Data science can improve risk assessment in blood cancer
Danish researchers suggest that data science and machine learning can give patients with lymphoma a more precise risk assessment. The goal is to help more patients get the right treatment at the right time and improve their chances of survival.
“Doctors today use a standard score to assess risk in patients with lymphoma, a type of blood cancer. It is based on surprisingly little information, even though it helps judge how serious the disease appears to be and how closely patients should be followed. In our data, the standard score identified only one patient over the age of 60 in the low-risk group with good treatment prospects. Our model identified 160. This suggests that age carries so much weight in the current score that older patients are almost never assessed as low risk, even when their prognosis is good.”
This is according to 29-year-old PhD student Mikkel Werling, whose PhD is funded by DDSA. He has just published an article together with Alexander D. Fuglkjær, Peter Brown, Carsten U. Niemann, and Rudi Agius in the European medical journal HemaSphere.
In the article, the researchers investigate whether machine learning can improve risk assessment for patients with diffuse large B-cell lymphoma, the most common aggressive form of lymphoma. They compared the clinical standard score used in hospitals around the world with machine learning models that could draw on far more information from the Danish national health registries.
Data from almost 15,000 Danish patients used in research
The research group trained the models on data from 14,832 Danish patients with different types of lymphoma treated over the past 20 years. This allowed the models to learn from multiple diseases.
“In medicine, diseases are often seen as isolated entities. That makes sense if you want to understand the mechanisms driving each disease. But when you are trying to predict patient outcomes, it can be an advantage to think more broadly and let related diseases inform one another,” says Mikkel Werling.
The researchers point to three findings in particular:
- Machine learning provided more precise risk assessment than the model currently used in the clinic. At the same level of false alarms, the model identified 22 percent more of the patients who went on to experience treatment failure. At the same time, it was also more precise in which patients it flagged.
- Age is a clear limitation of the current standard score. In the dataset, the standard score identified only one patient over the age of 60 in the low-risk group with good treatment prospects. The research group’s model identified 160.
- The more information, the better. The models improved as the researchers gave them access to more of the patients’ medical histories and allowed them to learn across related diseases.
“With the current score, you are in practice asking five questions. With our model, you are asking 87. That makes it possible to detect patterns that would otherwise be easy to miss, and gives doctors a better basis for making decisions,” says Mikkel Werling.
The doctor must always make the final decision
Mikkel Werling is a PhD student at the University of Copenhagen and works together with specialists from Rigshospitalet and the Danish Cancer Society. He stresses that the model presented in the study should be seen as decision support, not as a replacement for clinical judgment.
“There are already models built directly into clinical systems and used as support in everyday practice. The question is not whether models should replace the doctor, but how they can best support the doctor’s decisions,” he says.
According to Mikkel Werling, the results also point toward broader uses for the same way of thinking.
“We have just shown that models perform better when they can learn across related types of lymphoma. That is a new way of thinking within blood disorders and blood cancers, where many diseases have such small patient groups that it is difficult to develop strong models for each one. The idea is that related diseases can help models learn patterns that are hard to spot in small datasets. The next step is to take that exact same idea and scale it up fully: first letting the models learn from the health data of the whole population and then adapting them to blood disorders and blood cancers. If that succeeds, it could lead to better predictions and better decision support. In the longer term, it could mean that one and the same model can be used for many different clinical questions, including in areas where patient groups are currently too small to develop good models,” he says.