Selected work

Representative Papers

Deterministic Integrity Gates for LLM-Assisted Clinical Manuscript Preparation: An Auditable Biomedical Informatics Architecture

Jun 08, 2026

Current large language model (LLM)-generated clinical research manuscripts commonly suffer from fabricated citations, data drift, and omissions of reporting guidelines, yet existing tools lack effective validation mechanisms. This work proposes an integrated generation-and-verification architecture that decomposes the writing process into 43 skill modules—including 21 deterministic detectors—orchestrated by a unified coordinator. It introduces a “maximally deterministic” completeness gating mechanism to enforce structured, traceable audits and re-execution checks at each stage. Evaluated on the STARD, PRISMA, and STROBE benchmark datasets, the approach successfully identified all 27 injected defects with zero false positives, substantially outperforming general-purpose LLM-based review methods.

0 citationsRead paper

How does downsampling affect needle electromyography signals? A generalisable workflow for understanding downsampling effects on high-frequency time series

Jan 15, 2026

This study addresses the challenge of high computational cost and poor real-time performance in automated classification of neural muscular disorders using high-sampling-rate needle electromyography (nEMG) signals, where the impact of downsampling on diagnostic information retention remains unclear. The authors propose a unified evaluation framework that, for the first time, integrates shape-aware downsampling, feature space analysis, and classification performance to systematically quantify how different downsampling strategies affect waveform integrity and diagnostic capability in high-frequency temporal biosignals. Evaluated on a three-class neuromuscular disorder classification task, the framework identifies an optimal downsampling configuration that balances computational efficiency with preservation of diagnostic information. Results demonstrate that shape-aware downsampling significantly outperforms conventional methods, offering a generalizable analytical paradigm for efficient processing of high-frequency temporal biosignals.

0 citationsRead paper

CLUE: Controllable Latent space of Unprompted Embeddings for Diversity Management in Text-to-Image Synthesis

Nov 14, 2025

In few-shot domains such as medical imaging, text-to-image generation faces a fundamental trade-off between diversity and stability. To address this, we propose CLUE—a novel framework that achieves controllable, diverse, and stable generation without requiring additional training data. CLUE constructs a prompt-free, controllable latent embedding space and employs KL divergence to enforce prompt-agnostic continuous feature disentanglement. Built upon Stable Diffusion, it introduces a style encoder to generate style embeddings and adds a second attention layer in the U-Net for joint style-content modeling. Evaluated on an otitis media dataset, CLUE achieves an FID of 9.30 and a recall of 70.29%. A classifier trained solely on synthetic data attains an F1 score of 83.21%, rising to 94.76% when fused with real data—substantially outperforming the real-data-only baseline. Our key contribution is the first demonstration of controllable diversity generation and cross-domain generalization without introducing any new training data.

0 citationsRead paper

MedRegion-CT: Region-Focused Multimodal LLM for Comprehensive 3D CT Report Generation

Jun 29, 2025

Existing CT report generation methods overly rely on global image features, often neglecting fine-grained lesion details, thereby compromising clinical relevance and interpretability. To address this, we propose a fine-grained multimodal report generation framework: (1) a region-representative token pooling mechanism guided by pseudo-masks enables lesion-level localization; (2) a patient-specific attribute textualization module improves semantic alignment between imaging findings and clinical semantics; and (3) a unified architecture integrates pretrained 2D vision models, general-purpose segmentation models, mask encoders, and multimodal large language models, jointly modeling multi-scale visual and linguistic features via region pooling and text-prompt fusion. Evaluated on the RadGenome-Chest CT dataset, our method achieves state-of-the-art performance, significantly improving report fluency, clinical accuracy, and lesion interpretability—demonstrating superior fidelity to radiological reasoning and diagnostic utility.

0 citationsRead paper

Comparisons between a Large Language Model-based Real-Time Compound Diagnostic Medical AI Interface and Physicians for Common Internal Medicine Cases using Simulated Patients

May 27, 2025

This study addresses the low diagnostic accuracy, prolonged duration, and high cost of initial diagnoses in primary care. We propose a real-time composite diagnostic AI interface powered by large language models (LLMs). Methodologically, we introduce a novel real-time interactive architecture that integrates multi-stage clinical reasoning with dynamic information assimilation, grounded in USMLE Step 2 CS–style case modeling and simulated patient assessment. Experimental results demonstrate that the AI achieves 80% top-1 diagnostic accuracy—surpassing physicians’ 50–70%—and 100% top-2 accuracy—exceeding physicians’ 70–90%. Diagnostic time is reduced by 44.6%, and per-case cost drops by 98.1%; patient satisfaction matches that of physicians. To our knowledge, this is the first LLM-based system achieving clinical-grade real-time composite diagnosis, delivering concurrent advances in diagnostic accuracy, operational efficiency, and economic viability—establishing a deployable technical paradigm for AI-augmented primary-care triage.

0 citationsRead paper

Recent publications

Latest Papers

Deterministic Integrity Gates for LLM-Assisted Clinical Manuscript Preparation: An Auditable Biomedical Informatics Architecture

Jun 08, 2026

Current large language model (LLM)-generated clinical research manuscripts commonly suffer from fabricated citations, data drift, and omissions of reporting guidelines, yet existing tools lack effective validation mechanisms. This work proposes an integrated generation-and-verification architecture that decomposes the writing process into 43 skill modules—including 21 deterministic detectors—orchestrated by a unified coordinator. It introduces a “maximally deterministic” completeness gating mechanism to enforce structured, traceable audits and re-execution checks at each stage. Evaluated on the STARD, PRISMA, and STROBE benchmark datasets, the approach successfully identified all 27 injected defects with zero false positives, substantially outperforming general-purpose LLM-based review methods.

0 citationsRead paper

How does downsampling affect needle electromyography signals? A generalisable workflow for understanding downsampling effects on high-frequency time series

Jan 15, 2026

This study addresses the challenge of high computational cost and poor real-time performance in automated classification of neural muscular disorders using high-sampling-rate needle electromyography (nEMG) signals, where the impact of downsampling on diagnostic information retention remains unclear. The authors propose a unified evaluation framework that, for the first time, integrates shape-aware downsampling, feature space analysis, and classification performance to systematically quantify how different downsampling strategies affect waveform integrity and diagnostic capability in high-frequency temporal biosignals. Evaluated on a three-class neuromuscular disorder classification task, the framework identifies an optimal downsampling configuration that balances computational efficiency with preservation of diagnostic information. Results demonstrate that shape-aware downsampling significantly outperforms conventional methods, offering a generalizable analytical paradigm for efficient processing of high-frequency temporal biosignals.

0 citationsRead paper

CLUE: Controllable Latent space of Unprompted Embeddings for Diversity Management in Text-to-Image Synthesis

Nov 14, 2025

In few-shot domains such as medical imaging, text-to-image generation faces a fundamental trade-off between diversity and stability. To address this, we propose CLUE—a novel framework that achieves controllable, diverse, and stable generation without requiring additional training data. CLUE constructs a prompt-free, controllable latent embedding space and employs KL divergence to enforce prompt-agnostic continuous feature disentanglement. Built upon Stable Diffusion, it introduces a style encoder to generate style embeddings and adds a second attention layer in the U-Net for joint style-content modeling. Evaluated on an otitis media dataset, CLUE achieves an FID of 9.30 and a recall of 70.29%. A classifier trained solely on synthetic data attains an F1 score of 83.21%, rising to 94.76% when fused with real data—substantially outperforming the real-data-only baseline. Our key contribution is the first demonstration of controllable diversity generation and cross-domain generalization without introducing any new training data.

0 citationsRead paper

MedRegion-CT: Region-Focused Multimodal LLM for Comprehensive 3D CT Report Generation

Jun 29, 2025

Existing CT report generation methods overly rely on global image features, often neglecting fine-grained lesion details, thereby compromising clinical relevance and interpretability. To address this, we propose a fine-grained multimodal report generation framework: (1) a region-representative token pooling mechanism guided by pseudo-masks enables lesion-level localization; (2) a patient-specific attribute textualization module improves semantic alignment between imaging findings and clinical semantics; and (3) a unified architecture integrates pretrained 2D vision models, general-purpose segmentation models, mask encoders, and multimodal large language models, jointly modeling multi-scale visual and linguistic features via region pooling and text-prompt fusion. Evaluated on the RadGenome-Chest CT dataset, our method achieves state-of-the-art performance, significantly improving report fluency, clinical accuracy, and lesion interpretability—demonstrating superior fidelity to radiological reasoning and diagnostic utility.

0 citationsRead paper

Comparisons between a Large Language Model-based Real-Time Compound Diagnostic Medical AI Interface and Physicians for Common Internal Medicine Cases using Simulated Patients

May 27, 2025

This study addresses the low diagnostic accuracy, prolonged duration, and high cost of initial diagnoses in primary care. We propose a real-time composite diagnostic AI interface powered by large language models (LLMs). Methodologically, we introduce a novel real-time interactive architecture that integrates multi-stage clinical reasoning with dynamic information assimilation, grounded in USMLE Step 2 CS–style case modeling and simulated patient assessment. Experimental results demonstrate that the AI achieves 80% top-1 diagnostic accuracy—surpassing physicians’ 50–70%—and 100% top-2 accuracy—exceeding physicians’ 70–90%. Diagnostic time is reduced by 44.6%, and per-case cost drops by 98.1%; patient satisfaction matches that of physicians. To our knowledge, this is the first LLM-based system achieving clinical-grade real-time composite diagnosis, delivering concurrent advances in diagnostic accuracy, operational efficiency, and economic viability—establishing a deployable technical paradigm for AI-augmented primary-care triage.

0 citationsRead paper