Data Augmentation for Radiology Report Simplification

Findings of the Association for Computational Linguistics: EACL 2023, pages 1922–1932 May 2-6, 2023 ©2023 Association for Computational Linguistics Data Augmentation for Radiology Report Simplification Ziyu Yang CIS, Temple University zyyang@temple.edu Santhosh Cherian Temple University Hospital santhosh.cherian@tuhs.temple.edu Slobodan Vucetic CIS, Temple University vucetic@temple.edu Abstract This work considers the development of a text simplification model to help patients better un- derstand their radiology reports. This paper proposes a data augmentation approach to ad- dress the data scarcity issue caused by the high cost of manual simplification. It prompts a large foundational pre-trained language model to generate simplifications of unlabeled radiol- ogy sentences. In addition, it uses paraphrasing of labeled radiology sentences. Experimental results show that the proposed data augmenta- tion approach enables the training of a signifi- cantly more accurate simplification model than the baselines. 1 Introduction Radiology reports are unstructured documents writ- ten by radiologists to communicate imaging find- ings to another physician or a qualified medical professional (Goldberg-Stein and Chernyak, 2019). Radiology reports have been increasingly available to patients through portals (Lourenco and Baird, 2020), which has been generally welcomed by pa- tients (Cooper et al., 2020). However, the health literacy of most patients is insufficient to fully com- prehend radiology reports (Lalor et al., 2018) be- cause such reports rely on complex medical jar- gon and use explanations that imply highly special- ized medical knowledge (Delbanco et al., 2012). Several studies even identified adverse effects of sharing radiology reports with patients, including dissatisfaction with care (Rosenkrantz and Flagg, 2015) and undue anxiety and stress (Arora, 2013). There is an increasing need for patient-friendly radiology reporting that can communicate results clearly and be understandable by a diverse patient population. However, asking a radiologist to sup- plement a traditional report with a patient-friendly summary would negatively impact their cognitive load and productivity. This problem motivated re- cent research on the automatic simplification of health records. The proposed approaches include both lexical simplification that paraphrases text (Chen et al., 2018; Biran et al., 2011; Weng et al., 2018) and semantic simplification that seeks to sim- plify grammatically complex text (Shardlow, 2014; Leroy et al., 2016) which recently included deep learning approaches (Lewis et al., 2019; Zhang et al., 2020). However, training deep learning mod- els for medical text simplification requires the col- lection of costly labeled data. To alleviate the data scarcity issue in simplify- ing health reports, particularly radiology reports, this paper proposes a novel approach for data aug- mentation. It augments manually-created labeled data with simplifications generated by a large pre- trained language model such as GPT-3 (Brown et al., 2020). To improve the quality of data aug- mentation, the approach develops a separate deep learning model that evaluates the quality of gen- erated simplifications. Furthermore, the approach also provides data augmentation through paraphras- ing the originally labeled radiology sentences. The proposed data augmentation approach is ex- perimentally evaluated on a unique corpus of man- ually generated labeled data for radiology report simplification. The evaluation includes both auto- matic …

Content truncated to 500 words