Relapse prediction in multiple myeloma patients treated with isatuximab, carfilzomib, and dexamethasone

Multiple myeloma (MM) patients experience repeated cycles of treatment response and relapse, yet despite close monitoring of disease status through M protein measurements, no standard model exists for relapse prediction in MM. We investigate the feasibility of predicting relapse using a hierarchical Bayesian model of subpopulation dynamics by training and testing the model on 229 patients from the IKEMA trial. After observing between 11 and 18 treatment cycles, the model predicted relapse within six cycles with an average sensitivity between 60 and 80 %, and an average specificity between 60 and 90 %. A model of linear extrapolation is preferable when patients have been observed for less than 6 cycles, but for longer observation windows the hierarchical Bayesian model is preferred. Including available baseline and longitudinal covariate information did not improve predictive accuracy. A survival analysis showed that two model parameters separated patients into groups with significantly different PFS (p < 0.001). Statement of SignificanceCurrently, no standard model exists for relapse prediction in multiple myeloma. A personalized model of M protein development could guide the frequency of follow-up measurements, reduce uncertainty for patients, and give clinicians more time to choose the best subsequent treatment for each patient. Furthermore, models that predict relapse are required to study the effect of changing treatment in advance of relapse rather than in response to it. Our work addresses this need by developing a hierarchical Bayesian model of subpopulation dynamics for prediction of future M protein values. We validate the model on a patient cohort treated with state-of-theart CD38 inhibitor therapy and show that it can accurately predict relapse within the next six treatment cycles, highlighting the promise of mathematical modeling in multiple myeloma and for personalized medicine in general. Declaration of InterestsF.S. received honorarium from Sanofi, Janssen, BMS, Oncopeptides, Abbvie, GSK, and Pfizer. The authors declare that they have no other conflicts of interest.