Machine Learning Reveals the Value of Unconventional T Lymphocytes in Sepsis and Prognosis of Elderly Patients With Severe Lower Respiratory Tract Infections

Abstract

Objective: This study enrolled 119 elderly patients with severe lower respiratory tract infections (LRTIs) and used machine learning (ML) to evaluate the predictive value of unconventional T lymphocytes (uT cells) in sepsis and 90-day prognosis.

Methods: We used random forest (RF) and LASSO analyses to screen model uT cells (identified by RF-LASSO overlapping). The ML models, including LR, LDA, RandomForest, XGBoost, KNN, QDA, NaiveBayes, and ANN, were developed. These models were evaluated and compared based on accuracy, precision, recall, F1 score, sensitivity, specificity, area under the ROC curve (AUROC), and Brier score.

Results: Two T cells were identified as factors of sepsis diagnosis: CD3⁺ and CD4⁺CD25⁺CD127dim. The LDA model demonstrated superior performance, achieving an accuracy of 0.806, AUROC of 0.771, F1 score of 0.720, and a low Brier score of 0.182. Four T cells were identified for predicting the 90-day prognosis: CD3⁺, CD3⁺CD4⁺, CD4⁺CD28⁺, and CD4⁺CD25⁺CD127dim. For the 90-day prognosis, the LDA model again performed best, with an accuracy of 0.972, F1 score of 0.952, AUROC of 0.935, and a low Brier score of 0.059.

Conclusions: The LDA model is optimal for both diagnosing sepsis and predicting the 90-day prognosis in elderly patients with severe LRTIs. Key T-cell markers identified for sepsis include CD3⁺ and CD4⁺CD25⁺CD127dim, while the 90-day prognosis model includes CD3⁺, CD3⁺CD4⁺, CD4⁺CD28⁺, and CD4⁺CD25⁺CD127dim T cells. These markers should be prioritized for clinical testing.

Trial registration: Not applicable.

Keywords: T lymphocytes; lower respiratory tract infections; machine learning; prognosis; sepsis.

FIGURE 1

Baseline characteristics of 119 subjects with severe LRTIs. (A) Comorbidities and co‐acute infections. (B) Types of infections. (C) The identified 285 pathogens are associated with severe LRTIs.

FIGURE 2

Flow chart of this study. Initially, 166 subjects aged 65 years or older with severe LRTIs met the Diagnostic Criteria, but 30 cases were excluded due to incomplete clinical data, and 17 cases were excluded due to the presence of malignancies, connective tissue diseases, or long‐term use of immunosuppressive agents. Ultimately, 119 patients were included in the study. The study used RF and LASSO analysis to identify candidate T cells for both the sepsis group and the 90‐day prognosis group. Eight machine learning models were developed: logistic regression (LR), linear discriminant analysis (LDA), RandomForest, extreme gradient boosted trees (XGBoost), k‐nearest neighbor (KNN), quadratic discriminant analysis (QDA), NaiveBayes, and artificial neural network (ANN). The ROC curves were drawn to calculate the area under the ROC curve (AUC). Calibration plots were utilized for further assessment with Brier score. The model performances were assessed based on sensitivity, specificity, accuracy, precision, recall, and F1 score in both training and testing datasets.

FIGURE 3

Differential T‐cell parameters in nonsepsis cohort versus sepsis cohort. (A) Univariable boxplot of 14 T‐cell parameters. (B) Univariable ROC curve and AUC value (95% confidence interval) of 14 T‐cell parameters. (C) Ranking of features from random forest (RF) algorithm based on standardized variable importance (VIMP) scores. (D) Distribution of LASSO coefficients and penalty plot of 14 cells. (E) Candidate T cells identified by RF and LASSO overlapping. (F) Representative dot plots of two candidate T cells showing distribution in nonsepsis and sepsis cases.

FIGURE 4

Differential T‐cell parameters in 90‐day survival cohort versus 90‐day death cohort. (A) Univariable boxplot of 14 T‐cell parameters. (B) Univariable ROC curve and AUC value (95% confidence interval) of 14 T‐cell parameters. (C) Ranking of features from random forest (RF) algorithm based on standardized variable importance (VIMP) scores. (D) Distribution of LASSO coefficients and penalty plot of 14 cells. (E) Candidate T cells identified by RF and LASSO overlapping. (F) Representative dot plots of four candidate T cells showing distribution in 90‐day survivor and 90‐day nonsurvivor.

FIGURE 5

Eight Machine Learning Models of Sepsis Cohort and 90‐day Prognosis Cohort. ROC curve: The dashed diagonal line serves as an area under the curve (AUC) equal to 0.5; the solid lines of different colors represent eight different models, and each model's AUC was calculated. Calibration plot: The x‐axis represents the actual probability of the event, while the y‐axis represents the average predictive probability; the black solid line serves as a reference, and the other color solid lines are the different model fitting lines. The closer the fitting line is to the reference line, the smaller the Brier score is, and the more accurately the model predicted. (A) Training ROC curve of sepsis cohort. (B) Training calibration plot of sepsis cohort. (C) Testing ROC curve of sepsis cohort. (D) Testing calibration plot of sepsis cohort. (E) Training ROC curve of 90‐day prognosis cohort. (F) Training calibration plot of 90‐day prognosis cohort. (G) Testing ROC curve of 90‐day prognosis cohort. (H) Testing calibration plot of 90‐day prognosis cohort.