Robust Method to Predict Inspiratory Flowrate from the Acoustic Signature of Swirl-Based DPIs using Deep Learning
Lim K M1, Harris D S1, Sail P1 & Parry M2
1PA Consulting, Global Innovation and Technology Centre, Back Lane, Melbourn, Herts, SG8 6DP, UK
2Intertek Melbourn, Saxon Way, Melbourn, Herts, SG8 6DN, UK
We present a novel method to predict the inspiratory flowrate through swirl-based dry powder inhalers (DPIs), using deep learning techniques to analyse acoustic data. We collected extensive acoustic data from several DPIs using a Copley Scientific BRS3000 breath simulator producing three distinct flowrate profiles to represent a range of typical inspiratory manoeuvres: weak (~2 kPa), moderate (~4 kPa), and strong (~8 kPa). The data was then processed and fed into a bidirectional Recurrent Neural Network (RNN), with the output metrics being the predicted flowrate profile measured by the mean square error (MSE), a binary classification of whether the inhaler was used or not, and the time of detection deviation in milliseconds. A MSE of 5.34 LPM was achieved with 97.5% detection accuracy, and 3% deviation in duration. The preliminary trained RNN model shows that it is possible to predict the inspiratory flowrate solely from the sound emitted by the inhaler in representative use environments, even with high levels of background noise. It is also able to robustly detect usage and the total inspiratory duration. The use of deep learning techniques for the analysis of acoustic signals from inspiratory manoeuvres has great potential value, including: Characterisation of inhaler usage to provide valuable feedback to the patient to improve their inhalation technique, understanding the efficiency of drug delivery and usage patterns over time.