(198) Predicting survival using geriatric assessment (GA) in older adults with advanced cancer using least absolute selection and shrinkage operator (LASSO)

Background: Predicting survival for patients with advanced cancer, especially for older patients, is crucial in clinical decision-making but is challenging, due to the presence of comorbidities and geriatric syndromes. Geriatric assessment (GA) is a multi-dimensional evaluation of the health of older adults that holds promise for improving prognostication in older cancer patients. Developing a new prognostic tool that incorporates GA may help oncologists better predict survival for older cancer patients.

Objectives: To develop and internally validate a survival prediction model using GA in older adults with advanced cancer receiving first-line treatment.

Methods: We used data from patients receiving first-line treatment in two cluster-randomized trials of GA (NCT02107443, 10/2014-4/2017; NCT02054741, 2/2015-3/2019) and vital status linked via National Death Index through 12/31/2021. Both trials enrolled patients aged ≥70 with an advanced solid tumor or lymphoma. A total of 752 patients were eligible; 665 with complete data were included in the analysis. Outcomes were 6-month, 1-, 2-, and 4-year all-cause mortality. Two statistical models were constructed: (1) a model that incorporates routinely collected data, including age, gender, cancer stage, cancer prognosis based on cancer type, and physician-rated Karnofsky Performance Status (KPS), and (2) a model that incorporates GA measures using least absolute shrinkage and selection operator (LASSO)-based Cox regression using a random 80% training sample. Candidate predictors of the LASSO-based GA model included all routinely collected variables and 16 GA measures. Model discrimination was assessed using Harrell’s c-index and compared across models in the 20% validation sample.

Results: We included 665 patients with a mean age of 77 years and median KPS=80. Overall, 23% of patients died within 6 months, 41% died within 1 year, 54% died within 2 years, and 64% died within 4 years. In the validation sample, the c-index for the LASSO-based GA models was 0.70 for 6 months, 0.68 for 1 year, 0.69 for 2 years, 0.69 for 4 years, which was comparable to the routinely collected data model (C-index=0.70 for 6 months, 0.68 for 1 year, 0.69 for 2 years, 0.68 for 4 years).

Conclusions: The LASSO-based GA models showed only moderate discrimination for survival, similar to the routinely collected data model; future research directions should explore communicating prognostic uncertainty for older adults with advanced cancer.