Acute Care Events During Systemic Cancer Treatment: Moving From Risk Prediction to Clinical Decision Support Using a Two-Model Approach

PURPOSE Acute care events (ACEs, emergency department visits and hospitalizations) are burdensome among patients with cancer—many are preventable. Prognostic risk models have not been consistently deployed as a preventive strategy. We convened a Clinical Advisory Panel (CAP) to address this translational bottleneck and develop clinically and statistically valid prognostic models to enable risk-stratified intervention. METHODS We identified patients age 21+ years initiating cancer treatment at an academic center or affiliated sites. We extracted sociodemographic and clinical information from the EHR. Data were divided into training (50%), validation (25%), and test (25%) sets. Models were developed using constrained elastic-net logistic regression, with clinically informed coefficient constraints. RESULTS In all, 4,697 patients were included, the mean age was 64 years, 71.9% were White, 21.8% had GI cancers, 21.7% had hematologic malignancies, 46.0% were Medicare insured, 77.6% were receiving chemotherapy, and 25.4% were receiving immunotherapy. We developed two models with our CAP, a baseline model predicting risk using a planned anticancer regimen and a follow-up model updating with drug dispensing and clinical changes. ACE in the previous month was the strongest predictor in both models (odds ratio OR, 1.5, baseline model), with chemotherapy receipt (OR, 1.18), heart failure (OR, 1.22), abnormal international normalized ratio (OR, 1.30), and late-stage cancer (OR, 1.20) contributing. Both had acceptable statistical performance (C-statistic 0.71 and 0.70) and identified patients at the highest risk for ACE. CONCLUSION We present a novel approach to ACE prediction among patients receiving cancer treatment using two models to anticipate patients' risk before treatment starts and then update based on clinical trajectory, facilitated by engagement of a CAP. To prevent ACE, risk-stratified interventions should focus on the factors we observed—optimizing comorbidities, proactively managing symptoms from high-toxicity regimens, or advanced disease.