Further validation of computer-based prediction of chemical asthma hazard

Aims To iterate the external validation of a previously published quantitative structure-activity relationship (QSAR) model for the prediction of novel chemical respiratory sensitizers and to better characterize its predictive accuracy. Methods An external validation set of control chemicals was identified from the Australian Hazardous Substances Information System. An external validation set of asthmagenic chemicals was identified by a thorough search of the peer-reviewed literature from January 1995 onwards using the Medline database. The QSAR model was used to determine an 'asthma hazard index' (between 0 and 1) for each chemical. Results A total of 28 external validation asthmagens and 129 control chemicals were identified. The area under the receiver operating characteristic (ROC) curve for the model's ability to distinguish asthmagens from controls was 0.87 (95% CI 0.76-0.97). Using a cut-off hazard index of 0.5 resulted in sensitivity of 79% and specificity of 93%. For prior probability ranging from 1:300 to 1:100, the negative predictive value (NPV) was 1 and positive predictive value (PPV) 0.04-0.1 while for prior probability ranging from 1:20 to 1:3, the NPV was 0.91-0.99 and PPV 0.39-0.85. Conclusions The ROC curve for this QSAR demonstrates good global predictive power for distinguishing asthmagenic from non-asthmagenic LMW organic compounds. Potential for utilization by occupational and respiratory physicians is evident from its predictive values.