Instrumentation & Measurement Magazine 25-1 - 86

Table 3 - Performance of estimation of alcohol content for K-fold CV strategy, depending on the window size
CV Metric
Window size (s)
1
MSE
K-fold
(g/dl)
Corr.
0.001559 ±
0.000011
2
0.001625 ±
0.000016
3
0.001956 ±
0.000033
4
0.002087 ±
0.000039
5
0.002357 ±
0.000057
6
0.002694 ±
0.000083
0.66 p<0.0001 0.64 p<0.0001 0.58 p<0.0001 0.54 p<0.0001 0.51 p<0.0001 0.49 p<0.0001
approach was able to compute a robust motion biomarker
for alcohol consumption, obtaining the best results using this
dataset.
Experiments for Estimation of Alcohol Content
In this subsection, we show the results for estimation of alcohol
content. Table 3 summarizes the MSE and correlation obtained
depending on the window size.
Results suggest that shorter windows provide better performance
for estimation of alcohol content. MSE of alcohol
content estimation is around 0.002 g/dl, and that corresponds
to less than 3% of the threshold (TAC of 0.08 g/dl). The best
correlation is 0.66, demonstrating that estimation evolves in
a similar way to the ground truth. This means that the deep
architecture learns meaningful features from accelerometers
integrated in different smartphones models for estimation of
alcohol content using low frequency resolution in windows.
Conclusions and Future Work
Excessive alcohol consumption is an avoidable health risk
that is highly frequent among young people. This paper
contributes to the supervision of alcohol consumption by defining
a robust motion biomarker that uses accelerometer data
collected from different smartphones. This biomarker is computed
by a deep learning architecture based on CNNs, which
includes three convolutional layers for learning features and
additional fully connected layers to perform classification and
estimation of alcohol content tasks. This paper uses the public
dataset called Bar Crawl: Detecting Heavy Drinking Data Set,
where real drinking activities were performed by 13 young
participants. The proposed approach reported the best results
over this dataset, thanks to its robustness against the variety
of accelerometer characteristics presented in different smartphones.
This deep learning architecture reached an accuracy of
80.43 ± 0.21 % using 2-second windows for classification task
and a MSE of alcohol content estimation of 0.001559 ± 0.000011
g/dl (lower than a 3% of the TAC threshold of 0.08 g/dl) and a
correlation of 0.66 using 1-second windows.
For future work, it would be interesting to test other alcohol
consumption datasets, evaluate the datasets over Leave-OneSubject-Out
CV strategy and perform post-processing
techniques for combining information from consecutive
windows.
Acknowledgment
The work leading to these results is part of the projects AMIC
(TIN2017-85854-C4-4-R) and CAVIAR (TEC2017-84593-C2-1-R)
86
funded by MINECO/AEI/10.13039/501100011033 and by
" ERDF A way of making Europe. "
Authors also thank Mark Hallett for the English revision
of this paper and all the other members of the Speech Technology
Group for the continuous and fruitful discussion on these
topics. We gratefully acknowledge the support of NVIDIA
Corporation with the donation of the Titan X Pascal GPU used
for this research.
References
[1] " Global status report on alcohol and health, 2018, " World Health
Organization, 2018.
[2] A. M. Arria, K. M. Caldeira, H. K. Allen, K. B. Vincent, B. A.
Bugbee, and K. E. O'Grady, " Drinking like an adult? trajectories
of alcohol use patterns before and after college graduation, "
Alcoholism-Clinical and Experimental Res., vol. 40, no. 3, pp.
583-590, Mar. 2016.
[3] B. Suffoletto et al., " An interactive text message intervention to
reduce binge drinking in young adults: a randomized controlled
trial with 9-month outcomes, " Plos One, vol. 10, no. 11, Nov.
2015.
[4] C. Wright et al., " Mobile phone-based ecological momentary
intervention to reduce young adults' alcohol use in the event: a
three-armed randomized controlled trial, " JMIR Mhealth Uhealth,
vol. 6, no. 7, Jul. 2018.
[5] I. Nahum-Shani et al., " Just-in-Time Adaptive Interventions
(JITAIs) in mobile health: key components and design principles
for ongoing health behavior support, " Annals of Behavioral Med.,
vol. 52, no. 6, pp. 446-462, Jun. 2018.
[6] M. Gil-Martin, R. San-Segundo, F. Fernandez-Martinez, and J.
Ferreiros-Lopez, " Improving physical activity recognition using a
new deep learning architecture and post-processing techniques, "
Eng. Applications of Artificial Intelligence, vol. 92, Jun. 2020.
[7] M. Gil-Martín, R. San-Segundo, F. Fernández-Martínez, and R.
de Córdoba, " Human activity recognition adapted to the type
of movement, " Computers Electrical Eng., vol. 88, p. 106822,
2020.
[8] R. San-Segundo, J. David Echeverry-Correa, C. Salamea, and J.
Manuel Pardo, " Human activity monitoring based on hidden
Markov models using a smartphone, " IEEE Instrum. Meas. Mag.,
vol. 19, no. 6, pp. 27-31, Dec. 2016.
[9] M. Gil-Martin, R. San-Segundo, R. de Cordoba, and J. Manuel
Pardo, " Robust biometrics from motion wearable sensors using
a D-vector approach, " Neural Processing Letters, vol. 53, pp.
2109-2125, 2020.
[10] R. San-Segundo, J. D. Echeverry-Correa, C. Salamea-Palacios, S.
L. Lutfi, and J. M. Pardo, " I-vector analysis for gait-based person
IEEE Instrumentation & Measurement Magazine
February 2022

Instrumentation & Measurement Magazine 25-1

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 25-1