SoundCatcher: Acoustic Emission with Machine Learning for Root Cause Analysis of Complaint Devices

Johanna Fredriksson1, Stina Sjödin1, Charlie Dempster1, Mats Josefson1, Ellinor Nilsson1,
Gabriel Bjerner2, Roland Greguletz3, and Lars Karlsson1

1Inhalation Pharmaceutical Development, Pharmaceutical Technology & Development, Operations, AstraZeneca, Gothenburg, Sweden

2Technology Operations Science and Innovation, Pharmaceutical Technology & Development, Operations, AstraZeneca, Södertälje, Sweden

3Inhalation Pharmaceutical Development, Pharmaceutical Technology & Development, Operations, AstraZeneca, Bad Homburg, Germany


Acoustic Emission (AE) profiles, in combination with machine learning (ML) algorithms, have been shown to provide valuable quality related information for inhalation devices and formulations. Here, the technique applicability is extended to patient complaint analysis. A feasibility study was carried out using a dry powder inhaler (DPI) platform, designed to be highly resistant to powder clogging, as a model device. Test inhalers were mechanically manipulated to simulate patient failure modes associated with clogging. AE profiles from the manipulated devices were collected under a non-destructive analysis by purging air through the devices with various flows without actuation while using an automated delivered dose analytical platform. The acquired AE data was analysed using different ML algorithms including orthogonal projections to latent structures (OPLS) and deep learning algorithms such as convolutional neural networks (CNN). Predictions were made on the position of particles glued on internal surfaces of the inhaler device as well as on the related amount of artificially applied powder residue. Using this approach, AE profiles from 123 inhalers showed an accuracy of above 85% when predicting residue level. At higher residue levels, the model could also distinguish between residue position with an accuracy of above 85%. Flow rate was also assessed in relation to model predictability. Based on these results, the models developed were applied on a set of 6 devices with unknown and challenging simulated residue levels, giving an accuracy of around 75%.  The overall conclusion is that AE combined with ML is a non-destructive analytical approach with the potential of providing valuable information when assessing the root cause for complaint sample returns. Long term, the hope is for the methodology to be evaluated on other device platforms to generalise the approach.

Key Message

Acoustic Emission combined with Artificial Intelligence (AI) algorithms is a very powerful, sensitive, and non-destructive technology capable of contributing to finding the root cause of device complaint sample returns, thus potentially adding to the quality of information in response letters to pharmacies or patients.