Identifying and predicting the most costly patients in oncology.

Abstract

6633 Background: Quality based payment programs in medicine are currently being introduced nationally, aimed to improve care and reduce cost. This study aimed to evaluate the top spenders (TS) after cancer diagnosis and predict TS at two separate time points using predictive analytics. Methods: Patient characteristics, cancer details, treatments, adverse events, and outcomes were collected for patients treated for cancer at Mayo Clinic from 2007 - 2017. Standardized costs over a 2 year period after first treatment were obtained from the Mayo Clinic Cost Data Warehouse with Medicare reimbursements assigned to all services and adjusted to the 2017 GDP Implicit Price Deflator for inflation. TS were identified as those with greater than 93rd percentile costs (≥$113,158) due to a sharp rise in cost at that level. Descriptive statistics and univariate analysis were used for comparison. A prediction model with a training (80%) and validation set (20%) using multivariate selection was used to predict TS and was repeated using information available at 1) the time of consultation and 2) at last follow-up. Results: A total of 80,385 patients were included and 5,626 TS were identified. Mean cost (25th, 75th percentile) overall was $44,953 ($16,776, $51,889). Prediction models at time 1 and 2 had ROC AUC statistics of 0.82 and 0.89 in training and 0.82 and 0.88 in the validation indicating good prediction of high costs. Factors most predictive of TS included need for blood transfusions within 90 days of treatment (OR 5.3), bone marrow transplant (OR 4.0), mild liver disease (OR 3.5), hemiplegia (OR 3.4), weight loss > 10% within 90 days of treatment (OR 3.3), upper GI cancer (OR 3.0), ‘other’ cancer type (OR 2.8), immunotherapy use (OR 2.7), and hospitalizations within 90 days (OR 2.4), all p < 0.001, among others. The largest costs were due to hospital services in the TS and non-TS groups, mean costs $114,258 and $13,185 respectively. Conclusions: This is the first study to predict with high accuracy the top spenders in Oncology. Patient comorbidities and toxicities were among the strongest predictors of high costs, along with certain treatments (bone marrow transplant and immunotherapy). Our findings suggest that quality payment programs should adjust for comorbidities, and that reducing toxicity may be an effective method at reducing costs.