Application of Machine Learning methods to predict critical Multiple Myeloma events

Abstract

Multiple Myeloma (MM), a hematological malignancy, presents a complex challenge due to its multifaceted nature driven by genetic, molecular, and clinical factors. This thesis aimed to leverage comprehensive multi-source data from the Multiple Myeloma Research Foundation CoMMpass study to develop accurate and reliable predictive models addressing three critical events in MM. The case studies examined include prediction of relapse, treatment response, and mortality risk. This research constructed predictive models employing longitudinal clinical, phenotype, and transcriptomic data, to anticipate relapse at the first and fifth year since diagnosis and treatment initiation, since a key characteristic of MM is frequent relapses after initial therapeutic responses, signaling drug-resistant clones or disease progression. The models demonstrated robust predictive performance, enabling timely identification of patients at risk for relapse. These forecasting models could allow clinicians to implement preemptive strategies, such as salvage therapies or intensified treatment regimens. Furthermore, predicting treatment response is crucial in MM for guiding therapeutic decisions, optimizing dosages, and customizing surveillance strategies. This task was divided into three subtasks reflecting different contexts of treatment response. For the first and fifth year of first-line treatment, models were trained as binary classifiers using baseline cross-sectional data and longitudinal panel data as two different subtasks. The XGBoost model excelled at both prediction horizons with baseline input data, while the Attention-LSTM model effectively handled sequential data, suggesting further optimization. Additionally, a third subtask predicted response to any line therapy six months ahead, using six months of patient history, with LSTM-variant models yielding high performance results. Significant predictors for treatment response were also identified including treatment-related information, status of MM disease at diagnosis as well as some laboratory measurements. Lastly, this thesis integrated clinical and transcriptomic features to develop prognostic models for 5-year survival in MM, stratifying patients into distinct risk cohorts. The models exhibited strong predictive performance, aiding shared decision-making, resource allocation, and tailoring palliative care. Clustering transcriptomic data provided a grouped risk stratification tool, with SHAP analysis revealing significant predictors, including disease status at diagnosis and duration of treatment responsiveness. In summary, this thesis highlights the potential of advanced predictive modeling in enhancing MM patient management, offering actionable insights for relapse prediction, treatment response, and long-term survival, thereby overall improving quality of life and patient outcomes.