D4.2 Signal processing algorithms for extraction of asthma related indicators

This document describes the signal processing algorithms developed for the computation of the value of a set of physiological related indicators on the bases of measurements captured by the different types of sensors defined by WP2. A preliminary version of the deliverable has been submitted on M16.  In the preliminary version, a  set of relevant and informative indicators has been presented. These indicators have been identified from the state-of-the-art analysis and the experience of the participating partners. In addition, a set of supervised machine learning (ML) algorithms and knowledge based approaches are employed for the fusion of information extracted from the collected measurements. At this point it should be mentioned that the measurements of the quantification campaigns were not available at the time of this deliverable preparation and the presented indicators are expected to be slightly modified as soon as the quantification dataset will be available. A number of acoustic measurements were collected and used for monitoring drug usage and medication adherence. Therefore, the preliminary version of D4.2 presented the methodology for the extraction of informative indicators from acoustic measurements and the use of supervised ML approaches for the assessment of adherence of patients that use of pressurized Metered Dose Inhalers (pMDIs). This methodology has been applied to the available acoustic recordings and will be applied to additional data that will be available upon the completion of the Quantification Campaigns, end will be included in future version of this deliverable.
This deliverable initially provides a state-of-the-art analysis on the extraction of asthma related indicators, by presenting related works and the parameters recorded in the framework of the project. In section 2, the developed data processing algorithms for the extraction of asthma related indicators, are described. The application of these algorithms to audio signals is explained for identifying events that determine the pMDI usage adherence. In addition to the time-series analysis, the intelligent processing method used for the analysis of questionnaires is presented. Section 3 includes the classification used for evaluating critical parameters related to asthma, using state-of-the-art classifiers. In section 4 the monitoring process for the audio date set is described, along with processing algorithms, the cloud based server and the user interface. Section 5 describes the framework used for managing uncertainty of the system due to incompleteness and noise of wearable sensors, through the integration of the machine learning (ML) methods with compressed sensing and matrix completion approaches. The proposed solution enables the energy efficient monitoring of metered dose inhaler usage, by exploiting the specific characteristics of the reconstructed audio features at the receiver. Simulation studies, for indoor & outdoor data sets revealed high levels of accuracy (above 98%) by transmitting data streams in the range of 2% of the recorded audio series to the receiver, demonstrating the potential of this method for the development of novel energy efficient inhalers and medical devices in the area of respiratory medicine.