Abstract
Introduction: Pulmonary function testing (PFT) is the main tool to evaluate the function of the respiratory system. However used alone, it hardly leads to disease diagnosis. Based on artificial intelligence (AI) we aimed to develop a smart software which improves the clinical reading of a lung function and suggests a respiratory disease diagnosis if possible.
Methods: Data of 1430 subjects with respiratory symptoms were taken from 33 Belgian hospitals to develop the algorithm. The final diagnosis (healthy, asthma, COPD, ILD, neuromuscular disease, chest wall or pleural disease, pulmonary vascular disease, other obstructive disease) was obtained from the clinical history, lung function and all additional tests, and confirmed by an expert panel. A cloud-based solution was incorporated into the clinical setting to validate the accuracy of the algorithm on a random sample of 86 new subjects. Finally, the software diagnoses were compared with the diagnostic opinions of 16 pulmonologists provided with PFT and clinical data of 50 new subjects.
Results: Software presented a high accuracy of 74% after 10-fold cross-validation when detecting lung diseases (8 possible disease categories). The high accuracy was maintained in a real clinical setting (76%). The software-based automated diagnoses (76% accuracy) were superior over the suggested diagnoses of pulmonologists (53.5 ± 6 % mean accuracy), which were not affected by the experience of the reader.
Conclusions: AI can be used to identify different lung diseases. Due to its superiority and work consistency, such software can provide a powerful decision support system in daily clinical routine.
- Copyright ©the authors 2017
