Automated surveillance system for surgical site infections from hospital discharge letters

BACKGROUND: The surgical site infection (SSI) occurs in the part of the body where the surgery took place within 30 days. In Europe, 1-10% of surgical patients develop a SSI. Clinical reporting can play an important role in identifying SSIs, but these clinical data should be integrated into surveillance systems to increase their performances. We aim at developing an automated surveillance system to identify SSIs from unstructured text of hospital discharge letters (HDLs). METHODS: We extracted a sample of 2020-2021 HDLs from the University Hospital of Pisa (Italy) and we used a set of 60 keywords to select only HDLs including surgery and infection related terms. Using a positive wound swab test as a proxy for suspected SSI, we performed a record-linkage between HDLs and laboratory data for 2021. We compared the number of records with suspected SSIs before and after the filtering, to assess its sensitivity. Filtered HDLs have been manually labeled by qualified operators in 3 classes: SSI, No SSI, or SSI caused by a prior hospitalization. RESULTS: From 63,609 HDLs in our sample, 22,625 were filtered through keywords. In the pre-filtered dataset (only 2021), 255 patients have at least a positive wound swab test. 92% (235/255) of the patients with a suspected SSI were included after the filtering. In a preliminary analysis, after the labeling of 9,385 HDLs (41% of the dataset) we identified 340 SSIs (3,6%), of which 254 (85%) were caused by a prior hospitalization. CONCLUSIONS: The labeled dataset will be used to train a natural language processing classification algorithm, able to identify suspected SSIs from HDLs. The keyword filtering can be a useful support to reduce the number of HDLs that need to be manually labeled. KEY MESSAGES: • Keyword filtering is the first step for an automated surveillance system of surgical site infections from hospital discharge letters. • Labeling is time-consuming but necessary for the development of a natural-language-processing-based surveillance system for surgical site infections.