# Prediction of Blood Lactate Values in Critically Ill Patients: A Retrospective Multi-center Cohort Study

**Authors:** Behrooz Mamandipoor, Wesley Yeung, Louis Agha-Mir-Salim, David J. Stone, Venet Osmani, Leo Anthony Celi  
**arXiv:** [2107.07582](https://arxiv.org/abs/2107.07582)  
**Date:** 2021-07-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2107.07582v1/

## Key points

- Mamandipoor et al. conducted a retrospective multi-center cohort study on predicting blood lactate values in critically ill patients to assess mortality risk.
- Elevated serum lactate levels upon initial testing are known predictors of poor outcomes in critical care settings.
- The study aimed to determine if machine learning models could accurately predict subsequent changes in serum lactate levels to help clinicians adjust patient care accordingly.
- Three distinct subgroups were established based on initial lactate levels: normal (<2 mmol/L), mild (2-4 mmol/L), and severe (>4 mmol/L).
- LSTM models successfully predicted deterioration in serum lactate values among MIMIC-III patients, achieving AUC of 0.77 for the normal group, 0.77 for the mild group, and 0.85 for the severe group.
- While slightly lower performance was observed during external validation, the LSTM models still showed promising discrimination capabilities.
- The research suggests that machine learning models can effectively forecast changes in serum lactate levels among critically ill patients, potentially aiding clinicians in early intervention strategies and optimizing patient care decisions.

## Summary

In a retrospective multi-center cohort study, Mamandipoor et al. investigated the prediction of blood lactate values in critically ill patients to assess mortality risk. Elevated serum lactate levels upon initial testing are known predictors of poor outcomes in critical care settings. The study aimed to determine if machine learning models could accurately predict subsequent changes in serum lactate levels, providing valuable information for clinicians to adjust patient care accordingly. The researchers utilized the MIMIC-III and eICU-CRD datasets for internal validation and further validated their findings on an external eICU cohort within the MIMIC-III dataset. Three distinct subgroups were established based on initial lactate levels: normal (<2 mmol/L), mild (2-4 mmol/L), and severe (>4 mmol/L). Outcomes were defined by changes in serum lactate levels between these groups, with additional sensitivity analyses considering a >10% change threshold and evaluating the impact of time intervals between subsequent measurements on predictive performance. The results demonstrated that LSTM models successfully predicted deterioration in serum lactate values among MIMIC-III patients, achieving an area under the curve (AUC) of 0.77 for the normal group, 0.77 for the mild group, and 0.85 for the severe group. While slightly lower performance was observed during external validation, the LSTM models still showed promising discrimination capabilities. Overall, this research suggests that machine learning models can effectively forecast changes in serum lactate levels among critically ill patients, potentially aiding clinicians in early intervention strategies and optimizing patient care decisions. Further clinical studies are warranted to explore how implementing a clinical decision support tool based on these predictive models could positively influence patient outcomes and clinical decision-making processes.