OBJECTIVE: To develop a user-friendly visualization application for the automatic annotation of Human Phenotype Ontology (HPO) terms based on large language models and retrieval-augmented generation (RAG) technology, and to validate its performance in an authoritative case dataset. METHODS: By integrating the domestic open-source large language model DeepSeek-V3 with RAG technology, an interactive web application was deployed on the Streamlit cloud platform. Using only the latest official HPO dataset as the data source, the lightweight sentence-embedding model BAAI/bge-small-en-v1.5 was employed to construct a FAISS vector index. During the online phase, a four-step closed-loop process is automatically completed: multilingual translation, phenotype phrase extraction, RAG candidate retrieval, term mapping, and official database validation. 121 English case reports publicly released by BMJ Case Reports and Oxford Medical Case Reports (with a gold-standard HPO set of 1 794 terms) were selected for application validation. Precision, recall, and F1 score were calculated and compared horizontally with traditional dictionary tools, standalone large language models, and the similar application "RAG-HPO". Finally, replace the model with the more advanced ChatGPT-5 and evaluate its performance on the newly extracted dataset. RESULTS: An HPO term automatic annotation visualization application named PhenoRAG, based on large language models and RAG technology, was successfully developed. Users can access it directly via a web link. Across the 112 cases, a total of 2 150 HPO terms were generated; 2,064 (96.0%) were fully validated by the official database, with a hallucination rate of 1.3% and an HPO ID-name mismatch rate of 2.7%. After deduplication, 1,906 terms remained for testing. The overall precision was 63.65%, recall was 67.34%, and F1 was 65.44%, significantly outperforming traditional annotation tools (F1: 0.45-0.49, P < 0.001). Although PhenoRAG's F1 was lower than that of RAG-HPO (F1 = 0.78, P < 0.001), which relies on a manually constructed synonym database of 54 000 entries plus the HPO dataset, it requires no additional dictionary maintenance and can be used without any background in computer programming. Moreover, after switching to the GPT-5 model, PhenoRAG exhibited no hallucination rate on the new dataset, and its F1 score significantly increased (P = 0.038). CONCLUSION Without constructing a synonym database, the PhenoRAG achieved high-accuracy automatic mapping from clinical text to standard HPO terms. It features a low usage threshold, free access, and a Chinese-language interface, and can directly serve rare disease diagnosis, genetic counseling, and research scenarios in China and worldwide, warranting further clinical promotion and multicenter validation.
Humans ; Phenotype ; Biological Ontologies ; Language ; Software ; Large Language Models

OBJECTIVE: To explore the application value of artificial intelligence (AI)-assisted chromosomal karyotype analysis in the diagnosis of prenatal chromosomal mosaicism. METHODS: A retrospective analysis was conducted on 172 pregnant women who underwent amniocentesis at the Department of Medical Genetics and Prenatal Diagnosis, the Third Affiliated Hospital of Zhengzhou University between January 2019 and December 2024. All cases whose fetuses were diagnosed with chromosomal mosaicism via karyotype analysis and stratified into two groups based on the analytical software employed: the conventional analysis group (n = 70), which utilized Leica analysis software for karyotype image recognition and cell counting; and the AI-assisted analysis group (n = 102), which utilized AI-assisted software for the same procedures. The clinical performance of AI-assisted karyotype analysis in diagnosing chromosomal mosaicism was comprehensively evaluated by comparing the types of mosaic karyotypes, distribution of mosaic ratios, and verification outcomes of different detection modalities between the two groups. This study was approved by the Medical Ethics Committee of the Third Affiliated Hospital of Zhengzhou University (Ethics No.: 2024-406-01). RESULTS: No statistically significant difference was observed in baseline characteristics (maternal age, gestational week, and indications for prenatal diagnosis) between the two groups. Regarding the detection efficacy for numerical and structural mosaicisms, no significant difference was found in the detection of numerical mosaicism. However, the conventional analysis group exhibited a significantly higher detection rate of autosomal structural mosaicism compared to the AI-assisted group (11.43% vs. 0.98%, P < 0.05). Numerical mosaicism cases were further verified using copy number variation sequencing (CNV-seq) and/or fluorescence in situ hybridization (FISH). The AI-assisted group demonstrated a significantly lower inconsistency rate (5.56% vs. 20.41%, P < 0.05) compared to the conventional group. For low-proportion (< 10%) chromosomal mosaicism, the AI-assisted group had a significantly lower detection rate (13.25% vs. 29.69%, P < 0.05). Subsequent validation of low-proportion mosaicism by CNV-seq and/or FISH showed a higher consistency rate in the AI-assisted group (81.82% vs. 54.55%), though the difference did not reach statistical significance (P = 0.360). CONCLUSION For the karyotyping analysis of prenatal chromosomal mosaicism, AI-assisted karyotype analysis shows high accuracy and consistency in identifying numerical chromosomal mosaicism, particularly in reducing the detection of low-proportion (< 10%) mosaicism while improving verification accuracy. AI-assisted analysis can significantly improve the detection accuracy of numerical mosaicism and mitigate the risk of misclassification for low-proportion (< 10%) mosaicism, thereby providing more precise clinical evidence for the prenatal diagnosis of chromosomal mosaicisms.
Humans ; Female ; Mosaicism ; Pregnancy ; Karyotyping/methods* ; Artificial Intelligence ; Prenatal Diagnosis/methods* ; Adult ; Retrospective Studies ; Chromosome Disorders/genetics* ; Amniocentesis

Human ; Writing ; Artificial Intelligence

INTRODUCTION: Obstructive sleep apnoea (OSA) is a serious but underdiagnosed condition. Demand for the gold standard diagnostic polysomnogram (PSG) far exceeds its availability. More efficient diagnostic methods are needed, even in tertiary settings. Machine learning (ML) models have strengths in disease prediction and early diagnosis. We explored the use of ML with oximetry, demographic and anthropometric data to diagnose OSA. METHODS: A total of 2,996 patients were included for modelling and divided into test and training sets. Seven commonly used supervised learning algorithms were trained with the data. Sensitivity (recall), specificity, positive predictive value (PPV) (precision), negative predictive value, area under the receiver operating characteristic curve (AUC) and F1 measure were reported for each model. RESULTS: In the best performing four-class model (neural network model predicting no, mild, moderate or severe OSA), a prediction of moderate and/or severe disease had a combined PPV of 94%; one out of 335 patients had no OSA and 19 had mild OSA. In the best performing two-class model (logistic regression model predicting no-mild vs. moderate-severe OSA), the PPV for moderate-severe OSA was 92%; two out of 350 patients had no OSA and 26 had mild OSA. CONCLUSION Our study showed that the prediction of moderate-severe OSA in a tertiary setting with an ML approach is a viable option to facilitate early identification of OSA. Prospective studies with home-based oximeters and analysis of other oximetry variables are the next steps towards formal implementation.
Humans ; Oximetry/methods* ; Sleep Apnea, Obstructive/diagnosis* ; Male ; Female ; Middle Aged ; Machine Learning ; Polysomnography ; Adult ; Anthropometry ; ROC Curve ; Aged ; Algorithms ; Predictive Value of Tests ; Sensitivity and Specificity ; Neural Networks, Computer ; Demography

INTRODUCTION: Radiology plays an integral role in fracture detection in the emergency department (ED). After hours, when there are fewer reporting radiologists, most radiographs are interpreted by ED physicians. A minority of these interpretations may miss diagnoses, which later require the callback of patients for further management. Artificial intelligence (AI) has been viewed as a potential solution to augment the shortage of radiologists after hours. We explored the efficacy of an AI solution in the detection of appendicular and pelvic fractures for adult radiographs performed after hours at a general hospital ED in Singapore, and estimated the potential monetary and non-monetary benefits. METHODS: One hundred and fifty anonymised abnormal radiographs were retrospectively collected and fed through an AI fracture detection solution. The radiographs were re-read by two radiologist reviewers and their consensus was established as the reference standard. Cases were stratified based on the concordance between the AI solution and the reviewers' findings. Discordant cases were further analysed based on the nature of the discrepancy into overcall and undercall subgroups. Statistical analysis was performed to evaluate the accuracy, sensitivity and inter-rater reliability of the AI solution. RESULTS: Ninety-two examinations were included in the final study radiograph set. The AI solution had a sensitivity of 98.9%, an accuracy of 85.9% and an almost perfect agreement with the reference standard. CONCLUSION An AI fracture detection solution has similar sensitivity to human radiologists in the detection of fractures on ED appendicular and pelvic radiographs. Its implementation offers significant potential measurable cost, manpower and time savings.
Humans ; Singapore ; Emergency Service, Hospital ; Fractures, Bone/diagnostic imaging* ; Artificial Intelligence ; Retrospective Studies ; Adult ; Male ; Female ; Cost Savings ; Middle Aged ; Pelvic Bones/diagnostic imaging* ; Reproducibility of Results ; Aged ; Sensitivity and Specificity ; Radiography

INTRODUCTION: In this study, we aimed to compare the performance of a convolutional neural network (CNN)-based deep learning model that was trained on a dataset of normal and abnormal paediatric elbow radiographs with that of paediatric emergency department (ED) physicians on a binomial classification task. METHODS: A total of 1,314 paediatric elbow lateral radiographs (patient mean age 8.2 years) were retrospectively retrieved and classified based on annotation as normal or abnormal (with pathology). They were then randomly partitioned to a development set (993 images); first and second tuning (validation) sets (109 and 100 images, respectively); and a test set (112 images). An artificial intelligence (AI) model was trained on the development set using the EfficientNet B1 network architecture. Its performance on the test set was compared to that of five physicians (inter-rater agreement: fair). Performance of the AI model and the physician group was tested using McNemar test. RESULTS: The accuracy of the AI model on the test set was 80.4% (95% confidence interval [CI] 71.8%-87.3%), and the area under the receiver operating characteristic curve (AUROC) was 0.872 (95% CI 0.831-0.947). The performance of the AI model vs. the physician group on the test set was: sensitivity 79.0% (95% CI: 68.4%-89.5%) vs. 64.9% (95% CI: 52.5%-77.3%; P = 0.088); and specificity 81.8% (95% CI: 71.6%-92.0%) vs. 87.3% (95% CI: 78.5%-96.1%; P = 0.439). CONCLUSION The AI model showed good AUROC values and higher sensitivity, with the P-value at nominal significance when compared to the clinician group.
Humans ; Deep Learning ; Child ; Retrospective Studies ; Male ; Female ; Radiography/methods* ; ROC Curve ; Elbow/diagnostic imaging* ; Neural Networks, Computer ; Child, Preschool ; Elbow Joint/diagnostic imaging* ; Emergency Service, Hospital ; Adolescent ; Infant ; Artificial Intelligence

Hospitals worldwide recognise the importance of data and digital transformation in healthcare. We traced a smart hospital's data-driven journey to build an artificial intelligence and digital ecosystem (AIDE) to achieve healthcare excellence. We measured the impact of data and digital transformation on patient care and hospital operations, identifying key success factors, challenges, and opportunities. The use of data analytics and data science, robotic process automation, AI, cloud computing, Medical Internet of Things and robotics were stand-out areas for a hospital's data-driven journey. In the future, the adoption of a robust AI governance framework, enterprise risk management system, AI assurance and AI literacy are critical for success. Hospitals must adopt a digital-ready, digital-first strategy to build a thriving healthcare system and innovate care for tomorrow.
Artificial Intelligence ; Humans ; Delivery of Health Care ; Hospitals ; Cloud Computing ; Robotics ; Internet of Things ; Data Science

This study aims to develop an abdominal acupoint localization system based on computer vision and convolutional neural networks (CNNs). To address the challenge of abdominal acupoint localization, a multi-task CNNs architecture was constructed and trained to locate the Shenque (CV8) and human body boundaries. Based on the identified Shenque (CV8), the system further deduces key characteristics of four acupoints: Shangwan (CV13), Qugu (CV2), and bilateral Daheng (SP15). An affine transformation matrix is applied to accurately map image coordinates to an acupoint template space, achieving precise localization of abdominal acupoints. Testing has verified that this system can accurately identify and locate abdominal acupoints in images. The development of this localization system provides technical support for TCM remote education, diagnostic assistance, and advanced TCM equipment, such as intelligent acupuncture robots, facilitating the standardization and intelligent advancement of acupuncture.
Acupuncture Points ; Humans ; Deep Learning ; Abdomen/diagnostic imaging* ; Neural Networks, Computer ; Acupuncture Therapy ; Image Processing, Computer-Assisted

OBJECTIVE: To screen the population for acupuncture treatment of neck pain, using functional magnetic resonance imaging (fMRI) technology and based on machine learning algorithms. METHODS: Eighty patients with neck pain were recruited. Using FPX25 handheld pressure algometer, the tender points were detected in the areas with high-frequent onset of neck pain and high degree of acupoint sensitization. Acupuncture was delivered at 4 tender points with the lowest pain threshold, once every two days; and the treatment was given 3 times a week and for 2 consecutive weeks. The amplitude of low-frequency fluctuation (ALFF) of the brain before treatment was taken as a predictive feature to construct support vector machine (SVM), logistic regression (LR), and K-nearest neighbors (KNN) models to predict the responses of neck pain patients to acupuncture treatment. A longitudinal analysis of the ALFF features was performed before and after treatment to reveal the potential biological markers of the reactivity to the acupuncture therapy. RESULTS: The SVM model could successfully distinguish high responders (48 cases) and low responders (32 cases) to acupuncture treatment, and its accuracy rate reached 82.5%. Based on the SVM model, the ALFF values of 4 brain regions were identified as the consistent predictive features, including the right middle temporal gyrus, the right superior occipital gyrus, and the bilateral posterior cingulate gyrus. In the patients with high acupuncture response, the ALFF value in the left posterior cingulate gyrus decreased after treatment (P<0.05), whereas in the patients with low acupuncture response, the ALFF value in the right superior occipital gyrus increased after treatment (P<0.01). The longitudinal functional connectivity (FC) analysis found that compared with those before treatment, the high responders showed the enhanced FC after treatment between the left posterior cingulate gyrus and various regions, including the bilateral Crus1 of the cerebellum, the right insula, the bilateral angular gyrus, the left medial superior frontal gyrus, and the left middle cingulate gyrus (GRF: corrected, voxel level: P<0.05, mass level: P<0.05). In contrast, the low responders exhibited the enhanced FC between the left posterior cingulate gyrus and the left Crus2 of the cerebellum, the left middle temporal gyrus, the right posterior cingulate gyrus, and the left angular gyrus; besides, FC was reduced in low responders between the left posterior cingulate gyrus and the right supramarginal gyrus (GRF: corrected, voxel level: P<0.05, mass level: P<0.05). CONCLUSION This study validates the practicality of pre-treatment ALFF feature prediction for acupuncture efficacy on neck pain. The therapeutic effect of acupuncture on neck pain is potentially associated with its impact on the default mode network, and then, alter the pain perception and emotional regulation.
Humans ; Neck Pain/physiopathology* ; Acupuncture Therapy ; Female ; Male ; Adult ; Middle Aged ; Machine Learning ; Magnetic Resonance Imaging ; Young Adult ; Brain/physiopathology* ; Acupuncture Points ; Aged

OBJECTIVE: To develop an artificial intelligence-assisted system for the automatic detection of the features of common 21 auricular points based on the YOLOv8 neural network. METHODS: A total of 660 human auricular images from three research centers were collected from June 2019 to February 2024. The rectangle boxes and features of images were annotated using the LabelMe5.3.1 tool and converted them into a format compatible with the YOLO model. Using these data, transfer learning and fine-tuning training were conducted on different scales of pretrained YOLO neural network models. The model's performance was evaluated on validation and test sets, including the mean average precision (mAP) at various thresholds, recall rate (recall), frames per second (FPS) and confusion matrices. Finally, the model was deployed on a local computer, and the real-time detection of human auricular images was conducted using a camera. RESULTS: Five different versions of the YOLOv8 key-point detection model were developed, including YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x. On the validation set, YOLOv8n showed the best performance in terms of speed (225.736 frames per second) and precision (0.998). On the external test set, YOLOv8n achieved the accuracy of 0.991, the sensitivity of 1.0, and the F1 score of 0.995. The localization performance of auricular point features showed the average accuracy of 0.990, the precision of 0.995, and the recall of 0.997 under 50% intersection ration (mAP⁵⁰). CONCLUSION The key-point detection model of 21 common auricular points based on YOLOv8n exhibits the excellent predictive performance, which is capable of rapidly and automatically locating and classifying auricular points.
Humans ; Neural Networks, Computer ; Artificial Intelligence ; Acupuncture Points