Seite - 321 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

Figure4. Segmenting theacoustic signal (20L/min)using the sequential detectionalgorithm. Allof theacquiredsignalswereprocessedandanalyzedbyMATLABR2017asoft- ware. The signals were sampled at 48 KHz sampling rate. The inhalation starts once the air flow is generated using theDUSA.The inspiratory phase ends once the airflow stops.Each inhalationphase is followedbyapausephase.The recorded soundof three consecutive inhalationadjusted toa testflowrateof20L/min is illustrated in3. Inthisstudy,weappliedabandpassFIRfilterwithlowercutofffrequency2KHzand higher cutoff frequency8KHz to remove theundesired frequencycomponents.Second, theautomatic sequential detectionalgorithmproposedbyHubert et al. [26]wasapplied to thefilteredsignal (seeFigure4). The adopted detection algorithm can be described as an unsupervised learning methodwith theonly assumptionbeing that a segment involves a change in the signal’s energy.Based entirely on aBayesian inference, this algorithmoptimizes the likelihood function of entropy tofind themaximumaPosteriori estimate (MAP).Let I denote the numberofdetected segments in theprocessed signal.Thedetectionalgorithmestimates the start andendofeachsegment, i.e. for i=1,..,I,wehave that • mˆi: theMAPestimateof the starting timeof the ith segment; • qˆi: theMAPestimateof theending timeof the ith segment. This sequential detection is a recursivealgorithmthat returns theconcatenatedvec- tor [mˆ1mˆ1···mˆImˆI]. In example shown in Figure 4, we obtained the following vec- tor of six cutting points bounding each inhalation phase: [mˆ1 qˆ1 mˆ2 qˆ2 mˆ3 qˆ3]= [77401 194702 277803 371504 453405 546401]. 2.3. Estimationof inhalationparameters When using the inhaler, the recorded signal x(n), n∈A with A=[1,··· ,N], can be expressedas follows : Z.Jeddi etal. /Estimationof InhalationFlowParameters forAsthmaMonitoring 321