Objective Assessment of Patient Inhaler User Technique Using an Audio-Based Classification Approach

Many patients make critical user technique errors when using pressurised metered dose inhalers (pMDIs) which reduce the clinical efficacy of respiratory medication. Such critical errors include poor actuation coordination (poor timing of medication release during inhalation) and inhaling too fast (peak inspiratory flow rate over 90 L/min). Here, we present a novel audio-based method that objectively assesses patient pMDI user technique. The Inhaler Compliance Assessment device was employed to record inhaler audio signals from 62 respiratory patients as they used a pMDI with an In-Check Flo-Tone device attached to the inhaler mouthpiece. Using a quadratic discriminant analysis approach, the audio-based method generated a total frame-by-frame accuracy of 88.2% in classifying sound events (actuation, inhalation and exhalation). The audio-based method estimated the peak inspiratory flow rate and volume of inhalations with an accuracy of 88.2% and 83.94% respectively. It was detected that 89% of patients made at least one critical user technique error even after tuition from an expert clinical reviewer. This method provides a more clinically accurate assessment of patient inhaler user technique than standard checklist methods.

Does the patient exhale sufficiently (breathe out as far as comfortable) prior to inhalation?
Does the patient place their teeth around the Flo-Tone, without biting the mouthpiece, and close their lips around it correctly?
Does the patient inhale through their mouth?
Does the patient inhale at an appropriate flow rate suitable for metered dose inhalers? (Is the whistle too loud or not heard at all?) When hearing the Flo-Tone, does the patient press down on the canister to release the aerosol?
Does the patient continue to inhale steadily, slowly and deeply after they actuate the inhaler? (Does the Flo-Tone still generate a sound after the canister is pressed?) After a full inhalation, does the patient hold their breath? (up to 10 seconds) Does the patient exhale after holding their breath? Figure S1. Inhaler user technique checklist used for patient pMDI recordings. Figure S2. Graphical user interface for labelling inhaler audio frames. The graphical user interface (GUI) allows each inhaler audio signal to be manually labelled for the training process and for the evaluation of the inhaler sound event classification algorithm in the testing phase. Once an audio signal is loaded into the GUI, the audio time domain signal as well as the spectrogram of the audio signal are displayed. From here, each sound event can be manually labelled. Additionally, each labelled sound event can be played to assist the labelling process. The labelled signal (pink) is also shown. Once the labelling process is complete, the user can save a segmentation text file which contains the labelling information of sound events for the corresponding audio signal.

Supplementary Material D -Inhaler sound event classification
Comparison between quadratic discriminant analysis with artificial neural network: A feed forward artificial neural network (ANN) with three hidden layers was compared to the quadratic discriminant analysis (QDA) method described in this study. The ANN performance measures generated using 11 selected features from the QDA feature selection process are presented in comparison to the QDA results in Table S2 and Table S3. In order to assess whether the selected features were biased only to the QDA, a sequential forward feature selection process was also performed with the ANN. It was observed that the ANN feature selection selected eight optimum features which included: E, Z, f0, c9, c22, c30, a1, and a3. The performance measure results generated from the training and testing datasets using the eight ANN selected features with QDA and ANN classification methods are presented in Table S4 and Table S5.   Figure S3. pMDI audio-based flow estimation experimental setup.

Supplementary Material E -Audio-based inhaler inhalation flow estimation
Blue arrows indicate the airflow during the inhalation recordings. Airflow passes through the spirometer to measure peak inspiratory flow rate (PIFR) and volume during inhalation. Airflow also passes through the reed on the Flo-Tone which was accounted for in the experiment described in the supplementary material.

Method:
In order to ensure that the flow measurements obtained from the spirometer represented the true flow measurements of the participant's inhalation (and not influenced by the additional airflow at the Flo-Tone reed aperture), an in-vitro experiment was performed relating the PIFR recorded from the spirometer to the PIFR recorded from the mouthpiece of the Flo-Tone. This was performed by connecting the Flo-Tone mouthpiece to a high capacity vacuum pump [HCP4, Copley Scientific] and Critical Flow Controller (air valve) (TPK 2000, Copley Scientific). The flow pump simulated inhalations through the device at 10 flow rates between 20-100 L/min, which covers the clinically relevant inspiratory flow rates for pMDIs (the flow pump setup was calibrated up to 100 L/min also so it was not recommended to exceed 100 L/min). The setup for this in-vitro experiment is shown in Figure S4. The PIFR recorded from the spirometer was compared to that of the reading on the vacuum pump. Figure S4. Inhaler flow measurement setup comparing difference in measured PIFR between the spirometer and inhaler mouthpiece.

Result:
It was observed that the difference between the PIFR measurements obtained from the spirometer and the vacuum pump was almost negligible (4.65±2.54% error (mean ± SD)). There was a highly statistically significant linear relationship between the PIFR measured at the spirometer and at the mouthpiece (vacuum pump) (R 2 =0.99, p<0.0001). This linear relationship is presented in Figure S5. Therefore, this indicated that the flow measurements obtained from the pneumotachograph spirometer were reliable measurements of flow through the Flo-Tone mouthpiece.

Supplementary Material F -Audio-based assessment of patient inhaler user technique
Wilcoxon signed rank test analysis comparing inhaler PIFR and volume before and after tuition: If a patient did not inhale before tuition or did not inhale after tuition they were discarded from this statistical analysis as at least one inhalation was needed to compare before and after tuition. This was the case for 16 patients where they did not inhale either before or after tuition. A further four patients were removed from this analysis across labelled and detected audio data due to the presence of outliers in the PIFR and IC values according to Grubbs' outlier test (p<0.05).