Volume 06 · Issue 03 · 2026

Artificial intelligence (AI) in medicine is advancing steadily toward real clinical practice, not only through improved predictive performance but also through more decision-relevant modeling, evaluation, and interaction. The studies highlighted here illustrate three complementary directions. First, contemporary imaging models—exemplified by three-dimensional vision transformer-based analysis of preoperative chest computed tomography (CT)—are being used to infer clinically consequential phenotypes, advancing from image recognition toward decision support in high-stakes settings such as surgical planning. Second, image-to-biomarker pipelines such as automated quantification of retinal vascular fractal dimension demonstrate how medical images can be transformed into reproducible quantitative markers suitable for population-level analysis and risk stratification. Third, large language models (LLMs) are increasingly being evaluated and positioned as clinical communication and interpretation components, most notably structured assessments in telepharmacy and bilingual patient education move beyond fluency to safety, actionability, empathy, and readability, whereas emerging perspectives consider LLMs as interpretive interfaces for complex and temporally evolving health data, including wearable sensing. At the same time, these advances also expose persistent bottlenecks that limit real-world deployment, particularly the continued dominance of single-task and static formulations, fragmented systems driven by task-specific fine-tuning, limited reasoning over disease trajectories and evolving clinical contexts, and evaluation practices that remain insufficiently coupled to clinical workflows and downstream consequences. We argue that the next stage of medical AI should shift from accuracy-centered prediction toward decision-oriented, practice-ready systems that are robust across settings, clinically aligned in evaluation, and deployable at scale in routine care.

The eye is no longer confined to the study of isolated ophthalmic diseases; instead, it is emerging as a scalable, noninvasive source of biomarkers for systemic health assessment. Advances in ophthalmic imaging and artificial intelligence (AI) have revealed that retinal structure, microvasculature, and neurovascular coupling encode clinically meaningful information about cardiovascular, metabolic, and neurodegenerative disorders. This growing body of evidence challenges the traditional organ-centric view of ophthalmology and positions retinal biomarkers as a foundation for population-level risk stratification and longitudinal disease monitoring. In this Editorial, we argue that the true clinical impact of retinal AI will not be realized through incremental improvements in prediction accuracy, but from a paradigm shift toward clinically validated, human-trustworthy, and interpretable multimodal foundation models for ophthalmology. Despite promising results across multiple systemic diseases, translation remains hindered by substantial modality heterogeneity, insufficient clinical validation, and limited interpretability. These barriers have constrained retinal AI largely to retrospective studies, preventing its adoption as a reliable clinical decision-support tool. We highlight key priorities for the field, including multimodal and longitudinal modeling, biologically grounded feature learning, human-AI collaboration, and rigorous prospective validation. Addressing these challenges is essential for transforming retinal imaging from a diagnostic adjunct into a cornerstone of precision medicine. Ultimately, realizing the full potential of the eye as a window to systemic health will require aligning algorithmic innovation with clinical relevance, interpretability, and trust.

Artificial intelligence (AI) has shown significant promise in gynecological imaging, particularly in diagnosing gynecological cancers and benign diseases. However, the application of AI in this field faces several technical challenges, including data quality, model generalizability, and clinical interpretability. We have reviewed the current state of AI in gynecological imaging, identified key technical bottlenecks, and proposed future engineering solutions. Through case studies, we highlight the potential and challenges of AI in gynecological diagnosis, providing a roadmap for future research directions. We emphasize the importance of data sharing, model optimization, and clinical validation to overcome these challenges and enhance the integration of AI into clinical practice.

Computer vision (CV) and natural language processing (NLP) are 2 crucial subsets of artificial intelligence (AI). Large language models (LLMs) represent a significant application of deep learning in NLP. LLMs are AI systems with transformer architectures trained on large-scale datasets that are capable of understanding and generating natural language. LLMs have been widely applied in tasks such as text generation and translation. In dermatology, LLMs have been used for various purposes, such as diagnostic assistance, medical practice and decision support, patient communication, and professional education. However, dermatology differs from other medical specialties in that the diagnosis of diseases, selection of treatment methods, and prediction of prognoses rely heavily on the recognition of visual patterns. Therefore, in dermatology, LLMs must simultaneously process dermatological images and natural language. The aim of this study was to systematically review NLP and CV applications of LLMs in dermatology. In accordance with the preferred reporting items for systematic reviews and meta-analyses guidelines, the applications of LLMs in dermatological NLP and image processing tasks were systematically reviewed. After searching the MEDLINE (PubMed), Web of Science, and Scopus databases, a total of 37 original studies were included. The results showed that, in NLP tasks, LLMs demonstrated satisfactory performance in some studies, but most image processing results were not ideal. This review summarized the research findings on the applications of LLMs in dermatological natural language and image analysis, to identify LLMs capable of simultaneously processing text and images as a direction for future development.

Large language models (LLMs) represent a significant advancement in artificial intelligence (AI), offering versatile tools for natural language generation and understanding. The integration of LLMs into health and medicine is progressing rapidly, coinciding with the rising popularity and accessibility of wearable devices. Modern wearable devices such as smartwatches, heart rate monitors, activity trackers, and smart rings can now continuously monitor and measure various physiological and behavioral data, such as physical activity, motion, heart rate, skin temperature, and sleep. Wearable devices have several applications in biomedical research, health, and medicine. Some of these include self-monitoring, remote patient care, and out-of-clinic assessments. However, the continuous streams of data generated by wearables present significant challenges due to their sheer volume, unique characteristics, and inherent complexity. Although still in its infancy, there is a growing trend toward leveraging pre-trained LLMs to interpret and analyze these wearable-generated data streams for various applications. This opinion piece explores the expanding role of pre-trained LLMs in interpreting, analyzing, and leveraging physiological and behavioral data from wearables and discusses how LLMs offer a promising path forward in this evolving field.

Rectal cancer with lateral pelvic lymph node (LPLN) involvement presents significant challenges. Neoadjuvant chemoradiation (nCRT) became accepted as a standard preoperative approach in locally advanced rectal cancer, but its effectiveness in addressing LPLN metastasis remains debated, particularly when compared with prophylactic LPLN dissection (LPLND), due to patient and treatment heterogeneity. Current evidence supports reviewing pre- and post-treatment imaging to identify suspicious LPLNs, using a baseline short axis ≥7 mm and restaging short axis ≥4 mm as thresholds. The introduction of total neoadjuvant therapy has added complexity to LPLN management, with indications for LPLND evolving alongside advances in understanding LPLN size, location, and morphology. In recent years, artificial intelligence (AI) techniques including radiomics, image segmentation (e.g., conventional neural network), multimodal model, natural language processing, and surgery navigation have emerged as a promising solution to tackle the complex challenges in LPLN management. Accurate LPLN mapping and minimally invasive techniques enable more precise LPLND, reshaping the scale of surgical decision-making. Thus, optimizing LPLN evaluation methods and exploring AI applications in this process offer potential for tailoring treatment strategies, enhance tumor response, and improve patient outcomes.

Neurological disorders, encompassing a diverse array of conditions affecting the brain, spinal cord, and peripheral nerves, have emerged as significant contributors to global disability and mortality. Recent studies indicate a marked increase in the burden of these disorders, necessitating innovative, scalable solutions for diagnosis and management. This review explores the translational potential of advanced artificial intelligence (AI) and machine learning (ML) technologies in the assessment and treatment of neurological and neurodevelopmental disorders, including epilepsy, attention deficit hyperactivity disorder (ADHD), autism spectrum disorder, and neurodegenerative diseases like Alzheimer’s, Parkinson’s, Huntington’s, stroke, multiple sclerosis, etc. By integrating AI-driven toolsranging from diagnostic algorithms to wearable technologies,this research highlights their capabilities in early detection, personalized treatment, and improved patient outcomes. AI-powered technologies aid clinicians in understanding the complex, nonlinear neurological features of an individual for better diagnosis of neurological disorders. The review covers recent developments in learning algorithms and feature selection methodologies for better accuracy in disease prediction. The application of deep learning algorithms for brain image-based data analysis, feature extraction, and pattern learning for developing a diagnostic tool is highlighted. We have explored the integration of these AI models in wearables, ChatGPT, and gaming-based apps for real-time monitoring and patient-specific therapies. The paper discusses real-world applications, ethical considerations, and challenges associated with AI implementation in healthcare while providing a comprehensive overview of current methodologies and statistical insights. A critical analysis of the AI tools was performed using a rubric of seven domain grading system. We have discussed the importance of data volume and feature variablity for reducing the misdiagnosis due to bias or variance, which will improve diagnostic precision and user trust. The review serves as a resource for researchers, clinicians, and policymakers aiming to advance global brain health initiatives and optimize neurological care. Through systematic literature analysis, the findings highlight the critical role of AI in reshaping the landscape of neurological disorder management, emphasizing the need for continued research and development in this rapidly evolving field.

Background Pulmonary hypertension (PH) causes high mortality and poses diagnostic challenges. Current guidelines require invasive right heart catheterization (RHC) to confirm mean pulmonary artery pressure ≥25 mmHg. Delayed diagnosis impairs timely treatment. It is unknown whether standard chest X-rays can stratify PH severity. We aimed to develop and validate Exemplar MobileNet (ExMobileNet), an explainable artificial intelligence (AI) model that classifies PH into hemodynamic categories from routine chest X-ray images and thus supports non-invasive severity assessment. Methods We collected 1,293 de-identified chest X-rays obtained from 2018 to 2023. The cohort comprised 135 patients with PH confirmed using RHC and 551 healthy controls. We defined seven multi-class tasks for key hemodynamic parameters (such as mean pulmonary artery pressure, pulmonary vascular resistance, and cardiac index). The ExMobileNet workflow consists of: (1) Feature extraction via MobileNetV2, (2) feature selection by neighborhood component analysis and chi-square feature selectors, (3) classification with k-nearest neighbors and support vector machines and (4) decision fusion by majority vote and greedy optimization. Results Task-level accuracy ranged from 90.3% to 93.2%. Geometric mean scores ranged from 78.9% to 85.1%. Overall sensitivity and specificity were 88.5% and 91.3%, respectively. Mean accuracy across all tasks was 92.0% (±1.2%). Average inference time was 2.3 ± 0.4 second per image on CPU-only hardware. Conclusion ExMobileNet achieved high agreement with RHC-based assessments using routine chest X-rays. This AI tool may enable earlier, non-invasive PH screening in clinical practice.

Background Assessing the physiological discharge fitness of acute medical unit (AMU) patients is feasible using machine learning with electronic medical record (EMR) data, including intermittently measured vital signs. Wearable sensor data, including continuously monitored vital signs, may improve this assessment. Therefore, the aim was to investigate the value of sensor data in assessing the physiological discharge fitness of AMU patients. Methods EMR and sensor data of AMU patients (n = 145) who participated in a randomized controlled trial, conducted at a large teaching hospital in the Netherlands from December 2021 to October 2023, were used. EMR features were extracted from patient characteristics, vital signs, nursing assessments, and laboratory results. Statistical features and neural features, derived from convolutional neural networks, were extracted from sensor data. Logistic regression (LR), random forest (RF), and extreme gradient boosting (XGB) were used as algorithms. For each, three models were optimized and assessed using nested cross-validation: based on EMR, EMR with sensor, or only sensor data. The mean area under the receiver operator characteristic curve (AUROC) value with a 95% confidence interval (95% CI) was used as the primary metric. Results For LR, RF, and XGB, the mean AUROC value (shown with 95% CI), obtained by models based on EMR with sensor data were 0.73 (95% CI: 0.70-0.77), 0.73 (95% CI: 0.69-0.77), and 0.69 (95% CI: 0.64-0.73), respectively, which were slightly better than or as good as that obtained by models only considering EMR data: 0.69 (95% CI: 0.65-0.73), 0.70 (95% CI: 0.66-0.73), and 0.69 (95% CI: 0.66-0.71), respectively, and as that obtained by models only considering sensor data: 0.70 (95% CI: 0.66-0.75), 0.71 (95% CI: 0.67-0.75), and 0.67 (95% CI: 0.63-0.71), respectively. Conclusion Wearable sensor data could slightly improve the assessment of the physiological discharge fitness of AMU patients compared with only using EMR data.

Background Colorectal cancer (CRC) presents a significant global health challenge owing to its rising incidence. Early detection and precise prediction of molecular target expression, guided by key biomarkers, such as CK-20, Ki-67, P-53, and microsatellite instability, are essential for personalizing treatment strategies and improving patient outcomes. However, current approaches often struggle to capture intricate semantic and relational structures within vast medical records and corpus. Methods We developed CRC-Bidirectional Encoder Representations from Transformer (BERT)-graph convolutional network (GCN), a pre-trained deep learning framework that integrates graph neural network, to fully use the semantic information in predicting cancer phenotypes. We retrospectively collected data on 6,468 patients with CRC from a tertiary hospital in Beijing, from 2012 to 2022, electronic health records and radiology reports were collected and divided into training sets (70%), validation sets (10%), and tests (20%) stratified by molecular phenotypes. The model performance was evaluated on accuracy, precision, recall, and F1-Score. Ablation studies and attention analysis were performed to interpret the model’s performance and internal mechanisms. Results CRC-BERT-GCN outperformed all basic models and achieved the best performance across all evaluation metrics, with an F1 Score of 73.67%. Compared with the baseline models, it improved accuracy by 4.25% and F1 Score by 4.39%, demonstrating the effectiveness of integrating GCN for biomarker prediction. Moreover, the results were interpreted using attention heatmap. Conclusion CRC-BERT-GCN may demonstrate superior performance in predicting CRC biomarkers, and outperforming baseline models, highlighting its potential for advancing personalized treatment strategies through improved molecular target prediction.

Background Chronic psychological stress is an increasingly recognized risk factor for both renal and cardiovascular diseases, yet the mechanisms linking stress exposure to vascular remodeling remain poorly defined. We aimed to investigate whether renal pathological markers can predict abdominal aortic microstructural changes under chronic unpredictable mild stress (CUMS) and to evaluate the predictive value of these markers using machine learning models in both normal and metabolically compromised murine models. Methods A total of 60 male mice, including 30 wild-type (WT) C57BL/6J and 30 ApoE-deficient mice (ApoE-/-) were randomized into four groups (n =15 per group): WT+ normal diet (ND), WT + ND + CUMS, ApoE-/- + western diet (WD), and ApoE-/- + WD + CUMS. CUMS was administered for 12 weeks. Histopathological assessment of renal and aortic tissues was performed using hematoxylin and eosin (H&E), periodic acid-Schiff (PAS), reticular fiber, and immunohistochemical staining. Expression of sodium-glucose cotransporters (SGLT1 and SGLT2), edema, glycogen content, and abdominal aortic intima-media thickness (AA-IMT) were quantified. Correlation analyses were conducted between renal and vascular parameters. Two machine learning models—support vector machine (SVM) and backpropagation (BP) neural network—were used to evaluate the predictive relationship between renal biomarkers and aortic microstructural alterations. Results CUMS significantly increased renal SGLT1/SGLT2 expression, edema, and glycogen content in wild-type mice (all P < 0.01), along with increased AA-IMT and reduced aortic diameter (P < 0.05). ApoE-/- mice exhibited pronounced aortic remodeling under CUMS, despite no further renal deterioration. Correlation analysis showed a strong association between renal and vascular indicators, particularly in the CUMS + ApoE-/- group. Machine learning models demonstrated robust performance in predicting abdominal aorta diameter (AA-D) and AA-IMT based on SGLT1 and SGLT2 levels (SVM R2 for AA-D: 0.79, BP network correlation for AA-IMT: 0.99), identifying these markers as early indicators of stress-induced vascular remodeling. Conclusion Renal injury biomarkers, particularly SGLT1 and SGLT2, can serve as reliable predictors of abdominal aortic microstructural changes under chronic stress. Integration of machine learning models enhances early risk stratification and provides a novel approach for mechanistic insight into cardiorenal stress responses.

Background Regularly updating literature evidence for clinical practice guidelines (CPGs) is necessary but time intensive. Artificial intelligence (AI) tools, such as Covidence, may accelerate literature screening. We prospectively evaluated the effectiveness and accuracy of two Covidence functions: the auto-marking filter (adopted from Cochrane’s RCT classifier) for identifying randomized controlled trials (RCTs) and machine learning-assisted title and abstract (Stage I) screening in two cancer-related CPGs. Methods We updated the literature searches for the breast and lung cancer CPGs at the Program in Evidence-Based Care (PEBC), Ontario Health (Cancer Care Ontario). Two methodologists trained the machine learning model by indicating which of the first 25 references were tagged by Covidence’s RCT filter. Main outcomes included workload and time savings, work saved over sampling (WSS), and the impact on whether the original guideline recommendations required changes. Results A total of 1,270 (breast cancer) and 1,734 (lung cancer) references were imported into Covidence. Among them, 633 breast cancer and 656 lung cancer references were tagged using Covidence’s RCT filter as possible RCTs and required review at Stage I. Covidence’s RCT filter excluded non-RCTs with high accuracy, reducing manual screening by 50.2% (breast cancer) and 62.2% (lung cancer). For Stage I screening, Covidence reduced workload by 48.3% and 55.2% at 95% sensitivity, and by 44.2% and 16.9% at 100% sensitivity, respectively. WSS values aligned with these reductions. Owing to the small number of references screened and time needed to become initially trained to use Covidence, time savings was minimal. At 95% sensitivity, one included reference per CPG was missed, but guideline recommendations remained unchanged. Conclusion Covidence AI-assisted screening may effectively support updating literature reviews for oncology CPGs by reducing workload without compromising the integrity of the final recommendations. A 95% detection threshold may provide a practical balance between efficiency and accuracy.