An explainable prognostic prediction panel for sepsis based on serum amino acid profiles.

Background: Sepsis is a life-threatening syndrome requiring aggressive management, and novel noninvasive biomarkers to enable high-confidence risk stratification of sepsis and to predict sepsis-related outcomes are urgently needed. Methods: Quantitative (liquid chromatography) mass spectrometry was used to compare the abundance of serum amino acids between patients with sepsis and healthy controls (HC) to characterize alterations in amino acid profiles associated with sepsis. In addition, multiple machine learning (ML) methods were applied to construct a prognostic prediction model for patients with sepsis. The predictive performance was assessed, and the feature contributions were screened, followed by the development of an explainable prognostic prediction panel for sepsis. Results: Sixty participants in the HC group and 172 patients in the sepsis group (82 patients with septic shock) were included in this study. A discernible difference in the amino acid profiles between the HC and sepsis groups was detected, and the abundance of amino acids differed significantly among the HC, septic shock, and non-septic shock groups, indicating that amino acids could differentiate patients with sepsis from the HC group with good diagnostic performance. Then, 172 patients in the sepsis group were assigned to training and validation sets, and 130 patients with sepsis in the external test set were included. Five ML models (deephit, piecewise constant hazard, probability mass function, resource selection function, and extreme gradient boosting) were subsequently used, and the deephit model was selected according to the greater area under the curve and clinical benefits in the training, validation, and test sets. In the process of reducing features on the basis of feature importance rankings, the Deephit model, which is based on five features, had the best ability to predict survival probability, and an optimized Deephit model, which involves screening five features, namely, glutamine, glycine, lysine, pyroglutamic acid and proline, was successfully developed to predict the prognostic risk probability for patients with sepsis. Conclusions: Collectively, our findings demonstrated that alterations in serum amino acid profiles distinguish patients with sepsis from HCs and are strongly associated with clinical outcomes, supporting their utility as candidate features for prognostic prediction models in sepsis.