Journal of Medical Systems最新文献

Knowledge acquisition tools facilitate the creation and maintenance of decision support content, but thus far there is little formal investigation of the requirements and desiderata for such tools. This leads to researchers re-inventing the wheel when building knowledge acquisition tools. This study aims to draw up a requirements list for manual knowledge acquisition tools. A systematic review of the literature combined with the opinion of experts on the research team was used to write the initial requirements list. This requirements list was presented to a group of experts and revised according to their feedback. Embase, MEDLINE, IEEE, and ACM were searched, leading to a total of 1628 records. After screening, 199 records were sought for retrieval and 3 studies were included in the review. The initial requirements list consisted of 24 requirements and 15 sub-requirements, covering multiple aspects of knowledge acquisition tools. The expert consensus sessions resulted in 13 new requirements and 4 new sub-requirements, but no changes to the major categories. The final requirements list contained 37 requirements and 21 sub-requirements, divided over 4 categories. This study has shown that much is still unknown about requirements for KA tools , making it complex for developers to design effective and efficient tools. While our results are preliminary, they might prove to be a valuable starting point for developers and researchers.

Deep Learning methods have become a powerful tool in medical imaging, with great potential to improve diagnostic accuracy and support early disease detection. This is especially critical for breast cancer, one of the most common cancers among women, where early detection of abnormal tissue is crucial to improving survival rates. In this paper, we explore the application of Deep Learning techniques to segment and classify breast masses as malignant or benign using ultrasound images, aiming to support breast cancer diagnosis. We propose a modular dual-stage pipeline that first segments suspicious regions and then classifies them into benign or malignant categories. The framework is designed to flexibly integrate different backbone architectures, allowing adaptation to task- or dataset-specific requirements. Experimental results show that, within this pipeline, DeepLabV3+ with a ResNet34 encoder provided the most accurate segmentation, while lightweight classifiers such as MobileNetV3-Small and EfficientNet-B0 yielded the best classification performance. Moreover, an ablation study was conducted to tune parameters and determine their optimal configuration. Finally, our approach was tested on two breast ultrasound datasets, and the results show promising improvements in diagnostic accuracy, demonstrating the potential of our method to enhance early breast cancer detection.

Data scarcity remains a major obstacle to the application of deep learning techniques in medical imaging, particularly for rare neurodegenerative diseases. This study investigates the use of denoising diffusion probabilistic models (DDPMs) to generate synthetic 3D T1-weighted brain MRI images in this context. Addressing the dual challenges of limited training data and structural fidelity, we propose a generative pipeline trained on a multicenter dataset of healthy subjects. The model suggests the potential to produce anatomically coherent synthetic scans with realistic variability. Quantitative evaluation based on Maximum Mean Discrepancy confirms the similarity between real and generated data distributions, while visual assessments highlight the preservation of global and local brain structures. Despite limitations in high-frequency detail reconstruction, the results suggest that DDPMs hold promise as a tool for augmenting neuroimaging datasets and supporting downstream tasks such as classification and segmentation. This work lays the foundation for future research aimed at improving resolution and adapting generative models to the specific challenges of rare disease imaging.

This study explored the integration of supervised, semi-supervised, and unsupervised learning strategies to classify ophthalmic images under label-scarce conditions. Given the high cost of annotations in medical imaging, the goal was to improve diagnostic performance using minimal labeled data and robust feature representations. A dataset of 18,767 multimodal ophthalmic images was collected - 1,877 labeled and 16,890 unlabeled. Three transformer-based architectures -ViT-Base, DeiT-Base, and MaxViT-L-were used for supervised learning. Semi-supervised learning employed pseudo-labeling with a confidence threshold ≥ 0.98. For unsupervised learning, SimCLR-based contrastive learning and K-means clustering were implemented on extracted features. Performance was evaluated using classification accuracy, AUC, F1-score, clustering indices (Silhouette Score, DBI, CH Index), and computational metrics. In supervised learning, ViT-Base achieved the highest accuracy (92.47%), followed by DeiT-Base (89.38%) and MaxViT-L (85.27%). After pseudo-labeling, MaxViT-L achieved the best accuracy (97.49%) and AUC (0.9982). Contrastive learning significantly improved feature clustering, with MaxViT-L reaching a Silhouette Score of 0.556 and a reduced DBI of 0.541. However, computational analysis revealed that MaxViT-L exhibited the highest computational complexity (81,713 MFLOPs) and longest inference (~ 102 ms), while ViT-Base and DeiT-Base showed considerably lower FLOPs (39,120.6 MFLOPs) and faster inference (~ 52 ms). On external validation set, MaxViT demonstrated the best overall performance. Although ViT-Base achieved the highest accuracy in supervised training, MaxViT-L demonstrated the most favorable trade-off between performance and model generalization in semi- and unsupervised settings, Despite its higher computational complexity and longer inference time, MaxViT-L consistently achieved strong accuracy and clustering performance. This approach minimizes dependence on expert annotations, supporting scalable and automated ophthalmic diagnosis.

Chronic disease management requires continuous monitoring, lifestyle modification and therapy adherence, thus requiring constant support from healthcare professionals. Chatbots have proven to be a promising approach for engaging patients in managing their health condition at home and for offering continuous assistance by being readily available to answer questions. While large language models offer an impressive solution for chatbot implementation, third-party systems raise privacy concerns, and computational requirements limit small-scale deployment. We address these challenges by developing a chatbot for hypertensive patients based on open-source small language models (SLMs), specifically designed for running on personal resource-constrained devices and for providing assistance in QA tasks. In order to guarantee comparable conversational performances with respect to larger language models, we exploited retrieval-augmented generation (RAG) with a local knowledge base. This ensures data privacy by deploying models locally while achieving competitive accuracy and maintaining low computational costs suitable for end-user devices. We experimented with eight SLMs, two prompt configurations, and different RAG strategies - both in the embedding and retrieval components - to identify the most effective solution. The evaluation of our solution grounds on both reference metrics and expert evaluation. Our findings suggest that RAG-enhanced SLMs can improve response clarity and content accuracy. However, our results also indicate that newer SLMs like Qwen3 demonstrate strong performance even without RAG, suggesting a potential shift in the necessity for complex retrieval mechanisms with rapidly evolving model architectures.

The cost-effective open-source artificial intelligence (AI) model DeepSeek-R1 in China holds significant potential for healthcare applications. As a health education tool, it could help patients acquire health science knowledge and improve health literacy. Low back pain (LBP), the most common musculoskeletal problem globally, has seen increasing use of large language model (LLM)-based AI chatbots by patients to access health information, making it critical to further examine the quality of such information. This study aimed to evaluate the response quality and readability of answers generated by DeepSeek-R1 to common patient questions about LBP. Ten questions were formulated using inductive methods based on literature analysis and Baidu Index data, which were presented to DeepSeek-R1 on March 10, 2025. The evaluation spanned readability, understandability, actionability, clinician assessment, and reference assessment. Readability was measured using the Flesch-Kincaid Grade Level, Flesch Reading Ease Scale, Gunning Fog Index, Coleman-Liau Index, and Simple Measure of Gobbledygook (SMOG Index). Understandability and actionability were assessed via the Patient Education Materials and Assessment Tool for Printable Materials (PEMAT-P). Clinicians evaluated accuracy, completeness, and correlation. A reference evaluation tool was used to assess reference quality and the presence of hallucinations. Readability analysis indicated that DeepSeek's responses were overall "difficult to read", with Flesch-Kincaid Grade Level (mean 12.39, SD 1.91), Flesch Reading Ease Scale (mean 19.55, Q1 12.94, Q3 29.78), Gunning Fog Index (mean 13.95, SD 2.61), Coleman-Liau Index (mean 17.46, SD 2.30), and SMOG Index (mean 11.04, SD 1.37). PEMAT-P revealed good understandability but weak actionability. Consensus among five clinicians confirmed satisfactory accuracy, completeness, and relevance. References Assessment identified 9 instances (14.8%) of hallucinated references, while Supporting was rated as moderate, with most references sourced from authoritative platforms. Our study demonstrates the potential of DeepSeek-R1 in the educational content for patients with LBP. It can be employed as a supplement to patient education tools rather than substituting for clinical judgment.

Zero-shot anomaly detection (ZSAD) is gaining traction in medical imaging as a way to identify abnormalities without task-specific supervision. In this work, we benchmark state-of-the-art CLIP-based ZSAD models -originally developed for industrial inspection -on brain metastasis detection using the BraTS-MET dataset. We evaluate both general-purpose and medical-adapted variants across multiple training paradigms with little to no supervision, emulating real-world scenarios with scarce labeled imaging data. While the models can apply general knowledge to medical images, we show that their accuracy remains limited, especially in peripheral brain regions, and that substantial but still suboptimal performance gains are achieved only via domain-specific fine-tuning. Our findings highlight current limitations in spatial consistency when using 2D-based approaches for 3D problems, and suggest that adaptation is required to make CLIP-based ZSAD viable for clinical use.

Background: Assessing Left Ventricular Ejection Fraction (LVEF) is crucial for diagnosing reduced systolic function, yet echocardiography (ECHO) may not always be readily available, potentially delaying treatment. Electrocardiography (ECG) offers a cost-effective and accessible alternative for estimating LVEF. However, specialized AI models for this purpose are often complex and costly to develop.

Objective: This study uniquely evaluates GPT-4o Fine-Tuned Vision (GPT-4o-FTV), a general-purpose AI model, for detecting LVEF ≤ 35% from ECG images, comparing its performance with a Convolutional Neural Network (CNN) model and clinician assessments.

Methods: We analyzed ECGs from 202 patients (42.6% women, mean age 64.5 ± 16.3 years) at a tertiary center, excluding those with pacemakers and including only high-quality ECGs. LVEF ≤ 35% was present in 11.9% (n = 24). GPT-4o-FTV, trained on 20 labeled ECGs, was tested using a structured prompt across four runs. Accuracy, sensitivity, specificity, and positive predictive value (PPV) were compared to a CNN model and four clinicians.

Results: GPT-4o-FTV achieved 79.9% accuracy, 72.9% sensitivity, 80.8% specificity, an F1-score of 46.4%, and a PPV of 34%, outperforming clinicians (74.9% accuracy, 65.6% sensitivity, 76.1% specificity, 39% F1-score, PPV 27.9%). The CNN model had the highest performance (89.1% accuracy, 79.2% sensitivity, 90.4% specificity, 63.3% F1-score, PPV 52.8%).

Conclusions: GPT-4o-FTV demonstrates strong potential as an accessible tool for cardiac diagnostics, particularly in resource-limited settings. While CNN models remain superior in accuracy, the ease of fine-tuning GPT-4o-FTV highlights its practical utility. Future research should focus on larger datasets, additional optimization, and exploring its ability to detect early predictors of LVEF decline.

Public health increasingly relies on digital infrastructures, yet data remains fragmented across clinical, behavioral, and social domains. Customer Data Platforms (CDPs), originally created in marketing to unify diverse information into dynamic individual profiles, could provide a new approach for person-centered public health. This article explores the strategic potential of applying CDP principles, such as data unification, identity resolution, segmentation, and timely intervention, to enhance surveillance, prevention, and chronic disease management. A conceptual framework is presented and demonstrated through a breast cancer screening scenario, illustrating how CDPs could enable personalized outreach and integration with artificial intelligence (AI). Although promising, there are significant challenges related to privacy, interoperability, fairness, and governance. Responsible deployment requires socio-technical strategies that emphasize transparency, ethical oversight, and person involvement.