.","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"warning","value":"Study is from four Chinese tertiary centers only; performance in community or non-Asian settings is unknown.","value_numeric":null},{"paper_doi":"10.1186/s12876-026-04848-9","tag_type":"warning","value":"The dataset imbalance (benign images ~5.7× more than malignant) could bias recall estimates; mitigation strategies are not described in the abstract.","value_numeric":null}],"summaries":{"limitations":"Retrospective, image-based dataset from four centers without reported patient-level outcomes. Performance was reported only on the study’s validation set.","what_studied":"A multicenter retrospective study developed and validated an improved YOLOv8 AI to distinguish benign from malignant gastric ulcers in real time during endoscopy. The dataset contained 27,338 images (9,820 benign ulcers; 1,727 malignant ulcers; 15,791 normal mucosa) split 8:1:1, with performance assessed by precision, recall, specificity, and latency.","key_result":"On the validation set, overall precision, recall, and specificity were 0.91, 0.91, and 0.95; for malignant ulcers, precision, recall, and specificity were 0.90, 0.91, and 0.99. Real-time performance showed 8.84 ms/frame latency and 113 frames/second processing speed.","practice_takeaway":"In this multicenter dataset, the AI achieved high accuracy and real-time speed for classifying gastric ulcers, suggesting potential to assist endoscopic decision-making and biopsy strategy selection."}},{"doi":"10.1002/acr.80054","pmid":"41940478","title":"Artificial Intelligence–Aided Lung Ultrasound Detection of Interstitial Lung Disease in Systemic Sclerosis and Inflammatory Myopathy","journal_title":"Arthritis Care & Research","journal_issn":"2151-464X","published_at":"2026-04-29T00:00:00.000Z","url":"https://doi.org/10.1002/acr.80054","source_score":27,"is_retracted":false,"is_open_access":false,"raw_metadata":{"pubmed":true,"crossref":true,"unpaywall":true},"created_at":"2026-05-07T22:49:23.777Z","updated_at":"2026-05-07T22:49:23.777Z","tags":[{"paper_doi":"10.1002/acr.80054","tag_type":"age_group","value":"adult","value_numeric":null},{"paper_doi":"10.1002/acr.80054","tag_type":"category","value":"Rheumatology","value_numeric":null},{"paper_doi":"10.1002/acr.80054","tag_type":"condition","value":"Idiopathic Inflammatory Myopathy","value_numeric":null},{"paper_doi":"10.1002/acr.80054","tag_type":"condition","value":"Interstitial Lung Disease","value_numeric":null},{"paper_doi":"10.1002/acr.80054","tag_type":"condition","value":"Systemic Sclerosis","value_numeric":null},{"paper_doi":"10.1002/acr.80054","tag_type":"impact_flag","value":"practice_changing","value_numeric":null},{"paper_doi":"10.1002/acr.80054","tag_type":"level_of_evidence","value":"3","value_numeric":null},{"paper_doi":"10.1002/acr.80054","tag_type":"modality","value":"Artificial Intelligence / Deep Learning","value_numeric":null},{"paper_doi":"10.1002/acr.80054","tag_type":"modality","value":"Lung Ultrasound","value_numeric":null},{"paper_doi":"10.1002/acr.80054","tag_type":"sample_size","value":"n=140","value_numeric":140},{"paper_doi":"10.1002/acr.80054","tag_type":"sex","value":"mixed","value_numeric":null},{"paper_doi":"10.1002/acr.80054","tag_type":"study_design","value":"Diagnostic Accuracy Study","value_numeric":null},{"paper_doi":"10.1002/acr.80054","tag_type":"warning","value":"Sex distribution was not explicitly reported in the abstract; 'mixed' was inferred from the patient population (SSc and IIM typically include both sexes).","value_numeric":null},{"paper_doi":"10.1002/acr.80054","tag_type":"warning","value":"The impact_flag 'practice_changing' was applied cautiously given promising but preliminary single-cohort results.","value_numeric":null},{"paper_doi":"10.1002/acr.80054","tag_type":"warning","value":"The study design is a retrospective diagnostic accuracy study, not an RCT; 'practice_changing' flag reflects the high AUC and novel AI application but prospective validation is still needed.","value_numeric":null}],"summaries":{"what_studied":"This study evaluated whether convolutional neural networks (CNNs) could detect interstitial lung disease (ILD) and its severity on lung ultrasound (LUS) in 140 patients with systemic sclerosis (SSc) or idiopathic inflammatory myopathy (IIM), using chest CT and pulmonary function tests as reference standards.","key_result":"VGG-16 achieved the best patient-level performance (AUC 0.972, sensitivity 97.4%, specificity 92.6%), matching or exceeding expert LUS-ILD-24 interpretation; Grad-CAM identified pleural features — not B-lines — as the primary drivers of model predictions.","limitations":"- Small sample size (n=140; independent test set n=66) from what appears to be a single center, limiting generalizability.\n- Models were trained on labeled images using CT results and human expert interpretation, so errors in labeling could propagate.\n- Real-world deployment performance and prospective validation have not been assessed.","practice_takeaway":"AI-assisted LUS analysis may offer a radiation-free, high-accuracy screening tool for ILD in connective tissue diseases like SSc and IIM. Clinicians should note that pleural irregularities — even without prominent B-lines — may be key diagnostic features worth evaluating."}},{"doi":"10.1002/art.70196","pmid":"42011795","title":"autoscoRA: Deep Learning to Automate Sharp/van der Heijde Scoring of Radiographic Damage in Rheumatoid Arthritis","journal_title":"Arthritis & Rheumatology","journal_issn":"2326-5191","published_at":"2026-04-21T00:00:00.000Z","url":"https://doi.org/10.1002/art.70196","source_score":17,"is_retracted":false,"is_open_access":true,"raw_metadata":{"pubmed":true,"crossref":true,"unpaywall":true},"created_at":"2026-05-07T22:49:29.450Z","updated_at":"2026-05-07T22:49:29.450Z","tags":[{"paper_doi":"10.1002/art.70196","tag_type":"age_group","value":"adult","value_numeric":null},{"paper_doi":"10.1002/art.70196","tag_type":"category","value":"Rheumatology","value_numeric":null},{"paper_doi":"10.1002/art.70196","tag_type":"condition","value":"Rheumatoid Arthritis","value_numeric":null},{"paper_doi":"10.1002/art.70196","tag_type":"impact_flag","value":"practice_changing","value_numeric":null},{"paper_doi":"10.1002/art.70196","tag_type":"level_of_evidence","value":"3","value_numeric":null},{"paper_doi":"10.1002/art.70196","tag_type":"modality","value":"Artificial Intelligence / Deep Learning","value_numeric":null},{"paper_doi":"10.1002/art.70196","tag_type":"sample_size","value":"n=769","value_numeric":769},{"paper_doi":"10.1002/art.70196","tag_type":"sex","value":"mixed","value_numeric":null},{"paper_doi":"10.1002/art.70196","tag_type":"study_design","value":"Diagnostic / Validation Study","value_numeric":null},{"paper_doi":"10.1002/art.70196","tag_type":"warning","value":"The 'practice_changing' flag is aspirational — the paper itself notes routine clinical care use is a future goal pending further validation. No external validation cohort was used. The study design is best classified as a diagnostic/validation study rather than a classic RCT or cohort study; level of evidence set to 3 accordingly.","value_numeric":null}],"summaries":{"what_studied":"autoscoRA, a deep learning system, was trained and tested on the largest reported dataset of adult RA patients (769 total; 155 in test set; 3,437 visits; 12,144 radiographs) to fully automate Sharp/van der Heijde (SvdH) scoring of joint space narrowing and bone erosion on hand and foot radiographs.","key_result":"The model achieved excellent agreement with a human scorer (ICC ~0.9) for joint space narrowing, erosion, and combined SvdH scores at both the joint and total-score levels; it outperformed a second human reader in agreement with the primary reader, and showed ~70% average agreement for detecting longitudinal progression across multiple SvdH cut-offs.","limitations":"- Agreement for longitudinal progression detection averaged only 70%, which may limit use in individual patient monitoring.\n- A single primary human scorer was used as the reference standard, introducing potential anchor bias.\n- Generalizability to external cohorts, different X-ray equipment, or diverse ethnic populations was not tested.","practice_takeaway":"autoscoRA can reliably replicate expert SvdH scoring (ICC ~0.9) and even outperform a second human reader, making it a strong candidate for automating radiographic damage assessment in RA clinical trials and registries. Routine clinical deployment still warrants validation in independent, diverse cohorts before adoption."}}],"initialOffset":9,"fetchConfig":{"kind":"tags","tags":[{"tag_type":"modality","value":"Artificial Intelligence / Deep Learning"}]},"breadcrumbItems":[{"name":"Home","href":"/"},{"name":"Research Hub","href":"/hub"},{"name":"Modalities","href":"/hub/tag/modality"},{"name":"Artificial Intelligence / Deep Learning"}],"hrefConfig":{"basePath":"/hub/papers","webViewQueryParams":{"hub_ctx":"tag","hub_tag_type":"modality","hub_tag_value":"Artificial Intelligence / Deep Learning"}}}]]

Papers