Use of Machine Learning in 2074 Cases of Acute Myeloid Leukemia for Genetic Risk Profiling

Background: Prognosis of Acute Myeloid Leukaemia (AML) is guided by chromosomal and molecular profiling. Standard regression approaches may not be optimal for identifying complex gene-gene interactions.

Aims: We investigated use of machine-learning (ML) by recursive partitioning (RP) and random forest (RF) analysis on two large AML datasets to develop a hierarchical prognostic risk model.

Methods: This study included 2074 non-APL AML cases aged ≤70 from National Taiwan University Hospital (NTUH; n=759) and the German-Austrian AML Study Group (AMLSG; n=1315)(Table 1). Statistical analysis performed in R (version 3.5.1) with RP performed using the rpart (version 4.1-13); RF analysis used randomForestSRC (version 2.7.0) with complexity parameter 0.0007 and minimum subgroup 20. Missing data imputed, branch-points without prognostic significance trimmed. Kaplan-Meier analysis utilized for survival with patients with incomplete data excluded from survival analysis. Overall survival (OS) outcomes were classified into Good (5-year OS >55%), Intermediate (5-year OS >40 - 55%), Poor (5-year OS >20 - 40%) and Very Poor (5-year OS ≤20%).

Results: Median age 48 (range 15 - 70) years. Cohorts were evenly matched except higher rate of allograft in complete remission (CR) or CR with incomplete blood count recovery (CRi) for AMLSG (31% vs 18%)(Table 1). CR or CRi achieved in 70%. Median OS for the entire cohort 2.3 years. Machine-learning to develop a prognostic decision-tree for OS based on random 2:1 selection of cases (training-set) and validated on remaining cases (validation-set). Each cohort produced similar RP-trees (Figure 1). Through ML, a decision tree was developed that hierarchically categorized cytogenetic and molecular factors into groupings that most accurately predicted survival. AML features used in the tree included complex karyotype (CK), inv(16), CEBPA^dmut, inv(3)/t(3;3), FLT3-ITD, spliceosome mutations (U2AF1, SRSF2 or SF3B1), NPM1^mut (in the absence of FLT3-ITD), t(8;21), and other AML identified by single genetic lesions including MLL translocations, NRAS^mut, isolated TP53^mut or isolated ASXL1^mut (Table 2). 261 patients (12.5%) were not classified by the model due to incomplete cytogenetic or mutational data and were excluded from survival analysis. In the CK group, a negative effect on prognosis (CR and OS) was conferred by inclusion of high-risk monosomies or structural chromosomal alterations (-7, del(7q), chromosome 17 abnormalities, -5 or del(5q)) or TP53^mut, whereas CK cases lacking these lesions had a better prognosis (median OS 5.5 vs 25.1 months, p<0.001 and CR 36 vs 70%, p<0.001). In CEBPA^dmut AML, ML highlighted the favorable impact of NRAS^mut (present in 27%) on survival. In patients with FLT3-ITD, very poor prognosis seen with co-occurrence of spliceosome mutations. High FLT3-ITD allelic ratio (AR >0.5) also associated with very poor prognosis in the absence of both spliceosome lesions and NPM1^mut (median OS 11.6 months)_. A poor prognostic category comprising triple mutant FLT3-ITD, NPM1^mut and DNMT3A^mut was identified (median OS 13.2 vs 41.7 months, p=0.002). Cases with spliceosome mutation (SRSF2, U2AF1, SF3B1) had a very poor prognosis if accompanied by either ASXL1^mut or potential heterozygous loss of ASXL1 on 20q (either by del(20q) or -20)(median OS 11.8 vs 21.6 months, p=0.003). In NPM1^mut disease, ML identified good prognosis for patients who with NRAS^mut, whereas co-mutation with IDH1^mut conferred an intermediate prognosis. In patients with t(8;21) AML, KIT mutations identified patients with intermediate prognosis (median OS 36.6 months vs NR, p=0.02). There were 49 patients (2.4%) with RUNX1^mut not categorized to other groups and RUNX1^mutper se did not emerge as a prognostic discriminator in this analysis.

In comparison to ELN2017 classification, ML reclassified 45% of cases - 30% good, 26% intermediate, 26% poor and 18% very poor (compared to 39% good, 31% intermediate and 30% poor). In establishing the accuracy of predicting death at any time for individual cases the error rate was 35% for our ML model compared to 45% for the ELN2017 classification and 51% for ELN based cytogenetic analysis alone.

Conclusion: By combining large datasets from two AML groups using machine-learning, an AML decision-tree was able to refine the accuracy of assessing survival prognosis compared to ELN 2017 classification by providing further prognostic granularity.

View largeDownload slide

View largeDownload slide

Close modal