Conventional and unconventional T-cell responses contribute to the prediction of clinical outcome and causative bacterial pathogen in sepsis patients

Abstract

Sepsis is characterized by a dysfunctional host response to infection culminating in life-threatening organ failure that requires complex patient management and rapid intervention. Timely diagnosis of the underlying cause of sepsis is crucial, and identifying those at risk of complications and death is imperative for triaging treatment and resource allocation. Here, we explored the potential of explainable machine learning models to predict mortality and causative pathogen in sepsis patients. By using a modelling pipeline employing multiple feature selection algorithms, we demonstrate the feasibility of identifying integrative patterns from clinical parameters, plasma biomarkers, and extensive phenotyping of blood immune cells. While no single variable had sufficient predictive power, models that combined five and more features showed a macro area under the curve (AUC) of 0.85 to predict 90-day mortality after sepsis diagnosis, and a macro AUC of 0.86 to discriminate between Gram-positive and Gram-negative bacterial infections. Parameters associated with the cellular immune response contributed the most to models predictive of 90-day mortality, most notably, the proportion of T cells among PBMCs, together with expression of CXCR3 by CD4+ T cells and CD25 by mucosal-associated invariant T (MAIT) cells. Frequencies of Vδ2+ γδ T cells had the most profound impact on the prediction of Gram-negative infections, alongside other T-cell-related variables and total neutrophil count. Overall, our findings highlight the added value of measuring the proportion and activation patterns of conventional and unconventional T cells in the blood of sepsis patients in combination with other immunological, biochemical, and clinical parameters.

Keywords: cytokines; endotoxin shock; inflammation; sepsis; unconventional T cells.

Graphical Abstract

Figure 1.

Proportion of T cells, monocytes, and neutrophils, and conventional and unconventional T-cell subsets in patients after sepsis diagnosis. Comparisons shown are between (A) survivors and non-survivors 30 days after sepsis diagnosis; (B) survivors and non-survivors 90 days after sepsis diagnosis; (C) those without and with a microbiologically confirmed infection; and (D) those with a Gram-positive and Gram-negative infection, among those with a positive bacterial culture. P values were generated using two-tailed Mann–Whitney U tests with Bonferroni–Holm corrections for multiple comparisons

Figure 2.

Mean fluorescence intensity (MFI) of HLA-DR, CD86, CD64, CD40, and CD62L on circulating monocytes in sepsis patients. Comparisons between survivors and non-survivors 30 (top) and 90 (bottom) days following a diagnosis of sepsis are shown. p values were generated using two-tailed Mann–Whitney U tests with Bonferroni–Holm corrections for multiple comparisons

Figure 3.

Mean fluorescence intensity (MFI) of CXCR3, CD161, HLA-DR, CD69 and CD25 on Vδ2⁺ γδ T cells, and MFI of CXCR3, HLA-DR, CD69 and CD25 on MAIT cells, with comparisons between sepsis patients with a Gram-positive versus a Gram-negative infection. P values were generated using two-tailed Mann–Whitney U tests with Bonferroni–Holm corrections for multiple comparisons

Figure 4.

Complete case analysis for an Extra Random Forest model tasked with predicting 90-day mortality in sepsis. Performance is documented by a receiver–operating characteristic (ROC) curve (left) and a bar plot (right) showing balanced accuracy, macro F1 score, and ROC area-under-curve (AUC) score. The dotted diagonal line accompanying the ROC curves represents a model with a random performance level

Figure 5.

SHAP (SHapely Additive exPlanations) values for an Extra Random Forest model to predict 90-day mortality. The beeswarm plot (top) shows each observation as a single data point coloured by the value of the feature for that instance, and ranked from the most impactful on the model outcome to the least impactful. The x-axis shows the SHAP value, with a lower value corresponding to an instance having a more significant impact on the negative case for the model (i.e. prediction of survival), and a positive value corresponding to having a more significant impact on the positive case for the model (i.e. prediction of death). The bar plot on the right-hand side of the beeswarm plot shows the imputation error (with a maximum value of 1) and the percentage of missing values observed in the original data. The heatmap (bottom) shows the SHAP values for each patient. The bar plot on the right-hand y-axis shows each feature’s mean absolute SHAP value as a measure of a feature’s impact on model prediction. The line plot above the heatmap displays each patient’s predicted outcome (black line) and the actual outcome (coloured line). The dotted line between the possible outcomes is the expected value, equivalent to the observed mortality. Note that predictions reflect performance on the complete training data and do not reflect how the model would perform when exposed to new data

Figure 6.

SHAP (SHapely Additive exPlanations) values for a Random Forest model to predict Gram-negative infection. A lower SHAP value corresponds to an instance having a more significant impact on the negative case for the model (i.e. prediction of Gram-positive sepsis), and a positive value corresponds to having a more significant impact on the positive case for the model (i.e. prediction of Gram-negative sepsis). The dotted line between the possible outcomes in the heatmap is the expected value, equivalent to the observed incidence of Gram-negative sepsis

Figure 7.

Proportion of Vδ2⁺ T cells plotted against corresponding SHAP (SHapely Additive exPlanations) values that explain the impact on a Random Forest model to predict Gram-negative infection. Each data point represents a unique patient, coloured by the causative pathogen of their acute infection