Using Machine Learning for Rapid M-Spike Prediction

Myeloma patients may not have to wait the standard 3-7 days for monoclonal protein test results to be processed to predict how their disease is progressing (or not) with a new machine learning model. 

Recent research and the development of a new machine learning (ML) model by Roswell Park Comprehensive Cancer Center, Case Western Reserve University, and University Hospitals Seidman Cancer Center experts can predict if the monoclonal protein will go up or down based on other, more immediate lab values. This new ML model was validated with real-world data provided by HealthTree Foundation. 

The monoclonal protein is the key marker that indicates the rise or lowering of myeloma tumor burden, but the electrophoresis test is not always accurate, and it takes several days to process. Delays in testing results can lead to postponed treatment adjustments and uncertainty for patients about their condition. 

Results from this research were recently presented at the 66th ASH Annual Meeting, demonstrating that faster monitoring can lead to faster clinical decisions and improved patient outcomes. 

Developing the Machine Learning Model

The initial phase of the study involved building a predictive model using a Random Forest machine learning algorithm. Random Forest is a popular and highly effective method for regression and classification tasks, known for its ability to handle complex datasets and select important features. The researchers reviewed 171 patient charts from a single institution, extracting 43 independent variables from electronic health records (EHR). These variables included clinical and laboratory data, such as prior M-spike levels, serum protein levels, and other relevant biomarkers.

The dataset was divided into training and test sets to evaluate the model’s predictive performance. Statistical analysis showed that the model had a high level of accuracy in predicting M-spike values. The most significant predictors of the m-spike were previous m-spike measurements and serum total protein. 

Once the model was developed and the machine learning trained, the team tested the model on a larger, independent dataset to confirm its reliability.

External Validation with the HealthTree Cure Hub Dataset

To validate their findings, the researchers applied the model to an external dataset of 619 multiple myeloma patients from the HealthTree Foundation’s Cure Hub Registry. HealthTree Cure Hub is a national patient registry that collects real-world data from multiple myeloma patients, providing a rich resource for research and validation studies.

The external dataset contained de-identified clinical and laboratory parameters, preserving the structure of the original model. The researchers limited the Random Forest model to the three most significant predictors identified in the initial study: serum total protein and the first and second lagged M-spike values. 

The HealthTree data showed that the machine learning model performed well. It would predict m-spike values with a high level of accuracy, with the model showing that 78% of the predicted m-spike levels matched the actual test results. (The model achieving an R² of 0.779)  Another measure called RMSE, showed that the average difference between predicted and actual values was quite small—indicating the predictions were close to reality.

Importantly, the three key predictors remained highly significant, reinforcing the validity of the simplified model and its potential for broader clinical use.

Clinical Implications

The validation of this machine learning model represents a significant advancement in multiple myeloma care. By providing same-day predictions of M-spike values, the model could enable faster treatment decisions, improving outcomes for patients and reducing anxiety associated with long wait times for test results. This could speed treatment adjustments, provide faster and accurate information to patients and streamline the lab testing and healthcare delivery process. 

The study also highlights the potential of patient-driven data platforms like HealthTree Cure Hub to advance medical research. By providing a large, real-world dataset for external validation,  the HealthTree Cure Hub Registry played a crucial role in confirming the model’s reliability and applicability.

Future research will focus on integrating the model into decision-support systems and adding additional data like genetics or treatment history to improve its accuracy. The researchers also recommend exploring the model’s ability to predict treatment outcomes and disease progression under the International Myeloma Working Group (IMWG) criteria.

As myeloma care continues to embrace data-driven solutions and new technological advances, this study serves as a reminder of the transformative potential of machine learning. With further refinement and real-world validation, the model could become a standard component of multiple myeloma care. 

Innovations like this bring us one step closer to reducing the burden on laboratory services, enhancing clinical workflows, and improving care.

To continue reading about research at HealthTree, follow the link below:

Read More HealthTree Research