Comparison of anomaly detection and supervised learning for acute myeloid leukemia blast identification: A proof-of-concept study
Background. Flow cytometry is the established standard for identifying leukemic blast populations in AML, but relies predominantly on subjective, time-consuming manual gating. Machine learning approaches are being explored to support automated blast detection, yet the relative performance of anomaly detection versus supervised classification remains incompletely characterized. Methods. We compared six computational approaches for blast identification: four anomaly detection algorithms (Isolation Forest, One-Class SVM, Local Outlier Factor, Half-Space Trees), a supervised Random Forest classifier, and an unsupervised UMAP-HDBSCAN clustering approach. The models were developed on the BLAST110 cohort (110 samples) using nested cross-validation with patient-level grouping, and externally validated on two independent LAIP29 samples spanning high (about 41%) and low (about 1%) blast prevalence. Results. Random Forest achieved the strongest external validation performance (F1 = 0.965 diagnostic sample; F1 = 0.587 follow-up sample). Anomaly detection methods showed weaker, more variable performance, with substantial further decline at low blast prevalence (F1 as low as 0.0006); One-Class SVM performed best among anomaly detectors (F1 = 0.731, diagnostic sample). Several models showed a marked drop in performance between cross-validation and external validation. UMAP-HDBSCAN clustering failed to separate blast and non-blast populations into distinct clusters. Conclusions. Supervised classification substantially outperformed the anomaly detection and unsupervised clustering approaches evaluated here, contrasting with prior work favoring anomaly detection for AML blast identification. This divergence may reflect differences in anomaly detection design; label-efficient approaches may still hold practical value where labeled data collection is limited.