每日医学AI论文简报 v0.3

[日报] 2026-08-13

💡 今日学习推荐 CS231n 卷积神经网络（Stanford） — 一、卷积神经网络（CNN）架构（进度0%） ⏭ 下一步：CS229 机器学习讲义（Andrew Ng）

1. Generalizable cancer detection from ultra-low-pass WGS via deep contextual modeling of cfDNA sequences. 🆕

Mol Biomed (IF: N/A) | 2026 Aug 12 | PubMed ⚠️ | DOI | ⚠️ 中置信 | 研究设计: 未识别分析来源： PMC全文
领域：液体活检与肿瘤学；基于超低深度全基因组测序（ULP-WGS）的游离DNA（cfDNA）片段组学癌症检测
方法：开发并验证Fragmentia-AI™ WGS框架，该框架不依赖突变检出，采用基于Transformer的多示例学习架构，并跨肿瘤分数（TF）分层进行顺序微调；在多个独立队列中评估性能，包括17种癌症类型的泛癌测试集、不同测序平台生成的外部公共数据集，以及包含异质性前处理与实验条件的技术变异队列；通过比较模型预测与晚期非小细胞肺癌患者接受化学免疫治疗后的无进展生存期（PFS）评估临床相关性。
核心发现：跨TF分层的顺序微调显著提升低TF样本性能，相比仅用高TF训练，AUC相对提高35.6%（0.884 vs. 0.652）。独立测试队列总体AUC为0.930，且在不同TF分层和癌症类型中表现一致。外部验证确认了跨平台稳健性（AUC：0.929；敏感性：0.78；特异性：0.92）。前处理和技术变量导致的评分波动未破坏分类稳定性。多变量调整后，模型阴性状态（预测分数低于训练衍生阈值）与模型阳性状态相比，仍显著关联更优的PFS（HR=0.49，95% CI：0.29–0.82）。
相关度（医学AI研究者）：高。该研究展示了将Transformer架构应用于ULP-WGS低深度cfDNA数据实现稳健癌症检测与风险分层的可行路径，并强调了跨平台泛化、固定操作阈值及临床结局关联的重要性，对开发可落地的医学AI液体活检模型具有直接参考价值。
一句话：基于Transformer的Fragmentia-AI™框架可从极稀疏的ULP-WGS cfDNA数据中提取癌症信号，在实现高准确性检测的同时提供具有预后意义的模型评分。

🔬 新增AI临床试验

近1日新增： 90 项 | 数据源： ClinicalTrials.gov

NCT03986034 Early Clonal Dynamics During Venetoclax Treatment for Relapsed or Refractory Chronic Lymphocytic Leukemia (CLL)

状态： RECRUITING | 期： PHASE2 | 赞助： | 招募: 75 条件： Chronic Lymphocytic Leukemia

NCT07760675 Evaluating in Person Curriculum for Educating Obstetrical Residents on Better Caring for 2 Spirit Metis Patients Using and AI Assisted OSCE.

状态： NOT_YET_RECRUITING | 期： NA | 赞助： | 招募: 90 条件： Cultural Safety; Medical Education; Cultural Competency

NCT05847894 Assisting Pulmonary Disease Diagnosis With Ophthalmic Artificial Intelligence Technology

状态： RECRUITING | 期： N/A | 赞助： | 招募: 10000 条件： Pulmonary Diseases; Ophthalmological Diagnostic Techniques; Artificial Intelligence

NCT05633550 Flexion-Extension Radiograph Imaging Protocol Reliability Study

状态： COMPLETED | 期： N/A | 赞助： | 招募: 50 条件： Lumbar Spinal Stenosis; Spinal Instability

NCT07637656 Discovering Determinants of Food Intake by Application of Artificial Intelligence to Complex, High-Dimensional Data

状态： NOT_YET_RECRUITING | 期： N/A | 赞助： | 招募: 800 条件： Healthy Volunteer

📄 医学AI预印本速览（arXiv 近3天）

1. MultiModal Code-Switching: Interleaving Visual Objects into Language for Explicit Object-Level Alignment Changhao Xiang, Shangyu Xing, Zhen Wu et al. | 2026-08-11 | arXiv Existing Multimodal Large Language Models (MLLMs) predominantly rely on image-text pairs for modality alignment pretraining, mapping global image representations to long textual descriptions. However, this image-level alignment suffers from referential ambiguity: models struggle …

2. Scheduling Mixed RL Rollouts Beyond Prefix Locality Zetao Hong, Song Yuan, Yuanhao Ding et al. | 2026-08-11 | arXiv Modern reinforcement learning (RL) post-training pipelines for large language models (LLMs) increasingly combine rollout workloads across multiple domains and feedback paradigms. Prefix-aware routing improves inference efficiency through cache reuse and load balancing, but it doe…

3. The Illusion of Cross-Lingual Safety in Low-Resource Languages Abigail Oppong, P Sam Sahil, Tadesse Destaw Belay et al. | 2026-08-11 | arXiv Safety alignment in large language models (LLMs) is largely developed in English, assuming these safeguards generalize across multilingual settings. However, this assumption remains underexplored and exposes a vulnerability in low-resource languages. We investigate cross-lingual …

4. Attention-Path Fragility as an Uncertainty Signal in Large Language Models Minsoo Kim, Sungyoung Ji, Kisung Moon et al. | 2026-08-11 | arXiv We propose that a model’s uncertainty about a token is reflected not only in the breadth of its output distribution but also in whether a confident prediction is \emph{fragile} under perturbation of its attention pathways. We instantiate this as ASMI (Attention-Subnetwork Mutual …

5. 3D Weighted Geometric Graph Neural Networks for Sheep Facial Pain Assessment Alam Noor, Luis Almeida, Mohamed Daoudi | 2026-08-11 | arXiv Deep learning systems perform mainly within the 2D for a single image domain and take the face as a single-dimension representation, losing sight of the 3D anatomy of sheep and cross-landmark spatial relationships that are intrinsic to the clinically proven Sheep Pain Facial Expr…

6. TEAMMix: Taxonomy Enrichment Augmentation and Minority-augmented Mixing Strategy for LLM-enhanced Weak-Supervised Hierarchical Text Classification Jian Zhang, Zhuohao Yang, Songlin Lei et al. | 2026-08-11 | arXiv Hierarchical Text Classification (HTC), as a critical text mining task, faces challenges such as complex label hierarchies and class imbalance. Existing methods based on large language models (LLMs) struggle to be efficiently applied to this task due to issues like lengthy prompt…

7. myMediWhisper: Construction of Burmese Medical Speech Corpus and Whisper Fine-Tuning for Clinical Dialogue ASR Ye Kyaw Thu, Ye Bhone Lin, Thura Aung et al. | 2026-08-11 | arXiv Although Whisper models benefit from large-scale multilingual pre-training, their performance on Burmese medical speech remains limited. This work presents a Burmese medical speech recognition framework built on a high-quality 28-hour corpus recorded and validated by native speak…

8. SCOUT: Symmetric Consensus Outlier Detection for Failure Localization in LLM Pre-Training Zhuang Wang | 2026-08-11 | arXiv In LLM pre-training, synchronization propagates rank-local stalls, slowdowns, and numerical errors into job-wide symptoms, obscuring their origin. Existing diagnosis often relies on in-process monitors that cannot report after the trainer blocks or terminates, or on post-mortem l…