Basic characteristics
A total of 328 kidney transplant recipients underwent 955 tests of the immune monitoring panel from November 1st, 2017 to December 31th, 2019 in our center. A sizable part of the them were perioperative patients. Because the induction treatment had a significant impact on lymphocytes, only the patients more than or equal to three months post kidney transplantation were enrolled. 46 eligible kidney transplant recipients diagnosed with pneumonia underwent the immune monitoring test during the first week after admission to hospital. Among them, 17 (37.0%) patients developed into severe pneumonia. As control, 100 eligible stable kidney transplant recipients with the data of 100 tests were randomly selected. The study flow was shown in figure 1.
The clinical characteristics of the pneumonia group and the stable group showed no significant difference in age, gender, donor source, time since transplant and CNI regimen. Obviously, the stable group had a better allograft function than the pneumonia group. Because 16 patients in the pneumonia group received transplants in other hospitals, the induction treatment was not available. As a result, the induction treatment showed significant difference in these two groups. The details were shown in table 1.
Immune status characterized by the panel
Compared with the stable group, the pneumonia group showed poor immune status, which was characterized by significant lower cell counts of total T cells (CD3+ T cells), T cell subsets (CD4+ T cells and CD8+ T cells), B cells and NK cells (table 2). Although the percentages of total T cells and NK cells showed statistical difference, they were not clinically significant (pneumonia versus stable, 76.79 ± 11.71versus 73.35 ± 10.28 for total T cells, P = 0.015; 12.78 ± 8.81 versus 17.11 ± 9.68 for NK cells, P = 0.003). The percentages of T cell subsets and B cells showed no significant difference. Notably, the CD4/CD8 ratio, which was reported as an immune biomarker, also showed no significant difference (pneumonia versus stable, 1.21 ± 0.61versus 1.12 ± 0.59, P = 0.320).
The remaining two parameters also provided meaningful information. The expression of HLA-DR on monocytes was significantly lower in the pneumonia group (931.17 ± 671.15 versus 1392.53 ± 764.37, P < 0.001), while the expression of CD64 on neutrophils were much higher in the pneumonia group (589.20 ± 605.44 versus 101.11 ± 54.08, P < 0.001).
Machine learning models based on immune monitoring
To study whether the parameters of the immune monitoring panel were associated with pneumonia in kidney transplant recipients, univariate LR was performed to assess each parameter (supplement table 2). Several parameters including monocyte HLA-DR, neutrophil CD64 and cell counts of T cells, B cells and NK cells showed significance, but the performance was not ideal (data not shown).
To improve the performance, machine learning models including SVM, LR, MLP and RF were developed as described in the methods to evaluate the risk of pneumonia. After five rounds of training/validation rotation, the average sensitivity, specificity, PPV, NPV and AUC of these modes were shown in table 3. All the models had good results with AUC (figure 2), of which the SVM model had the highest AUC of 0.940. Notably, the SVM model also had good clinical practicality, with sensitivity of 81.7%, specificity of 92.0%, PPV of 83.6% and NPV of 91.3%. Monocyte HLA-DR (MFI), neutrophil CD64 (MFI), CD8+ T cells (cells/μl), NK cells (cell/μl) and TBNK (T cells, B cells and NK cells, cells/μl) were selected to build the SVM and LR models. The parameter coefficients were shown in table 4. The MLP model, as one of the techniques of artificial neural network (ANN), calculated the probability of each category. The average AUC was 0.923, and the sensitivity, specificity, PPV and NPV were 71.8%, 92.0%, 82.7% and 87.9%, respectively. As an example, one tree of the RF model was shown in figure 3. A total of ten trees were developed. The final result was obtained through majority voting from the ten trees. The average AUC was 0.895, and the sensitivity, specificity, PPV and NPV were 73.6%, 95.0%, 88.0% and 89.2%, respectively.
Compared with mild pneumonia, severe pneumonia had a worse impact on allograft and patient survival. Among the 46 pneumonia patients, 17 cases progressed to severe pneumonia. Three patients died with functioning allografts, and one patient lost allograft. All of them were from the severe pneumonia group. Because all pneumonia patients received the immune monitoring tests early after admission, we also studied whether the result of the immune monitoring panel could predict the prognosis of pneumonia. The comparison between the two groups was shown in table 5. Only the cell count of NK cells showed significance (135.60 ± 108.79 versus 59.28 ± 39.50, P = 0.027); the mild pneumonia group had higher monocyte HLA-DR, but not statistically significant (1068.59 ± 758.07 versus 696.76 ± 410.57, P = 0.127).
The RF model was used to predict the prognosis based on the immune monitoring panel. A ten-tree RF model was developed, and one tree of the final algorithm was shown (figure 4a). Similarly, after five rounds of training/validation rotation, the average sensitivity, specificity, PPV, NPV and AUC were 53.3%, 80.0%, 68.0%, 75.3% and 0.760, respectively (figure 4b).