Background: There is no gold standard approach for delirium diagnosis, making the assessment of its epidemiology difficult. Delirium can only be inferred though observation of behavioural disturbance and described with relevant nouns or adjectives. Objective: We aimed to use natural language processing (NLP) and its identification of words descriptive of behavioural disturbance to study the epidemiology of delirium in critically ill patients. Study design: Retrospective study using data collected from the electronic health records of a university-affiliated intensive care unit (ICU) in Melbourne, Australia. Participants: 12 375 patients Intervention: Analysis of electronic progress notes. Identification using NLP of at least one of a list of words describing behavioural disturbance within such notes. Results: We analysed 199 648 progress notes in 12 375 patients. Of these, 5108 patients (41.3%) had NLP-diagnosed behavioural disturbance (NLP-Dx-BD). Compared with those who did not have NLP-Dx-DB, these patients were older, more severely ill, and likely to have medical or unplanned admissions, neurological diagnosis, chronic kidney or liver disease and to receive mechanical ventilation and renal replacement therapy (P < 0.001). The unadjusted hospital mortality for NLP-Dx-BD patients was 14.1% versus 9.6% for patients without NLP-Dx-BD. After adjustment for baseline characteristics and illness severity, NLP-Dx-BD was not associated with increased risk of death (odds ratio [OR], 0.94; 95% CI, 0.80–1.10); a finding robust to multiple sensitivity, subgroups and time of observation subcohort analyses. In mechanically ventilated patients, NLP-Dx-BD was associated with decreased hospital mortality (OR, 0.80; 95% CI, 0.65–0.99) after adjustment for baseline severity of illness and year of admission. Conclusions: NLP enabled rapid assessment of large amounts of data identifying a population of ICU patients with typical high risk characteristics for delirium. Moreover, this technique enabled identification of previously poorly understood associations. Further investigations of this technique appear justified.
Natural language processing and recurrent network models for identifying genomic mutation-associated cancer treatment change from patient progress notes
