Classification methods for the identification of `case' in epidemiological diagnosis of asthma

Abstract

The identification of the asthmatic `case' in epidemiological research is a controversial issue. This study was aimed at classifying asthmatic subjects using a statistical decision rule that minimised the misclassification rate with respect to the clinicians' diagnosis. The rule was defined by a combination of predictors that are easily observed in epidemiological studies (asthma-like questions, physiological tests) without necessarily including the clinical opinion of expert physicians. From pooled data on 1103 subjects at the three Italian centres of the European Community Respiratory Health Survey (ECRHS) a post-consensus clinicians' diagnosis of asthma was obtained, and seven predictors were selected from among 18 potential candidates (specificity ranged from 64 to 99%, but sensitivity ranged from 22 to 62%). This data set was processed with tree-structured classifier techniques (the Classification And Regression Trees, CART), classical discriminant analysis (Fisher's Linear Discriminant Function, LDF), and the neural network method (Multi-Layer Perceptron, MLP model). The results suggest that modifications of the `classification tree' provide a more useful decision rule, sensitive (93%) and specific (85%), than either LDF or MLP. The decision tree is readily interpretable from a clinical perspective and uses five out of the seven predictors (in descending hierarchical order: ever had asthma, current asthma, shortness of breath, atopy and wheezing and breathless). The findings seem to indicate a considerable success with respect to previous epidemiological studies and await repetition in other ECHRS populations.