A Danish team has developed a prognostic model for diffuse large B-cell lymphoma, using machine learning to incorporate a more detailed evaluation of risk factors. They found that the algorithm outperforms standard clinical prognostic models currently in use.
In the study, published in JCO Clinical Cancer Informatics, the authors described the design of their model. They trained the model on data from the Danish Lymphoma Registry, a nationwide dataset that includes nearly all patients diagnosed with malignant lymphoma in Denmark. The model uses a machine learning technique called stacking, which combines multiple models that may each have different strengths.
“Quite often, one of these models might be good at predicting, for example, the survival of old patients, or then another might be good for predicting the survival of young patients,” explains study first author Jorne Biccler, a PhD student in the department of hematology at Aalborg University. “Instead of picking one model that’s good at predictions for some patients, you try to combine all the different models in such a way that you get very good predictions for most of the patients.”
The current risk stratification model for DLBCL is the International Prognostic Index (IPI), which assigns one point to each of several risk factors, such as patient age older than 60 years, or the presence of extranodal sites. Each risk factor is weighted the same, producing a score between 0-5.
The IPI, introduced in 1993, works well and is simple for clinicians to calculate. “It’s very easy to just look at a patient and see, is he older or younger than 60?” Biccler says.
But sometimes risks change along a continuum. “The prognosis of a 60-year-old patient and a 95-year-old patient are going to be very different,” Biccler noted. And, as increasingly powerful computers become more common, it could become easier for doctors to incorporate more complex predictive algorithms into their routines.
However, the algorithm’s modest boost in predictive power may not yet be enough to supplant the current clinical model, said Christopher Flowers, MD, from the Winship Cancer Institute of Emory University, who was not involved in the study. “The simple clinical model is easy to use and can readily be applied in practice by clinicians and understood by patients and their care providers,” Dr Flowers wrote in an email to Cancer Therapy Advisor. For machine learning approaches to gain value in the clinic, he said, “more precise models involving factors such as gene-expression profiling, computational immunohistochemical or other analyses of the tumor microenvironment, and/or tumor-mutation profiling” would be required.