Objective
Use Machine Learning (ML) to identify patterns and trajectories in symptom data in a cohort of older adults receiving chemo in community settings.
Materials & Methods
•The GAP trial enrolled 718 pts aged ≥70 with incurable cancers starting chemo
•708 pts had baseline symptom data
•Patients completed Patient-Reported Outcomes version of Common Terminology Criteria for Adverse Events (PRO-CTCAE) for 26 symptoms at baseline, 4-6 weeks, and 3 and 6 months
•At each timepoint, patients rated symptom severity on a 5-point Likert scale (“None” to “Very Severe”)
•Ward hierarchical clustering was applied to symptom severity data. Number and composition of clusters were obtained at each timepoint by minimizing cluster variance.
•Sankey diagrams were generated to display trajectories.
Results
-
Symptoms seemed to cluster by severity
-
Heatmap examples below​
-
-
At baseline, 3 clusters corresponded to low, medium, and high symptom burden as measured by summing severity scores​
-
Violin plots​
-
-
Similar clusters emerged at each timepoint, with smaller clusters (​e.g., "hand-foot" cluster, with high hand-foot symptom severity but low symptoms otherwise, and "skin" cluster with high skin and mucosal symptoms but low otherwise) emerged at 4-6 weeks and 6 months
-
Violin plots​
-
-
A higher proportion of the medium and high clusters dropped out of the study at each timepoint, compared to the low cluster​
-
Sankey diagram​
-
Conclusions
-
ML permits exploratory analysis of symptom phenotypes in older adults starting chemo.
-
Future work includes:
-
Further understanding symptom trajectory
-
Cluster association with outcomes
-
​
With thanks to: the Geriatric Oncology Research Program, NCI U01 “Understanding Treatment Tolerability in Older Patients with Cancer” (site PI: Mohile), and of course the patients