Comprehensive Summary
The study, done by Lipi Mishra and colleagues, looks at how AI and natural language processing models can detect sepsis from unstructured triage notes in the emergency department. Data was obtained from 134266 patients between the years 2016 and 2021 and a variety of machine learning models were tested including random forest, BERT, and neural networks. The random forest model performed the best with an AUPRC of 0.789, with key predictors being age, triage scale, treatment teams and descriptions of symptoms. Right behind it was the BERT model with an AUPRC of 0.7542, highlighting the potential of unstructured triage notes for sepsis prediction.
Outcomes and Implications
Early detection of sepsis in the emergency department is crucial, especially given the prevalence of the disease globally. This paper shows that AI and NLP tools are promising in helping early detection but several limitations need to be addressed before implementing them in a clinical setting. In this particular study, temperature and heart rate were not used and data augmentation strategies are needed to give more artificial training data.