Utilising routinely collected clinical data through time series deep learning to improve identification of bacterial bloodstream infections: a retrospective cohort study

Abstract

Background

Blood cultures are the gold standard for diagnosing bacterial bloodstream infections, but test results are only available 24–48 h after sampling. We aimed to develop and evaluate models using health-care data to predict bloodstream infections in patients admitted to hospital.

Methods

In this retrospective cohort study, we used routinely collected blood biomarkers and demographic data from patients who underwent blood sample collection for testing via culture between March 3, 2014, and Dec 1, 2021, at Imperial College Healthcare NHS Trust (London, UK) as model features. Data up to 14 days before blood sample collection were provided to long short-term memory (LSTM) or static logistic regression models. The primary outcome was prediction of blood culture results, defined as a pathogenic bloodstream infection (ie, isolation of pathogenic bacteria of interest) or no bloodstream infection (ie, no growth or contamination). Data collected up to Feb 28, 2021 (n=15 212) comprised the training set and were evaluated against a temporal hold-out test set comprising patients who were sampled after March 1, 2021 (n=5638).

Findings

Among 20 850 patients with available data, pathogenic bacteria were observed in the cultured blood samples of 3866 (18·5%) patients. 2920 (62·2%) of 4897 patients who had their blood samples taken more than 48 h after admission to hospital had pathogenic bloodstream infections, and so were defined as having hospital-acquired bloodstream infections. Including data from the 7 days before admission (7-day window approach) and using five-fold cross validation in the training set gave an area under receiver operator curve (AUROC) of 0·75 (IQR 0·68–0·82) and an area under the precision recall curve (AUPRC) of 0·58 (0·46–0·77) for static models and an AUROC of 0·92 (0·91–0·93) and AUPRC of 0·75 (0·72–0·76) for the LSTM model. In the hold-out test set performances were: AUROC of 0·74 (95% CI 0·70–0·78) and AUPRC of 0·48 (0·43–0·53) for static models and AUROC of 0·97 (0·96–0·97) and AUPRC of 0·65 (0·60–0·70) for LSTM. Removal of time series information resulted in lower model performance, particularly for hospital-acquired bloodstream infections. Dynamics of C-reactive protein concentration, eosinophil count, and platelet count were important features for prediction of blood culture results.

Interpretation

Deep learning models accounting for longitudinal changes could support individualised clinical decision making for patients at risk of bloodstream infections. Appropriate implementation into existing diagnostic pathways could enhance diagnostic stewardship and reduce unnecessary antimicrobial prescribing.

Funding

UK Department of Health and Social Care, the National Institute for Health and Care Research, and the Wellcome Trust.