Using a Natural Language Processing and Machine Learning Algorithm Program to Analyze Inter-Radiologist Report Style...

To use a natural language processing and machine learning algorithm to evaluate inter-radiologist report variation and compare variation between radiologists using highly structured versus more free text reporting.

28,615 radiology reports were analyzed for 4 metrics: verbosity, observational terms only, unwarranted negative findings, and repeated language in different sections. Radiology reports for two imaging examinations were analyzed and compared - one which was more templated (ultrasound - appendicitis) and one which relied on more free text (chest radiograph - single view). For each metric, the mean and standard deviation for defined outlier results for all dictations (individual and group mean) was calculated. The mean number of outlier metrics per reader per study was calculated and compared between radiologists and between the two report types. Wilcoxon rank test and paired Wilcoxon signed rank test were applied. The radiologists were also ranked based on the number of outlier metrics identified per study.

There was great variability in radiologist dictation styles - outlier metrics per report varied greatly between radiologists with the maximum 10 times higher than the minimum score. Metric values were greater (P < 0.0001) on the standardized reports using free text than the more structured reports.

The algorithm successfully evaluated metrics showing variability in reporting profiles particularly when there is free text. This variability can be an obstacle to providing effective communication and reliability of care.