Introduction. For newly diagnosed multiple myeloma (NDMM) patients, median overall survival (OS) and progression free survival (PFS) have dramatically improved during recent years due to the introduction of novel agents. Unfortunately, a subset of patients with NDMM does not benefit from newer therapies reflected in persisting poor outcomes. In contrast, another subset has favorable outcomes despite receiving limited therapy. Though having demonstrated that several genomic events contribute to clinical outcomes, yet the "one-size-fits-all" treatment paradigm remains dominant for NDMM. Here, we present the first artificial intelligence individualized prediction model for NDMM to facilitate individually tailored therapeutic decisions. Methods. We included 1840 patients with available clinical and genomic data from the following cohorts: MMRF CoMMpass (n=1062), MGP (n=492), Moffit AVATAR (n=177), and MSKCC (n=109). The median follow up was 42 months. To build a treatment-adjusted predictive model of individualized risk for patients with NDMM, we considered 160 variables across clinical (e.g., age, ECOG, sex, ISS), therapeutics, genomics, and time-dependent treatments such as autologous stem cell transplant (ASCT) and continuous treatment (Palumbo et al. JCO 2015). A multi-state model was designed across two phases: induction (phase 1), and post-induction (phase 2). Phase 1 included patients that: 1) completed the induction without progression (PD); 2) PD of failed to respond during induction remaining alive after; 3) PD during induction and subsequently died. Phase 2 have patients that: 4) PD after induction and were alive; 5) PD after induction and died; 6) reached remission after induction and were alive; 7) responded after induction and died due to other causes. We leveraged and compared survival methods based on deep neural networks (Neural Cox Non-proportional-hazards;NCNPH), Random Survival Forest (RSF), and Cox proportional-hazard(CPH). Results. NCNPH showed the best cross-validated prognostic performance for OS with median Uno's concordance (C=0.67), followed by RSF (C=0.65) and CPH (C=0.64; Fig. 1a-b). Overall, the model significantly outperformed R2-ISS (C=0.6), ISS (C=0.59) and R-ISS (C=0.57; Fig. 1b). Additionally, the model found 28 genomic features to increase concordance accuracy for PFS by (3% in phase 1 and 1.5% in phase 2), and with greater effect on OS (10% in phase 1 and 5% in phase 2). The 14% of patients who did not respond in phase 1 were enriched for ISS3, age >75 y, 1q amp, NSD2 translocation, TP53 mutations, and deletions on 17p13 and 1p. Varying therapies emerged as a key determinant of risk, in particular, in phase 2, supporting the idea that effective combinations can have a different impact in individual patients and have the potential to significantly change the clinical outcome despite poor prognostication. The model not only predicts patient risk but also the impact of various treatment strategies (i.e., treatment variance). We identified 9 distinct clusters (C#1-9) based on predicted risk and treatment variance. C#8 and C#9 show favorable outcome, especially with ASCT and continuous treatment. C#4 and C#5 included patients with favorable outcome but low treatment variance where ASCT had marginal impact. C#6 had a substantial presence of high-risk patients associated with low variance, age >75y, 1q gain, and NSD2 translocations. Interestingly, NSD2 translocated patients not included in C#6 had an intermediate favorable outcome (p<0.0001). Early PD was observed in C#1 and C#7, both enriched for low treatment variance, and high risk genomic and clinical features. Finally, we identified a group of high-risk patients (C#2 and C#3) with high treatment variance whose poor outcome was mostly driven by a suboptimal treatment (i.e., use of doublet, no ASCT, no continuous treatment) rather than presence of distinct clinical and genomic features, highlighting the need for treatment to be considered to build more robust and accurate prognostic models. Conclusion: This work shows the first comprehensive model integrating new and historical features to train a neural network for individualized risk prediction in NDMM. Utilizing data from 1840 patients that received a variety of therapies, the model captures the interaction of genomic, clinical and therapy factors, enabling personally-tailored therapeutic decisions in NDMM. Figure 1View largeDownload PPTFigure 1View largeDownload PPT Close modal