Tag
该论文提出了一种针对小型临床表格数据的混合注意力网络+逻辑回归（HAN+LR）集成模型，用于在鲜胚移植前预测IVF/ICSI活产，在616个周期的数据集上取得0.710的AUC-ROC，优于TabPFN和文本序列化语言模型基线，并提出对滋养层细胞差异化加权的胚胎形态编码方法。
The study identifies 'ECG Mirage' in vision-language models for clinical prediction, where models underutilize ECG data despite apparent multimodal capability, and proposes visual prompt tuning as an efficient mitigation strategy.
This paper introduces M2G-LLM, a framework that enhances Large Language Models with multimodal graph reasoning via Graph Neural Networks to improve clinical prediction tasks using electronic health records.
This paper proposes a reinforcement learning fine-tuning framework to enhance clinical reasoning in EHR foundation models by treating them as generative policies over patient trajectories, showing that RL fine-tuning improves performance and enables smaller models to surpass larger ones in data-limited scenarios.
DR-LabStack is a React–Flask web system that integrates four pretrained diabetic retinopathy prediction models into a common clinician-facing interface, demonstrating a reusable workflow for heterogeneous ML models.
PhenoBench is an executable benchmark for evaluating AI models on clinical tasks using a deeply phenotyped human cohort, demonstrating limited aggregate improvements of foundation models over standard methods.
This paper introduces a framework to audit clinical prediction models by distinguishing between learner inefficiencies and measurement limitations, validated on multiple cohorts to guide whether to improve models or data collection.
ATHENA is a knowledge-guided agentic neural architecture search framework that automates Transformer-based electronic health record modeling by reusing architecture knowledge across hospitals to reduce manual tuning.
This paper addresses the problem of treatment-induced label indeterminacy in clinical prediction, using post-cardiac-arrest neurological prognostication as a case study. The authors propose a framework that incorporates expert annotations of counterfactual outcomes and highlight tradeoffs between accuracy on certain versus uncertain cases.
This paper introduces MiGHT-EHR, a multi-task graph transformer for heterogeneous temporal EHR data, jointly modeling clinical entities, temporal trajectories, and task dependencies. It outperforms state-of-the-art methods on MIMIC-III and MIMIC-IV across drug recommendation, length-of-stay, mortality, and readmission prediction.
This paper proposes Patients-like-me (PLM), a unified LM–GNN framework that integrates local patient semantics with global cohort structure for explainable clinical prediction. It introduces a Variational Expectation-Maximization algorithm and demonstrates state-of-the-art results on MIMIC-III and MIMIC-IV with reference-patient explanations.
This arXiv paper evaluates federated training of tokenized generative event models (GEMs) on ICU EHR data from three health systems, showing that federated learning preserves most centralized performance and improves cross-site transportability compared to conventional supervised models.
Introduces CT-HEG, a continuous-time heterogeneous EHR graph schema for ICU mortality prediction, with an ablation study showing bidirectional connectivity and time-attentive edge features matter; surprisingly, a simplified homogeneous graph outperformed the full heterogeneous model on the MIMIC-IV cohort.
This paper introduces xMICD, a method that combines ICD code groupings with pre-trained embedding similarities to create low-dimensional, clinically interpretable patient representations, achieving predictive performance comparable to embedding-based approaches.
A scoping review of 190 studies characterizing end-to-end machine learning pipelines for surgical risk stratification and outcome prediction using EHR data, identifying methodological gaps in preprocessing, evaluation, and explainability.
Presents a lightweight knowledge-injection framework for zero-shot ICU delirium prediction that augments structured EHR data summaries with external clinical knowledge at inference time, improving AUROC by up to 8.57 percentage points on LLaMA models without fine-tuning.
The paper proposes converting multimodal patient data (text, labs, vitals) into a single natural language sequence and fine-tuning LLMs for clinical prediction, achieving comparable or better performance than specialized fusion architectures across three tasks.
The paper studies interaction-aware mixture-of-experts for post-stroke rigidity prediction using multi-level views of structured health records, showing that while performance gains are minimal, routing attribution reveals systematic importance differences across views, highlighting view construction as key to interpretability.
This paper proposes an explicit multimodal routing framework for clinical prediction using EHR data, enabling interpretable, robust, and auditable reasoning across structured variables, clinical notes, and chest X-rays via discrete unimodal, bimodal, and trimodal routes with inference-time route masking for missing modality simulation.
This paper systematically compares discrete, continuous, and hybrid value encoding strategies for transformers in electronic health record data, finding that hybrid token-based approaches with binning provide robust performance and are recommended as a practical default.