NCI-funded researchers used a machine learning (ML) approach to identify patients who were most likely to benefit, or have adverse effects, from cancer treatment late in life.

The researchers took a simple, unsupervised algorithm (i.e., “k means,” a type of analysis that clusters observations into subgroups without prior knowledge) to look for patterns and to make predictions. Using this predictive ML tool, they were able to assess symptoms and assign risk for adverse consequences from treatment, including unplanned hospitalization and unexpected death.  

The researchers used the algorithm to sort through patient-reported symptom data from 706 older adults who had participated in an earlier NCI clinical trial. The patients reported pre-treatment symptoms ranging in severity, including fatigue, insomnia, memory problems, headaches, and more.

The patients who fell within the moderate-to-severe symptom category were more likely to have unplanned hospitalization and poorer overall survival, but not more toxic effects from treatment. Patients appeared to be most likely to experience problems if they had a greater number of severe symptoms (before treatment) and were suffering from more aggressive cancers (i.e., pancreatic).

Corresponding author, Dr. Huiwen Xu, University of Texas Medical Branch, noted, “Our machine learning approach let us classify patients based on their symptom severity before starting cancer treatments, and to use those classifications to predict how the patients from each group would handle treatment.” He added, “Decision-making support like this could be an essential part of the clinician’s toolkit in the future, enabling oncologists to tailor the care for each patient for the best possible outcome.”