**Background:** Older adults with advanced cancer often present with multiple physical and psychological symptoms prior to treatment, which may be associated with adverse outcomes such as poor treatment tolerability, functional decline, and worse quality of life. The PRO-CTCAE (Patient-Reported Outcomes version of the Common Terminology Criteria for Adverse Events) is a validated tool for capturing patient-reported symptom burden, but no standard summative score exists. Unsupervised machine learning offers a data-driven approach to identify patient subgroups based on symptom patterns without requiring predefined cutoffs or outcome data.
**Methods:** This secondary analysis used data from the GAP70+ trial, a nationwide, multicenter, cluster-randomized study that enrolled 718 older adults with advanced cancer starting a new treatment regimen between July 2014 and March 2019. A total of 706 participants who completed baseline PRO-CTCAE were included. The analysis used 24 PRO-CTCAE severity items (each scored 0-4, total severity score range 0-96). K-means clustering with Euclidean distance was applied to group patients by symptom severity patterns. The number of clusters was determined by visual examination of the reduction in sum of squared distances. Sociodemographic and clinical variables were compared across clusters using ANOVA and chi-square tests. Generalized linear mixed models (for hospitalization), Cox shared frailty models (for mortality), and multivariable logistic regression (for toxic effects) were used to examine associations between clusters and outcomes, adjusting for age, sex, cancer type, cancer treatment, number of geriatric assessment domain impairments, and Karnofsky performance status.
**Key Results:** The k-means algorithm identified three clusters: low-severity (n=310, 43.9%), moderate-severity (n=295, 41.8%), and high-severity (n=101, 14.3%). Mean total severity scores were 6.33 (SD 3.44), 16.57 (SD 4.32), and 29.80 (SD 7.80), respectively (P<.001). All 24 individual symptom items were significantly higher in the moderate and high clusters. Patients in the high-severity cluster were more likely to receive multiple chemotherapy agents (59.4% vs 40.6% in low cluster), have poorer KPS scores (21.8% with KPS 20-60 vs 6.8% in low cluster), and have more GA impairments (mean 5.49 vs 3.80, P<.001). Unplanned hospitalization occurred in 18.1% (low), 29.5% (moderate), and 34.7% (high) of patients (P<.001). One-year mortality was 37.7%, 51.9%, and 66.3%, respectively (P<.001). In adjusted models, the moderate-severity cluster was associated with higher risk of hospitalization (adjusted RR 1.36, 95% CI 1.01-1.84, P=.046) and mortality (adjusted HR 1.31, 95% CI 1.01-1.69, P=.04). The high-severity cluster was associated with higher mortality (adjusted HR 2.00, 95% CI 1.43-2.78, P<.001) but not significantly with hospitalization (adjusted RR 1.44, 95% CI 0.99-2.10, P=.05). Neither moderate nor high clusters were significantly associated with grade 3-5 toxic effects.
**Clinical Implications:** Unsupervised machine learning using k-means clustering can effectively stratify older adults with advanced cancer into distinct symptom severity groups based on baseline PRO-CTCAE data. The strong association between higher symptom burden and increased risks of hospitalization and mortality suggests that routine symptom assessment prior to treatment initiation provides clinically useful prognostic information beyond traditional geriatric assessment and performance status. These findings support the integration of patient-reported symptom data and machine learning-based risk stratification into clinical decision-making for older adults with advanced cancer, potentially guiding treatment modifications and supportive care referrals. However, the algorithm requires external validation, and the study population was predominantly non-Hispanic White and well-educated, limiting generalizability.