Introduction. Chronic obstructive pulmonary disease (COPD) is a the leading causes of death worldwide. Nowadays, conventional spirometry-based diagnosis entails many error-prone steps/stages involving patients or well-trained technicians. In addition, the early diagnosis of COPD is challenging.

Aims and objectives. To offer a rigorous alternative, we raised two hypotheses: (1) the physiological signals relevant to COPD have a multi- fractal nature and their fractional-order dynamics specifically characterize the COPD pathogenic mechanisms; (2) we can capture the fingerprints of the COPD-related physiological processes with the coupling matrix in our fractional dynamics (FD) modeling of the physiological dynamics [1]. 

Methods. We generated two novel COPD physiological signals datasets and implemented our FD model by extracting the FD signatures (coupling matrixes) from the relevant physiological signals recorded with an Internet of Medical Things infrastructure. The coupling matrixes capture the LRM of the interdependencies among the recorded signals and are used to train a deep neural network to diagnose COPD stages (see Fig. 1).

Results. We developed the fractional dynamic deep learning model, which uses FD signatures to predict the suspected patients? COPD stages. The model was trained by 18256 medical records from 588 patients from four Pulmonology Clinics in Western Romania. We show that our COPD diagnostics methods achieve a high prediction accuracy (98.66% ± 0.45%) and can serve as an robust alternative to traditional medical diagnosis. 

Conclusion. Fractional dynamics deep learning approach can serve as a robust alternative to traditional spirometry-based medical diagnosis.

1. C. Yin, Adv. Sci. 2023, 2203485