AFFILIATIONS
1 Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States, Mass General Brigham, Boston, Massachusetts, United States, The Broad Institute of Harvard and MIT, Cambridge, Massachusetts, United States
2 Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States, The Broad Institute of Harvard and MIT, Cambridge, Massachusetts, United States
3 Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States, The Broad Institute of Harvard and MIT, Cambridge, Massachusetts, United States
4 Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States, The Broad Institute of Harvard and MIT, Cambridge, Massachusetts, United States
5 Harvard Medical School, Boston, Massachusetts, United States, Mass General Brigham, Boston, Massachusetts, United States, The Broad Institute of Harvard and MIT, Cambridge, Massachusetts, United States
6 Harvard Medical School, Boston, Massachusetts, United States, Mass General Brigham, Boston, Massachusetts, United States, The Broad Institute of Harvard and MIT, Cambridge, Massachusetts, United States
7 Mass General Brigham, Boston, Massachusetts, United States
8 Dana-Farber Cancer Institute, Boston, Massachusetts, United States
9 Harvard Medical School, Boston, Massachusetts, United States, Massachusetts General Hospital, Boston, Massachusetts, United States, The Broad Institute of Harvard and MIT, Cambridge, Massachusetts, United States
10 Harvard Medical School, Boston, Massachusetts, United States, Mass General Brigham, Boston, Massachusetts, United States, The Broad Institute of Harvard and MIT, Cambridge, Massachusetts, United States
11 Harvard Medical School, Boston, Massachusetts, United States, Mass General Brigham, Boston, Massachusetts, United States, The Broad Institute of Harvard and MIT, Cambridge, Massachusetts, United States
12 Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States, The Broad Institute of Harvard and MIT, Cambridge, Massachusetts, United States
ABSTRACT
INTRODUCTION
While monoclonal gammopathy of undetermined significance (MGUS) and smoldering multiple myeloma (SMM) have overall low rates of progression to multiple myeloma (MM), stratification of patient progression risk is clinically valuable, justifying closer surveillance and potentially earlier therapeutic intervention in high-risk patients. Established risk models, including IMWG 20/2/20 and PANGEA, mainly rely on clinical, laboratory, and cytogenetic variables while only incorporating a single bone marrow histologic parameter: pathologist-reported overall plasma cell percentage. We hypothesized that AI-based digital pathology models could identify histologic features that improve risk assessment of MGUS/SMM progression to MM.
METHODS
A total of 4,176 whole-slide bone marrow core biopsy images (including H&E and immunohistochemistry) were obtained from a retrospective cohort of 752 MGUS and SMM patients with longitudinal follow-up. Spatial integration of H&E and IHC slides was performed via a custom digital pathology pipeline for tissue segmentation and alignment of bone marrow biopsies. Mask R-CNN-based cell detection models trained on manual cell-level annotations were used to create whole-biopsy cell network graphs from H&E and IHC images, and to accurately quantify plasma cell and immune cell burden and spatial clustering. Additionally, local tissue patch encoder models with a vision transformer architecture were trained via self-supervised learning on H&E and IHC images. Combinations of cell- and patch-level histologic features were incorporated into Cox proportional hazards models to predict time to MM progression among MGUS and SMM patients, evaluated using 5-fold cross-validation.
RESULTS
A CD138 IHC-based cell detection model showed strong performance in plasma cell detection (F1 score: 89.8%), capturing a mean of 37,457 cells per case. Model-derived overall bone marrow plasma cell percentages were found to be systematically lower than pathologist-reported plasma cell burden, consistent with independently verified correlations using spatial -omics technologies and with known upward bias in pathologist visual estimation. Substituting model-derived plasma cell percentages for pathologist-reported values improved prediction of MM progression (mean c-index 0.764 vs 0.710), suggesting that automated enumeration better captures true bone marrow plasma cell burden. In addition, plasma cell clustering quantified by Moran’s I spatial autocorrelation statistic showed systematic increases across MGUS, SMM, and high-risk SMM patients, as well as among MGUS/SMM patients who progressed to MM compared to non-progressors (p=1.6×10-2), demonstrating that plasma cell clustering is a useful adjunctive histologic biomarker.
Incorporating model quantification of kappa and lambda light chains from in situ hybridization slides and T cells from CD3 immunohistochemistry slides further improved discrimination of progressors from non-progressors (mean c-index 0.807). In particular, a systematic increase in total CD3+ T-cell bone marrow involvement was observed across MGUS, SMM, and high-risk SMM patients. Unbiased screening of patch-level histologic features identified the presence of CD3+ T-cell-rich lymphoid aggregates, seen in a subset of MGUS/SMM bone marrow biopsies, as strongly associated with lack of progression to MM (p=1.5×10-4). Finally, substituting AI model-derived histologic biomarkers for pathologist-reported plasma cell burden significantly improved the performance of existing risk stratification models, including 20/2/20 (mean c-index 0.837 vs 0.726; p=2.1×10-4) and PANGEA (mean c-index 0.836 vs 0.755; p=3.6×10-5).
CONCLUSIONS
AI-based digital pathology modeling enables more accurate prediction of MGUS/SMM progression to MM from bone marrow H&E and immunohistochemistry slides, both alone and in combination with clinical and laboratory features used in existing risk stratification models. Additionally, these digital pathology models identify novel histologic features associated with progression to MM, including increased plasma cell clustering and decreased T-cell-rich lymphoid aggregates. These findings support the generalized use of digital pathology models in the routine bone marrow biopsy workup of MGUS and SMM patients for improved risk stratification.