Machine Learning Model Defines Higher Risk of Venous Thromboembolism in Young Adults with Multiple Myeloma

Introduction: Venous thromboembolism (VTE) is a common cause of morbidity and mortality in patients with multiple myeloma (MM). Up to 10% of them may develop a thrombotic complication, being the risk higher during the first year of diagnosis. Currently, the use of thromboprophylaxis is based on risk scales according to logistic regression that generally are not validated and present modest predictive values. Machine learning (ML) algorithms identifies patterns that help us predict risk factors in a given disease. Numerous studies have shown that ML improves outcome prediction over traditional multivariate analysis models. To optimize VTE management in MM patients, thrombotic risk assessment models (RAM) will be generated in order to predict VTE development in MM patients (VTE-MM). To test our objective, we retrospectively analyzed a total of 133 patients with MM from 5 Spanish hospitals between January 2014 and December 2018. MM cohort was enriched with a total of 45 VTE patients in order to help machine learning algorithms and solving the imbalance problem. To build our predictive models, 131 clinical and biological variables were included. Variables with more than 25% of missing data were removed. Feature pre-filtering was carried out using unsupervised Boruta algorithm. Subsequently, the supervised Artifitial Neural Network (ANN) algorithm (80% training and 20% test) was applied on an input layer of 10 dimensions. To evaluate our ML model, we performed 10-fold cross validation. A confusion matrix was also used in the testing sets to evaluate the accuracy of the model. Parallelly, a univariate logistic regression (ULR) analysis was performed to identify significant variables (P<0.05) to entering into a multiple logistic regression (MLR) model. To determine a score for each independent variable, we used beta coefficients. The model was validated and evaluated applying leave-one-out method and ROC curve respectively. All analyses were carried out in R 3.6.3 and SPSS 26.0. Among VTE-MM (median age 67y, RIQ=60-74), 53% were men, mostly with ECOG >1, BMI 28 kg/m² (RIQ 24-32), 53% had IgG and 45% of VTE-MM had revised international staging system (R-ISS) high at diagnosis. PE was the most frequent event (55%), occurring after diagnosis or 1st line of treatment. (Table 1). Boruta significantly selected 10 features (Fig.1A) and ANN algorithm demonstrated an AUC of 0.82 (Fig.1B and 1C). ML model selected young patients with previous surgeries, CVC before VTE, acute infections and with a high R-ISS as risk predictor factors. Multivariate-derived model were constructed with 8 ULR-derived variables (Fig 1D) and showed an AUC of 0.82 (Fig 1E). MLR model highlighted age (> 65y) as a protective factor (score of -4, CI: 0.003-0.308) (Fig 1D). Deep analysis of this group showed women as the higher risk patients (p=0.02). Conclusion: Both RAM models are good predictors of VTE in MM. They share 4 variables. Among them, young adult MM women seem to have more risk to develop VTE.