Feedback-Driven Radiology Exam Report Retrieval with Semantics

Clinical documents are vital resources for radiologists to have a better understanding of patient history. The use of clinical documents can complement the often brief reasons for exams that are provided by physicians in order to perform more informed diagnoses. With the large number of study exams that radiologists have to perform on a daily basis, it becomes too time-consuming for radiologists to sift through each patient's clinical documents. It is therefore important to provide a capability that can present contextually relevant clinical documents, and at the same time satisfy the diverse information needs among radiologists from different specialties. In this work, we propose a knowledge-based semantic similarity approach that uses domain-specific relationships such as part-of along with taxonomic relationships such as is-a to identify relevant radiology exam records. Our approach also incorporates explicit relevance feedback to personalize radiologists information needs. We evaluated our approach on a corpus of 6,265 radiology exam reports through study sessions with radiologists and demonstrated that the retrieval performance of our approach yields an improvement of 5% over the baseline. We further performed intra-class and inter-class similarities using a subset of 2,384 reports spanning across 10 exam codes. Our result shows that intra-class similarities are always higher than the inter-class similarities and our approach was able to obtain 6% percent improvement in intra-class similarities against the baseline. Our results suggest that the use of domain-specific relationships together with relevance feedback provides a significant value to improve the accuracy of the retrieval of radiology exam reports.