Predicting Outcomes for Patients With Mantle Cell Lymphoma Using Integrative Prognostic Machine Learning Models

Holly Hill, PhD, MD Anderson Cancer Center and Rice University, Houston, Texas, shares findings from a study on the predictive value and clinical utility of integrative prognostic machine learning models for determining disease outcomes among patients with mantle cell lymphoma (MCL).

“Our model is the first to integrate a dynamic algorithm with multiple clinical and molecular features, allowing for accurate predictions of MCL disease outcomes in a large patient cohort,” Hill and coauthors added. 

Transcript: 

Hello, I'm Holly Hill from the University of Texas, MD Anderson Cancer Center and Rice University in Houston, Texas. I'm pleased to summarize our recent study, which examined 862 patients with mantle cell lymphoma (MCL), who were diagnosed at our institution from 2014 to 2022. 

Clinical pathological data, along with cytogenetics and DNA sequencing data from the patient's tumors were compiled and integrated into a large machine learning model, which utilized [extreme gradient boosted] (XGBoost), which is a gradient-boosted algorithm. From this large patient database, we identified 121 features to predict whether a patient would have an aggressive disease course, or alternatively a more indolent or responsive prognosis.

Our large XGBoost model predicted the binary disease course with an area under the curve of 0.83, which is a measure of performance. We then used the most important features that were identified in the machine learning model to construct smaller parsimonious models with 10 and 20 features respectively. These were able to predict disease almost as well as the model with 121 features. 

The top 10 features important in determining mantle cell lymphoma prognosis were lactate dehydrogenase (LDH), Beta-2 microglobulin and hemoglobin levels, Ki-67% platelet count, bone marrow involvement, [Eastern Cooperative Oncology Group] (ECOG) performance status, the total number of observed somatic mutations, TP53 mutational status, and cytomorphological variant type.

We also used these features in traditional statistical linear and survival models, which performed well but were not as sensitive as the machine learning models. However, these statistical modeling techniques are more explainable to clinicians and other researchers, and show how machine learning can be integrated with more familiar inferential methods. 

We then launched our smaller machine learning model as a [representational state transfer application programming interface] (REST API), or internet-hosted application interface, as proof of concept of how machine learning models could be used in a clinical setting. This interface can be linked to a dynamic machine learning model that can be updated as new research in MCL is presented.

In our study, we also highlight the opportunities provided by novel machine learning methods in understanding rare cancers. This algorithm that we used aptly dealt with missing data and heterogeneous data types. This is common in large clinical databases. In conclusion, machine learning techniques hold a great amount of promise in integrating multiomics and precision oncology. Thank you.

Source: 

Hill HA, Jain P, Ok CY, et al; Integrative prognostic machine learning models in mantle cell lymphoma. Cancer Research Communications, August 1, 2023; 3 (8): 1435–1446. doi: 10.1158/2767-9764.CRC-23-0083