medical-ai

topic9 events

papersTODAY 04:00 UTC

Attention-Enhanced Deep Learning Classifies Autism from 3D Gait Data

A new arXiv preprint describes a deep learning pipeline that uses 3D gait recordings to help identify autism spectrum disorder, aiming to sidestep the subjectivity and cost of standard clinical assessments. The approach adds attention mechanisms to the model and evaluates results across multiple data folds to test how stable the performance is. The work is presented as a step toward more objective, non-invasive screening tools, though the abstract only covers the motivation and method rather than deployment.

autism/deep-learning/healthcare/gait-analysis

papersTODAY 04:00 UTC

EMR: Self-Evolving Medical Multi-Agent System via Experience Mining and Reuse

Researchers propose EMR, a medical multi-agent framework built on large language models that improves over time by mining and reusing experience from earlier cases. The work targets a common limitation of clinical AI agents, which typically rely on fixed strategies and cannot retain a persistent memory of past diagnostic successes. By storing and reusing prior reasoning, the system aims to support self-evolution rather than static performance.

agents/medical-ai/llm/research

papersTODAY 04:00 UTC

MARCUS: Agentic Multimodal Vision-Language Model for Cardiac Diagnosis

Researchers present MARCUS, a vision-language system designed to assist with cardiac diagnosis and management. Unlike earlier medical AI models that handle only one type of input and cannot interact with users, it is described as agentic and multimodal. The work targets cardiovascular disease, the leading cause of death worldwide, where interpreting complex cardiac tests remains a bottleneck.

agents/multimodal/medical-ai/vision-language

papersTODAY 04:00 UTC

Image-swap audit tests whether medical vision-language models rely on scans

A new arXiv paper introduces ModaLens, an audit method that swaps paired images to check how much a medical vision-language model actually depends on the image versus the accompanying radiology report. Because a report may already contain the answer to a clinical question, the technique measures how image sensitivity shifts when report text is available. The authors report results for MedGemma-27B among other models.

medical-ai/vision-language/evaluation/arxiv

papersTODAY 04:00 UTC

Study Probes How Misleading Context Skews Medical Question Answering

A new arXiv paper investigates why large language models can give wrong medical answers when the context they receive is misleading, even though their standalone medical accuracy is high. The authors analyze the internal mechanisms behind this susceptibility, aiming to explain how flawed context overrides a model's medical knowledge. The work is a revised cross-listing on arXiv and falls under AI and machine learning research.

medical-ai/llm-safety/research/context

papersSEP 12 04:00 UTC

Instance segmentation models support automated multi-class wound assessment

A new arXiv paper presents an approach to automated wound care that combines dedicated instance segmentation models for detecting wound boundaries with multi-class classification. The authors argue that existing AI systems for wound analysis tend to be narrow in scope, and propose handling boundary detection and wound typing as separate, specialized tasks. The method targets clinical decision support in both chronic and acute wound management.

medical-imaging/segmentation/computer-vision/healthcare

papersSEP 12 04:00 UTC

VeriSim framework tests medical AI robustness against noisy patient communication

Researchers released VeriSim, a patient simulation framework that introduces controllable, clinically grounded noise into medical AI evaluations. The tool aims to move testing beyond idealized patient cases toward more realistic communication conditions. The work is described in an arXiv paper (2604.10441v2).

medical-ai/evaluation/safety/benchmarks

papersSEP 10 04:00 UTC

MedDeID: on-premises de-identification of clinical text using real and synthetic training data

Clinical notes often contain personal identifiers that block their reuse in research and medical AI, especially where rules prevent data from leaving a hospital's systems. A new arXiv paper introduces MedDeID, a framework that runs entirely on local infrastructure and is trained on a combination of in-house annotations and synthetic examples to strip such information from notes.

privacy/healthcare/nlp/de-identification

papersSEP 10 04:00 UTC

Study examines when clinical AI agents should stop testing and commit to a diagnosis

A new arXiv paper addresses the stopping problem for AI agents in clinical diagnosis, which must decide when to request another test, when to commit to a diagnosis, and when to defer. The authors note that current agent benchmarks typically measure accuracy under fixed or unconstrained interaction, leaving the reliability of autonomous stopping untested. They propose a risk-constrained framework for evaluating and controlling these stopping decisions in sequential diagnosis settings.

agents/healthcare/safety/benchmarks