PP499 Topic: AS22–Quality and Safety/Errors/Data Management/Other: SURVEILLANCE OF CENTRAL VENOUS LINE OCCLUSION: HARNESSING BOTH ARTIFICIAL INTELLIGENCE AND HUMAN INSIGHT

Le résumé fourni par la source

Aims & Objectives: Central Venous Line (CVL) occlusion is associated with the risk of infection. Surveillance for CVL occlusion is often limited to self-reporting logs, and thus the incidence is underestimated. We sought to assess whether the nursing note narrative combined with plausible structure data elements, can be modeled using Artificial Intelligence (AI) to complement our current surveillance. Methods: Retrospective chart review of all nursing notes at a large tertiary care Pediatric Intensive Care unit between 2018-2023. We abstracted plausible orders associated with the occlusion of a CVL (for example, ordering a thrombolytic agent). We used Natural Language Processing (NLP) to identify notes describing an occlusion, by manually reviewing 480 notes: using 60% of them to train a model, 20% for validation and 20% for testing. We used Regular Expressions, Support Vector machine and Large Language model to assess the clinical notes, and finally Random Forest modeling to combine it with structure data. The model was then applied to all 5 years of our data. Results: During the study period there were 27,362 notes representing 1347 unique admissions and 1065 unique patients. To achieve a model with a sensitivity of 94% (95% C.I. 86-100%) we had a specificity of 79.3% (95% C.I. 69-90%) (see figure 1 for receiver operating curves). We identified 168 CVL occlusions, of which only 87 were identified by orders of thrombolytic agents. Conclusions: AI modeling with clinical narrative is sensitive for surveillance of CVL occlusion and may enhance compliance with self-reporting logs for granular data. Keywords: Natural Language Processing, artificial intelligence, CLABSI, occlusion

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.