Background: PRO-CTCAE captures symptomatic adverse events (e.g. pain, fatigue) and may indicate poor treatment tolerability in older patients (pts) with advanced cancer. Using unsupervised machine learning which can detect unknown patterns in data, we aimed to evaluate if clusters identified based on PRO-CTCAE severity were associated with hospitalization and survival. Methods: We included pts randomized to the control arm of GAP 70+ (URCC 13059; PI: Mohile), which enrolled pts aged ≥70, with incurable solid tumors or lymphoma, and ≥1 geriatric assessment (GA) domain impairment starting a new treatment regimen. Measures included 24 PRO-CTCAE items (v1.0) with severity attributes (item 0-4; total score 0-96, higher score = greater severity). The unsupervised algorithm (K-means with Euclidean Distance) clustered pts at baseline based on similarities of severities of the 24 items. We examined if the clusters were associated with treatment-related hospitalization within 3 months and lower survival at 6 months using Logistic and Cox regressions. Results: Of the 369 control pts, 366 completed GA and PRO-CTCAE at baseline (mean age 77.2, 94.3% white, 30.9% with GI and 31.4% with lung cancer; mean number of impaired GA 4.4). By PRO-CTCAE, the most prevalent symptoms were fatigue (82.7%), pain (60.9%), and decreased appetite (58.7%). Greater GA impairment was associated with 20 PRO-CTCAE items (fatigue, pain, and decreased appetite having the strongest associations; all Pearson's r > 0.33). Three clusters were identified: Low Severity (51.4%); Moderate Severity (34.4%), and High Severity (14.2%). Mean total severity score was 6.9 (low), 16.9 (moderate), and 28.7 (high), respectively (p < 0.01). No difference in demographics was found among clusters. Percent of pts hospitalized were 21.3% (low), 36.5% (moderate), and 38.5% (high) (p < 0.01); survival rates were 81.9% (low), 71.4% (moderate), and 55.3% (high) (p < 0.01). Controlling for cancer type and GA, compared to pts in Low Severity cluster, pts in Moderate and High Severity were more likely to be hospitalized (odds ratio = 1.77, p = 0.03); pts in High Severity cluster were more likely to die (hazard ratio = 2.23, p = 0.01). Conclusions: Unsupervised machine learning was able to partition pts into different PRO-CTCAE severity clusters; pts with higher baseline severity were more likely to be hospitalized or die.