Joeky Senders

118 Chapter 6 TABLE 3. Model performance per radiographic feature according to the area under the receiving operating characteristic curve and accuracy. Domain Characteristics AUC (±SD) Accuracy (±SD) Laterality left-sided involvement 0.984±0.017 93.6±3.4 right-sided involvement 0.973±0.024 94.1±3.2 multifocality 0.816±0.048 78.6±4.4 Location frontal lobe 0.960±0.015 89.1±3.5 temporal lobe 0.965±0.022 90.9±3.0 parietal lobe 0.966±0.022 91.3±3.3 occipital lobe 0.982±0.019 96.6±2.2 corpus callosum 0.945±0.031 93.1±3.3 Tumor aspect necrosis 0.962±0.023 92.2±2.0 cystic 0.956±0.056 94.0±2.1 ring enhancement 0.926±0.029 89.0±3.8 heterogenous enhancement 0.853±0.049 82.9±4.8 Other characteristics hemorrhage 0.901±0.052 84.0±6.9 edema 0.949±0.026 89.0±4.1 mass effect 0.899±0.037 82.7±4.8 Abbreviations: AUC=area under the receiver operating characteristic curve SD=standard deviation Discussion The aim of this study was to develop an NLP pipeline that allows for automated variable extraction from narratively-written clinical reports. In the current application, NLP was able to extract 15 radiological characteristics from free-text radiology reports of brain MRI studies in glioblastoma patients with high to excellent performance. Model performance was correlated with the interrater agreement of the manually provided labels rather than the frequency distribution of the variables of interest. Several studies have already developed NLP frameworks for text mining of medical information. Characterizing neoplastic lesions by parsing free-text radiology and pathology reports has been the primary focus in this field of research; 13–18,18–27 however, operative notes, discharge summaries, and outpatient notes are increasingly being analyzed to examine clinical symptoms, postoperative complications, adverse (drug) events, and post-discharge patient follow-up as well. 28–32 Although model performance reported in these studies already exceeds human performance in terms speed and consistency, very few study groups have actually made their code publicly-