(PA-340) Modeling Trajectories to Organ Recovery in AL Amyloidosis Using an Ensemble Machine Learning Approach

Introduction: Organ recovery is a major determinant of clinical outcomes in light chain (AL) amyloidosis. We aimed to better define trajectories to organ recovery in AL amyloidosis using logistic regression (LR) and XGBoost (ML), a machine learning model based on gradient boosting.

Methods: This analysis included 376 patients (pts) with newly diagnosed AL amyloidosis from 2017-2023, at a quaternary care center. Hematologic and organ responses were assessed at 6, 12, 18, and 24 months using International Society of Amyloidosis criteria. Predictive models were developed using LR and ML to incorporate 37 variables assessed at the time of diagnosis and during treatment. Hyperparameters for ML were optimized using Bayesian search. The dataset was split into training (70%) and testing (30%) subsets. Shapley additive explanation (SHAP) values were used to interpret feature importance.

Results: In describing clinical trajectories, pts were classified as early responders or non-responders based on organ recovery within 6 months. For pts with renal involvement (n = 238), early response occurred in 67% of pts. At 12, 18, and 24 months respectively, 86%, 75%, and 71% enjoyed sustained responses. For early non-responders, 12, 18, and 24 month response rates were 8.4%, 19%, and 19%. Early response occurred in 59% (n=273) of pts with cardiac involvement and improved to 85%, 68%, and 70% at 12, 18, and 24 months. For early non-responders, 21%, 26%, and 29% responded at these time points. For pts achieving an early VGPR or better, 64.5% achieved early cardiac response and 70.8% achieved early renal response.  

In our analysis, ML consistently outperformed LR in the prediction of organ responses at all time points along these trajectories. As an example, for 6-month cardiac response, ML was 85% accurate for response classification with 94% sensitivity and 73% specificity. LR was 63% accurate, 67% sensitive, and 56% specific. SHAP summary plots identified 24-hour urine protein (SHAP = 0.43), serum M-protein (0.33), age (0.32), dFLC (0.20), and NT-proBNP (0.16) as the most impactful predictors of response. ML illustrated distinct non-linear relationships for each variable that were not captured in LR. SHAP values for 24-hour urine protein decreased within a narrow range, indicating a threshold beyond which proteinuria is associated with lower likelihood of response. Detailed model structure, inputs, outputs, and performance for both LR and ML at all time points will be presented.

Conclusions: This analysis demonstrates that patients with AL amyloidosis follow distinct, time-variable clinical trajectories toward organ recovery. ML provides an accurate and explainable platform for predicting response trajectories that outperforms conventional regression. Our model provides a novel framework for capturing non-linear and dynamic patterns in disease behavior, offering a foundation for earlier risk stratification and more personalized treatment strategies.