Robust Digitization of Perioperative Surgical Flowsheets From Low and Middle Income Countries

In many low- and middle-income countries, healthcare providers continue to rely heavily on paper-based medical records for data collection. This reliance stems from multiple barriers, including the high costs and infrastructure demands associated with implementing and maintaining electronic medical record systems and automated data capture technologies. Consequently, clinical data documented on paper remains largely inaccessible in digital form, in contrast to the availability of real-time digital data common in high-income settings. This lack of digitization presents a significant obstacle for clinicians, researchers, and quality improvement initiatives, impeding efforts to leverage data for improving patient care and outcomes. To address these challenges, the ChartExtractor program was created to automate the digitization of paper anesthesia charts. In this study, we introduce several enhancements to the system through the integration of advanced image processing and machine learning techniques. Unlike previous methods, this work combines thin plate splines and nonrigid point registration to enhance image alignment. Furthermore, we leverage multiple clustering techniques to accurately label the axes of intraoperative vital sign charts, making it easier to extract the corresponding values. Lastly, we used hyperparameter tuning to select smaller, more accurate object detection models, reducing inference time by $54 \%$ and deployment size by 183 MB. These advancements lay the groundwork for the EQUAL Anaesthesia mobile health platform, a low-cost solution designed to support digital data collection and clinical decision-making in resourceconstrained healthcare settings.