Application Evaluation of the Temporal Action Segmentation and Action Quality Assessment Algorithm in Thoracentesis
Author(s)
Yu Qing, MD, Min Zhang, PHD, Zhang Wen, MD.
Zhongshan Hospital，Fudan University, shanghai, China.
Zhongshan Hospital，Fudan University, shanghai, China.
OBJECTIVES: This study evaluates the application effectiveness and cost of an action analysis system in thoracentesis based on computer vision and artificial intelligence technology.
METHODS: For the first time, this study constructs a temporal medical action knowledge graph of thoracocentesis, dividing the thoracocentesis into 12 stages and 39 sub actions, and defining 47 possible incorrect actions. Based on the knowledge graph, we recruited 9 volunteers to collect video data of thoracocentesis and created a dataset. A model based on diffusion mechanism and temporal clustering attention is proposed. In experiments, we divided the experimental group and the control group for a comparison of incorrect action recognition accuracy and time consumption. The experimental results confirm the effectiveness of the proposed framework.
RESULTS: We adopt the universal evaluation metrics on temporal segmentation datasets to evaluate the performance: the frame-wise accuracy (Acc), the Edit Score (Edit), and the F1 scores under different overlap thresholds 10%, 25%, 50% (F1@{10, 25, 50}). Our framework achieved 73.51% Acc, 76.08% Edit Score, 78.36% F1@10, 72.79% F1@25, and 61.54% F1@50, respectively. In the control experiment, the control group was evaluated using traditional assessment methods, while the experimental group used the proposed model as an auxiliary tool. The results show that the experimental group was able to save nearly 50% of time without compromising evaluation performance.
CONCLUSIONS: This study constructs a temporal medical action knowledge graph and a video dataset of thoracocentesis. Additionally, we propose a framework based on temporal clustering attention mechanism and diffusion model. The experimental results and actual deployment confirm the effectiveness of the method.
METHODS: For the first time, this study constructs a temporal medical action knowledge graph of thoracocentesis, dividing the thoracocentesis into 12 stages and 39 sub actions, and defining 47 possible incorrect actions. Based on the knowledge graph, we recruited 9 volunteers to collect video data of thoracocentesis and created a dataset. A model based on diffusion mechanism and temporal clustering attention is proposed. In experiments, we divided the experimental group and the control group for a comparison of incorrect action recognition accuracy and time consumption. The experimental results confirm the effectiveness of the proposed framework.
RESULTS: We adopt the universal evaluation metrics on temporal segmentation datasets to evaluate the performance: the frame-wise accuracy (Acc), the Edit Score (Edit), and the F1 scores under different overlap thresholds 10%, 25%, 50% (F1@{10, 25, 50}). Our framework achieved 73.51% Acc, 76.08% Edit Score, 78.36% F1@10, 72.79% F1@25, and 61.54% F1@50, respectively. In the control experiment, the control group was evaluated using traditional assessment methods, while the experimental group used the proposed model as an auxiliary tool. The results show that the experimental group was able to save nearly 50% of time without compromising evaluation performance.
CONCLUSIONS: This study constructs a temporal medical action knowledge graph and a video dataset of thoracocentesis. Additionally, we propose a framework based on temporal clustering attention mechanism and diffusion model. The experimental results and actual deployment confirm the effectiveness of the method.
Conference/Value in Health Info
2025-09, ISPOR Real-World Evidence Summit 2025, Tokyo, Japan
Value in Health Regional, Volume 49S (September 2025)
Code
RWD217
Topic Subcategory
Distributed Data & Research Networks
Disease
SDC: Respiratory-Related Disorders (Allergy, Asthma, Smoking, Other Respiratory)