Tag
The paper presents MedPAO-Assist, an edge-deployable, modality-aware agentic AI framework that structures radiology reports from dictation using a ~0.6B-parameter small language model, running with ~3.3GB RAM and ~31s latency on a CPU-only laptop.
This paper studies hindsight-guided rationale distillation on the ZebraMap rare disease benchmark, where a 1.5B student is fine-tuned on chain-of-thought traces from an 8B teacher that sees the ground-truth diagnosis. It finds that a contamination-filtered variant slightly but significantly beats the teacher, identifying 'GT hallucination'—label phrases leaking into reasoning chains—as the key artifact that must be filtered.
The paper introduces a method using corpus characterization and inverse constitutional fine-tuning to improve the stylistic alignment of AI-generated radiology reports with authentic radiologist writing. This approach achieves significant gains in text alignment metrics, demonstrating effectiveness for style-aware report generation.
This paper evaluates neural, workflow, and agentic systems for ICD-10-CM coding, identifies failure modes on rare and complex codes, and shows that tool-augmented agentic configurations can recover performance on specific subsets.
KREL is a framework that uses large language models for automatic medical coding by integrating structured ICD coding guidelines to improve reasoning and reduce hallucinations. It outperforms existing methods on benchmark datasets.
The paper introduces a knowledge-guided agentic framework that identifies missing patient context in health queries and asks targeted follow-up questions to improve the accuracy and consistency of downstream language model responses.
This paper presents an empirical study of inference-time gating in a production-scale clinical NLP pipeline using Llama and MMed-Llama models, showing that learning filtering rules from verifier rejections fails at scale, while ontology-based and evidence-testing filters are effective.
This paper investigates LLM-based metrics for evaluating clinical significance in radiology report generation. It identifies discrimination bias in existing LLM evaluators and proposes training lightweight interpretable metrics to improve the balance between error detection and tolerance of harmless variations.
This paper introduces ChristBERT, a family of domain-specific RoBERTa-based language models for German clinical NLP, and evaluates three domain adaptation strategies (continued pre-training, pre-training from scratch, and vocabulary adaptation) on medical named entity recognition and text classification tasks, achieving state-of-the-art results.
This paper presents a specialty-specific medical language model for extracting information from clinical narratives about immune-mediated and infectious diseases, using a BiLSTM-CNN-Char architecture trained on a curated corpus of 371 case reports, achieving an F1 score of 0.89.
Leanly_AI is a specialized large language model developed by Fuzhou University hospitals to provide evidence-informed psychological support for patients undergoing clinical weight management. The model integrates population health data to address obesity-related emotional challenges while maintaining clinical interpretability and safety.