Background: Emergency department (ED) visits and unplanned hospitalizations account for one of the largest drivers of cancer care costs, but targeted measures can help reduce preventable acute care use. Medicare has focused on this, making OP-35 – Admissions and Emergency Department Visits for Patients Receiving Outpatient Chemotherapy – a quality measure that affects reimbursement. We hypothesized that structured Electronic Health Record (EHR) data combined with patient reported outcomes (PRO) could be used to model risk of preventable ED and inpatient (IP) visits. Methods: We selected patients who met inclusion criteria for OP-35 and had one or more Patient-Reported Outcome Measurement Information System (PROMIS) surveys prior to starting chemotherapy in our EHR. Pre-chemotherapy data on patient demographics, diagnoses, procedures, medications, laboratory values, vital signs, prior healthcare utilization and PROs were obtained. Patients with preventable care visits that met OP-35 criteria were identified at several timepoints following chemotherapy initiation. The cohort was randomly split for model training and testing: 80% were selected to train machine learning models, while the remaining 20% were used to measure performance by the Area Under the Receiver Operator Curve (AUROC). Results: 2149 patients were included, of whom 23% had at least one preventable care event within 180 days of starting chemotherapy. Logistic regression with elastic net regularization performed best for predicting preventable ED or IP visits within 180 days of starting chemotherapy (AUROC: 0.799). For this model, predictors of ED or IP visits included liver or brain metastases (ORs 1.12 and 1.07, respectively), prior ED visits (OR 1.09), white race (OR 0.90), worse global physical health (PROMIS) scores (OR 1.04) and higher self-reported pain scores (OR 1.10). Events further from the start of chemotherapy were generally predicted less accurately and models more accurately predicted IP visits compared to ED visits for all timepoints. Conclusions: Chemotherapy treatment has known side effects and complications that result in preventable acute care visits, which are now measured by OP-35. We demonstrate that machine learning models using PROs as well as more traditional clinical variables can predict these events with accuracies that are actionable. Implementation at the point of care is necessary to determine the impact of these predictions on patient outcomes.