Use of artificial intelligence to support surgical education personnel shortages in low- and middle-income countries: developing a safer surgeon

Purpose In many low- and middle-income countries (LMICs), non-surgeon physicians are responsible for urgent and emergent surgical procedures due to the paucity of formally trained surgeons. The insufficiency of practicing surgeons also limits the mentorship and educational opportunities for these physicians. To assist with the limited formal surgical education for non-surgeon physicians, our team created an appendicitis curriculum with an associated task-trainer. The aim of this study was to develop an artificial intelligence (AI) model to automatically identify the tasks of the open appendectomy procedure from recorded videos for assessment and feedback. Methods Fifty videos were collected via cell phone cameras of expert surgeons and novices performing open appendectomies on a low-cost task-trainer. Video frames were extracted at one frame-per-second intervals, and then two investigators made annotations using a python-based video annotation software. A deep learning model was designed to identify which steps of the procedure were shown in the video frames. The AI model consists of a deep convolutional neural network for capturing the visual features and an attention-augmented bi-directional recurrent neural network to incorporate temporal dependencies among multiple randomly sampled frames from each video. Results Videos were used to train the AI model and evaluate its performance. Videos that did not include all procedural steps ( n = 7) were excluded. The current accuracy and per-class F1 score (harmonic mean of precision and sensitivity) are 60.74% and 42.43% respectively. The average throughput of the model is > 100 frames per second, which substantiates that the AI model is suitable for real-time applications. Conclusion AI can be used to discriminate performed steps of an open appendectomy on our task-trainer. Collecting additional videos for AI model training will help increase its accuracy and efficiency. This technology can be further developed to determine if each step is performed correctly, which can facilitate providing feedback to learners. Such future AI developments may help assess and provide objective feedback to LMIC physicians who have limited access to expert feedback.