PREDICTION OF DYSPNOEA FOLLOWING LUNG RESECTION SURGERY: POST-HOC ANALYSIS OF ‘PROFILES’ STUDY

Introduction Lung cancer is the leading cause of cancer death in Europe. Surgical resection is often the preferred treatment but is associated with morbidity and mortality. Survival with a meaningful quality of life is important; however, the prediction of post-operative dyspnoea (POD) is often difficult and innaccurate.1 The European Society of Thoracic Surgeons (ESTS) and the (UK) National Institute of Clinical Excellence (NICE) advocate studies concerning operative risk for surgical resection. Conventional prediction uses pulmonary function; predicted post-operative FEV1%(ppoFEV1%) and predicted post-operative DLCO%(ppoDLCO%) with <40% in either domain being ‘high risk’. The aim is to improve conventional prediction of the risk of POD and identify a sub-population for targeted recruitment (prognostic enrichment) to interventional studies seeking to mitigate the risk of breathlessness Methods With informed consent and ethics approval, we prospectively recruited 250 patients undergoing lung resection in four UK centres. Dyspnoea was measured pre-operatively and 3 months post-operatively using the Medical Research Council (MRC) score. The primary outcome was patients with a post-operative MRC>2, excluding those with an MRC>2 pre-operatively. Two conventional models were derived (n=93, 1 site), before external validation (n=85, 3 sites) using the variables age, gender and ppoFEV1%/ppoDLCO%. Model 1(M1) incorporates ppoFEV1%/ ppoDLCO% with conventional cut offs and Model 2(M2) treats them continuously. Using similar internal derivation and external validation, two new models were explored. Univariate analysis identified risk predictors (p<0.1) for candidates with and without the primary outcome. Variables with significance were then used in logistic regression to create Model 3(M3) (M2 with the next-best additional variable- pre-operative EQ-5DL index score) & Model 4(M4) (not pre-defined and selected from all significant variables- ppoFEV1%, BMI, Diabetes status and pre-operative brief pain inventory score). Models were compared using sensitivity, specificity, positive predicted value (PPV), negative predictive value (NPV) and Net Reclassification Indexing (NRI) Results New models improved prediction within the internal dataset: M2 Vs M4 (AUROC comparison, p=0.03, NRI 0.26). (Fig.1) The best conventional and new models (M2 & M4) performed similarly within the external population: Sensitivity (55% vs 50%), Specificity (68% Vs 73%), PPV (38% Vs 39%), NPV (81% Vs 81%), respectively. Discussion This study demonstrates conventional risk prediction for POD using pulmonary function is poor. It also highlights challenges in creating new scoring tools: at external validation conventional models performed equally to new models with similar sensitivity/specificity/NPV and PPV. Using ppoFEV1%/ ppoDLCO% as continuous variables rather than dichotomised at 40%, may increase predictive strength. Future work should explore new variables to predict POD, such as pre-operative quality of life and biomarkers. For prognostic enrichment, models should have high sensitivity & high NPV, targeting those who would benefit most from low-risk interventions