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

Summary

Using 20,850 admissions, an LSTM model leveraging 14-day longitudinal labs predicted pathogenic bloodstream infections with AUROC 0.97 in a temporal hold-out set, outperforming static models. Temporal dynamics of CRP, eosinophils, and platelets were key features, suggesting feasibility for earlier, individualized decision-making.

Key Findings

• LSTM using up to 14 days of prior labs achieved AUROC 0.97 and AUPRC 0.65 in a temporal hold-out test set, outperforming static logistic models (AUROC 0.74).
• Time-series information was critical, especially for hospital-acquired bloodstream infections; removing temporal dynamics degraded performance.
• CRP, eosinophil, and platelet trajectories were consistently important predictors of culture outcomes.

Clinical Implications

Integrating time-series predictive models into sepsis workups could triage high-risk patients for expedited diagnostics and targeted therapy, while curbing empiric antibiotic use in low-risk cases.

Limitations

• Single health system; external multi-center prospective validation is needed.
• Outcome labeling depends on culture classification (pathogen vs contamination), which may introduce misclassification.

Future Directions

Prospective impact studies integrating the model into clinical workflows, assessment of clinician-in-the-loop strategies, and external validation across diverse settings.