Figure 2.
Machine-learning algorithms identify serum CXCL9 level and absolute T cell conventional cell numbers of aGVHD development following posttransplant cyclophosphamide-based myeloblative allogeneic BMT. (A) Using classification tree nodal selection at day 28 for prediction of aGVHD, the first tree branch was formed by CXCL9 levels <6.7, which made up 52% of recipients for whom the incidence of grade 2 to 4 aGVHD was 11%, which is less than the overall incidence after day 28 for the cohort of 28%. The second tree branch occurred within patients who had CXCL9 levels ≥6.7; those with conventional T cells (Tconv) <38 cells/µL represented 20% of patients and had an incidence of grade 2 to 4 aGVHD of 24%, whereas those with higher levels represented 28% of patients and had a 60% incidence of grade 2 to 4 aGVHD. The stopping rule was a minimum of 30 observations in a node and a minimum of 15 observations in the terminal node with a maximum of 2 depths of any node in the final tree. (B) Cumulative incidence curves based on the factors identified in the classification tree demonstrated that patients with high CXCL9 levels and high Tconv levels had the highest incidence of grade 2 to 4 and grade 3 to 4 aGVHD at 1 year (63% and 26%), whereas those with high CXCL9 levels and low Tconv counts at day 28 had intermediate rates of grade 2 to 4 and grade 3 to 4 aGVHD at 1 year (26% and 10%), and those with low CXCL9 levels had the lowest levels of grade 2 to 4 and grade 3 to 4 aGVHD at 1 year (11% and 1%). P values in Figure 2B were based on Gray’s test without Bonferroni adjustment. Patients who experienced grade 2 to 4 or grade 3 to 4 aGVHD within 28 days after BMT were not included in this analysis due to the potential for overlap with engraftment syndrome and due to occurrence of GVHD before the biomarker collection at day 28. (C) Random forest regression analyses were then performed to identify the top 10 factors associated with aGVHD development. Similar to classification tree analyses, random forest identified CXCL9, followed by Tconv, as the variables most associated with aGVHD. ST2, memory B cells, CD4+ T cells, BMT year, Reg3α, CD38+ CD4+ T cells, pre-BMT disease status, and age at BMT were also associated with aGVHD development.

Machine-learning algorithms identify serum CXCL9 level and absolute T cell conventional cell numbers of aGVHD development following posttransplant cyclophosphamide-based myeloblative allogeneic BMT. (A) Using classification tree nodal selection at day 28 for prediction of aGVHD, the first tree branch was formed by CXCL9 levels <6.7, which made up 52% of recipients for whom the incidence of grade 2 to 4 aGVHD was 11%, which is less than the overall incidence after day 28 for the cohort of 28%. The second tree branch occurred within patients who had CXCL9 levels ≥6.7; those with conventional T cells (Tconv) <38 cells/µL represented 20% of patients and had an incidence of grade 2 to 4 aGVHD of 24%, whereas those with higher levels represented 28% of patients and had a 60% incidence of grade 2 to 4 aGVHD. The stopping rule was a minimum of 30 observations in a node and a minimum of 15 observations in the terminal node with a maximum of 2 depths of any node in the final tree. (B) Cumulative incidence curves based on the factors identified in the classification tree demonstrated that patients with high CXCL9 levels and high Tconv levels had the highest incidence of grade 2 to 4 and grade 3 to 4 aGVHD at 1 year (63% and 26%), whereas those with high CXCL9 levels and low Tconv counts at day 28 had intermediate rates of grade 2 to 4 and grade 3 to 4 aGVHD at 1 year (26% and 10%), and those with low CXCL9 levels had the lowest levels of grade 2 to 4 and grade 3 to 4 aGVHD at 1 year (11% and 1%). P values in Figure 2B were based on Gray’s test without Bonferroni adjustment. Patients who experienced grade 2 to 4 or grade 3 to 4 aGVHD within 28 days after BMT were not included in this analysis due to the potential for overlap with engraftment syndrome and due to occurrence of GVHD before the biomarker collection at day 28. (C) Random forest regression analyses were then performed to identify the top 10 factors associated with aGVHD development. Similar to classification tree analyses, random forest identified CXCL9, followed by Tconv, as the variables most associated with aGVHD. ST2, memory B cells, CD4+ T cells, BMT year, Reg3α, CD38+ CD4+ T cells, pre-BMT disease status, and age at BMT were also associated with aGVHD development.

Close Modal
This Feature Is Available To Subscribers Only

or Create an Account

Close Modal
Close Modal