Text Report Analysis to Identify Opportunities for Optimizing Target Selection for Chest Radiograph Artificial Intelligence Models

• DOI: 10.1007/s10278-023-00927-5
• Publication Date: 2024-01-12T18:02:28Z
• Authors: Carl Sabottke; Jason Lee; Alan Chiang; Bradley Spieler; Raza Mushtaq
• Keywords: Lasso; Chest radiograph

Our goal was to analyze radiology report text for chest radiographs (CXRs) identify imaging findings that have the most impact on length and complexity. Identifying these can highlight opportunities designing CXR AI systems which increase radiologist efficiency. We retrospectively analyzed from 210,025 MIMIC-CXR reports 168,949 our local institution collected 2019 2022. Fifty-nine categories of finding keywords were extracted using natural language processing (NLP), their assessed linear regression with without LASSO regularization. Regression also used assess additional factors contributing length, such as signing use terms perception. For modeling word counts regression, mean coefficient determination, R2, 0.469 ± 0.001 0.354 0.002 when considering only keyword features. Mean R2 significantly less at 0.067 0.086 MIMIC-CXR, a combined model data accounting radiologist, keywords, perception, 0.570 0.002. With LASSO, highest value coefficients pertained endotracheal tubes pleural drains masses, nodules, cavitary cystic lesions MIMIC-CXR. Natural analysis textual targets models offer bolster