医学算法前沿简报

研究用大规模纵向眼底照片训练时间感知基础模型，再对心肌梗死、脑卒中、高血压、糖尿病、肾病和眼病的 5 年发病或进展进行微调。相较横断面 RETFound，模型在多个全身疾病任务上提高了风险区分度，并在多地区外部队列中保持总体趋势。优势来自对同一眼随时间变化的显式建模；局限是回顾性队列、临床数据受限，以及尚无前瞻性决策获益验证。

This study trains a time-aware retinal foundation model on repeated fundus photographs from the same eye. A teacher–student objective combines masked reconstruction with temporal representation consistency, followed by pseudo-survival fine-tuning for five-year disease outcomes. The model improves discrimination for myocardial infarction, stroke, hypertension, and diabetes across several external cohorts. Its main contribution is trajectory-aware representation learning; prospective evaluation and accessible clinical datasets remain missing.

研究在多种眼病和全身疾病检测任务上比较 RETFound 与三种 ImageNet 预训练模型，并系统缩减微调标签量。标签充足时，各模型在眼病任务上的差异有限；标签降至数百张且任务涉及高血压、糖尿病或慢性肾病时，RETFound 更具优势。结果支持把视网膜基础模型用于标注稀缺场景，但外部 AUC 仍可能偏低，不能直接推导出临床筛查价值。

This retrospective benchmark compares a retina-specific foundation model with standard ImageNet-pretrained backbones across ocular and systemic disease detection. Performance is similar on several ocular tasks when full labels are available, while RETFound is more label-efficient for systemic disease detection with only hundreds of images. The work supports selective rather than universal use of retinal pretraining. External discrimination remains modest, and the endpoints are cross-sectional rather than future clinical events.

研究把完整 OCT 体积视作视频序列，用 V-JEPA 编码切片间关系，并与只使用中心 B-scan 的多种图像基础模型及 3D CNN 比较。体积范式在多数数据集上取得更高判别性能，说明单张切片会丢失有价值的三维结构。研究开放了实验代码，并使用多个公开数据集；但部分临床队列需申请，且视频模型的计算负担和设备域偏移仍是部署障碍。

The study treats a three-dimensional OCT scan as a video and applies V-JEPA to learn interactions across neighboring B-scans. It benchmarks this volumetric approach against central-slice retinal foundation models, DINOv2, and a 3D CNN on multiple AMD and glaucoma datasets. Volumetric representation generally improves discrimination by preserving inter-slice context. Reproduction is supported by released code and public datasets, although some cohorts require permission and compute demands are higher.

ECG-R1 面向多模态心电解读，把协议化指令生成、模态解耦和基于诊断证据的强化学习结合起来。交替模态丢弃让模型在波形或图像缺失时仍能维持一致性，证据奖励则鼓励输出与可测量心电特征相匹配的解释。项目同时公开代码、9B 权重和训练数据，复现条件较好；但大模型可能产生幻觉，评价结果不能替代前瞻性临床验证。

ECG-R1 is a multimodal reasoning model for electrocardiogram interpretation. It combines protocol-guided instruction construction, modality-decoupled learning, interleaved modality dropout, and reinforcement learning based on diagnostic evidence. The design aims to remain robust when either ECG signals or rendered images are missing and to ground explanations in measurable findings. Code, 9B-parameter weights, and instruction data are released, but hallucination risk and the absence of prospective clinical validation remain important limitations.