A deep learning approach to automated assessment of peripheral intravenous catheterization skills

Background Peripheral intravenous catheterization is a fundamental, yet high‑risk, nursing procedure that requires precision and repeated practice. However, opportunities for safe, hands‑on training are limited, and conventional assessment relies on subjective, resource‑intensive instructor judgment. Recent advances in computer vision and deep learning offer the potential for objective and scalable evaluation of clinical procedures in nursing education. Methods This study aimed to develop and evaluate a deep learning-based framework for the automated assessment of IV insertion performance. The model was designed to recognize procedural stages from video sequences and determine overall pass/fail outcomes. A total of 300 videos were recorded from six nurses performing IV insertion on mannequin arms. Using a pose estimation algorithm, 42 hand keypoints were extracted from each frame. These sequential data were analyzed by a bidirectional long short-term memory (Bi-LSTM) model to classify six procedural stages and predict overall competency. Results The keypoint detection model demonstrated robust tracking even under partial occlusion and varying illumination. The stage classification achieved an accuracy of 92.9% (F1-score = 0.84; mIoU=0.83), while procedure-level evaluation reached 84.6% accuracy. Misclassifications primarily occurred at stage transitions, notably during catheter insertion and tourniquet release. Conclusions This deep learning framework enables objective, automated, and scalable evaluation of IV insertion performance. It reduces faculty workload, enhances feedback efficiency, and promotes student confidence. Integrating AI-driven assessment into nursing education may advance evidence-based teaching practices and improve clinical readiness and patient safety.