This paper introduces LEA-Dialog, a multi-turn diagnostic dialogue dataset for lower-extremity arteriovenous diseases, together with a carefully developed diagnostic handbook and a process-aligned agentic framework for structured outpatient diagnosis. The dataset provides stage annotations for each turn and guideline-grounded probability trends, enabling evaluation beyond final diagnostic accuracy. Experiments show that the framework improves reasoning stability and reduces drift across both online and offline LLMs, with particularly large gains for smaller offline models.

In this study, we introduce an MLP approach for extracting multimodal cause utterances in conversations, utilizing the multimodal conversational emotion causes from the ECF dataset. Our research focuses on evaluating a bi-modal framework that integrates video and audio embeddings to analyze emotional expressions within dialogues. The core of our methodology involves the extraction of embeddings from pre-trained models for each modality, followed by their concatenation and subsequent classification via an MLP network. We compared the accuracy performances across different modality combinations including text-audio-video, video-audio, and audio only.