Handwritten Text and Digit Classification on Rwandan Perioperative Flowsheets via YOLOv5

The African Surgical Outcomes Study, a seven-day, prospective, observational cohort study across 25 countries in Africa reported a rate of serious postoperative complications of 18% and mortality of 2% [1]. 95% of these deaths occurred in the postoperative period and were considered preventable. There are many factors that contribute to postoperative outcomes, but a key approach to decreasing complications is the ability to predict patient outcome trajectories from perioperative parameters [2]. In order to efficiently predict these outcomes, electronic medical record systems are needed. As compared to handwritten paper records, these systems offer profound advantages, including automated transfer of medical information, dynamic search queries, and improved resilience for data backups. In this paper, we implement the digitization of the drug and physiological indicator portions of 363 handwritten perioperative flowsheets sourced from the University Teaching Hospital of Kigali in Rwanda. In both sections, the detection of handwritten words and digits is accomplished using a YOLOv5 model trained on a single class. The bounding boxes are then cropped and classified by a convolutional neural network (CNN). Our experimental results suggest that our proposed method can successfully detect handwritten digits and words as evaluated on object mean average precision (mAP).