AutoMedTS: Automated Modeling of Physiological Time Series for Surgical Suturing Action Recognition

In laparoscopic surgical training and evaluation, real-time recognition of surgical actions with transparency outputs is crucial for automated, objective, and immediate instructional feedback to support skills improvement. However, we face challenges due to limited dataset sizes and variability in surgical environments. This study presents AutoMedTS, an end-to-end automated machine learning framework customized for medical time-series data, enabling rapid deployment using surgical suturing trajectories collected from both expert and novice surgeons. The proposed method features key improvements including: (i) a novel temperature-scaled Softmax resampling technique effectively addressing severe class imbalance, and (ii) an uncertainty-aware ensemble selection mechanism ensuring robust predictions across surgeons with varying skill levels. Additionally, the approach emphasizes model transparency to meet the high standards of reliability and transparency required in medical applications. Compared to deep learning methods, traditional machine learning models not only facilitate efficient rapid deployment but also offer significant transparency advantages. Experimental results demonstrate that our method provides fast, stable, and reliable real-time surgical action recognition in clinical training environments. Code and data are publicly available at https://github.com/baobingzhang/AutoMedTS.