(PA-389) Uncovering Hidden Prognostic Patterns in Multiple Myeloma: A Novel Unsupervised Deep-learning Method for Risk Factor Discovery

Introduction: Current survival analysis models often fail to provide actionable clinical stratification criteria. While the Revised International Staging System (R-ISS) for Multiple Myeloma offers clear decision boundaries, existing machine learning approaches have limitations - relying on metrics unrelated to survival or requiring expert interpretation.
We introduce a novel optimization approach derived from the multivariate logrank statistic that enables neural networks to directly learn to derive clusters of distinct survival outcomes from patient features alone. This method successfully stratifies multiple myeloma patients into sharp risk categories while revealing clinically meaningful risk factors in an unsupervised manner.

Methods: We developed a differentiable adaptation of the multivariate logrank statistic compatible with neural network training that optimizes for survival heterogeneity across patient clusters. We applied this to multiple myeloma laboratory data from the CoMMpass study (n=722), using standard clinical biomarkers (hemoglobin, calcium, creatinine, WBC, M-protein, free light chains, LDH, albumin, β2-microglobulin). Explainability analyses quantified each parameter's contribution to risk stratification.

Results: Our approach stratified patients into three distinct risk groups with a C-index of 0.647, outperforming the R-ISS (typical C-index 0.61-0.63). Survival differences were highly significant (p=7.3×10^-17), with the high-risk group showing median PFS of approx. 3 years while the low-risk group maintained >70% progression-free survival at 11 years.
The model independently identified established risk factors (β2-microglobulin, creatinine, M-protein) while also highlighting the importance of WBC count and specific free light chain patterns not explicitly incorporated in current staging systems.

Conclusions: Our novel methodology for optimization of survival heterogeneity can outperform established clinical staging systems, even while relying on standard laboratory parameters alone. The method's ability to discover known risk factors without prior knowledge validates its potential to also identify novel prognostic signatures. This approach is model-agnostic and could yield even greater stratification power when applied to genetic, molecular, or imaging data. By enabling more precise risk assessment, this methodology promises to enhance treatment personalization and resource allocation in multiple myeloma management.