Fig. 4: Automated AML-MRD assessment using GMM-based novelty detection.

a Theoretical example of how log-likelihood values extracted from a reference model separate leukemic (LAIP+) and non-leukemic (LAIP-) cells. b Trade-off between mean precision and recall across different GMM percentile thresholds in the semi and fully automated setting in 47 Dx and 35 FU measurements from LAIP29 cohort. c Correlation between MRD% and cMRD% for GMM-based novelty detection in semi- and fully automated settings (47 Dx, 35 FU measurements from LAIP29 cohort). d Correlation between MRD% and cMRD% for UMAP-HDBSCAN in semi- and fully automated settings (47 Dx, 35 FU measurements from LAIP29 cohort). e Spearman correlation between MRD% and cMRD% for different methods in semi- and fully automated settings in 47 Dx and 35 FU measurements from the LAIP29 cohort. f ROC curve for GMM-based cMRD% and manual MRD positivity (>0.1% of WBCs) in 35 FU measurements from the LAIP29 cohort. g Concordance between cMRD and MRD positivity (>0.1% of WBCs) in the semi- and fully automated setting in LAIP29 FU measurements (n = 35). Optimal thresholds (dotted lines) are defined by ROC-AUC statistics. MRD measurable residual disease, cMRD computational measurable residual disease, GMM Gaussian mixture model, LOF local outlier factor, nuSVM one-class support vector machine, LAIP leukemia-associated immunophenotype, Dx diagnosis, FU follow-up, LAIP29 test cohort receiver operating characteristic, AUC area under the curve, ROC, WBC white blood cell.