Development of a predictive model for chronic obstructive pulmonary disease onset with identification and ranking of risk factors

Abstract

Several risk factors for chronic obstructive pulmonary disease (COPD) onset are known (e.g. smoking), however only a few studies tried to combine them to predict COPD onset and none of them ranked risk factors according to their predictive ability. The aims of this work are to identify and rank risk factors for COPD onset and to develop a predictive model of COPD onset based on them.

Data of 7341 subjects aged 62.9±9.1 years, 46.2% males, free of COPD at baseline, were extracted from the English Longitudinal Study of Ageing dataset. Over a follow-up of up to 12 years, 305 subjects had a new diagnosis of COPD. A Cox model was trained on a training set (n=5841) for predicting COPD onset, considering 38 variables collected at baseline as predictors (demographics, medical history, lifestyle, spirometry parameters, blood biomarkers, socio-economic and psychologic factors). The variables were ranked by their predictive ability using a method based on recursive feature addition and bootstrap sampling. The top predictive ones were selected for the final model, that was validated on the test set (n=1500).

The set of top-ranked risk factors contains 12 variables related to lifestyle factors, spirometry parameters, blood biomarkers and comorbidities, that are, in order of importance: FEV1, wheezing, gender, smoking, FEV1/FVC, LDL cholesterol, plasma haemoglobin, diabetes, CESD depression score, education level, cigarettes smoked per day, physical activity. The predictive model of COPD onset based on such variables performed well on the test set, with concordance index equal to 0.83 (0.02).

Future work includes the model validation on different populations.