Validation and enhancement of the PROACCT model for predicting 30-day acute care events following outpatient systemic therapy using granular health care utilization variables.

Abstract

e23243 Background: Acute care events within 30 days of outpatient systemic therapy (ACE30)are common, costly, and potentially preventable. Stein et al. (JOP, 2023) developed thePROACCT Model #2 using nine demographic and treatment variables, achieving apositive predictive value (PPV) of 0.23. Our prior work (He, ASCO 2024) demonstratedminimal improvement using advanced machine-learning methods with the same inputs.We hypothesized that incorporating granular pre-chemotherapy healthcare utilization,vital signs, and laboratory data would improve identification of patients at high risk forACE30. Methods: We analyzed outpatient systemic therapy administrations among 12,231 uniqueadult patients treated at Orlando Health between February 2012 and April 2021. Theoriginal PROACCT Model #2 was validated in this cohort. We then evaluated 133candidate variables and developed enhanced models using 56 variables, includingdetailed acute care utilization history (prior hospitalizations with or without emergencydepartment [ED] visits, prior direct admissions, etc.), pre-chemotherapy vital signs, andlaboratory values. L1-penalized logistic regression, XGBoost, and neural networks weretrained using a 70/30 split. For comparability, patients in the top decile of predicted riskwere classified as high risk. Results: Validation of the PROACCT Model #2 achieved AUC 0.62, sensitivity 0.16,specificity 0.92, and PPV 0.37 in the high-risk decile. The enhanced L1-penalizedlogistic regression model demonstrated superior performance with AUC 0.68, sensitivity0.21, specificity 0.93, and PPV 0.47. Using identical risk thresholds, the enhancedmodel identified 172 ACE30 events among 367 high-risk administrations, comparedwith 135 events using the original model. The strongest predictors of ACE30 includedprior hospitalization with ED visits, prior ED visits without hospitalization, prior directadmissions, metastatic disease, and head and neck cancer, while endocrine therapyclass and breast cancer diagnoses were associated with lower risk. Conclusions: Incorporating granular healthcare utilization history substantiallyimproved precision for identifying patients at high risk for acute care events aftersystemic therapy. These findings suggest that prior acute care utilization captures clinicallymeaningful vulnerability beyond demographic and treatment variables, enabling moreefficient targeting of supportive interventions and care navigation resources. Futurestudies should evaluate whether utilization-informed risk stratification reduces avoidableacute care use.