What question did this study set out to answer?

This research evaluates a machine learning framework for automated minimal residual disease detection in multiple myeloma using raw flow cytometry data.

April 19, 2026Open Access

P74 | Automated Panel-Agnostic Flow Cytometric Minimal Residual Disease Detection in Multiple Myeloma Using Quality Control-Aware Machine Learning

Key Points

This research evaluates a machine learning framework for automated minimal residual disease detection in multiple myeloma using raw flow cytometry data.
Analyzed bone marrow samples via multiparametric flow cytometry.
Applied machine learning on raw list-mode event data exported as CSV files.
Developed models using two tiers: Tier-A for development and Tier-B for stress testing.
Used XGBoost and logistic regression for classification of MRD status.
Analyzed 125 samples with 63.2% MRD-positive results.
Stage-1 XGBoost achieved a PR-AUC of 0.85 and ROC-AUC of 0.80 with 93% sensitivity.
Tier-B stress testing maintained discrimination with PR-AUC of 0.95; ROC-AUC of 0.85.
Dual thresholds enabled high sensitivity and specificity for clinical decision-making.

Abstract

Introduction. Flow cytometric minimal residual disease (MRD) assessment in multiple myeloma is highly operator-dependent, relying on expert-driven manual gating that varies with acquisition depth, panel composition, and fluorochrome configuration. The aim of this study was to evaluate automated MRD detection using a quality control (QC)–aware, panel-agnostic machine learning (ML) framework applied directly to raw flow cytometry data.Methods. Bone marrow samples were analyzed by multiparametric flow cytometry. Raw, ungated, list-mode event data were exported as CSV files following matrix-based fluorescence compensation. All panels contained core plasma cell markers (CD38, CD138, CD19, CD56, CD200, CD27, CD81, CD45). MRD status was reassessed by two experts and finalized by consensus at a limit of quantification (LOQ) of 10-4. Samples were stratified by acquisition depth into Tier-A (≥500,000 events) for model development and Tier-B (<500,000 events) reserved for independent stress testing. Feature generation followed a fully scripted pipeline applied independently per sample, approximating manual plasma cell and aberrant plasma cell (aPC) gating using distribution-based rules. Acquisition stability, singlet integrity, event depth, distributional features, phenotype correlations, and gate concordance measures were quantified and included as model features. ML analysis followed a staged framework. Stage-1 targeted binary classification of LOQ-positive MRD (≥50 aPC) versus MRD-negative (<20 aPC), excluding positive-not-quantifiable (PNQ; 20–49 aPC) samples from training. Subsequent stages evaluated negative reliability and PNQ behavior. L2-regularized logistic regression and XGBoost models were evaluated using nested cross-validation, with probability thresholds derived in Tier-A and applied unchanged to Tier-B.Results. A total of 125 samples were analyzed: 79 (63.2%) MRD-positive, 36 (28.8%) MRD-negative, and 10 (8.0%) PNQ. Ninety-three samples (74.4%) were Tier-A and 32 (25.6%) Tier-B. Across all configurations, Stage-1 XGBoost showed the best performance. Tier-A out-of-fold evaluation yielded a PR-AUC of 0.85 and a ROC-AUC of 0.80, with sensitivity 0.93 at a probability threshold of 0.5 (Brier score 0.17). Discrimination was preserved in Tier-B stress testing (PR-AUC 0.95; ROC-AUC 0.85). Dual probability thresholds enabled clinically interpretable rule-out (sensitivity 0.96 at probability 0.36) and rule-in (specificity 0.96 at probability 0.89) strategies. Predicted probability distributions showed substantial overlap between PNQ (median 0.83; IQR 0.61–0.86) and LOQ-positive cases (median 0.88; IQR 0.80–0.90), whereas MRD-negative samples clustered at lower probabilities (median 0.27; IQR 0.23–0.36). Logistic regression features achieved lower but competitive performance (Tier-A PR-AUC 0.76). Multiclass analyses (Stages 3–4) with either model showed poor PNQ discrimination with zero recall.Conclusions. This study demonstrates that MRD assessment can be automated using a panel-agnostic ML framework applied directly to raw flow cytometry data. Incorporation of acquisition-level QC metrics and biologically anchored, in-sample feature generation reduces operator dependence and supports probabilistic MRD interpretation. Despite limitations related to retrospective design, single-center data, and sample size, this framework provides a reproducible foundation for scalable, operator-independent MRD assessment and warrants external validation.

Bookmark

View Full Paper