BMC Medical Informatics and Decision Making最新文献

{"title":"Remote clinical decision support tool for Parkinson's disease assessment using a novel approach that combines AI and clinical knowledge.","authors":"Harel Rom, Ori Peleg, Yovel Rom, Anat Mirelman, Gaddi Blumrosen, Inbal Maidan","doi":"10.1186/s12911-025-03104-6","DOIUrl":"10.1186/s12911-025-03104-6","url":null,"abstract":"<p><strong>Background: </strong>Early diagnosis of Parkinson's disease (PD) can assist in designing efficient treatments. Reduced facial expressions are considered a hallmark of PD, making advanced artificial intelligence (AI) image processing a potential non-invasive clinical decision support tool for PD detection. This study aims to determine the sensitivity of image-to-text AI, which matches facial frames recorded in home settings with descriptions of PD facial expressions, in identifying patients with PD.</p><p><strong>Methods: </strong>Facial image of 67 PD patients and 52 healthy-controls (HCs) were collected via standard video recording. Using clinical knowledge, we compiled descriptive sentences detailing facial characteristics associated with PD. The facial images were analyzed with OpenAI's CLIP model to generate probability scores, indicating the likelihood of each image matching the PD-related descriptions. These scores were used in an XGBoost model to identify PD patients based on the total, motor, and facial-expression item of the MDS-UPDRS, a common scale for assessing disease severity.</p><p><strong>Results: </strong>The image-to-text AI technology showed the best results in identifying PD patients based on the facial expression item (AUC = 0.78 ± 0.05), especially for those with 'mild' facial symptoms (AUC = 0.87 ± 0.04). The motor MDS-UPDRS score followed (AUC = 0.69 ± 0.05), while the total MDS-UPDRS score showed the lowest performance in identifying PD patients (AUC = 0.59 ± 0.05). PD matching probabilities between facial images and sentences revealed significant correlations across all MDS-UPDRS components (r > 0.23, p < 0.0001).</p><p><strong>Conclusions: </strong>Our results demonstrate the feasibility of using advanced AI in a clinical decision support tool for PD diagnosis, suggesting a novel approach for home-based screening to identify PD patients. This method represents a significant innovation, transforming clinical knowledge into practical algorithms that can serve as effective screening tools.</p><p><strong>Clinical trial number: </strong>MOH_2023-04-16_012535.</p>","PeriodicalId":9340,"journal":{"name":"BMC Medical Informatics and Decision Making","volume":"25 1","pages":"294"},"PeriodicalIF":3.8,"publicationDate":"2025-08-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12329937/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144798216","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Dynamic frailty risk prediction in elderly hip replacement: a deep learning approach to personalized rehabilitation.","authors":"Xujing Lv, Hongmei Li, Yue Li, Ruibing Zhuo, Yiting Yue, Ying Wang, Xiaoyun Zheng, Huanling Gao","doi":"10.1186/s12911-025-03143-z","DOIUrl":"10.1186/s12911-025-03143-z","url":null,"abstract":"<p><strong>Background: </strong>Osteoarthritis and related degenerative conditions in the elderly often necessitate hip replacement surgery. Frailty is common in this population and significantly increases the risk of postoperative complications and delayed recovery. Accurate prediction of postoperative frailty risk and its temporal progression is essential for guiding personalized rehabilitation strategies.</p><p><strong>Methods: </strong>This study prospectively included 647 patients aged 60 years or older who underwent hip replacement surgery at the Affiliated Hospital of Shanxi Medical University between June 2021 and December 2023. Clinical, biochemical, demographic, and surgical data were collected at preoperative and postoperative stages. To mitigate sample size limitations, data augmentation was applied, expanding the dataset to approximately 2,500 cases for model training. Seven survival analysis models-Cox-Time, DeepHit, DeepSurv, MP-RSF, MP-AdaBoost, MP-LogitR-were employed to dynamically predict frailty risk over time. Model performance was evaluated using the C-index and Brier score. Model interpretability was assessed using SHAP analysis.</p><p><strong>Results: </strong>DeepSurv demonstrated the highest predictive performance (C-index = 0.95, Brier score = 0.03), while MP-RSF performed less optimally (C-index = 0.77). The predicted frailty risk peaked around postoperative day 30 and declined by day 90. SHAP analysis identified low-density lipoprotein cholesterol (LDL-C), age, body mass index (BMI), and surgical indication as key contributors to frailty prediction across models.</p><p><strong>Conclusion: </strong>The findings of this study suggest that the DeepSurv model may more accurately predict the postoperative frailty trajectory than other models. Identifying high-risk periods and key clinical predictors enables clinicians to implement timely, individualized interventions that may reduce frailty risk and improve functional recovery.</p>","PeriodicalId":9340,"journal":{"name":"BMC Medical Informatics and Decision Making","volume":"25 1","pages":"292"},"PeriodicalIF":3.8,"publicationDate":"2025-08-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12330151/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144793547","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Hybrid CNN-Transformer-WOA model with XGBoost-SHAP feature selection for arrhythmia risk prediction in acute myocardial infarction patients.","authors":"Li Li, Wenjun Ren, Yuying Lei, Lixia Xu, Xiaohui Ning","doi":"10.1186/s12911-025-03127-z","DOIUrl":"10.1186/s12911-025-03127-z","url":null,"abstract":"<p><strong>Background: </strong>Arrhythmia is a frequent and serious complication of acute myocardial infarction (AMI), leading to higher mortality. Early prediction is critical for timely intervention, but existing methods are limited by poor accuracy and low clinical applicability.</p><p><strong>Methods: </strong>We developed a novel hybrid model integrating convolutional neural network (CNN), Transformer, and Whale Optimization Algorithm (WOA) for arrhythmia prediction in AMI patients. A two-stage feature selection using XGBoost and SHAP identified the top 10 clinical predictors from 45 variables. The model was trained and validated using stratified 10-fold cross-validation on a retrospective cohort of 2,084 patients. Performance was compared with traditional machine learning and deep learning baselines using accuracy, AUC-ROC, F1-score, MCC, and G-Mean.</p><p><strong>Results: </strong>The CNN-Transformer-WOA model achieved an accuracy of 92.4%, an AUC-ROC of 0.96, and an F1-score of 0.91, outperforming all baseline models (p < 0.01). Ablation studies showed that combining CNN and Transformer improved predictive power and that WOA-based hyperparameter tuning further enhanced robustness. The model maintained stable performance across subgroups and demonstrated low inference latency (<8 ms per case). SHAP-based analysis provided interpretable clinical insights.</p><p><strong>Conclusion: </strong>This study presents an accurate, interpretable, and robust deep learning solution for arrhythmia prediction in AMI patients. The framework enables real-time, evidence-based risk stratification, and is suitable for integration into clinical decision support systems, offering practical value for improving patient care in real-world hospital environments.</p><p><strong>Clinical trial number: </strong>(No.: ChiCTR2100041960).</p>","PeriodicalId":9340,"journal":{"name":"BMC Medical Informatics and Decision Making","volume":"25 1","pages":"291"},"PeriodicalIF":3.8,"publicationDate":"2025-08-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12330184/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144793548","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Toward a fair, gender-debiased classifier for the diagnosis of attention deficit/hyperactivity disorder- a Machine-Learning based classification study.","authors":"Susanne Neufang, Feifei Li, Atae Akhrif, Oya D Beyan","doi":"10.1186/s12911-025-03126-0","DOIUrl":"10.1186/s12911-025-03126-0","url":null,"abstract":"<p><strong>Background: </strong>Attention deficit/hyperactivity disorder (ADHD) is the most common neurodevelopmental disorder. Gender disparities in the diagnosis of ADHD have been reported, suggesting that females tend to be diagnosed later in life than males are. The delayed diagnosis in females has been attributed to an inequality in the diagnostic criteria, failing to focus on the gender differences regarding symptomatology, comorbidity, and societal factors contributing to this disparity.</p><p><strong>Methods: </strong>In this study, we introduced debiased classifiers for the diagnosis of ADHD via different bias mitigation algorithms of the AI Fairness 360 toolbox on a training dataset of 400 children and adolescents with and without ADHD (98 females, 25 ADHD patients, 73 typically developing females), a subsample of the Child Mind Institute dataset. Test data were acquired in an earlier study. Two datasets were used, one including personal characteristic features, scores of the clinical questionnaire Child Behavior Checklist, and wavelet variance coefficients as quantifiers of neural dynamics (fMRI), a second dataset included personal characteristic features, scores of the clinical questionnaire Child Behavior Checklist, and radiomic features of neural structure (sMRI).</p><p><strong>Results: </strong>We found that the reweighed XGBoost model achieved the best accuracy and highest fairness in both datasets. Using model explanation, we showed how reweighing influenced feature importance at the global and local levels.</p><p><strong>Conclusion: </strong>Based on methodological characteristics and insights from global and local model explana-tion, we discuss the reasons of these findings and conclude, that using the combination of bias mitigation and model explanation, improved classification models can be achieved.</p>","PeriodicalId":9340,"journal":{"name":"BMC Medical Informatics and Decision Making","volume":"25 1","pages":"290"},"PeriodicalIF":3.8,"publicationDate":"2025-08-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12326834/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144788363","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Evaluation of large language models in patient education and clinical decision support for rotator cuff injury: a two-phase benchmarking study.","authors":"Yi-Lin Wang, Li-Chao Tian, Jing-Yuan Meng, Jie-Chao Zhang, Zhi-Xing Nie, Wen-Rui Wei, Dao-Fang Ding, Xiao-Ye Tang, Qian Zhang, Yong He","doi":"10.1186/s12911-025-03105-5","DOIUrl":"10.1186/s12911-025-03105-5","url":null,"abstract":"<p><strong>Objective: </strong>This study evaluates the accuracy of ChatGPT-4o, ChatGPT-o1, Gemini, and ERNIE Bot in answering rotator cuff injury questions and responding to patients. Results show Gemini excels in accuracy, while ChatGPT-4o performs better in patient interactions.</p><p><strong>Methods: </strong>Phase 1: Four LLM chatbots answered physician test questions on rotator cuff injuries, interacting with patients and students. Their performance was assessed for accuracy and clarity across 108 multiple-choice and 20 clinical questions. Phase 2: Twenty patients questioned the top two chatbots (ChatGPT-4o, Gemini), with responses rated for satisfaction and readability. Three physicians evaluated accuracy, usefulness, safety, and completeness using a 5-point Likert scale. Statistical analyses and plotting used IBM SPSS 29.0.1.0 and Prism 10; Friedman test compared evaluation and readability scores among chatbots with Bonferroni-corrected pairwise comparisons, Mann-Whitney U test compared ChatGPT-4o versus Gemini; statistical significance at p < 0.05.</p><p><strong>Results: </strong>Gemini achieved the highest average accuracy. In the second part, Gemini showed the highest proficiency in answering rotator cuff injury-related queries (accuracy: 4.70; completeness: 4.72; readability: 4.70; usefulness: 4.61; safety: 4.70, post hoc Dunnett test, p < 0.05). Additionally, 20 rotator cuff injury patients questioned the top two models from Phase 1 (ChatGPT-4o and Gemini). ChatGPT-4o had the highest reading difficulty score (14.22, post hoc Dunnett test, p < 0.05), suggesting a middle school reading level or above. Statistical analysis showed significant differences in patient satisfaction (4.52 vs. 3.76, p < 0.001) and readability (4.35 vs. 4.23). Orthopedic surgeons rated ChatGPT-4o higher in accuracy, completeness, readability, usefulness, and safety (all p < 0.05), outperforming Gemini in all aspects.</p><p><strong>Conclusion: </strong>The study found that LLMs, particularly ChatGPT-4o and Gemini, excelled in understanding rotator cuff injury-related knowledge and responding to patients, showing strong potential for further development.</p>","PeriodicalId":9340,"journal":{"name":"BMC Medical Informatics and Decision Making","volume":"25 1","pages":"289"},"PeriodicalIF":3.8,"publicationDate":"2025-08-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12323112/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144783565","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Partial proportional odds model for predicting multiple lower extremity amputation among T2DM patients.","authors":"Sabiha Khan, Karuna Reddy, Momtaz Ahmed, Donald Wilson, Bibhya Sharma","doi":"10.1186/s12911-025-03112-6","DOIUrl":"10.1186/s12911-025-03112-6","url":null,"abstract":"<p><strong>Background or introduction: </strong>Multiple Lower extremity amputation (MLEA) is an unfortunate outcome following a lower extremity amputation (LEA) in individuals with diabetes. The challenges faced by individual with MLEA are significantly higher than those who have undergone a single amputation. Therefore, developing a reliable and accurate method for determining risk factors associated with MLEA is essential for reducing the incidence of this outcome among diabetic patients.</p><p><strong>Objectives: </strong>This study aimed to explore the demographic and clinical characteristics of diabetic inpatients with foot ulcers. The goal was to develop a statistical model to determine the risk factors of MLEA among patients type 2 diabetic mellitus (T2DM).</p><p><strong>Methods: </strong>Data for statistical model development were collected from patients' folders involving 1,972 patients with T2DM who were hospitalized for acute diabetic foot ulcers (DFU) at three tertiary care hospitals in Fiji from 2016 to 2019. This cross-sectional study was conducted in accordance with the STROBE guidelines focusing on patients who experienced MLEA at the hospitals. Patients were categorized into three ordinal outcomes: no-amputation, primary amputation, and multiple amputations. A partial proportional odds model was developed to fit the ordinal outcome and determine the risk factors associated with MLEA. The proposed model was validated by comparing it to a proportional odds model and a multinomial logistics regression model.</p><p><strong>Results: </strong>The proposed partial proportional odds model (PPOM) identified several risk factors for MLEA, including age, gender, ethnicity, hypertension, anemia, leukocytosis, and thrombocytosis.</p><p><strong>Conclusions: </strong>The analytical findings reveal that the PPOM is appropriate for determining the risk factors associated with MLEA in T2DM patients in Fiji.</p>","PeriodicalId":9340,"journal":{"name":"BMC Medical Informatics and Decision Making","volume":"25 1","pages":"287"},"PeriodicalIF":3.8,"publicationDate":"2025-08-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12323144/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144783566","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Performance of CAC-prob in predicting coronary artery calcium score: an external validation study in a high-CAC burden population.","authors":"Pakpoom Wongyikul, Phichayut Phinyo, Pannipa Suwannasom, Apichat Tantraworasin, Chanikan Srikuenkaew, Pichyapa Jira, Kempiya Pornpipatsakul, Arachaporn Ngachuea, Phansa Chanthanom, Kanogphol Prayongkul, Warisara Chavalitjiraphan, Thitirat Rattananalin, Surasak Saokaew","doi":"10.1186/s12911-025-03128-y","DOIUrl":"10.1186/s12911-025-03128-y","url":null,"abstract":"<p><strong>Background: </strong>Although CAC screening is gaining recognition in developing countries such as Thailand, official guidelines for using the CAC score in cardiovascular risk assessment remain lacking. This study aims to externally validate CAC-prob, a recently developed prediction model that can estimate the probability of CAC > 0 and CAC ≥ 100, to confirm its robustness.</p><p><strong>Method: </strong>This study externally validated the CAC-prob model using retrospective data from a tertiary care centre in northern Thailand. Patients who underwent CAC screening between 2019 and 2022 were included. CAC-prob consists of two models: one predicting the probability of CAC > 0 (Model 1) and another predicting the probability of CAC ≥ 100 (Model 2). Model performance was assessed in terms of discrimination (Ordinal C-index), calibration slope, and diagnostic indices for each model.</p><p><strong>Results: </strong>A total of 329 patients were included. The patient characteristics observed in this study indicated a higher prevalence of DM, hypertension, dyslipidaemia, CKD, and CAC ≥ 100 compared to the development study. The ordinal C-index derived from the validation study showed a slight decline (0.78). The calibration slope for Model 1 and Model 2 was 1.28 (95% CI 0.95-1.63) and 1.06 (95% CI 0.78-1.36), respectively. In Model 1, CAC-prob demonstrated comparable diagnostic performance. However, in Model 2, it showed slightly better performance, with significantly improved sensitivity compared to the development study.</p><p><strong>Conclusion: </strong>This external validation study confirms the predictive performance of CAC-prob in Northern Thai patients. The findings support the integration of CAC-prob into routine clinical practice to aid physicians in making recommendations for CAC screening.</p>","PeriodicalId":9340,"journal":{"name":"BMC Medical Informatics and Decision Making","volume":"25 1","pages":"288"},"PeriodicalIF":3.8,"publicationDate":"2025-08-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12323013/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144783567","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Advanced multi-label brain hemorrhage segmentation using an attention-based residual U-Net model.","authors":"Xinxin Lin, Enmiao Zou, Wenci Chen, Xinxin Chen, Le Lin","doi":"10.1186/s12911-025-03131-3","DOIUrl":"10.1186/s12911-025-03131-3","url":null,"abstract":"<p><strong>Objective: </strong>This study aimed to develop and assess an advanced Attention-Based Residual U-Net (ResUNet) model for accurately segmenting different types of brain hemorrhages from CT images. The goal was to overcome the limitations of manual segmentation and current automated methods regarding precision and generalizability.</p><p><strong>Materials and methods: </strong>A dataset of 1,347 patient CT scans was collected retrospectively, covering six types of hemorrhages: subarachnoid hemorrhage (SAH, 231 cases), subdural hematoma (SDH, 198 cases), epidural hematoma (EDH, 236 cases), cerebral contusion (CC, 230 cases), intraventricular hemorrhage (IVH, 188 cases), and intracerebral hemorrhage (ICH, 264 cases). The dataset was divided into 80% for training using a 10-fold cross-validation approach and 20% for testing. All CT scans were standardized to a common anatomical space, and intensity normalization was applied for uniformity. The ResUNet model included attention mechanisms to enhance focus on important features and residual connections to support stable learning and efficient gradient flow. Model performance was assessed using the Dice Similarity Coefficient (DSC), Intersection over Union (IoU), and directed Hausdorff distance (dHD).</p><p><strong>Results: </strong>The ResUNet model showed excellent performance during both training and testing. On training data, the model achieved DSC scores of 95 ± 1.2 for SAH, 94 ± 1.4 for SDH, 93 ± 1.5 for EDH, 91 ± 1.4 for CC, 89 ± 1.6 for IVH, and 93 ± 2.4 for ICH. IoU values ranged from 88 to 93, with dHD between 2.1- and 2.7-mm. Testing results confirmed strong generalization, with DSC scores of 93 for SAH, 93 for SDH, 92 for EDH, 90 for CC, 88 for IVH, and 92 for ICH. IoU values were also high, indicating precise segmentation and minimal boundary errors.</p><p><strong>Conclusions: </strong>The ResUNet model outperformed standard U-Net variants, achieving higher multi-label segmentation accuracy. This makes it a valuable tool for clinical applications that require fast and reliable brain hemorrhage analysis. Future research could investigate semi-supervised techniques and 3D segmentation to further enhance clinical use.</p><p><strong>Clinical trial number: </strong>Not applicable.</p>","PeriodicalId":9340,"journal":{"name":"BMC Medical Informatics and Decision Making","volume":"25 1","pages":"286"},"PeriodicalIF":3.8,"publicationDate":"2025-07-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12315353/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144759246","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Thyroid disease classification using generative adversarial networks and Kolmogorov-Arnold network for three-class classification.","authors":"Aysel Topşir, Ferdi Güler, Ecesu Çetin, Mehmet Furkan Burak, Melih Ağraz","doi":"10.1186/s12911-025-03014-7","DOIUrl":"10.1186/s12911-025-03014-7","url":null,"abstract":"<p><p>Thyroid disease classification is a critical challenge in medical diagnostics, requiring accurate differentiation between hyperthyroidism, hypothyroidism, and normal thyroid function. This study introduces an advanced machine learning approach that integrates generative adversarial networks (GANs) for data augmentation and Kolmogorov-Arnold networks (KANs) for classification. Various machine learning models including logistic regression, random forest, support vector machines, multilayer perceptrons, and KANs were trained and evaluated. The results indicate that the application of GAN-based data augmentation has significantly improved classification accuracy, particularly for minority classes. Specifically, the KAN model achieved an accuracy of 98.68% and random forest (RF) F1-score of 98.00%, outperforming traditional neural network applications. The results demonstrate that GAN-augmented datasets significantly improve classification accuracy, and the KAN model achieves superior performance and generalization capabilities compared to traditional neural networks. Additionally, the SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) were employed to ensure model transparency and interpretability. These explainability methods highlight thyroid stimulating hormone as the most prominent feature in classification, further supporting its clinical utility in the diagnosis of thyroid diseases. The findings underscore the potential of advanced AI-driven techniques in improving thyroid disease classification, addressing class imbalance, and enhancing explainability in healthcare applications. By leveraging synthetic data generation, this study provides a feasible framework for actual clinical application, particularly in situations where clinical data are limited or imbalanced. The integration of GANs and KANs enhances diagnostic accuracy while preserving robustness and generalizability to diverse patient populations. Besides, the approach fosters the deployment of explainable AI models in clinical decision support systems so that healthcare practitioners can make improved and more reliable decisions, thus leading to better patient outcomes and resource allocation.</p>","PeriodicalId":9340,"journal":{"name":"BMC Medical Informatics and Decision Making","volume":"25 1","pages":"284"},"PeriodicalIF":3.8,"publicationDate":"2025-07-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12315378/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144759248","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Enhanced stroke risk prediction in hypertensive patients through deep learning integration of imaging and clinical data.","authors":"Hui Li, Tianyu Zhang, Guochao Han, Zonghui Huang, Huiyu Xiao, Yunzhe Ni, Bo Liu, Wennan Lin, Yuan Lin","doi":"10.1186/s12911-025-03120-6","DOIUrl":"10.1186/s12911-025-03120-6","url":null,"abstract":"<p><strong>Background: </strong>Stroke is one of the leading causes of death and disability worldwide, with a significantly elevated incidence among individuals with hypertension. Conventional risk assessment methods primarily rely on a limited set of clinical parameters and often exclude imaging-derived structural features, resulting in suboptimal predictive accuracy.</p><p><strong>Objective: </strong>This study aimed to develop a deep learning-based multimodal stroke risk prediction model by integrating carotid ultrasound imaging with multidimensional clinical data to enable precise identification of high-risk individuals among hypertensive patients.</p><p><strong>Methods: </strong>A total of 2,176 carotid artery ultrasound images from 1,088 hypertensive patients were collected. ResNet50 was employed to automatically segment the carotid intima-media and extract key structural features. These imaging features, along with clinical variables such as age, blood pressure, and smoking history, were fused using a Vision Transformer (ViT) and fed into a Radial Basis Probabilistic Neural Network (RBPNN) for risk stratification. The model's performance was systematically evaluated using metrics including AUC, Dice coefficient, IoU, and Precision-Recall curves.</p><p><strong>Results: </strong>The proposed multimodal fusion model achieved outstanding performance on the test set, with an AUC of 0.97, a Dice coefficient of 0.90, and an IoU of 0.80. Ablation studies demonstrated that the inclusion of ViT and RBPNN modules significantly enhanced predictive accuracy. Subgroup analysis further confirmed the model's robust performance in high-risk populations, such as those with diabetes or smoking history.</p><p><strong>Conclusion: </strong>The deep learning-based multimodal fusion model effectively integrates carotid ultrasound imaging and clinical features, significantly improving the accuracy of stroke risk prediction in hypertensive patients. The model demonstrates strong generalizability and clinical application potential, offering a valuable tool for early screening and personalized intervention planning for stroke prevention.</p><p><strong>Clinical trial number: </strong>Not applicable.</p>","PeriodicalId":9340,"journal":{"name":"BMC Medical Informatics and Decision Making","volume":"25 1","pages":"285"},"PeriodicalIF":3.8,"publicationDate":"2025-07-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12315267/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144759247","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}