Abstract Background: Clinical outcomes for newly diagnosed multiple myeloma (NDMM) patients are heterogenous with survival ranging from months to > 10 years. Though several clinical and genomic features predict outcomes, the “one-size-fits-all” treatment paradigm remains dominant for NDMM. Hypothesis: By integrating clinical, genomic and therapeutic data, using artificial intelligence, an individualized risk-prediction model for NDMM (IRM) can facilitate individually-tailored therapeutic decisions. Methods: We included 1933 patients with clinical and genomic data from 5 cohorts: MMRF CoMMpass (n=1062), MGP (n=492), Moffit AVATAR (n=177), UAMS (n=93), and MSKCC (n=109). The median follow-up was 43 months. Overall, we considered 160 clinical (e.g., age, ECOG, race), therapeutics, and genomic variables. To correct for time-dependent variables such as autologous stem cell transplant (ASCT) and continuous treatment, a multi-state model was designed across two phases: induction (phase 1), and post-induction (phase 2). Neural Cox Non-proportional-hazards (NCNPH) was used to integrate the data and build the model. Results: Overall, the 5-year overall survival (OS) c-index for IRM was 0.73, significantly higher than all existing prognostic models: R2-ISS (0.62), ISS (0.61) and R-ISS (0.56). The overall model accuracy was significantly improved by the inclusion of 12 genomic features, including 1q21 gain/amp, TP53 loss, t(4;14)(NSD2;IGH), complex copy number signatures, APOBEC mutational signature contribution, and del1p. Prescribed therapy emerged as a key determinant of risk, suggesting that effective combinations may have a different impact in the context of individual patient features, with the potential to significantly change clinical outcomes despite poor historical prognostication (i.e., treatment variance). Leveraging these concepts, we interrogated the clinical impact of ASCT and continuous treatment in the context of NDMM treated with bortezomib, lenalidomide and dexamethasone (VRd). Integrating predicted outcomes and treatment variance for all 4 possible treatment combinations (i.e., VRd +/- ASCT +/- continuous treatment) we identified 3 patient groups. In the first group (n=632), patients were characterized by complex genomic features, older age, high ISS, poor outcomes and limited treatment variance, reflecting aggressive and refractory myeloma. The second group (n=571) was characterized by high treatment variance, with favorable outcomes if ASCT and continuous treatment are provided. The last group (n=730) included patients with favorable clinical and genomic profiles, achieving good outcomes, with minimal advantage from ASCT. Conclusion: Integrating historical and emerging genomic features with clinical and therapeutic data, we developed the first individualized risk-prediction model for personally-tailored therapeutic decisions in NDMM. Citation Format: Arjun Raj Rajanna, Francesco Maura, Andriy Derkach, Bachisio Ziccheddu, Niels Weinhold, Kylee Maclachlan, Benjamin Diamond, Faith Davies, Eileen Boyle, Brian Walker, Alexandra Pos, Malin Hulcrantz, Ariosto Silva, Oliver Hampton, Jamie K. Teer, Niccolò Bolli, Graham Jackson, Martin Kaiser, Charlotte Pawlyn, Gordon Cook, Dennis Verducci, Dickran Kazandjian, Fritz Van Rhee, Saad Usmani, Kenneth H. Shain, Marc S. Raab, Gareth Morgan, Ola Landgren. Individualized risk stratification in newly diagnosed multiple myeloma. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5453.
Read full abstract