(PA-388) Predicting Multiple Myeloma Mortality Using Long Short-term Memory Networks (LSTM)

Introduction:  Multiple Myeloma (MM) causes significant morbidity due to complications such as anemia, renal failure, bone lesions, and an increased risk of infections.  These clinical manifestations affect patients' quality of life and negatively impact survival. Predicting mortality risk in MM patients remains challenging due to the complex interplay of disease and treatment-related factors that evolve over time. Some studies have shown the utility of using Machine Learning (ML) models in predicting the incidence and relapse trends of Acute Lymphocytic Leukemia. However, little has been published on using LSTM to predict outcomes in MM patients. This study explores the application of LSTM to predict mortality outcomes in MM patients, leveraging sequential data to enhance prognostic accuracy.

Methods:   We extracted mortality data from the Wide-Ranging Online Data for Epidemiologic Research (WONDER) Centers for Disease Control and Prevention (CDC) database.  Patients from 1999-2021 were evaluated using established ICD-10 codes for Multiple Myeloma, normalized the data, and encoded demographic variables such as age, gender, and race numerically. A total of 12,654 patients were analyzed using this model. We created an LSTM model capable of identifying and leveraging long-term data patterns for mortality predictions using RStudio. The model was trained using a 20-year sequence length to forecast future data points based on historical trends.

Results:   Mortality analysis demonstrated significant variability across demographic groups. The LSTM model effectively captured nuanced variations in mortality trends based on demographic factors such as age, sex, and geographic location. Mortality increased exponentially with age, as evidenced by a strong correlation (R² = 0.98), and male mortality rates were consistently higher than female rates by approximately 38% (95% CI: 36–40%). The model's predictive accuracy was evaluated using Mean Squared Error (MSE) and Mean Absolute Error (MAE), achieving a training MSE of 0.0055 and a validation MAE of 0.053, confirming its reliability and precision. Furthermore, integrated performance metrics (MAE = 1.28 deaths per 1.000 patients, R² = 0.986) reinforce the model's exceptional predictive accuracy and highlight distinct mortality variations across demographic strata.  


Conclusions:   This study demonstrates the LSTM model's strong predictive capability in accurately forecasting mortality outcomes in patients with Multiple Myeloma using demographic and temporal data. The model's high accuracy indicates its potential as a valuable clinical decision-making tool to guide personalized interventions and optimize patient management strategies. Implementing LSTM-based predictive models in clinical settings could facilitate early intervention strategies, ultimately enhancing patient outcomes.  Future research should focus on integrating additional clinical biomarkers and treatment-related variables to enhance predictive performance and clinical applicability.