Overview

Minimal residual disease (MRD) assessment by flow cytometry is among the most powerful prognostic tools in multiple myeloma, yet its clinical utility is constrained by operator dependency and lack of standardization. Manual gating strategies vary across institutions, operators, and antibody panel configurations — introducing systematic variability that undermines reproducibility and cross-center comparability. This study presents a fully automated, panel-agnostic machine learning framework designed to address these limitations by processing raw flow cytometry data without manual intervention.

Methodology

The pipeline was developed around three core principles. First, feature generation is entirely within-sample and deterministic, anchoring all phenotypic descriptors to biologically defined plasma cell compartments — making the approach robust to inter-panel variability and fluorochrome-dependent signal scaling. Second, acquisition-level quality control metrics are explicitly incorporated into the modeling framework, allowing the model to account for technical uncertainty rather than assuming uniform data quality. Third, samples are stratified by acquisition depth into primary (Tier-A, ≥500,000 events) and stress-test (Tier-B) cohorts, enabling independent validation under reduced acquisition conditions.

The analytical framework employs a multi-stage classification strategy distinguishing MRD-positive, MRD-negative, and low-abundance borderline cases. Two model families — L2-regularized logistic regression and gradient-boosted decision trees (XGBoost) — were evaluated across two feature sets. Performance was estimated using nested 5-fold cross-validation, and model interpretability was assessed through SHAP-based feature attribution applied exclusively to out-of-fold predictions.

Key findings

• The primary model (EXTENDED feature set, XGBoost, Stage-1) achieved a Tier-A out-of-fold PR-AUC of 0.85 and ROC-AUC of 0.81 for binary MRD classification.
• Independent stress testing in Tier-B samples yielded PR-AUC of 0.95 and ROC-AUC of 0.85, suggesting robust generalization under reduced acquisition depth.
• A dual-threshold framework separates high-confidence rule-out and rule-in zones from an intermediate uncertainty band.
• SHAP analysis identified aberrant plasma cell CD138 expression heterogeneity as the most influential feature, followed by CD38 and CD200 distributional metrics.
• Reliable separation of low-abundance borderline cases (PNQ) remained analytically challenging, reflecting inherent biological and sampling variability of near-threshold disease states.

Significance

This work represents a proof-of-concept toward reproducible, automated MRD reporting in multiple myeloma — with direct implications for treatment response assessment, clinical trial endpoint standardization, and broader access to high-quality MRD evaluation across institutions with varying technical resources.