University of British Columbia
Towards agentic AI for point-of-care echocardiography reasoning and interpretation
Abstract
dc:descriptionCardiovascular disease remains the leading cause of death worldwide, and point-of-care ultrasound (POCUS) has expanded cardiac assessment beyond the specialist echocardiography laboratory. However, interpreting a focused bedside examination requires more than visual recognition. Clinicians must determine which measurements are supported by available cine clips, interpret ndings according to clinical guidelines, and integrate evidence to reach conclusions. Existing arti cial intelligence models for echocardiography have primarily focused on perception, while reasoning and interpretation remain largely unexplored. This thesis introduces EchoAgent, a guideline-centric agentic framework for video-level echocardiographic measurement and interpretation. The framework separates the scope of visual evidence, spanning frame, video, and study levels, from the function performed on that evidence, namely perception or reasoning. EchoAgent orchestrates specialized tools under a large language model, including cardiac phase detection, measurement-feasibility prediction, quantitative measurement extraction, and guideline retrieval. Measurements are grounded in speci c frames and anatomical structures, while interpretations are supported by retrieved clinical guidelines, producing an auditable reasoning process. EchoAgent is evaluated against monolithic baselines on a curated benchmark of sonographer-veri ed clinical question-answer pairs and externally validated on a subset of MIMIC-IV-EchoQA. Ablation studies analyze tool contributions, error propagation, and the role of temporal video evidence. EchoAgent outperforms ultrasound-speci c and general medical vision language models as well as echocardiography foundation models, with the largest gains on measurement-intensive and multi-step reasoning tasks. These results support guideline-grounded tool orchestration as a promising approach for echocardiographic reasoning and motivate future work toward study-level synthesis across multiple clips.
Degree
thesis:*- Name thesis:degree_name
- Master of Applied Science - MASc
- Level thesis:degree_level
- master's
- Discipline thesis:degree_discipline
- Electrical and Computer Engineering
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Daghyani, Matin
- Contributors dc:contributor
-
- Abolmaesumi, Purang
Rights
dc:rights- Statement dc:rights
-
- Attribution-NonCommercial-NoDerivatives 4.0 International
- Language dc:language
- eng
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/2429/94900