Visual summarisation of text for surveillance and situational awareness in hospitals

Abstract

Nosocomial infections (NIs, any infection that a patient contracts in a healthcare institution) cost 100, 000 lives and five billion dollars per year for 300 million Americans alone. Surveillance in hospitals holds the potential of reducing NI rates by more than thirty per cent but performing this task by hand is impossible at scale of every appointment, examination, intervention, and other event in healthcare. Narratives in patient records can indicate NIs and their automated processing could scale out surveillance. This paper describes a text summarisation system for NI surveillance and situational awareness in hospitals. The system is a cascaded sentence, report, and patient classifier. It generates three types of visual summaries for an input of patient narratives and ward maps: cross-sectional statuses at the same point of time, longitudinal trends in time, and highlighted text to see the textual evidence leading to a given status or trend. This gives evidence for and against a given NI in the levels of hospitals, wards, patients, reports, and sentences. The system has excellent recall and precision (e.g., 0.95 and 0.71 for reports) in summarisation for the subset of NIs from fungal species on 1,880 authentic records of 527 patients from 3 hospitals. To demonstrate the system design, we have developed a mobile iPad compatible web-application and a simulation with eighteen patients on three medical wards in one hospital during one month with 61 records in total. The design is extendable to other summarisation tasks.