Automating chart review for identification of infected patients using natural language processing

Expert clinician chart review is the most accurate method to identify patients with infection but is laborious and costly. We aimed to develop and externally validate machine-learning models using structured and unstructured data to identify infection in hospitalized patients compared to chart review.

Adult inpatient encounters at 2 health systems (2006-2016) were manually annotated for suspected infection on admission. Data from one site were used to develop gradient-boosted machine-learning models. Structured features included labs, vital signs, and nursing assessments. Unstructured clinical notes were converted to concept unique identifiers (CUIs) using the clinical Text Analysis and Knowledge Extraction System. Unimodal and multimodal models were tuned to maximize the area under the receiver operating curve (AUROC) in the derivation site through cross-validation. Model discrimination and calibration were assessed in the external validation site.

Of the 3014 included admissions, 677 (22.5%) were classified as having suspected infection on admission. A multimodal model combining n-gram TF-IDF and structured data demonstrated the best performance, significantly outperforming the structured data only model (AUROC 0.88 vs 0.82, P < .001, area under the precision-recall curve 0.71 vs 0.58, P < .001). The most important features came from unstructured data.

Our results highlight the value of unstructured text data for developing models that accurately replicate chart reviews in identifying patients with infection. By expediting the annotation of infection cases, our models can facilitate epidemiological research and the curation of datasets necessary for developing clinical decision support systems for early sepsis detection.