Prediction of progression in MGUS using transformer-based survival time-series modeling.

Abstract

7518 Background: Monoclonal gammopathy of undetermined significance (MGUS) is common, yet only a minority progress to malignant plasma-cell disorders. Mayo risk stratification uses baseline factors and is commonly implemented/validated in a Cox proportional hazards (CoxPH) framework, which may miss prognostic longitudinal patterns. Methods: We assembled 15,158 MGUS patients with longitudinal laboratory data and progression outcomes (smoldering multiple myeloma, multiple myeloma, AL amyloidosis, or Waldenström macroglobulinemia). We compared a Mayo-factor Cox proportional hazards (CoxPH) baseline versus STRaTS, a transformer-based survival model, to predict progression risk at 3-, 5-, and 10-year horizons. STRaTS leveraged longitudinal trajectories of routine labs including hemoglobin, platelets, differential counts (neutrophils/lymphocytes/monocytes), immunoglobulins (IgG/IgA/IgM), free light chains (kappa/lambda), M-spike, and creatinine. To minimize bias, we used laboratory results collected from 1–3 years after diagnosis and started follow-up at year 3; patients progressing before year 3 or lacking follow-up beyond year 3 were excluded. A total of 7,305 patients met eligibility and were split into train/validation/test cohorts of 4,383/1,461/1,461. Discrimination was assessed by concordance index (C-index). Clinical operating performance was evaluated on a held-out test cohort by comparing sensitivity and specificity at the same proportion of patients selected for intensified surveillance (same workload). Results: Among MGUS patients with complete baseline Mayo factors(N=3,388), Kaplan–Meier–estimated 3-/5-/10-year cumulative progression risk was 0.1/0.4/3.4% (Low), 3.0/4.3/7.9% (Low-Intermediate), 4.4/6.4/14.1% (High-Intermediate), and 8.9/12.5/16.9% (High). On a held-out test cohort, STRaTS demonstrated superior discrimination versus the Mayo-factor CoxPH (C-index 0.75 vs 0.538; p<0.0001). Both models were trained and evaluated on identical train/test splits. At the same surveillance workload, STRaTS achieved higher sensitivity with comparable specificity; when 40.3% of patients were selected for closer surveillance, 3-year sensitivity was 81.0% for STRaTS versus 47.6% for CoxPH (Table). Conclusions: Among MGUS patients who remain progression-free at 3 years, transformer-based survival modeling of longitudinal laboratory trajectories provides more accurate individualized conditional progression risk estimates and improved workload-matched identification of progressors compared with the conventional Mayo/Cox approach. Sensitivity/specificity (%). Operating Point Flagged % STRaTS Sens / Spec (%) CoxPH Sens / Spec (%) Int+High, 3y 40.3 81.0 / 58.5 47.6 / 58.9 Int+High, 5y 40.3 81.2 / 60.4 46.9 / 58.8 Int+High, 10y 40.3 25.0 / 56.3 23.8 / 58.7 Two-Group High (≥2 factors), 5y 5.6 37.4 / 93.4 35.4 / 93.1