Computer vision reimages rigor of APPT pain site and surface area measurement

Abstract

The Adolescent Pediatric Pain Tool (APPT) contains a body outline diagram (BOD) to facilitate self-report of pain sites and surface area. Inter-rater reliability for manually coded BOD drawings is only fair to moderate. Irregular drawings are prone to manual measurement error. The clinical utility, trending, and interpretability of BOD drawings are significant limitations of APPT. A rigorous method to code BODs and extract data for further analysis is needed. Aim: To digitize BOD drawings and using computer vision improve accuracy of BOD measures to facilitate further analysis. Open source computer vision library, OpenCV, was used to digitize BODs from adolescents with scoliosis (n=42) or pectus excavatum (n=24) obtained preoperatively, daily postoperatively during hospitalization, and at postoperative visits. Pain sites and areas were drawn by adolescents on paper BODs, then scanned and cleaned through grayscaling, median blurring, and thresholding techniques to reduce noise. Scanned images were aligned with reference diagram using homography matrix. Differences between images and reference were extracted. Identified body parts and surface areas were labeled for further analysis. The model was able to extract BOD drawings and identify pain sites that corresponded to APPT scoring template. Correct classification rate for each body location for each patient was 97.2%. Overall sensitivity was 95.5% and specificity 97.3%. This computer vision method of coding BODs is accessible, convenient, and provides better sensitivity and specificity for pain site and surface area than manual coding. Extracted data can be further analyzed for trends and modeled for pain trajectory prediction. While the use of digital applications to prospectively collect BOD data may overcome current APPT limitations, loss of historic data, expense, data access, and electronic medical record (EMR) integration present new limitations. Paper BODs are convenient, inexpensive, easily scanned into EMRs, and now easily analyzed.