Unsupervised learning to identify symptom clusters in older adults undergoing chemotherapy

Area of Science:

• Oncology
• Geriatrics
• Data Science

Background:

• Unsupervised machine learning (ML), specifically clustering, is underutilized for analyzing patient-reported outcomes.
• Patient-reported symptoms are crucial for understanding treatment experiences in older adults receiving chemotherapy.

Purpose of the Study:

• To apply ML methods to explore and describe patient-reported symptom trajectories in older adults undergoing chemotherapy.
• To identify distinct patterns and clusters of symptoms over time using unsupervised learning.

Main Methods:

• Secondary analysis of prospectively collected data from the GAP 70+ Trial.
• Application of multiple clustering algorithms (e.g., hierarchical agglomerative clustering) to patient-reported symptom data at baseline, 6 weeks, 3 months, and 6 months.
• Utilized gap statistic, elbow plots, silhouette scores, and t-SNE for algorithm selection and validation.

Main Results:

• Clustering analyses revealed significant variability depending on the algorithm used.
• Longitudinal analysis showed most patients maintained low symptom levels, but heterogeneous trajectories existed.
• Small patient clusters exhibited specific toxicities, such as hand-foot/neuropathy symptoms or mucosal toxicity, at different time points.

Conclusions:

• Unsupervised ML offers a data-driven approach to explore longitudinal patient-reported symptom data.
• While consistent patterns were challenging to uncover, symptom composite scores warrant further investigation.
• Further hypothesis-driven validation is necessary, and clinicians should understand ML approach limitations.