(A-146) Development of Model for Predicting Early Mortality in Older Cancer Patients Treated with Chemotherapy

Background: Several prediction models have been developed to identify geriatric oncology patients receiving chemotherapy with a high risk of early mortality who may benefit from additional care. Nevertheless, existing models are either burdensome to implement in practice, have a high risk of bias, or were inadequately reported to allow external validation and implementation. Consequently, a prediction model that builds on prior models and could be easily implemented in electronic health record (EHR) systems may be useful for personalizing care decisions for geriatric oncology patients receiving chemotherapy.

Objectives: We aimed to develop and internally validate a model to predict the risk of 6-month all-cause mortality after chemotherapy initiation among geriatric patients with solid cancers.

Methods: We used institutional registry and EHR data from an urban community cancer center in North Texas, USA. The eligible population included patients aged ≥65 years, diagnosed with a first primary histologically confirmed American Joint Committee on Cancer stage I – IV solid cancer between 2013 and 2023, and received first course chemotherapy at our institution. We fit a prespecified logistic regression model with 6-month all-cause mortality as the outcome and two predictors used in prior models that are routinely measured: stage IV and Geriatric Nutritional Risk Index (GNRI, calculated using serum albumin concentration, body weight, and height). We assessed model discrimination using the area under the receiver operating characteristic curve (AUC). In addition, we used bootstrap internal validation with 500 replicates to estimate optimism-corrected AUC, calibration, and net benefit.

Results: Our study population comprised 429 geriatric cancer patients. Median age was 68 years (interquartile range=66 – 73), 73% were racial/ethnic minorities, 60% were females, 41% had stage IV cancer, and 6-month mortality was 17%. AUC of the model was 0.74 (95% confidence limits [CL]: 0.69, 0.80). The optimism-corrected AUC was 0.74 (95% CL: 0.69, 0.79), the optimism-corrected mean calibration was 1.0 (95% CL: 0.80, 1.2), and the model had net benefit between decision thresholds of 1% and 40% compared with the strategies of intervening on all or none.

Conclusions: We developed a prediction model using routinely collected EHR data that may be clinically useful for identifying geriatric cancer patients receiving chemotherapy with a high risk of early mortality. The simplicity of the model and straightforward automation within EHR offers a practical alternative to more complex models that require predictors not routinely measured across all settings or time-intensive data collection. Nevertheless, model validation is recommended before use in other settings.