P-112: Determining the impact of multimorbidity in older patients initiating treatment for newly-diagnosed multiple myeloma using artificial intelligence/machine learning methods

Introduction: Older patients with multiple myeloma (MM) often present with multiple other comorbid conditions (multimorbidity) in addition to their MM. Due to the strict exclusion criteria of clinical trials, the nature and impact of these chronic conditions alongside myeloma treatment and measures of disease risk are not well-understood. Here, we used artificial intelligence and machine learning (AI/ML) methods to understand the incremental impact of multimorbidity on overall survival of MM patients as compared to MM disease characteristics alone.

Methods: We analyzed the impact of multimorbidity on prognosis using U.S. Veterans Affairs (VA) electronic health record and cancer registry data from the 132 VA Medical Centers that treat veterans with cancer. We included all veterans aged 65+ with newly diagnosed MM initiating treatment at VA between 2004 and 2020. We measured pretreatment predictors including sociodemographic variables, MM disease parameters, labs, and treatment. We further measured 33 comorbidities tracked by the CMS Chronic Conditions Data Warehouse, as well as the prescription of 15 clinically-important classes of chronic medication. We trained random forest AI/ML models to predict three-year mortality using 3/4 of the population and validated the best-performing model on the remaining 1/4. Model performance was evaluated using area under the receiver operating characteristic curve (AUC, higher numbers = better discrimination) to measure discrimination and the Brier score to measure calibration (lower numbers = better calibration). Final model performance was compared between models containing MM-specific predictors and models containing all comorbidities and medications in addition to MM-specific predictors.

Results: We identified 4416 patients with a median age of 74 years (interquartile range [IQR] 69-80) meeting the inclusion criteria. Median number of chronic conditions per patient was 8 (IQR 6-10), and median number of different drug classes per patient was 5 (IQR 3-7). The final model built using only MM-specific predictors demonstrated poor predictive performance (AUC 0.65, Brier score 0.23). The final model built using comorbidities and medications in addition to MM-specific predictors demonstrated superior predictive performance (AUC 0.87, Brier score 0.15). Being prescribed a diuretic, a beta-blocker, an opioid, or having chronic kidney disease were among the top predictors contributing to mortality risk.

Conclusions: Our findings in a real-world population of older adults with MM initiating treatment highlight that including information on multimorbidity alongside MM-specific information yields far superior prediction of mortality compared to a focus only on MM-specific information. Further investigation into the disease-disease, disease-drug, and drug-drug interactions that mediate this risk will yield important clinical insights into the mechanisms of mortality in patients treated outside of clinical trials.