Journal of Biomedical Informatics最新文献_第5页

Clinical pathway-aware large language models for reliable and transparent medical dialogue 临床路径感知大语言模型可靠和透明的医疗对话。

IF 4.5 2区医学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Biomedical Informatics

Pub Date : 2025-10-31 DOI: 10.1016/j.jbi.2025.104942

Jiageng Wu , Xian Wu , Yefeng Zheng , Jie Yang

Objective:

Large language models (LLMs) offer promising potential in answering real-time medical queries, but they often produce lengthy, generic, and even hallucinatory responses. We aim to develop a reliable and interpretable medical dialogue system that incorporates clinical reasoning and then mitigates the risk of hallucination.

Methods:

Two large datasets of real-world online consultation, MedDG and KaMed, were used for evaluation. We proposed a Medical Dialogue System with Knowledge Enhancement and Clinical Pathway Encoding (MedKP), which integrates an external medical knowledge graph and encodes internal clinical pathways to model physician reasoning. Performance was compared with state-of-the-art baselines, including GPT-4o and LLaMA3.1-70B. A multi-dimensional evaluation framework assessed (1) clinical relevance (medical entity-based), (2) textual similarity (ROUGE, BLEU), (3) semantic alignment (BERTScore), and (4) hallucination and consistency via an external LLM-based judge, as well as parallel human evaluation.

Results:

Across both datasets, MedKP (6B) achieved the best overall performance, outperforming other advanced baselines and producing responses that align more closely with those of human physicians. For clinical relevance, MedKP reached a macro F1-score of medical entity at 31.41 on MedDG (previous best DFMed: 24.76, improved 30.41%) and 26.62 on KaMed (previous best LLaM-A3.1-70B: 20.67, improved 25.62%). Consistent improvements were observed across other metrics. Ablation studies further validated the effectiveness of each model component.

Conclusion:

Our results highlight the critical role of clinical reasoning in advancing trustworthy AI for digital healthcare. By enhancing the reliability, coherence, and transparency of AI-generated responses, this pathway-aware approach bridges the gap between LLMs and real-world clinical workflows, improving the accessibility of high-quality telemedicine services, particularly benefiting underserved populations.

目的：大型语言模型（llm）在回答实时医疗查询方面提供了很好的潜力，但它们经常产生冗长、通用甚至是幻觉的响应。我们的目标是开发一个可靠的和可解释的医疗对话系统，结合临床推理，然后减轻幻觉的风险。方法：使用MedDG和KaMed两大真实在线咨询数据集进行评估。我们提出了一个具有知识增强和临床路径编码的医学对话系统（MedKP），该系统集成了外部医学知识图和编码内部临床路径来模拟医生推理。性能比较了最先进的基准，包括gpt - 40和LLaMA3.1-70B。多维评估框架评估了(1)临床相关性（基于医学实体），(2)文本相似性（ROUGE， BLEU），(3)语义一致性（BERTScore），以及(4)幻觉和一致性，通过外部基于llm的判断，以及并行的人类评估。结果：在两个数据集中，MedKP （6B）取得了最佳的总体表现，优于其他先进的基线，并产生与人类医生更接近的反应。临床相关性方面，MedKP在MedDG上达到宏观医疗实体f1评分31.41分（前最佳DFMed: 24.76分，提高30.41%），在KaMed上达到26.62分（前最佳LLaM-A3.1-70B: 20.67分，提高25.62%）。在其他指标中观察到一致的改进。消融研究进一步验证了各模型组成部分的有效性。结论：我们的研究结果强调了临床推理在推进数字医疗领域值得信赖的人工智能方面的关键作用。通过提高人工智能生成响应的可靠性、一致性和透明度，这种路径感知方法弥合了法学硕士与现实世界临床工作流程之间的差距，提高了高质量远程医疗服务的可及性，特别是使服务不足的人群受益。

{"title":"Clinical pathway-aware large language models for reliable and transparent medical dialogue","authors":"Jiageng Wu , Xian Wu , Yefeng Zheng , Jie Yang","doi":"10.1016/j.jbi.2025.104942","DOIUrl":"10.1016/j.jbi.2025.104942","url":null,"abstract":"<div><h3>Objective:</h3><div>Large language models (LLMs) offer promising potential in answering real-time medical queries, but they often produce lengthy, generic, and even hallucinatory responses. We aim to develop a reliable and interpretable medical dialogue system that incorporates clinical reasoning and then mitigates the risk of hallucination.</div></div><div><h3>Methods:</h3><div>Two large datasets of real-world online consultation, MedDG and KaMed, were used for evaluation. We proposed a Medical Dialogue System with Knowledge Enhancement and Clinical Pathway Encoding (MedKP), which integrates an external medical knowledge graph and encodes internal clinical pathways to model physician reasoning. Performance was compared with state-of-the-art baselines, including GPT-4o and LLaMA3.1-70B. A multi-dimensional evaluation framework assessed (1) clinical relevance (medical entity-based), (2) textual similarity (ROUGE, BLEU), (3) semantic alignment (BERTScore), and (4) hallucination and consistency via an external LLM-based judge, as well as parallel human evaluation.</div></div><div><h3>Results:</h3><div>Across both datasets, MedKP (6B) achieved the best overall performance, outperforming other advanced baselines and producing responses that align more closely with those of human physicians. For clinical relevance, MedKP reached a macro F1-score of medical entity at 31.41 on MedDG (previous best DFMed: 24.76, improved 30.41%) and 26.62 on KaMed (previous best LLaM-A3.1-70B: 20.67, improved 25.62%). Consistent improvements were observed across other metrics. Ablation studies further validated the effectiveness of each model component.</div></div><div><h3>Conclusion:</h3><div>Our results highlight the critical role of clinical reasoning in advancing trustworthy AI for digital healthcare. By enhancing the reliability, coherence, and transparency of AI-generated responses, this pathway-aware approach bridges the gap between LLMs and real-world clinical workflows, improving the accessibility of high-quality telemedicine services, particularly benefiting underserved populations.</div></div>","PeriodicalId":15263,"journal":{"name":"Journal of Biomedical Informatics","volume":"172 ","pages":"Article 104942"},"PeriodicalIF":4.5,"publicationDate":"2025-10-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145431672","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Multimodal large language models and mechanistic modeling for glucose forecasting in type 1 diabetes patients 1型糖尿病患者血糖预测的多模态大语言模型和机制模型。

IF 4.5 2区医学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Biomedical Informatics

Pub Date : 2025-10-31 DOI: 10.1016/j.jbi.2025.104945

J.C. Wolber , M. E. Samadi , J. Sellin , A. Schuppert

Introduction:

Management of type 1 Diabetes remains a significant challenge as blood glucose levels can fluctuate dramatically and are highly individual. We introduce an innovative approach that combines multimodal Large Language models (mLLMs), mechanistic modeling of individual glucose metabolism and machine learning (ML) for forecasting blood glucose levels.

Methods:

This study uses the D1NAMO dataset (6 patients with meal images) to demonstrate mLLM integration for glucose prediction. An mLLM (Pixtral Large) was employed to estimate macronutrients from meal images, providing automated meal analysis without manual food logging. We compare three distinct approaches: (1) Baseline using only glucose dynamics and basic insulin features, (2) LastMeal providing additional information about the last meal ingested by the patient, and (3) Bézier incorporating mechanistically modeled temporal features using optimized cubic Bézier curves to model temporal impacts of individual macronutrients on blood glucose. The modeled feature impacts served as input features for a LightGBM model. We also validate the mechanistic modeling component on the AZT1D dataset (24 patients with structured carbohydrate and correction insulin logs).

Results:

The Bézier approach achieved the best performance across both datasets: D1NAMO RMSE of 15.06 at 30 min and 28.15 at 60 min; AZT1D RMSE of 16.61 at 30 min and 24.58 at 60 min. One-way ANOVA revealed statistically significant differences across prediction horizons of 45 to 120 min for the AZT1D dataset. Patient-specific Bézier curves revealed distinct metabolic response patterns: simple sugars peaked at 0.74 h, complex sugars at 3.07 h, and proteins at 4.36 h post-ingestion. Feature importance analysis showed temporal evolution from glucose change dominance to macronutrient prominence at longer horizons. Patient-specific modeling uncovered individual metabolic signatures with varying nutritional sensitivity and circadian influences.

Conclusion:

This study demonstrates the potential of combining mLLMs with mechanistic modeling for personalized diabetes management. The optimized Bézier curve approach provides superior temporal mapping while patient-specific models reveal individual metabolic signatures essential for personalized care.

导论：1型糖尿病的管理仍然是一个重大的挑战，因为血糖水平可以剧烈波动，并且高度个体化。我们介绍了一种结合多模态大语言模型（mLLMs）、个体葡萄糖代谢机制建模和机器学习（ML）预测血糖水平的创新方法。方法：本研究使用D1NAMO数据集（6例患者膳食图像）来证明mLLM集成用于血糖预测。采用mLLM （Pixtral Large）从膳食图像中估计宏量营养素，提供自动化膳食分析而无需手动食物记录。我们比较了三种不同的方法：(1)仅使用葡萄糖动力学和基本胰岛素特征的基线，(2)LastMeal提供关于患者最后一餐摄入的额外信息，以及(3)bsamzier结合机械建模的时间特征，使用优化的立方bsamzier曲线来模拟个体宏量营养素对血糖的时间影响。建模的特征影响作为LightGBM模型的输入特征。我们还在AZT1D数据集（24例结构化碳水化合物患者和校正胰岛素日志）上验证了机制建模组件。结果：bsamzier方法在两个数据集上都取得了最好的性能：D1NAMO在30分钟和60分钟时的RMSE分别为15.06和28.15；AZT1D在30分钟和60分钟的RMSE分别为16.61和24.58。单因素方差分析显示AZT1D数据集在45至120分钟的预测范围内存在统计学上的显著差异。患者特异性bsamzier曲线显示了不同的代谢反应模式：单糖在摄入后0.74小时达到峰值，复合糖在3.07小时达到峰值，蛋白质在4.36小时达到峰值。特征重要性分析表明，在较长的时间尺度上，从葡萄糖变化主导向宏量营养素突出演化。患者特异性模型揭示了具有不同营养敏感性和昼夜节律影响的个体代谢特征。结论：本研究证明了将mllm与机制建模相结合用于个性化糖尿病管理的潜力。优化的bembrozier曲线方法提供了优越的时间映射，而患者特定的模型揭示了个性化护理所必需的个人代谢特征。

{"title":"Multimodal large language models and mechanistic modeling for glucose forecasting in type 1 diabetes patients","authors":"J.C. Wolber , M. E. Samadi , J. Sellin , A. Schuppert","doi":"10.1016/j.jbi.2025.104945","DOIUrl":"10.1016/j.jbi.2025.104945","url":null,"abstract":"<div><h3>Introduction:</h3><div>Management of type 1 Diabetes remains a significant challenge as blood glucose levels can fluctuate dramatically and are highly individual. We introduce an innovative approach that combines multimodal Large Language models (mLLMs), mechanistic modeling of individual glucose metabolism and machine learning (ML) for forecasting blood glucose levels.</div></div><div><h3>Methods:</h3><div>This study uses the D1NAMO dataset (6 patients with meal images) to demonstrate mLLM integration for glucose prediction. An mLLM (Pixtral Large) was employed to estimate macronutrients from meal images, providing automated meal analysis without manual food logging. We compare three distinct approaches: (1) <em>Baseline</em> using only glucose dynamics and basic insulin features, (2) <em>LastMeal</em> providing additional information about the last meal ingested by the patient, and (3) <em>Bézier</em> incorporating mechanistically modeled temporal features using optimized cubic Bézier curves to model temporal impacts of individual macronutrients on blood glucose. The modeled feature impacts served as input features for a LightGBM model. We also validate the mechanistic modeling component on the AZT1D dataset (24 patients with structured carbohydrate and correction insulin logs).</div></div><div><h3>Results:</h3><div>The <em>Bézier</em> approach achieved the best performance across both datasets: D1NAMO RMSE of 15.06 at 30 min and 28.15 at 60 min; AZT1D RMSE of 16.61 at 30 min and 24.58 at 60 min. One-way ANOVA revealed statistically significant differences across prediction horizons of 45 to 120 min for the AZT1D dataset. Patient-specific Bézier curves revealed distinct metabolic response patterns: simple sugars peaked at 0.74 h, complex sugars at 3.07 h, and proteins at 4.36 h post-ingestion. Feature importance analysis showed temporal evolution from glucose change dominance to macronutrient prominence at longer horizons. Patient-specific modeling uncovered individual metabolic signatures with varying nutritional sensitivity and circadian influences.</div></div><div><h3>Conclusion:</h3><div>This study demonstrates the potential of combining mLLMs with mechanistic modeling for personalized diabetes management. The optimized Bézier curve approach provides superior temporal mapping while patient-specific models reveal individual metabolic signatures essential for personalized care.</div></div>","PeriodicalId":15263,"journal":{"name":"Journal of Biomedical Informatics","volume":"172 ","pages":"Article 104945"},"PeriodicalIF":4.5,"publicationDate":"2025-10-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145431702","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Vision-language model-based semantic-guided imaging biomarker for lung nodule malignancy prediction 基于视觉语言模型的语义引导的肺结节恶性肿瘤预测成像生物标志物

IF 4.5 2区医学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Biomedical Informatics

Pub Date : 2025-10-27 DOI: 10.1016/j.jbi.2025.104947

Luoting Zhuang , Seyed Mohammad Hossein Tabatabaei , Ramin Salehi-Rad , Linh M. Tran , Denise R. Aberle , Ashley E. Prosper , William Hsu

Objective:

Machine learning models have utilized semantic features, deep features, or both to assess lung nodule malignancy. However, their reliance on manual annotation during inference, limited interpretability, and sensitivity to imaging variations hinder their application in real-world clinical settings. Thus, this research aims to integrate semantic features derived from radiologists’ assessments of nodules, guiding the model to learn clinically relevant, robust, and explainable imaging features for predicting lung cancer.

Methods:

We obtained 938 low-dose CT scans from the National Lung Screening Trial (NLST) with 1,261 nodules and semantic features. Additionally, the Lung Image Database Consortium dataset contains 1,018 CT scans, with 2,625 lesions annotated for nodule characteristics. Three external datasets were obtained from UCLA Health, the LUNGx Challenge, and the Duke Lung Cancer Screening. For imaging input, we obtained 2D nodule slices in nine directions from

50 \times 50 \times 50 mm

nodule crop. We converted structured semantic features into sentences using Gemini. We fine-tuned a pretrained Contrastive Language-Image Pretraining (CLIP) model with a parameter-efficient fine-tuning approach to align imaging and semantic text features and predict the one-year lung cancer diagnosis.

Results:

Our model outperformed the state-of-the-art (SOTA) models in the NLST test set with an AUROC of 0.901 and AUPRC of 0.776. It also showed robust results in external datasets. Using CLIP, we also obtained predictions on semantic features through zero-shot inference, such as nodule margin (AUROC: 0.807), nodule consistency (0.812), and pleural attachment (0.840).

Conclusion:

By incorporating semantic features into the vision-language model, our approach surpasses the SOTA models in predicting lung cancer from CT scans collected from diverse clinical settings. It provides explainable outputs, aiding clinicians in comprehending the underlying meaning of model predictions. The code is available at https://github.com/luotingzhuang/CLIP_nodule.

目的：机器学习模型利用语义特征、深度特征或两者同时使用来评估肺结节恶性肿瘤。然而，它们在推理过程中依赖于手动注释，有限的可解释性和对成像变化的敏感性阻碍了它们在现实世界临床环境中的应用。因此，本研究旨在整合来自放射科医生对结节评估的语义特征，指导模型学习临床相关的、稳健的、可解释的影像学特征，以预测肺癌。方法：我们从国家肺筛查试验（NLST）中获得938个低剂量CT扫描，其中有1261个结节和语义特征。此外，肺图像数据库联盟数据集包含1,018个CT扫描，其中2,625个病变注释了结节特征。从加州大学洛杉矶分校健康中心、LUNGx挑战和杜克大学肺癌筛查获得了三个外部数据集。作为成像输入，我们从50×50×50mm结节作物中获得了9个方向的二维结节切片。我们使用Gemini将结构化语义特征转换为句子。我们对预训练的对比语言-图像预训练（CLIP）模型进行了微调，采用参数有效的微调方法来对齐成像和语义文本特征，并预测一年的肺癌诊断。结果：我们的模型在NLST测试集上的AUROC为0.901，AUPRC为0.776，优于最先进（SOTA）模型。它还在外部数据集中显示了稳健的结果。使用CLIP，我们还通过零射推理获得了语义特征的预测，如结节边缘（AUROC: 0.807）、结节一致性（0.812）和胸膜附着（0.840）。结论：通过将语义特征整合到视觉语言模型中，我们的方法在从不同临床环境收集的CT扫描中预测肺癌方面优于SOTA模型。它提供了可解释的输出，帮助临床医生理解模型预测的潜在含义。代码可在https://github.com/luotingzhuang/CLIP_nodule上获得。

{"title":"Vision-language model-based semantic-guided imaging biomarker for lung nodule malignancy prediction","authors":"Luoting Zhuang , Seyed Mohammad Hossein Tabatabaei , Ramin Salehi-Rad , Linh M. Tran , Denise R. Aberle , Ashley E. Prosper , William Hsu","doi":"10.1016/j.jbi.2025.104947","DOIUrl":"10.1016/j.jbi.2025.104947","url":null,"abstract":"<div><h3>Objective:</h3><div>Machine learning models have utilized semantic features, deep features, or both to assess lung nodule malignancy. However, their reliance on manual annotation during inference, limited interpretability, and sensitivity to imaging variations hinder their application in real-world clinical settings. Thus, this research aims to integrate semantic features derived from radiologists’ assessments of nodules, guiding the model to learn clinically relevant, robust, and explainable imaging features for predicting lung cancer.</div></div><div><h3>Methods:</h3><div>We obtained 938 low-dose CT scans from the National Lung Screening Trial (NLST) with 1,261 nodules and semantic features. Additionally, the Lung Image Database Consortium dataset contains 1,018 CT scans, with 2,625 lesions annotated for nodule characteristics. Three external datasets were obtained from UCLA Health, the LUNGx Challenge, and the Duke Lung Cancer Screening. For imaging input, we obtained 2D nodule slices in nine directions from <span><math><mrow><mn>50</mn><mo>×</mo><mn>50</mn><mo>×</mo><mn>50</mn><mspace></mspace><mi>mm</mi></mrow></math></span> nodule crop. We converted structured semantic features into sentences using Gemini. We fine-tuned a pretrained Contrastive Language-Image Pretraining (CLIP) model with a parameter-efficient fine-tuning approach to align imaging and semantic text features and predict the one-year lung cancer diagnosis.</div></div><div><h3>Results:</h3><div>Our model outperformed the state-of-the-art (SOTA) models in the NLST test set with an AUROC of 0.901 and AUPRC of 0.776. It also showed robust results in external datasets. Using CLIP, we also obtained predictions on semantic features through zero-shot inference, such as nodule margin (AUROC: 0.807), nodule consistency (0.812), and pleural attachment (0.840).</div></div><div><h3>Conclusion:</h3><div>By incorporating semantic features into the vision-language model, our approach surpasses the SOTA models in predicting lung cancer from CT scans collected from diverse clinical settings. It provides explainable outputs, aiding clinicians in comprehending the underlying meaning of model predictions. The code is available at <span><span>https://github.com/luotingzhuang/CLIP_nodule</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":15263,"journal":{"name":"Journal of Biomedical Informatics","volume":"172 ","pages":"Article 104947"},"PeriodicalIF":4.5,"publicationDate":"2025-10-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145398392","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

TransDiffECG: Semantically controllable ECG synthesis via transformer-based diffusion modeling TransDiffECG：基于变压器扩散建模的语义可控心电合成。

IF 4.5 2区医学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Biomedical Informatics

Pub Date : 2025-10-27 DOI: 10.1016/j.jbi.2025.104948

Yuxin Lin , Jing Ma , Suyu Dong , Chaoyu Sun , Wanting Cong , Kuanquan Wang , Gongning Luo , Wei Wang

Objective:

Existing generative models for electrocardiogram (ECG) synthesis often lack fine-grained, interpretable control, limiting their utility for addressing data scarcity and imbalance. This study aims to develop a model capable of producing diverse and semantically controllable synthetic ECGs to fill this critical gap.

Methods:

We propose TransDiffECG, a novel Transformer-based diffusion model that integrates semantic information injection and global temporal modeling to enable fine-grained control over ECG synthesis. The model allows user-controllable generation of ECG signals with customized physiological details. We establish a comprehensive evaluation protocol, including downstream segmentation and classification tasks, to rigorously assess the authenticity and utility of the generated signals. Extensive experiments are conducted on both single-lead (QTDB) and multi-lead (LUDB) ECG datasets.

Results:

TransDiffECG significantly outperforms state-of-the-art baselines. On the multi-lead LUDB dataset, it achieved superior signal quality (MMD:

3.21 \times 1 0^{- 2}

; Pearson Correlation: 0.6177). The utility of the synthetic data was confirmed in downstream tasks, where data augmentation improved atrial fibrillation classification to an AUROC of 0.9451. Moreover, a segmentation model trained solely on our synthetic data rivaled one trained on real data (e.g.,

\sim 98 %

precision/recall on QTDB).

Conclusion:

TransDiffECG represents a significant advancement in synthetic medical signal generation by bridging the gap between clinical interpretability and generative flexibility. Its ability to generate semantically controllable and clinically valid ECGs greatly expands the application potential of generative models in healthcare research and practice.

目的：现有的心电图合成生成模型往往缺乏细粒度、可解释的控制，限制了它们在解决数据稀缺性和不平衡性方面的应用。本研究旨在开发一种能够产生多样化和语义可控的合成心电图的模型来填补这一关键空白。方法：我们提出了一种新的基于变压器的扩散模型TransDiffECG，该模型集成了语义信息注入和全局时间建模，可以对ECG合成进行细粒度控制。该模型允许用户可控地生成具有定制生理细节的心电信号。我们建立了一个全面的评估协议，包括下游分割和分类任务，以严格评估生成信号的真实性和实用性。在单导联（QTDB）和多导联（LUDB）心电数据集上进行了广泛的实验。结果：TransDiffECG显著优于最先进的基线。在多导联LUDB数据集上，它获得了优越的信号质量（MMD: 3.21×10-2; Pearson Correlation: 0.6177）。合成数据的效用在下游任务中得到证实，其中数据增强将房颤分类提高到AUROC为0.9451。此外，仅在我们的合成数据上训练的分割模型可以与在真实数据上训练的模型相媲美（例如，在QTDB上的精确度/召回率为98%）。结论：TransDiffECG通过弥合临床可解释性和生成灵活性之间的差距，代表了合成医学信号生成的重大进步。生成语义可控且临床有效的心电图的能力极大地拓展了生成模型在医疗保健研究和实践中的应用潜力。

{"title":"TransDiffECG: Semantically controllable ECG synthesis via transformer-based diffusion modeling","authors":"Yuxin Lin , Jing Ma , Suyu Dong , Chaoyu Sun , Wanting Cong , Kuanquan Wang , Gongning Luo , Wei Wang","doi":"10.1016/j.jbi.2025.104948","DOIUrl":"10.1016/j.jbi.2025.104948","url":null,"abstract":"<div><h3>Objective:</h3><div>Existing generative models for electrocardiogram (ECG) synthesis often lack fine-grained, interpretable control, limiting their utility for addressing data scarcity and imbalance. This study aims to develop a model capable of producing diverse and semantically controllable synthetic ECGs to fill this critical gap.</div></div><div><h3>Methods:</h3><div>We propose TransDiffECG, a novel Transformer-based diffusion model that integrates semantic information injection and global temporal modeling to enable fine-grained control over ECG synthesis. The model allows user-controllable generation of ECG signals with customized physiological details. We establish a comprehensive evaluation protocol, including downstream segmentation and classification tasks, to rigorously assess the authenticity and utility of the generated signals. Extensive experiments are conducted on both single-lead (QTDB) and multi-lead (LUDB) ECG datasets.</div></div><div><h3>Results:</h3><div>TransDiffECG significantly outperforms state-of-the-art baselines. On the multi-lead LUDB dataset, it achieved superior signal quality (MMD: <span><math><mrow><mn>3</mn><mo>.</mo><mn>21</mn><mo>×</mo><mn>1</mn><msup><mrow><mn>0</mn></mrow><mrow><mo>−</mo><mn>2</mn></mrow></msup></mrow></math></span>; Pearson Correlation: 0.6177). The utility of the synthetic data was confirmed in downstream tasks, where data augmentation improved atrial fibrillation classification to an AUROC of 0.9451. Moreover, a segmentation model trained solely on our synthetic data rivaled one trained on real data (e.g., <span><math><mrow><mo>∼</mo><mn>98</mn><mtext>%</mtext></mrow></math></span> precision/recall on QTDB).</div></div><div><h3>Conclusion:</h3><div>TransDiffECG represents a significant advancement in synthetic medical signal generation by bridging the gap between clinical interpretability and generative flexibility. Its ability to generate semantically controllable and clinically valid ECGs greatly expands the application potential of generative models in healthcare research and practice.</div></div>","PeriodicalId":15263,"journal":{"name":"Journal of Biomedical Informatics","volume":"172 ","pages":"Article 104948"},"PeriodicalIF":4.5,"publicationDate":"2025-10-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145400897","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Enhancing AI-based diabetic retinopathy screening in low- and middle-income countries with synthetic data 利用合成数据加强中低收入国家基于人工智能的糖尿病视网膜病变筛查。

IF 4.5 2区医学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Biomedical Informatics

Pub Date : 2025-10-24 DOI: 10.1016/j.jbi.2025.104938

Zitao Shuai , Chenwei Wu , Zhengxu Tang , David Restrepo , Michael Morley , Luis Filipe Nakayama

Objective:

AI-based DR screening is promising in low- and middle-income countries (LMICs), where limited human resources constrain access to specialist-led programs. However, current systems often degrade under real-world image-quality variations, especially with portable devices that are vital for low- and middle-income countries. This study aims to develop Retsyn, a synthetic-data augmentation framework that improves screening robustness across devices and imaging conditions.

Methods:

RetSyn leverages advanced diffusion models to generate synthetic retinal images with diverse device and imaging quality characteristics. To address the challenges of (1) portable device data scarcity, (2) disease and quality distribution imbalance, and (3) varying image quality, RetSyn uses class and quality-conditioned diffusion for controllable synthesis, a group-balanced loss to increase coverage of minority (quality, disease) pairs, and a Direct Preference Optimization alignment step with a small paired smartphone–tabletop set. The synthesized images are then used to augment classifier training.

Results:

The effectiveness of RetSyn-generated images was evaluated by training retinal diagnosis models on a combination of real and synthetic data. RetSyn yields consistent gains in-domain and out-of-domain. On low-quality tabletop images, F1 improves from 0.781 to 0.874 (binary) and 0.607 to 0.703 (three-class), while AUROC reaches 0.982 and 0.951, respectively. On out-of-domain portable images, RetSyn attains AUROC 0.813/F1 0.703 (binary) and AUROC 0.804/F1 0.609 (three-class), exceeding group-robustness baselines such as GroupDRO (binary: AUROC 0.786/F1 0.626; three-class: AUROC 0.789/F1 0.544).

Conclusion:

RetSyn presents an effective and scalable synthetic data framework that significantly enhances the robustness and generalizability of AI-based DR screening models in LMICs. By addressing the critical challenges posed by varying image quality and device characteristics, RetSyn facilitates more reliable deployment of AI diagnostics in underserved regions. Additionally, the release of the first publicly available paired smartphone-tabletop retinal image dataset will support further research into cross-device DR screening solutions.

目的：基于人工智能的DR筛查在低收入和中等收入国家（LMICs）是有希望的，在这些国家，有限的人力资源限制了获得专家主导的项目。然而，目前的系统在真实世界的图像质量变化下往往会退化，特别是对于低收入和中等收入国家至关重要的便携式设备。本研究旨在开发Retsyn，这是一种综合数据增强框架，可提高设备和成像条件下的筛选稳健性。方法：RetSyn利用先进的扩散模型生成具有不同设备和成像质量特征的合成视网膜图像。为了解决以下挑战：(1)便携式设备数据稀缺；(2)疾病和质量分布不平衡；(3)图像质量变化，RetSyn使用类别和质量条件扩散进行可控合成，使用群体平衡损失来增加少数（质量，疾病）对的覆盖率，并使用小型配对智能手机-桌面集的直接偏好优化校准步骤。然后将合成的图像用于增强分类器训练。结果：retsyn生成图像的有效性通过训练视网膜诊断模型在真实和合成数据的组合进行评估。RetSyn在域内和域外产生一致的增益。在低质量桌面图像上，F1从0.781提高到0.874（二值），从0.607提高到0.703（三类），AUROC分别达到0.982和0.951。在域外便携式图像上，RetSyn达到了AUROC 0.813/F1 0.703（二进制）和AUROC 0.804/F1 0.609（三级），超过了GroupDRO（二进制：AUROC 0.786/F1 0.626；三级：AUROC 0.789/F1 0.544）等组鲁棒性基线。结论：RetSyn提供了一个有效且可扩展的综合数据框架，显著增强了基于人工智能的低收入国家DR筛选模型的鲁棒性和泛化性。通过解决不同图像质量和设备特性带来的关键挑战，RetSyn有助于在服务不足的地区更可靠地部署人工智能诊断。此外，第一个公开可用的配对智能手机桌面视网膜图像数据集的发布将支持跨设备DR筛查解决方案的进一步研究。

{"title":"Enhancing AI-based diabetic retinopathy screening in low- and middle-income countries with synthetic data","authors":"Zitao Shuai , Chenwei Wu , Zhengxu Tang , David Restrepo , Michael Morley , Luis Filipe Nakayama","doi":"10.1016/j.jbi.2025.104938","DOIUrl":"10.1016/j.jbi.2025.104938","url":null,"abstract":"<div><h3>Objective:</h3><div>AI-based DR screening is promising in low- and middle-income countries (LMICs), where limited human resources constrain access to specialist-led programs. However, current systems often degrade under real-world image-quality variations, especially with portable devices that are vital for low- and middle-income countries. This study aims to develop Retsyn, a synthetic-data augmentation framework that improves screening robustness across devices and imaging conditions.</div></div><div><h3>Methods:</h3><div>RetSyn leverages advanced diffusion models to generate synthetic retinal images with diverse device and imaging quality characteristics. To address the challenges of (1) portable device data scarcity, (2) disease and quality distribution imbalance, and (3) varying image quality, RetSyn uses class and quality-conditioned diffusion for controllable synthesis, a group-balanced loss to increase coverage of minority (quality, disease) pairs, and a Direct Preference Optimization alignment step with a small paired smartphone–tabletop set. The synthesized images are then used to augment classifier training.</div></div><div><h3>Results:</h3><div>The effectiveness of RetSyn-generated images was evaluated by training retinal diagnosis models on a combination of real and synthetic data. RetSyn yields consistent gains in-domain and out-of-domain. On low-quality tabletop images, F1 improves from 0.781 to 0.874 (binary) and 0.607 to 0.703 (three-class), while AUROC reaches 0.982 and 0.951, respectively. On out-of-domain portable images, RetSyn attains AUROC 0.813/F1 0.703 (binary) and AUROC 0.804/F1 0.609 (three-class), exceeding group-robustness baselines such as GroupDRO (binary: AUROC 0.786/F1 0.626; three-class: AUROC 0.789/F1 0.544).</div></div><div><h3>Conclusion:</h3><div>RetSyn presents an effective and scalable synthetic data framework that significantly enhances the robustness and generalizability of AI-based DR screening models in LMICs. By addressing the critical challenges posed by varying image quality and device characteristics, RetSyn facilitates more reliable deployment of AI diagnostics in underserved regions. Additionally, the release of the first publicly available paired smartphone-tabletop retinal image dataset will support further research into cross-device DR screening solutions.</div></div>","PeriodicalId":15263,"journal":{"name":"Journal of Biomedical Informatics","volume":"172 ","pages":"Article 104938"},"PeriodicalIF":4.5,"publicationDate":"2025-10-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145370266","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

ADENER: A syntax-augmented grid-tagging model for Adverse Drug Event extraction in social media 社交媒体中不良药物事件提取的语法增强网格标记模型。

IF 4.5 2区医学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Biomedical Informatics

Pub Date : 2025-10-22 DOI: 10.1016/j.jbi.2025.104944

Weiru Fu , Hao Li , Ling Luo, Hongfei Lin

Objective:

Adverse Drug Event (ADE) extraction from social media is a critical yet challenging task due to the semantic similarity between adverse effects and therapeutic indications, as well as the prevalence of overlapping and discontinuous mentions often caused by comorbid conditions. This study aims to develop a robust model for accurate ADE extraction from noisy and irregular social media texts.

Methods:

We propose ADENER, a grid-tagging architecture that models ADE extraction as multi-label word-pair classification. ADENER incorporates two core encoding mechanisms: the convolutional capture layer fuses multi-dimensional textual features, captures long-range word-pair dependencies via dilated convolutions, and enhances interactions through semantic association matrices for social media text irregularities; the syntactic affine layer integrates path-level dependency information to enhance global logic understanding, enabling the model to distinguish between therapeutic symptom entities and ADE entities through syntactic cues. The decoding stage uses four-type relational labels to uniformly decode flat, overlapping, and discontinuous ADE mentions.

Results:

We evaluated ADENER on three widely used ADE extraction datasets: CADEC, CADECv2, SMM4H. The model achieved F1 scores of 74.64%, 77.97%, 61.73% on these datasets, respectively, outperforming all compared baseline models while maintaining competitive computational efficiency. The results demonstrate the effectiveness of our model in addressing the challenges posed by irregular and noisy social media data.

Conclusion:

ADENER offers a unified and effective solution for ADE extraction from social media, capable of handling flat, overlapping, and discontinuous entity mentions and correctly distinguishing ADE entities from therapeutic symptom entities. By incorporating convolutional capture layers for semantic word-pair interactions and syntactic affine layers for dependency-based logic understanding, our approach significantly improves extraction accuracy, providing a valuable tool for pharmacovigilance research and real-world drug safety monitoring.

目的：由于不良反应和治疗指征之间的语义相似性，以及通常由合并症引起的重叠和不连续提及的普遍存在，从社交媒体中提取药物不良事件（ADE）是一项关键但具有挑战性的任务。本研究旨在开发一个强大的模型，用于从嘈杂和不规则的社交媒体文本中准确提取ADE。方法：我们提出了ADENER，这是一种网格标记架构，将ADE提取建模为多标签词对分类。ADENER包含两种核心编码机制：卷积捕获层融合多维文本特征，通过扩展卷积捕获远程词对依赖关系，并通过语义关联矩阵增强社交媒体文本不规则性的交互；句法仿射层整合路径级依赖信息，增强整体逻辑理解，使模型能够通过句法线索区分治疗症状实体和ADE实体。解码阶段使用四种类型的关系标签来统一解码平坦的、重叠的和不连续的ADE提及。结果：我们在CADEC、CADECv2、SMM4H三个广泛使用的ADE提取数据集上对ADENER进行了评估。该模型在这些数据集上的F1得分分别为74.64%、77.97%和61.73%，在保持有竞争力的计算效率的同时，优于所有比较基线模型。结果表明，我们的模型在应对不规则和嘈杂的社交媒体数据带来的挑战方面是有效的。结论：ADENER为社交媒体的ADE提取提供了统一有效的解决方案，能够处理平坦、重叠、不连续的实体提及，并正确区分ADE实体与治疗症状实体。通过结合语义词对交互的卷积捕获层和基于依赖的逻辑理解的句法仿射层，我们的方法显着提高了提取精度，为药物警戒研究和现实世界的药物安全监测提供了有价值的工具。

{"title":"ADENER: A syntax-augmented grid-tagging model for Adverse Drug Event extraction in social media","authors":"Weiru Fu , Hao Li , Ling Luo, Hongfei Lin","doi":"10.1016/j.jbi.2025.104944","DOIUrl":"10.1016/j.jbi.2025.104944","url":null,"abstract":"<div><h3>Objective:</h3><div>Adverse Drug Event (ADE) extraction from social media is a critical yet challenging task due to the semantic similarity between adverse effects and therapeutic indications, as well as the prevalence of overlapping and discontinuous mentions often caused by comorbid conditions. This study aims to develop a robust model for accurate ADE extraction from noisy and irregular social media texts.</div></div><div><h3>Methods:</h3><div>We propose ADENER, a grid-tagging architecture that models ADE extraction as multi-label word-pair classification. ADENER incorporates two core encoding mechanisms: the convolutional capture layer fuses multi-dimensional textual features, captures long-range word-pair dependencies via dilated convolutions, and enhances interactions through semantic association matrices for social media text irregularities; the syntactic affine layer integrates path-level dependency information to enhance global logic understanding, enabling the model to distinguish between therapeutic symptom entities and ADE entities through syntactic cues. The decoding stage uses four-type relational labels to uniformly decode flat, overlapping, and discontinuous ADE mentions.</div></div><div><h3>Results:</h3><div>We evaluated ADENER on three widely used ADE extraction datasets: CADEC, CADECv2, SMM4H. The model achieved F1 scores of 74.64%, 77.97%, 61.73% on these datasets, respectively, outperforming all compared baseline models while maintaining competitive computational efficiency. The results demonstrate the effectiveness of our model in addressing the challenges posed by irregular and noisy social media data.</div></div><div><h3>Conclusion:</h3><div>ADENER offers a unified and effective solution for ADE extraction from social media, capable of handling flat, overlapping, and discontinuous entity mentions and correctly distinguishing ADE entities from therapeutic symptom entities. By incorporating convolutional capture layers for semantic word-pair interactions and syntactic affine layers for dependency-based logic understanding, our approach significantly improves extraction accuracy, providing a valuable tool for pharmacovigilance research and real-world drug safety monitoring.</div></div>","PeriodicalId":15263,"journal":{"name":"Journal of Biomedical Informatics","volume":"171 ","pages":"Article 104944"},"PeriodicalIF":4.5,"publicationDate":"2025-10-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145355020","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Pseudo-labeling and knowledge-guided contrastive learning for radiology report generation 伪标记与知识导向对比学习在放射学报告生成中的应用。

IF 4.5 2区医学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Biomedical Informatics

Pub Date : 2025-10-22 DOI: 10.1016/j.jbi.2025.104941

Fan Ye , Xuan Hu , Yihao Ding , Feifei Liu

Objective:

Radiology report generation (RRG) is a transformative technology in the field of radiology imaging that aims to address the critical need for consistency and comprehensiveness in diagnostic interpretation. Although recent advances in graph-based representation learning have demonstrated excellent performance in disease progression modeling, their application in radiology report generation still suffers from three inherent limitations: (i) semantic separation between local image features and free-text descriptions, (ii) inherent noise in automated medical concept annotation, and (iii) lack of anatomical constraints in cross-modal attention mechanisms.

Method:

This study proposes a pseudo-label and knowledge-guided comparative learning (PKCL) framework, which addresses the above issues through a novel fusion of dynamic query learning and knowledge-guided contrastive learning. The PKCL framework employs a trainable cross-modal query matrix (QM) to learn shared representations through parameter-sharing self-attention mechanisms between imaging and text encoders. The QM is used during training to query disease-related visual regions in reports and enables dynamic alignment between radiological features and textual descriptions during both training and inference. Additionally, this method combines pseudo labels with an adaptive top-k weighted feature fusion strategy to enhance learning from standard comparisons and leverages pre-built knowledge graphs via the XRayVision (Cohen et al., 2022) model to account for disease relationships and anatomical dependencies, thereby improving the clinical accuracy of generated reports.

Results:

Comprehensive evaluations on the IU-Xray and MIMIC-CXR datasets demonstrate that PKCL achieves state-of-the-art performance on both natural language generation metrics and clinical efficacy metrics. Specifically, it obtains 0.499 BLEU-1 and 0.374 RL on IU-Xray, and 0.346 BLEU-1 and 0.277 RL on MIMIC-CXR, outperforming prior methods such as R2GEN and CMCL.

Furthermore, PKCL exhibited robust generalization on the out-of-domain Montgomery County X-ray Set, effectively handling its low-resource conditions and brief, diagnostic-level textual supervision.

Conclusion:

The framework’s ability to maintain semantic consistency when generating clinically relevant reports represents a significant advancement over existing methods, particularly in capturing the subtle relationships between radiological findings and their textual descriptions.

目的：放射学报告生成（RRG）是放射学成像领域的一项变革性技术，旨在解决诊断解释中一致性和全面性的关键需求。尽管基于图的表示学习的最新进展在疾病进展建模方面表现出色，但其在放射学报告生成中的应用仍然存在三个固有限制：(i)局部图像特征和自由文本描述之间的语义分离，（ii）自动医学概念注释中的固有噪声，以及（iii）跨模态注意机制缺乏解剖约束。方法：本研究提出了一个伪标签和知识引导比较学习（PKCL）框架，该框架通过动态查询学习和知识引导对比学习的新颖融合来解决上述问题。PKCL框架采用可训练的跨模态查询矩阵（QM），通过图像和文本编码器之间的参数共享自注意机制来学习共享表示。QM在训练期间用于查询报告中与疾病相关的视觉区域，并在训练和推理期间实现放射特征和文本描述之间的动态对齐。此外，该方法将伪标签与自适应top-k加权特征融合策略相结合，以增强从标准比较中学习的能力，并通过XRayVision （Cohen等人，2022）模型利用预先构建的知识图来解释疾病关系和解剖依赖性，从而提高生成报告的临床准确性。结果：对iu - x射线和MIMIC-CXR数据集的综合评估表明，PKCL在自然语言生成指标和临床疗效指标上都达到了最先进的水平。具体而言，该方法在IU-Xray上获得0.499 BLEU-1和0.374 RL，在MIMIC-CXR上获得0.346 BLEU-1和0.277 RL，优于R2GEN和CMCL等先前的方法。此外，PKCL在域外蒙哥马利县x射线集上表现出鲁棒泛化，有效地处理了其低资源条件和简短的诊断级文本监督。结论：该框架在生成临床相关报告时保持语义一致性的能力代表了现有方法的重大进步，特别是在捕捉放射学发现与其文本描述之间的微妙关系方面。

{"title":"Pseudo-labeling and knowledge-guided contrastive learning for radiology report generation","authors":"Fan Ye , Xuan Hu , Yihao Ding , Feifei Liu","doi":"10.1016/j.jbi.2025.104941","DOIUrl":"10.1016/j.jbi.2025.104941","url":null,"abstract":"<div><h3>Objective:</h3><div>Radiology report generation (RRG) is a transformative technology in the field of radiology imaging that aims to address the critical need for consistency and comprehensiveness in diagnostic interpretation. Although recent advances in graph-based representation learning have demonstrated excellent performance in disease progression modeling, their application in radiology report generation still suffers from three inherent limitations: (i) semantic separation between local image features and free-text descriptions, (ii) inherent noise in automated medical concept annotation, and (iii) lack of anatomical constraints in cross-modal attention mechanisms.</div></div><div><h3>Method:</h3><div>This study proposes a pseudo-label and knowledge-guided comparative learning (PKCL) framework, which addresses the above issues through a novel fusion of dynamic query learning and knowledge-guided contrastive learning. The PKCL framework employs a trainable cross-modal query matrix (QM) to learn shared representations through parameter-sharing self-attention mechanisms between imaging and text encoders. The QM is used during training to query disease-related visual regions in reports and enables dynamic alignment between radiological features and textual descriptions during both training and inference. Additionally, this method combines pseudo labels with an adaptive top-k weighted feature fusion strategy to enhance learning from standard comparisons and leverages pre-built knowledge graphs via the XRayVision (Cohen et al., 2022) model to account for disease relationships and anatomical dependencies, thereby improving the clinical accuracy of generated reports.</div></div><div><h3>Results:</h3><div>Comprehensive evaluations on the IU-Xray and MIMIC-CXR datasets demonstrate that PKCL achieves state-of-the-art performance on both natural language generation metrics and clinical efficacy metrics. Specifically, it obtains 0.499 BLEU-1 and 0.374 RL on IU-Xray, and 0.346 BLEU-1 and 0.277 RL on MIMIC-CXR, outperforming prior methods such as R2GEN and CMCL.</div><div>Furthermore, PKCL exhibited robust generalization on the out-of-domain Montgomery County X-ray Set, effectively handling its low-resource conditions and brief, diagnostic-level textual supervision.</div></div><div><h3>Conclusion:</h3><div>The framework’s ability to maintain semantic consistency when generating clinically relevant reports represents a significant advancement over existing methods, particularly in capturing the subtle relationships between radiological findings and their textual descriptions.</div></div>","PeriodicalId":15263,"journal":{"name":"Journal of Biomedical Informatics","volume":"172 ","pages":"Article 104941"},"PeriodicalIF":4.5,"publicationDate":"2025-10-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145368032","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Discovering signature disease trajectories in pancreatic cancer and soft-tissue sarcoma from longitudinal patient records 从纵向患者记录中发现胰腺癌和软组织肉瘤的标志性疾病轨迹。

IF 4.5 2区医学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Biomedical Informatics

Pub Date : 2025-10-19 DOI: 10.1016/j.jbi.2025.104935

Liwei Wang , Rui Li , Andrew Wen , Qiuhao Lu , Jinlian Wang , Xiaoyang Ruan , Adriana Gamboa , Neha Malik , Christina L. Roland , Matthew H.G. Katz , Heather Lyu , Hongfang Liu

Background

Most clinicians have limited experience with rare diseases, making diagnosis and treatment challenging. Large real-world data sources, such as electronic health records (EHRs), provide a massive amount of information that can potentially be leveraged to determine the patterns of diagnoses and treatments for rare tumors that can serve as clinical decision aids.

Objectives

We aimed to discover signature disease trajectories of 3 rare cancer types: pancreatic cancer, STS of the trunk and extremity (STS-TE), and STS of the abdomen and retroperitoneum (STS-AR).

Materials and Methods

Leveraging IQVIA Oncology Electronic Medical Record, we identified significant diagnosis pairs across 3 years in patients with these cancers through matched cohort sampling, statistical computation, right-tailed binomial hypothesis test, and then visualized trajectories up to 3 progressions. We further conducted systematic validation for the discovered trajectories with the UTHealth Electronic Health Records (EHR).

Results

Results included 266 significant diagnosis pairs for pancreatic cancer, 130 for STS-TE, and 118 for STS-AR. We further found 44 2-hop (i.e., 2-progression) and 136 3-hop trajectories before pancreatic cancer, 36 2-hop and 37 3-hop trajectories before STS-TE, and 17 2-hop and 5 3-hop trajectories before STS-AR. Meanwhile, we found 54 2-hop and 129 3-hop trajectories following pancreatic cancer, 11 2-hop and 17 3-hop trajectories following STS-TE, 5 2-hop and 0 3-hop trajectories following STS-AR. For example, pain in joint and gastro-oesophageal reflux disease occurred before pancreatic cancer in 64 (0.5%) patients, pain in joint and “pain in limb, hand, foot, fingers and toes” occurred before STS-TE in 40 (0.9%) patients, agranulocytosis secondary to cancer chemotherapy and neoplasm related pain occurred after pancreatic cancer in 256 (1.9%) patients. Systematic validation using the UTHealth EHR confirmed the validity of the discovered trajectories.

Conclusion

We identified signature disease trajectories for the studied rare cancers by leveraging large-scale EHR data and trajectory mining approaches. These disease trajectories could serve as potential resources for clinicians to deepen their understanding of the temporal progression of conditions preceding and following these rare cancers, further informing patient-care decisions.

背景：大多数临床医生对罕见病的经验有限，使得诊断和治疗具有挑战性。电子健康记录（EHRs）等大型真实世界数据源提供了大量信息，可用于确定罕见肿瘤的诊断和治疗模式，从而作为临床决策辅助工具。目的：研究胰腺癌、躯干及四肢STS （STS- te）和腹部及腹膜后STS (STS- ar) 3种罕见肿瘤的特征发病轨迹。材料和方法：利用IQVIA肿瘤电子病历，我们通过匹配队列抽样、统计计算、右尾二项假设检验，确定了这些癌症患者在3 年内的显著诊断对，然后可视化了3个进展的轨迹。我们进一步用UTHealth电子健康记录（EHR）对发现的轨迹进行了系统验证。结果：结果包括266对胰腺癌，130对STS-TE， 118对STS-AR的显著诊断。我们进一步发现胰腺癌前44个2-跳（即2-进展）和136个3-跳轨迹，STS-TE前36个2-跳和37个3-跳轨迹，STS-AR前17个2-跳和5个3-跳轨迹。同时，我们发现胰腺癌后有54个2-跳和129个3-跳轨迹，STS-TE后有11个2-跳和17个3-跳轨迹，STS-AR后有5个2-跳和0个3-跳轨迹。例如，64例（0.5%）患者在胰腺癌前出现关节痛和胃食管反流病，40例（0.9%）患者在STS-TE前出现关节痛和“四肢、手、脚、手指和脚趾痛”，256例（1.9%）患者在胰腺癌后出现癌症化疗后继发粒细胞缺乏症和肿瘤相关疼痛。使用UTHealth电子病历系统验证了所发现轨迹的有效性。结论：通过利用大规模电子病历数据和轨迹挖掘方法，我们确定了所研究的罕见癌症的标志性疾病轨迹。这些疾病轨迹可以作为临床医生的潜在资源，加深他们对这些罕见癌症之前和之后病情的时间进展的理解，进一步为患者护理决策提供信息。

{"title":"Discovering signature disease trajectories in pancreatic cancer and soft-tissue sarcoma from longitudinal patient records","authors":"Liwei Wang , Rui Li , Andrew Wen , Qiuhao Lu , Jinlian Wang , Xiaoyang Ruan , Adriana Gamboa , Neha Malik , Christina L. Roland , Matthew H.G. Katz , Heather Lyu , Hongfang Liu","doi":"10.1016/j.jbi.2025.104935","DOIUrl":"10.1016/j.jbi.2025.104935","url":null,"abstract":"<div><h3>Background</h3><div>Most clinicians have limited experience with rare diseases, making diagnosis and treatment challenging. Large real-world data sources, such as electronic health records (EHRs), provide a massive amount of information that can potentially be leveraged to determine the patterns of diagnoses and treatments for rare tumors that can serve as clinical decision aids.</div></div><div><h3>Objectives</h3><div>We aimed to discover signature disease trajectories of 3 rare cancer types: pancreatic cancer, STS of the trunk and extremity (STS-TE), and STS of the abdomen and retroperitoneum (STS-AR).</div></div><div><h3>Materials and Methods</h3><div>Leveraging IQVIA Oncology Electronic Medical Record, we identified significant diagnosis pairs across 3 years in patients with these cancers through matched cohort sampling, statistical computation, right-tailed binomial hypothesis test, and then visualized trajectories up to 3 progressions. We further conducted systematic validation for the discovered trajectories with the UTHealth Electronic Health Records (EHR).</div></div><div><h3>Results</h3><div>Results included 266 significant diagnosis pairs for pancreatic cancer, 130 for STS-TE, and 118 for STS-AR. We further found 44 2-hop (i.e., 2-progression) and 136 3-hop trajectories before pancreatic cancer, 36 2-hop and 37 3-hop trajectories before STS-TE, and 17 2-hop and 5 3-hop trajectories before STS-AR. Meanwhile, we found 54 2-hop and 129 3-hop trajectories following pancreatic cancer, 11 2-hop and 17 3-hop trajectories following STS-TE, 5 2-hop and 0 3-hop trajectories following STS-AR. For example, pain in joint and gastro-oesophageal reflux disease occurred before pancreatic cancer in 64 (0.5%) patients, pain in joint and “pain in limb, hand, foot, fingers and toes” occurred before STS-TE in 40 (0.9%) patients, agranulocytosis secondary to cancer chemotherapy and neoplasm related pain occurred after pancreatic cancer in 256 (1.9%) patients. Systematic validation using the UTHealth EHR confirmed the validity of the discovered trajectories.</div></div><div><h3>Conclusion</h3><div>We identified signature disease trajectories for the studied rare cancers by leveraging large-scale EHR data and trajectory mining approaches. These disease trajectories could serve as potential resources for clinicians to deepen their understanding of the temporal progression of conditions preceding and following these rare cancers, further informing patient-care decisions.</div></div>","PeriodicalId":15263,"journal":{"name":"Journal of Biomedical Informatics","volume":"171 ","pages":"Article 104935"},"PeriodicalIF":4.5,"publicationDate":"2025-10-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145344968","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A non-interactive Online Medical Pre-Diagnosis system on encrypted vertically partitioned data 基于加密垂直分区数据的非交互式在线医疗预诊断系统。

IF 4.5 2区医学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Biomedical Informatics

Pub Date : 2025-10-17 DOI: 10.1016/j.jbi.2025.104940

Min Tang , Yuhao Zhang , Ronghua Liang , Guoqiang Deng

Objective:

In medical environments, patient records are stored as heterogeneous features across various institutions, prohibiting raw data sharing due to legal or institutional constraints. This fragmentation presents challenges for Online Medical Pre-Diagnosis (OMPD) systems. Existing methods (such as federated learning) require multiple rounds of interactions among all participating parties (hospitals and cloud servers), resulting in frequent communication. Moreover, due to the sharing of global gradients, they are vulnerable to inference attacks, leading to information leakage. In this paper, we propose a secure and efficient the OMPD system framework to address the problem of vertical data fragmentation, aiming to resolve the contradiction between medical data isolation and model collaboration.

Methods:

We propose PPNLR, a secure framework for building the OMPD systems. This framework combines functional encryption and blinding factors to design the sample-feature dimension encryption algorithm and the privacy-preserving vectorization training algorithm. Decoupling sample computation from model training enables cross-client data aggregation with only a single communication between hospitals and cloud servers.

Results:

Security analysis shows that PPNLR is resistant to semi-honest inference attacks and collusion attacks. Evaluation results based on six real-world medical datasets (text and images) show that: (i) The inference accuracy is close to that of the centralized plaintext training benchmark; (ii) The computational efficiency is at least 3.6

\times

higher than that of comparable approaches; (iii) The communication complexity is significantly reduced by eliminating dependencies on iteration count.

Conclusion:

PPNLR achieves data protection through cryptographic primitives, maintaining high diagnostic accuracy while ensuring the security of medical data and model parameters. Its single-communication architecture significantly reduces the deployment threshold in resource-constrained scenarios, providing a practical framework for building the privacy-friendly OMPD systems.

目的：在医疗环境中，患者记录作为异构特征存储在各个机构中，由于法律或制度的限制，禁止原始数据共享。这种碎片化给在线医疗预诊断（OMPD）系统带来了挑战。现有的方法（如联邦学习）需要在所有参与方（医院和云服务器）之间进行多轮交互，从而导致频繁的通信。此外，由于全局梯度的共享，它们容易受到推理攻击，导致信息泄露。本文提出了一种安全高效的OMPD系统框架来解决垂直数据碎片化问题，旨在解决医疗数据隔离与模型协作之间的矛盾。方法：提出了一种用于构建OMPD系统的安全框架PPNLR。该框架将功能加密和盲因子相结合，设计了样本特征维数加密算法和隐私保护矢量化训练算法。将样本计算与模型训练解耦，仅在医院和云服务器之间进行一次通信即可实现跨客户端数据聚合。结果：安全性分析表明，PPNLR能够抵抗半诚实推理攻击和串通攻击。基于6个真实医学数据集（文本和图像）的评估结果表明：(1)推理准确率接近集中式明文训练基准；（ii）计算效率至少比可比方法高3.6倍；（iii）通过消除对迭代计数的依赖，显著降低了通信复杂性。结论：PPNLR通过加密原语实现了数据保护，在保证医疗数据和模型参数安全的同时，保持了较高的诊断准确率。它的单通信体系结构显著降低了资源受限场景中的部署门槛，为构建隐私友好型OMPD系统提供了实用框架。

{"title":"A non-interactive Online Medical Pre-Diagnosis system on encrypted vertically partitioned data","authors":"Min Tang , Yuhao Zhang , Ronghua Liang , Guoqiang Deng","doi":"10.1016/j.jbi.2025.104940","DOIUrl":"10.1016/j.jbi.2025.104940","url":null,"abstract":"<div><h3>Objective:</h3><div>In medical environments, patient records are stored as heterogeneous features across various institutions, prohibiting raw data sharing due to legal or institutional constraints. This fragmentation presents challenges for Online Medical Pre-Diagnosis (OMPD) systems. Existing methods (such as federated learning) require multiple rounds of interactions among all participating parties (hospitals and cloud servers), resulting in frequent communication. Moreover, due to the sharing of global gradients, they are vulnerable to inference attacks, leading to information leakage. In this paper, we propose a secure and efficient the OMPD system framework to address the problem of vertical data fragmentation, aiming to resolve the contradiction between medical data isolation and model collaboration.</div></div><div><h3>Methods:</h3><div>We propose PPNLR, a secure framework for building the OMPD systems. This framework combines functional encryption and blinding factors to design the sample-feature dimension encryption algorithm and the privacy-preserving vectorization training algorithm. Decoupling sample computation from model training enables cross-client data aggregation with only a single communication between hospitals and cloud servers.</div></div><div><h3>Results:</h3><div>Security analysis shows that PPNLR is resistant to semi-honest inference attacks and collusion attacks. Evaluation results based on six real-world medical datasets (text and images) show that: (i) The inference accuracy is close to that of the centralized plaintext training benchmark; (ii) The computational efficiency is at least 3.6<span><math><mo>×</mo></math></span> higher than that of comparable approaches; (iii) The communication complexity is significantly reduced by eliminating dependencies on iteration count.</div></div><div><h3>Conclusion:</h3><div>PPNLR achieves data protection through cryptographic primitives, maintaining high diagnostic accuracy while ensuring the security of medical data and model parameters. Its single-communication architecture significantly reduces the deployment threshold in resource-constrained scenarios, providing a practical framework for building the privacy-friendly OMPD systems.</div></div>","PeriodicalId":15263,"journal":{"name":"Journal of Biomedical Informatics","volume":"171 ","pages":"Article 104940"},"PeriodicalIF":4.5,"publicationDate":"2025-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145329250","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Synthetic-to-real attentive deep learning for Alzheimer’s assessment: A domain-agnostic framework for ROCF scoring 用于阿尔茨海默氏症评估的综合到真实的专注深度学习：ROCF评分的领域不可知框架。

IF 4.5 2区医学 Q2 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Journal of Biomedical Informatics

Pub Date : 2025-10-17 DOI: 10.1016/j.jbi.2025.104929

Kassem Anis Bouali, Elena Šikudová

Objective:

Early diagnosis of Alzheimer’s disease depends on accessible cognitive assessments, such as the Rey-Osterrieth Complex Figure (ROCF) test. However, manual scoring of this test is labor-intensive and subjective, which introduces experimental biases. Additionally, deep learning models face challenges due to the limited availability of annotated clinical data, particularly for assessments like the ROCF test. This scarcity of data restricts model generalization and exacerbates domain shifts across different populations.

Methods:

We propose a novel framework comprising a data synthesis pipeline and ROCF-Net, a deep learning model specifically designed for ROCF scoring. The synthesis pipeline is lightweight and capable of generating realistic, diverse, and annotated ROCF drawings. ROCF-Net, on the other hand, is a cross-domain scoring model engineered to address domain discrepancies in stroke texture and line artifacts. It maintains high scoring accuracy through a novel line-specific attention mechanism tailored to the unique characteristics of ROCF drawings.

Results:

Unlike conventional synthetic medical imaging methods, our approach generates ROCF drawings that accurately reflect Alzheimer’s-specific abnormalities with minimal computational cost. Our scoring model achieves SOTA performance across differently sourced datasets, with a Mean Absolute Error (MAE) of 3.53 and a Pearson Correlation Coefficient (PCC) of 0.86. This demonstrates both high predictive accuracy and computational efficiency, outperforming existing ROCF scoring methods that rely on Convolutional Neural Networks (CNNs) while avoiding the overhead of parameter-heavy transformer models. We also show that training on our synthetic data generalizes as well as training on real clinical data, where the difference in performance was minimal (MAE differed by 1.43 and PCC by 0.07), indicating no statistically significant performance gap.

Conclusion:

Our work introduces four contributions: (1) a cost-effective pipeline for generating synthetic ROCF data, reducing dependency on clinical datasets; (2) a domain-agnostic model for automated ROCF scoring across diverse drawing styles; (3) a lightweight attention mechanism aligning model decisions with clinical scoring for transparency; and (4) a bias-aware framework using synthetic data to reduce demographic disparities, promoting fair cognitive assessment across populations.

目的：阿尔茨海默病的早期诊断依赖于可获得的认知评估，如Rey-Osterrieth Complex Figure （ROCF）测试。然而，手工评分是劳动密集型和主观的，这引入了实验偏差。此外，由于标注临床数据的可用性有限，深度学习模型面临挑战，特别是对于像ROCF测试这样的评估。这种数据的稀缺性限制了模型的泛化，并加剧了不同人群之间的领域转移。方法：我们提出了一个新的框架，包括一个数据合成管道和ROCF- net，一个专门为ROCF评分设计的深度学习模型。合成管道是轻量级的，能够生成真实的、多样的、带注释的ROCF图纸。另一方面，ROCF-Net是一个跨域评分模型，用于解决笔画纹理和线条伪像中的域差异。它通过针对ROCF图纸的独特特征量身定制的新颖的线特定注意机制保持高评分精度。结果：与传统的合成医学成像方法不同，我们的方法以最小的计算成本生成准确反映阿尔茨海默病特异性异常的ROCF图。我们的评分模型在不同来源的数据集上实现了SOTA性能，平均绝对误差（MAE）为3.53，Pearson相关系数（PCC）为0.86。这证明了高预测精度和计算效率，优于现有的依赖卷积神经网络（cnn）的ROCF评分方法，同时避免了重参数变压器模型的开销。我们还表明，在我们的合成数据上的训练与在真实临床数据上的训练一样一般化，其中性能差异很小（MAE差1.43，PCC差0.07），表明没有统计学上显著的性能差距。结论：我们的工作引入了四个贡献：(1)成本效益高的管道生成合成ROCF数据，减少对临床数据集的依赖；(2)跨不同画风的自动ROCF评分的领域不可知模型；(3)将模型决策与临床透明度评分相结合的轻量级注意机制；(4)利用综合数据构建偏见感知框架，减少人口差异，促进人群间的公平认知评估。

{"title":"Synthetic-to-real attentive deep learning for Alzheimer’s assessment: A domain-agnostic framework for ROCF scoring","authors":"Kassem Anis Bouali, Elena Šikudová","doi":"10.1016/j.jbi.2025.104929","DOIUrl":"10.1016/j.jbi.2025.104929","url":null,"abstract":"<div><h3>Objective:</h3><div>Early diagnosis of Alzheimer’s disease depends on accessible cognitive assessments, such as the Rey-Osterrieth Complex Figure (ROCF) test. However, manual scoring of this test is labor-intensive and subjective, which introduces experimental biases. Additionally, deep learning models face challenges due to the limited availability of annotated clinical data, particularly for assessments like the ROCF test. This scarcity of data restricts model generalization and exacerbates domain shifts across different populations.</div></div><div><h3>Methods:</h3><div>We propose a novel framework comprising a data synthesis pipeline and ROCF-Net, a deep learning model specifically designed for ROCF scoring. The synthesis pipeline is lightweight and capable of generating realistic, diverse, and annotated ROCF drawings. ROCF-Net, on the other hand, is a cross-domain scoring model engineered to address domain discrepancies in stroke texture and line artifacts. It maintains high scoring accuracy through a novel line-specific attention mechanism tailored to the unique characteristics of ROCF drawings.</div></div><div><h3>Results:</h3><div>Unlike conventional synthetic medical imaging methods, our approach generates ROCF drawings that accurately reflect Alzheimer’s-specific abnormalities with minimal computational cost. Our scoring model achieves SOTA performance across differently sourced datasets, with a Mean Absolute Error (MAE) of 3.53 and a Pearson Correlation Coefficient (PCC) of 0.86. This demonstrates both high predictive accuracy and computational efficiency, outperforming existing ROCF scoring methods that rely on Convolutional Neural Networks (CNNs) while avoiding the overhead of parameter-heavy transformer models. We also show that training on our synthetic data generalizes as well as training on real clinical data, where the difference in performance was minimal (MAE differed by 1.43 and PCC by 0.07), indicating no statistically significant performance gap.</div></div><div><h3>Conclusion:</h3><div>Our work introduces four contributions: (1) a cost-effective pipeline for generating synthetic ROCF data, reducing dependency on clinical datasets; (2) a domain-agnostic model for automated ROCF scoring across diverse drawing styles; (3) a lightweight attention mechanism aligning model decisions with clinical scoring for transparency; and (4) a bias-aware framework using synthetic data to reduce demographic disparities, promoting fair cognitive assessment across populations.</div></div>","PeriodicalId":15263,"journal":{"name":"Journal of Biomedical Informatics","volume":"171 ","pages":"Article 104929"},"PeriodicalIF":4.5,"publicationDate":"2025-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145329252","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0