A Low-cost Approach for Detecting Activities of Daily Living using

Audio Information: A Use Case on Bathroom Activity Monitoring

Georgios Siantikos, Theodoros Giannakopoulos and Stasinos Konstantopoulos

Institute of Informatics and Telecommunications, NCSR Demokritos, Aghia Paraskevi, Athens, Greece

Keywords:

Audio Analysis, Activities of Daily Living, Health Monitoring, Remote Monitoring, Audio Sensors, Raspber-

ryPI, Audio Event Recognition.

Abstract:

In this paper, we present an architecture for recognizing events related to activities of daily living in the context

of a health monitoring environment. The proposed approach explores the integration of a Raspberry PI single-

board PC both as an audio acquisition and analysis unit. A set of real-time feature extraction and classiﬁcation

procedures has been implemented and integrated on the Raspberry PI device, in order to provide continuous

and online audio event recognition. In addition, a tuning and calibration workﬂow is presented, according to

which the technicians installing the device in a fast ans user-friendly manner, without any requirements for

machine learning expertise. The proposed approach has been evaluated against a particular scenario that is

rather important in the context of any healthcare monitoring system for the elder, namely the ”bathroom sce-

nario” according to which a single microphone installed on a Raspberry PI device is used to monitor bathroom

activity in a 24/7 basis. Experimental results indicate a satisfactory performance rate on the classiﬁcation pro-

cess (around 70% for ﬁve bathroom-related audio classes) even when less than two minutes of annotated data

are used for training in the installation procedure. This makes the whole procedure non demanding in terms

of time and effort needed to be calibrated by the technician.

1 INTRODUCTION

Although fully autonomous artiﬁcial intelligence is

actively researched and advanced, the current state

of the art (and at the level of maturity required for

commodity electronics) has machine learning meth-

ods rely on delicate training and conﬁguration ses-

sions in order to adapt to different environments.

When embedding machine learning methods in com-

modity electronics this is typically worked around by

uploading the signal and receiving analysis results

from remote centralized services. Recent examples

include voice-operated personal assistant applications

and companion, toy, and ‘pet’ applications.

This model, however, suffers from its obvious pri-

vacy implications. These implications are further ex-

acerbated in the telemedicine domain for two reasons:

the data collected by the remote service is not only

more sensitive, but the users might also not be able to

make informed decisions or might not be offered rea-

sonable alternatives. Home monitoring for the elderly

is a prime example: the increase in life expectancy

and in the need for long-term care creates a pressure

to seek alternatives to institutional healthcare for the

aged population. Advancements in robotics and au-

tomation and in artiﬁcial intelligence and intelligent

monitoring are explored as a way to prolong inde-

pendent living at home while providing guarantees

of safety and adequate medical monitoring ((Barger

et al., 2005), (Hagler et al., 2010), (MOTS et al.,

2002)). The users of such solutions, however, might

be suffering from mild cognitive impairment or be un-

able to afford conventional monitoring, which makes

ethically questionable any consent they provide to up-

load and analyse raw content of their activities of

daily living (ADL) in order to extract medical mon-

itoring information. Several methods have been used

to detect activities of daily living in real home en-

vironments, focusing on elderly population ((Vacher

et al., 2013), (Vacher et al., 2010), (Costa et al., 2009),

(Botia et al., 2012)) and a wide range of modalities.

In this paper, we present an audio analysis system

(Section 2) that explores the integration of the audio

sensor and the processing unit as Raspberry PI

de-

vice. Such a unit is able to execute signal processing

and machine learning algorithms in order to eliminate

Please cf. https://www.raspberrypi.org

Siantikos, G., Giannakopoulos, T. and Konstantopoulos, S.

A Low-cost Approach for Detecting Activities of Daily Living using Audio Information: A Use Case on Bathroom Activity Monitoring.

In Proceedings of the International Conference on Information and Communication Technologies for Ageing Well and e-Health (ICT4AWE 2016), pages 26-32

ISBN: 978-989-758-180-9

the need to provide raw content: the only informa-

tion that leaves the conﬁnes of the integrated unit is

an abstract ADL log. Although such information still

needs to be managed in full accordance to guidelines

pertaining private data, the level of obtrusiveness is

greatly reduced by the assurance that no unwarranted

analysis or recording can conceivably be done.

Our system is designed to satisfy two key require-

ments: that the analysis algorithms are computation-

ally efﬁcient so that they can be implemented for

the Raspberry PI device; and that they can be tuned

and conﬁgured for different acoustic environments by

technicians without machine learning expertise. In

order to evaluate the proposed approach on these re-

quirements, we motivate and present a use case based

on bathroom usage (Section 3) and draw conclusions

(Section 3).

2 PROPOSED METHOD

2.1 Overall Architecture

The main part of the whole system is a microphone-

equipped Raspberry PI single-board PC that is used

for all data acquisition and processing. Its small-form

factor, low energy consumption and low overall cost

make it ideal for installing it in any room/area we

want to monitor and its processing power is enough

for running our algorithms in real time. In our ex-

periments we used a Raspberry PI model B with a

Wolfson audio card.

Communication to/from the PC is made using

the MQTT machine-to-machine communication pro-

tocol. MQTT is a lightweight messaging protocol

that implements the brokered publish/subscribe pat-

tern, created widely used in IoT applications. With-

out going into technical details, the main idea is that

when connected to a speciﬁed MQTT ‘broker’, vari-

ous machines/applications can send messages under a

certain topic and others can listen to these when ”sub-

scribed” to these topics. In our use case, it is used

both for sending commands to the Raspberry PI (for

example to start/stop recording) and for remotely re-

ceiving the processing results.

For this purpose, two MQTT clients were imple-

mented: The ﬁrst is installed in the Raspberry PI and

is subscribed to a ”command” topic in order to receive

requests for collecting training data, building audio

classes models and ﬁnally use them for real-time clas-

siﬁcation. The second one is bundled in an Android

application and is used for sending remotely the cor-

responding commands and listening to the classiﬁca-

tion results. The system is designed with ease of use

in mind and the only set-up needed is connecting the

two clients to the same broker. By having a dedicated

broker this step can be performed automatically, mak-

ing the whole system ‘plug-and-play’.

2.2 System Calibration

Once setup, the system has to go through a training

phase in order to be used for real-life scenarios. This

includes recording, feature extraction, manual anno-

tation of the recorded events and classiﬁer tuning /

training. Figure 1 shows the proposed calibration

procedure. During this phase, the various events are

recorded using the Android application as a remote

controller of the Raspberry PI device that makes the

actual recording and further processing. An audio ﬁle

is created on user’s demand and the user/technician

is informed about the categories and durations of al-

ready recorded data. He then provides the current

recording’s label (e.g. ”door bell”). When a reason-

ably large amount of data is gathered (typically about

1-2 minutes of recordings for each category), the tech-

nician uses the mobile application to trigger the train-

ing process (that is also executed on the Raspberry PI

device). After this process, the Raspberry PI is ready

to monitor and recognize sound in the ”learned” envi-

ronment.

2.3 Audio Event Recognition

2.3.1 Audio Features

In total, 34 audio features are extracted on a short-

term basis. This process results in a sequence of

34-dimensional short-term feature vectors. In addi-

tion, the processing of the feature sequence on a mid-

term basis is adopted. According to that the audio

signal is ﬁrst divided into mid-term windows (seg-

ments). For each segment, the short-term process-

ing stage is carried out and the feature sequence from

each mid-term segment, is used for computing feature

statistics (e.g. the average value of the ZCR). There-

fore, each mid-term segment is represented by a set of

statistics. In this Section we provide a brief descrip-

tion of the adopted audio features. For detailed de-

scription the reader can refer to the related bibliogra-

phy (Giannakopoulos and Pikrakis, 2014) (Theodor-

idis and Koutroumbas, 2008), (Hyoung-Gook et al.,

2005). The time-domain features (features 1 - 3) are

directly extracted from the raw signal samples, while

the frequency-domain features (features 4-34, apart

from the MFCCs) are based on the magnitude of the

Discrete Fourier Transform (DFT). The cepstral do-

main (e.g. used by the MFCCs) results after applying

A Low-cost Approach for Detecting Activities of Daily Living using Audio Information: A Use Case on Bathroom Activity Monitoring

/cmd/recording/start

{event: 'shower'}

Android application

Raspberry Pi

broker

shower/recording001.wav

/cmd/recording/stop

/response/recordings

{events: [

{ name: 'shower',

time: '30s'},

{ name: 'activity',

time: '90s'},

{ name: 'flush',

time: '20s'}, ...

]}

/cmd/classifier/create

{ events: [

'shower',

'flush',

'silence',

'tap' ],

name: 'test_02'

}

shower/recording001.wav

shower/recording002.wav

....

flush/recording001.wav

flush/recording002.wav

flush/recording003.wav

pyAudioAnalysis

SVM classifier

/response/classifiers

[ { name: 'test_01',

events: ['shower',

'silence',

'activity'] },

{ name: 'test_02',

events: ['shower', 'tap',

'silence',

'flush'] } ]

Figure 1: From left to right: User initiates an event recording and the corresponding ﬁle is created. When user stops, a

response is returned with information about the events recorded so far. When a reasonable amount of data is gathered, an

SVM classiﬁer for the desired events can be created using the pyAudioAnalysis library. In this case, the response contains

information about the classiﬁers available for future use.

the Inverse DFT on the logarithmic spectrum. The

complete list of features is presented in Table 1.

2.3.2 Classiﬁcation

As described in Section 2.3.1, the feature extraction

process leads to a 68-dimensional feature vector for

each 1-second audio segment, i.e. 2 statistics x 34

short-term features. Each unknown audio segment of

ﬁxed size (1 second in our case) is therefore repre-

sented by a 68-D feature vector. Each of these sam-

ples is classiﬁed using a Support Vector Machine with

probabilistic output. We have selected to use proba-

bilistic SVMs (Platt, 1999) due to their ability to gen-

eralize well especially in high dimensional classiﬁ-

cation problems (Chapelle et al., 1999). The model

is trained using a cross-validation procedure to select

the optimal SVM parameter, namely the soft margin

parameter C.

2.3.3 Audio Analysis Implementation

Audio feature extraction and classiﬁcation has been

implemented using the pyAudioAnalysis library (Gi-

annakopoulos, 15 ). This is an open-source Python li-

brary that implements a wide range of audio analysis

functionalities and can be used in several applications.

Using pyAudioAnalysis one can classify an unknown

audio segment to a set of predeﬁned classes, segment

an audio recording and classify homogeneous seg-

ments, remove silence areas from a speech record-

ing, estimate the emotion of a speech segment, ex-

tract audio thumbnails from a music track, etc. In this

work, pyAudioAnalysis has been used to extract au-

dio features, to train the classiﬁcation models and to

perform cross validation experimentation in order to

extract the respective performance measures. pyAu-

dioAnalysis achieves 2× realtime performance on the

Raspberry devices, which validates its usage in the

context of the particular setup.

3 BATHROOM USE CASE AND

EVALUATION

3.1 Use Case and Motivation

As discussed in the introduction, the motivating use

case for our approach is medical monitoring. Specif-

ically, we base our evaluation setup on allowing el-

derly people with mild cognitive impairment to main-

tain an independent life, at their own home, for longer

than what is safely possible today.

In order to have a guideline about what informa-

tion is used by medical doctors to assess such con-

ditions, we use the interRAI Long-Term Care Facil-

ities Assessment System (interRAI LTCF). interRAI

LTCF enables comprehensive, standardized evalua-

tion of the needs, strengths, and preferences of per-

sons receiving care. interRAI has been analysed pre-

viously in order to identify assessment items, such as

mood and ADL logs, that can be automatically rec-

ICT4AWE 2016 - 2nd International Conference on Information and Communication Technologies for Ageing Well and e-Health

Table 1: Adopted short-term audio features.

Index Name Description

1 Zero Crossing Rate Rate of sign-changes of the frame

2 Energy Sum of squares of the signal values,

normalized by frame length

3 Entropy of Energy Entropy of sub-frames’ normalized en-

ergies. A measure of abrupt changes

4 Spectral Centroid Spectrum’s center of gravity

5 Spectral Spread Spectrum’s second central moment of

the spectrum

6 Spectral Entropy Entropy of the normalized spectral en-

ergies for a set of sub-frames

7 Spectral Flux Squared difference between the nor-

malized magnitudes of the spectra of

the two successive frames

8 Spectral Rolloff The frequency below which 90% of the

magnitude distribution of the spectrum

is concentrated.

9-21 MFCCs Mel Frequency Cepstral Coefﬁcients: a

cepstral representation with mel-scaled

frequency bands

22-33 Chroma Vector A 12-element representation of the

spectral energy in 12 equal-tempered

pitch classes of western-type music

34 Chroma Deviation Standard deviation of the 12 chroma co-

efﬁcients.

ognized and are useful to medical personnel (RADIO

Project, 2015). Among the assessment items listed,

we identiﬁed those regarding bathroom use as being

closest to our concept: these items can be extracted

by processing very sensitive content, and being able

to provide guarantees about the management and pro-

cessing of this content would have signiﬁcant impact

on the acceptance of any relevant solution by users.

In this context, we have recorded and manually

annotated sounds from bathroom usage. In particular,

the following audio classes are trained and evaluated

by the proposed methodology:

• Silence - no sound

• Flushing water

• Shower

• Tap water

• Other activities

Note that the selected audio events are location-

speciﬁc and therefore the adopted calibration work-

ﬂow can be used in order during the installation phase,

as described in Section 2.2.

3.2 Dataset

In order to train and evaluate the proposed event

recognition methodology, we have recorded and

manually annotated (using the mobile app de-

scribed earlier in the paper) 4 recordings. The

total duration of the dataset is 7 minutes. The

extracted feature vector sequences and the respective

ground truth of each recording is openly available at

https://iit.demokritos.gr/ tyianak/bathroomScenario

Events.zip

3.3 Experimental Evaluation

3.3.1 Performance Measures

Let CM be the confusion matrix, i.e. a N

×N

matrix

is the total number of audio classes), whose rows

and columns refer to the true (ground truth) and pre-

dicted class labels of the dataset, respectively. In other

words, each element, CM(i, j), stands for the number

of samples of class i that were assigned to class j by

the adopted classiﬁcation method. The diagonal of

the confusion matrix captures the correct classiﬁca-

tion decisions (i = j). CM is normalized row-wise, in

order to discard the information that is related to the

A Low-cost Approach for Detecting Activities of Daily Living using Audio Information: A Use Case on Bathroom Activity Monitoring

size of each class:

(i, j) =

CM(i, j)

∑

n=1

CM(i, n)

(1)

Obviously, after the normalization process, the el-

ements of each row sum to unity.

Three useful performance measures are then ex-

tracted from the confusion matrix. The ﬁrst is the

overall accuracy, Acc, of the classiﬁer, which is de-

ﬁned as the fraction of samples of the dataset that have

been correctly classiﬁed:

Acc =

∑

m=1

CM(m, m)

∑

m=1

∑

n=1

CM(m, n)

(2)

Apart from the overall accuracy, we have adopted

two class-speciﬁc measures that describe how well

the classiﬁcation algorithm performs on each class.

The ﬁrst of these measures is the class recall, Re(i),

which is deﬁned as the proportion of data with true

class label i that were correctly assigned to class i:

Re(i) =

CM(i, i)

∑

m=1

CM(i, m)

(3)

where

∑

m=1

CM(i, m) is the total number of sam-

ples that are known to belong to class i. In addition,

we use the class precision (Pr(i)), i.e. the fraction

of samples that were correctly classiﬁed to class i if

we take into account the total number of samples that

were classiﬁed to that class:

Pr(i) =

CM(i, i)

∑

m=1

CM(m, i)

(4)

Finally, the F

-measure is also computed, which is

the harmonic mean of the precision and recall values:

(i) =

2Re(i)Pr(i)

Pr(i) + Re(i)

(5)

3.3.2 Results

Table 2 shows the (row-wise normalized) confusion

matrix and the respective precision, recall and F1

measures for each audio class. This is the result of the

evaluation process when only one recording is used

during the training phase.

These results correspond to the most realistic and

less demanding (in terms of calibration-training time).

In addition, Table 3 demonstrates the ability of the

classiﬁers to adopt to more data, if they can be avail-

able. In particular, 3 shows the same performance

measures if a ”leave one out” process is used in the

evaluation process, using the described dataset. That

is, if three whole recordings are used for each training

phase. Results indicate an almost 5% performance

boosting. However, using three recordings instead of

one means a 300% increase in the calibration time to

be carried out by the technicians.

4 CONCLUSIONS

We have presented an architectural scheme that em-

ploys a Raspberry PI device both as an audio acquisi-

tion and analysis unit, in the cotnext of a health mon-

itoring system that detects ADLs in the living envi-

ronments of elderly people. Real-time audio feature

extraction and classiﬁcation methods have been im-

plemented and integrated on the device. Also, we pro-

pose a procedure for a fast and easy-to-use calibration

procedure that actually trains the implemented classi-

ﬁers in the particular home’s sound conditions. Ex-

perimental evaluation has demonstrated a 70% classi-

ﬁcation performance even if a single recording (1 to

1.5 minute long) is used in the training process. The

complete software that was used for the experiments

can be found in the project’s Git repository

under an

open-source license.

The proposed system architecture satisﬁes three

vital requirements.

• First the implemented algorithms are computa-

tionally efﬁcient so that they can be implemented

for the Raspberry PI device, as they demonstrate

a 2× realtime performance on the Raspberry de-

vices. This validates the system’s usage in the

context of a low-cost health monitoring setup as it

does not require a workstation or a PC (e.g. (Chen

et al., 2005)), but a single Raspberry PI that serves

both as an acquisition and an analysis module. In

particular, the total cost of both the acquisition

and analysis modules is less than 100$.

• In addition, despite the low-cost characteristics

of the proposed approach, the system achieves a

satisfactory classiﬁcation performance, given (a)

the low cost and (b) the lack of demand for big

training data. Compared to other similar meth-

ods for ADL recognition in the context of a smart

home environment, our method does not outper-

form (in terms of overall classiﬁcation accuracy),

however, given the signiﬁcant differences in terms

of overall cost, the proposed approach is prefer-

able for real house applications. For instance, the

approach in (Vuegen et al., 2013) achieves a 85%

classiﬁcation accuracy in a ADL recognition task,

however the acquisition scenario requires multiple

microphone sensors and therefore much higher

cost.

https://bitbucket.org/radioprojectanalysis/ict4awe2016

ICT4AWE 2016 - 2nd International Conference on Information and Communication Technologies for Ageing Well and e-Health

Table 2: Single-recording training: Row-wise normalized confusion matrix, recall precision and F1 measures. Overall F1

measure: 68.1%.

Confusion Matrix (%)

Predicted

True ⇓ Shower Flush Tap Silence Activity

Shower 89.4 1.8 3.0 0.1 5.8

Flush 7.2 70.7 0.4 2.1 19.6

Tap 5.7 4.0 85.8 0.8 3.6

Silence 1.4 4.0 0.0 58.0 36.5

Activity 13.0 11.6 2.6 31.2 41.6

Performance Measurements (%, per class)

Recall: 89.4 70.7 85.8 58.0 41.6

Precision: 76.6 76.8 93.5 62.9 38.8

F1: 82.5 73.6 89.5 60.3 40.2

Table 3: Three-recording training: Row-wise normalized confusion matrix, recall precision and F1 measures. Overall F1

measure: 73.8%.

Confusion Matrix (%)

Predicted

True ⇓ Shower Flush Tap Silence Activity

Shower 85.6 1.6 2.6 0.2 10.0

Flush 5.7 86.5 0.0 1.5 6.3

Tap 5.2 3.7 85.5 0.7 4.9

Silence 0.1 3.4 0.0 72.5 24.0

Activity 5.8 9.7 1.5 29.0 53.9

Performance Measurements (%, per class)

Recall 85.6 86.5 85.5 72.5 53.9

Precision 83.6 82.5 95.4 69.8 54.4

F1: 84.6 84.5 90.2 71.1 54.2

• Finally, the whole system can be conﬁgured and

calibrated for different acoustic environments by

technicians without machine learning expertise.

Our ongoing and future research work focuses on

the following directions:

• Extend the calibration procedure so that it also

takes into account a ”base dataset”, i.e. an initial

classiﬁcation scheme that is tuned in the context

of the annotation process and not re-trained from

scratch.

• Use long-term temporal knowledge to smooth the

results of the classiﬁer, based on prior knowledge

regarding the events.

• The most important ongoing research direction

focuses on adopting more robust audio classiﬁ-

cation approaches. We are currently working to-

wards implementing deep learning methods for

audio event recognition, which have been proved

to achieve a signiﬁcant performance boosting

(more than 10%) (Lane et al., 2015), (Wang et al.,

2015).

ACKNOWLEDGEMENTS

This project has received funding from the European

Unions Horizon 2020 research and innovation pro-

gramme under grant agreement No 643892. Please

see http://www.radio-project.eu for more details.

REFERENCES

Barger, T. S., Brown, D. E., and Alwan, M. (2005). Health-

status monitoring through analysis of behavioral pat-

terns. Systems, Man and Cybernetics, Part A: Systems

and Humans, IEEE Transactions on, 35(1):22–27.

Botia, J. A., Villa, A., and Palma, J. (2012). Ambient as-

sisted living system for in-home monitoring of healthy

independent elders. Expert Systems with Applications,

39(9):8136–8148.

A Low-cost Approach for Detecting Activities of Daily Living using Audio Information: A Use Case on Bathroom Activity Monitoring

Chapelle, O., Haffner, P., and Vapnik, V. N. (1999). Sup-

port vector machines for histogram-based image clas-

siﬁcation. Neural Networks, IEEE Transactions on,

10(5):1055–1064.

Chen, J., Kam, A. H., Zhang, J., Liu, N., and Shue, L.

(2005). Bathroom activity monitoring based on sound.

In Pervasive Computing, pages 47–61. Springer.

Costa, R., Carneiro, D., Novais, P., Lima, L., Machado,

J., Marques, A., and Neves, J. (2009). Ambient as-

sisted living. In 3rd Symposium of Ubiquitous Com-

puting and Ambient Intelligence 2008, pages 86–94.

Springer.

Giannakopoulos, T. (2015–). pyAudioAnalysis: Python

audio analysis library: Feature extraction, classiﬁ-

cation, segmentation and applications. [Online; ac-

cessed 2015-04-27].

Giannakopoulos, T. and Pikrakis, A. (2014). Introduction to

Audio Analysis: A MATLAB

 Approach. Academic

Press.

Hagler, S., Austin, D., Hayes, T. L., Kaye, J., and Pavel,

M. (2010). Unobtrusive and ubiquitous in-home mon-

itoring: a methodology for continuous assessment of

gait velocity in elders. Biomedical Engineering, IEEE

Transactions on, 57(4):813–820.

Hyoung-Gook, K., Nicolas, M., and Sikora, T. (2005).

MPEG-7 Audio and Beyond: Audio Content Indexing

and Retrieval. John Wiley & Sons.

Lane, N. D., Georgiev, P., and Qendro, L. (2015). Deep-

ear: robust smartphone audio sensing in unconstrained

acoustic environments using deep learning. In Pro-

ceedings of the 2015 ACM International Joint Confer-

ence on Pervasive and Ubiquitous Computing, pages

283–294. ACM.

MOTS, T. M., Linda Fraas OTR, M., and Kathleen Stan-

ton MS, R. (2002). Elder acceptance of health moni-

toring devices in the home. Care Management Jour-

nals, 3(2):91.

Platt, J. C. (1999). Probabilistic outputs for support vec-

tor machines and comparisons to regularized likeli-

hood methods. In Advances in large margin classi-

ﬁers. Citeseer.

RADIO Project (2015). D2.2: Early detection meth-

ods and relevant system requirements. Available at

http://radio-project.eu/deliverables.

Theodoridis, S. and Koutroumbas, K. (2008). Pattern

Recognition, Fourth Edition. Academic Press, Inc.

Vacher, M., Portet, F., Fleury, A., and Noury, N. (2010).

Challenges in the processing of audio channels for

ambient assisted living. In e-Health Networking Ap-

plications and Services (Healthcom), 2010 12th IEEE

International Conference on, pages 330–337. IEEE.

Vacher, M., Portet, F., Fleury, A., and Noury, N. (2013).

Development of audio sensing technology for ambient

assisted living: Applications and challenges. Digital

Advances in Medicine, E-Health, and Communication

Technologies, page 148.

Vuegen, L., Van Den Broeck, B., Karsmakers, P., Vanrum-

ste, B., et al. (2013). Automatic monitoring of ac-

tivities of daily living based on real-life acoustic sen-

sor data: a preliminary study. In Fourth workshop

on speech and language processing for assistive tech-

nologies (SLPAT): Proceedings, pages 113–118. As-

sociation for Computational Linguistics (ACL).

Wang, H.-H., Liu, J.-M., You, M., and Li, G.-Z. (2015).

Audio signals encoding for cough classiﬁcation us-

ing convolutional neural networks: A comparative

study. In Bioinformatics and Biomedicine (BIBM),

2015 IEEE International Conference on, pages 442–

445. IEEE.

ICT4AWE 2016 - 2nd International Conference on Information and Communication Technologies for Ageing Well and e-Health