Comparison of clinician and model estimates of risk for hospitalization during systemic therapy for advanced cancer

1530 Background: Patients receiving treatment for advanced cancer are at substantial risk for unplanned hospitalization. A validated two-variable risk model can identify patients at increased risk for hospitalization. However, little is known about how model-based estimates of hospitalization risk compare with assessments of treating clinicians. Methods: We identified patients initiating a new line of systemic therapy for advanced non-hematologic cancer. For each patient, we assigned three categorical estimates of 30-day hospitalization risk. The first risk estimate was generated by a validated two-variable risk prediction model with inputs of pretreatment plasma sodium and albumin (PMID: 30995122); continuous risk scores were converted to risk tertiles. We solicited a second risk estimate by real-time survey of a treating oncology clinician; clinicians were instructed to estimate hospitalization risk as low, intermediate, or high, as compared with other patients. A third hybrid risk estimate retained the highest risk category from either the clinician or model risk assessment. We describe the agreement of clinician and model-based estimates of 30-day hospitalization risk, and we compare the sensitivity and specificity of clinician, model, and hybrid high-risk assessments, using McNemar’s test. We compared discrimination of the three risk estimates via the area under the ROC curve (AUC). Results: We identified 104 patients with valid clinician and model hospitalization risk estimates and complete 30-day follow-up. The most common cancer type was lung cancer (27%), the median age was 68 years, and 62% of patients were male. 30-day hospitalization occurred in 21 patients (20.2%). There was moderate to poor agreement between clinician and model categorical estimates of hospitalization risk (weighted kappa = 0.245). The proportion of patients identified as high-risk by the clinician, model, and hybrid assessments was 15.4%, 26.0%, and 33.7%. Sensitivity and specificity of the high-risk categorization for 30-day hospitalization were 38% and 90% for the clinician assessment, 57% and 82% for the model assessment (NSS for comparison with clinician assessment), and 76% and 77% for the hybrid assessment (greater sensitivity [p = 0.008] and lesser specificity [p = 0.001] than clinician assessment). The AUC values for the clinician, model, and hybrid assessments were 0.674, 0.757, and 0.764, respectively. Conclusions: Compared with the estimate of a treating clinician, a two-variable risk model exhibited similar sensitivity and specificity for 30-day hospitalization risk. A hybrid risk assessment incorporating information from the risk model significantly improved on the sensitivity of the clinician risk assessment. Future research should test strategies to prevent hospitalizations by targeting interventions to high-risk patients.