RESUMO
BACKGROUND: Presepsin is a soluble CD14 subtype that has been considered as a novel marker for patients with sepsis. This study explored the clinical value of presepsin for sepsis in Southern China, and further established models for diagnosis and prognosis of sepsis through using machine learning (ML), by combining presepsin and other laboratory parameters. METHODS: 269 subjects (105 infected patients, 164 sepsis and septic shock) and 198 healthy controls were enrolled. Laboratory parameters (hematological parameters, coagulation parameters, liver function indices, renal function indices, and inflammatory markers) were collected. Plasma presepsin was tested by chemiluminescence enzyme immunoassay. ML of DxAI™ Research platform was used to establish diagnostic and prognostic models. Sensitivity, specificity, and other performance indicators were used to evaluate the performance of each model. RESULTS: The level of presepsin was obviously increased in sepsis and sepsis shock, compared with that of infected and healthy group (all P < 0.0001). Presepsin concentration was positively correlated with positive blood culture and 30-day mortality in sepsis and septic shock patients. Through ROC curve analysis, Hb, UREA, APTT, CRP, PCT, and presepsin were incorporated into machine learning to construct diagnosis models. Ada Boost model had the best diagnostic efficiency (AUC: 0.94 (95% CI 0.919-0.968) in the training set and AUC: 0.86 (95% CI 0.813-0.900) in validation set). Furthermore, AST, APTT, UREA, PCT, and presepsin were included in the prognosis ML models, and the Bernoulli NB model had greater predictive ability for 30-day mortality risk of sepsis (AUC: 0.706), which was higher than that of PCT (AUC: 0.617) and presepsin (AUC: 0.634) alone. CONCLUSION: Machine-learning model based on presepsin and routinely laboratory parameters showed good performance of diagnostic and prognostic ability for sepsis patients.