Predicting the clinical effect of a short acting bronchodilator in individual patients using artificial neural networks.

Artificial neural networks were used in this study to model the relationships between in vitro data, subject characteristics and in vivo outcomes from N=18 mild-moderate asthmatics receiving monodisperse salbutamol sulphate aerosols of 1.5, 3 and 6 μm mass median aerodynamic diameter in a cumulative dosing schedule of 10, 20, 40 and 100 μg. Input variables to the model were aerodynamic particle size (APS), body surface area (BSA), age, pre-treatment forced expiratory volume in one-second (FEV(1)), forced vital capacity, cumulative emitted drug dose and bronchodilator reversibility to a standard salbutamol sulphate 200 μg dose MDI (REV(%)). These factors were used by the model to predict the bronchodilator response at 10 (T10) and 20 (T20) min after receiving each of the 4 doses for each of the 3 different particle sizes. Predictability was assessed using data from selected patients in this study, which were set aside and not used in model generation. Models reliably predicted ΔFEV(1)(%) in individual subjects with non-linear determinants (R(2)) of ≥ 0.8. The average error between predicted and observed ΔFEV(1)(%) for individual subjects was <4% across the cumulative dosing regimen. Increases in APS and drug dose gave improved ΔFEV(1)(%). Models also showed trends towards improved responses in younger patients and those having greater REV(%), whilst BSA was also shown to influence clinical effect. These data show that APS can be used to discriminate predictably between aerosols giving different bronchodilator responses across a cumulative dosing schedule, whilst patient characteristics can be used to reliably estimate clinical response in individual subjects.