Joint AI-driven event prediction and longitudinal modeling in newly diagnosed and relapsed multiple myeloma

Zeshan M. Hussain, Edward De Brouwer, Rebecca Boiarsky, Sama Setty, Neeraj Gupta, Guohui Liu, Cong Li, Jaydeep Srimani, Jacob Zhang, Rich Labotka, David A. SontagNpj Digit. Medicine 2024 · DOI

Introduction

Managing multiple myeloma, a complex blood cancer, requires clinicians to balance maximizing survival and preventing disease progression against the risk of toxic side effects from intensive therapies. Current clinical tools, such as the International Staging System (ISS), or specialized AI models often focus on isolated tasks—predicting either survival or specific biomarker trends. These fragmented approaches fail to capture the "big picture" of a patient's clinical history.

Researchers from MIT, Harvard, and Takeda have addressed this gap by developing SCOPE (Simultaneous Cancer Outcome Prediction Estimator), a transformer-based AI model published in npj Digital Medicine. SCOPE provides a unified, holistic view of a patient's journey by simultaneously predicting survival, forecasting future biomarkers, and estimating treatment effects.

Methods: The SCOPE Architecture

The SCOPE model is built on a transformer backbone, similar to the technology powering large language models, which allows it to capture complex, long-range dependencies in longitudinal patient data such as lab results, treatments, and vitals over time.

Overview of the SCOPE model architecture and training workflow, illustrating the joint learning approach for event prediction and biomarker forecasting.

The architecture relies on three key pillars:

  1. Joint Learning: The model is trained to perform three tasks simultaneously: event prediction (e.g., progression-free survival, overall survival, and adverse events), biomarker forecasting (e.g., M-protein and hemoglobin levels), and treatment effect estimation.
  2. Two-Step Training: SCOPE is first "pre-trained" on a forecasting task to learn the general disease dynamics and then "fine-tuned" for specific clinical event predictions.
  3. Counterfactual Simulation: By including treatment as an input variable, SCOPE can simulate different treatment plans for the same patient, providing predicted biomarker trajectories and survival curves for each scenario.

Datasets and Validation

The study utilized data from two major randomized controlled trials:

  • TOURMALINE-MM2 (N=703): Data from newly diagnosed multiple myeloma (NDMM) patients used for training and internal validation.
  • TOURMALINE-MM1 (N=720): Data from relapsed and refractory multiple myeloma (RRMM) patients used for external validation to ensure the model generalizes across different disease stages.

Key Results

SCOPE demonstrated significant improvements over existing clinical standards and specialized machine learning models.

Comparison of SCOPE performance (average c-statistic) against traditional clinical staging (CPH-ISS) and other machine learning models for progression-free survival (PFS) across different observation windows.
  • Superior Survival Prediction: SCOPE significantly outperformed the clinical standard of care (CPH-ISS) across both datasets ($p < 0.001$). For 6-month progression-free survival, SCOPE achieved a C-index of 0.67, compared to 0.60 for the ISS-based model.
  • Enhanced Forecasting: The model outperformed state-of-the-art deep learning models (RNNs and Deep Markov Models) in predicting future lab values. At a 6-month horizon, SCOPE’s Mean Squared Error (MSE) was 0.31, notably lower than the RNN (0.37) and DMM (0.92).
  • Adverse Event Tracking: SCOPE successfully predicted critical side effects like thrombocytopenia (C-index 0.85) and acute renal failure (C-index 0.62), which are vital for real-time dosage adjustments.

Key Insights

Beyond raw metrics, SCOPE revealed deeper clinical insights. A UMAP visualization of the model's "hidden state" showed that it naturally clustered patients by myeloma subtype (e.g., IgA vs. IgG) without explicit instruction. Furthermore, the model identified that patients with the IgA kappa subtype appeared to benefit significantly more from the IRd (ixazomib, lenalidomide, dexamethasone) regimen, suggesting a path toward more refined personalized prescribing.

Limitations and Future Directions

While promising, the researchers noted several areas for future work:

  • Real-World Data: The model was trained on "pristine" clinical trial data; its performance in messier, real-world hospital settings needs validation.
  • Treatment Diversity: Future iterations must include a broader range of therapies, such as quadruplet regimens and bone marrow transplants.
  • Interpretability and Trust: Bridging the gap between complex neural network predictions and clinical trust remains a challenge, despite the use of introspection techniques.
  • Prospective Studies: Live clinical trials are necessary to confirm if using SCOPE actually improves patient outcomes in practice.

Summary updated

BibTeX
@article{hussain2024joint,
  author = {Zeshan M. Hussain and Edward De Brouwer and Rebecca Boiarsky and Sama Setty and Neeraj Gupta and Guohui Liu and Cong Li and Jaydeep Srimani and Jacob Zhang and Rich Labotka and David A. Sontag},
  title = {{Joint AI-driven event prediction and longitudinal modeling in newly diagnosed and relapsed multiple myeloma}},
  journal = {Npj Digit. Medicine},
  year = {2024},
  doi = {10.1038/s41746-024-01189-3},
}
Figure