OBJECTIVES: There are multiple new treatment options available for relapsed/refractory Multiple Myeloma (rrMM) improving the outcomes in this incurable malignancy. Here, we investigated if Real World Data (RWD) on rrMM patients allow for characterizing and predicting clinical outcomes.

METHODS: We leveraged RWD database from 2400 rrMM patients who completed second line of treatment between 2004 and 2021, in representative sample of 107 German clinical sites. We developed Python-based data transformation scripts to handle multicollinearity, missing values, hidden missing values and outliers to optimize the dataset. Descriptive statistics, based on the chi-square test and interactive dashboards allowed to determine variables characterizing rrMM patient cohorts. We cross-validated several feature importance methods and submitted the datasets to an auto-ML platform to predict the Time to Next Treatment (TTNT) between the second and third line of treatment, or the remission period (refractory/relapsed patient).

RESULTS: From identified variables with impact on treatment outcomes, our script determined a series of variable subsets which allowed to keep the highest number of patients, while containing no missing values. A series of datasets were generated with specific numbers of variables and patients. For each dataset, Our models (notably XGBoost and neural network models) allowed the prediction of patient-specific TTNT, with a precision from 3 to 6 months, and remission duration as refractory/relapsed patient with an accuracy from 70 to 85%.

CONCLUSIONS: Our study demonstrates the feasibility to use RWD to predict the duration of remission in rrMM patients. Our work aims to develop digital tools to better identify potentially patients with poor outcomes and opens therapeutic perspectives for this patient population with a high treatment need.