Abstract 3634: A new model improves lung cancer prediction in the preoperative evaluation of patients with suspicious lung lesions.

Abstract

Abstract Lung cancer predictive models, such as the Mayo model (Swensen, 1997), that are used for screening and referral to surgery are based on only traditional variables and are outdated. Clinical information such as FEV1, FDG-PET and lesion growth increasingly are used in treatment decisions. This information needs to be incorporated into our clinical prediction models. We developed a lung cancer risk model utilizing the more extensive clinical information available at the point of a decision for surgery that predicts cancer better than existing models. We evaluated a lung cancer prediction model using multivariable logistic regression in a population being evaluated for lung surgery at a single academic institution in an IRB approved study. The model included non-linear relationships between continuous variables and lung cancer and used multiple imputation to handle missing data. Internal validation was conducted using the bootstrap method with 500 iterations. Area under the curve (AUC) and Brier score were calculated for the training model and internal bootstrap validation model. The characteristics of the new model were compared to the Mayo model. 492 individuals were recruited at who had been evaluated for surgery with known or suspected lung cancer. Lung cancer prevalence was 72%. Diagnosis was determined pathologically (92%) or by greater than 18 months of followup among those who didn't undergo surgery. Missing data occurred with FDG-PET scan (22%), growth on serial CT scans (13%), predicted FEV1 (10%) and pre-operative disease symptoms (7%). The remaining variables of interest had less than 5% missing data. Age (OR 1.05; 95%CI: 1.03-1.08), pack years (OR 1.03; 95%CI: 1.00-1.05), pre-operative lesion maximum diameter (OR 1.06; 95%CI: 1.04-1.08), lesion growth (OR 2.92 95%CI: 1.10-5.65), previous cancer (OR 1.86 95%CI: 1.05-3.32) and FDG-PET avidity (OR 4.56 95%CI: 2.17-9.57) predicted lung cancer (p<0.05). AUC for the initial model was 0.87 (95%CI: 0.84 - 0.91) and Brier score was 0.12. Bootstrap sampling estimated AUC of 0.85 and Brier score of 0.13 demonstrating internal validity of the model. The Mayo model was evaluated for those with complete data (93%) and had an AUC of 0.80 (95%CI: 0.75 - 0.85) which was significantly less (P<0.001) than that observed for the new model. The Mayo model generally underestimated risk and its Brier score was 0.17, showing poorer calibration than the new model. Our internally validated model performed better in distinguishing low risk from high risk patients than the Mayo model in a surgical population being evaluated for lung cancer. The Mayo model, with its more limited clinical information, underestimated risk. Future work will validate this model in an external dataset and prospectively evaluate the impact of the model on patient safety and resource utilization. Citation Format: Stephen A. Deppen, Melinda C. Aldrich, Ronald Walker, Catherine A. Necessary, Chiu-lan Chen, Huiyun Wu, Jeffery Blume, Pierre P. Massion, Theodore Speroff, Robert S. Dittus, Bill J. Putnam, Eric L. Grogan. A new model improves lung cancer prediction in the preoperative evaluation of patients with suspicious lung lesions. [abstract]. In: Proceedings of the 104th Annual Meeting of the American Association for Cancer Research; 2013 Apr 6-10; Washington, DC. Philadelphia (PA): AACR; Cancer Res 2013;73(8 Suppl):Abstract nr 3634. doi:10.1158/1538-7445.AM2013-3634