Page 78 - Fister jr., Iztok, Andrej Brodnik, Matjaž Krnc and Iztok Fister (eds.). StuCoSReC. Proceedings of the 2019 6th Student Computer Science Research Conference. Koper: University of Primorska Press, 2019
P. 78
o, they both mentioned that the main difficulty of clas- Figure 3: Example graph with colored background.
sifying recordings classified with the lowest confidence was Sections with green background are detected as nor-
the high amplitude noise and the irregular baseline changes mal rhythm sections, red indicates AF, and other
as it made P-wave detection very difficult. In case of hardly arrhythmias are highlighted with blue.
detectable P-waves, they both would rather look at RR in-
tervals to inspect whether they are regular or not. 5. CONCLUSION
In addition, Expert-2 noted that the model appeared to rec- While it is clear that current conditions demand expertise
ognize the regularity of RR-intervals and used it as a strong in both domains of cardiology and machine learning, the
evidence of normal class (doctors also consider it as a sign emergence of cheap hand-held devices creates a niche for
of healthy heart rhythm). Unfortunately, it was misleading approaches capable of utilizing larger amount of data, and
sometimes because some recordings exhibited both unusu- that gives rise to adaptive and scalable algorithms, such as
ally regular RR-intervals and also some clear signs of AF, Deep Neural Networks.
and thus the model predicted normal class (with low confi-
dence) instead of AF. Besides, Expert-1 noticed that all of We carried out an extensive architecture and hyperparame-
the confidently classified AF recordings have arrhythmia ab- ter search, and reported our findings on the Computing in
soluta (i.e., the RR intervals are always changing), and most Cardiology 2017 Challenge. To contribute to the field, we
of these recordings have high BPM value, while recordings have open-sourced our project providing a general training
of low confidence prediction have much lower BPM on aver- environment for practitioners to quickly evaluate baseline
age. She mentioned that arrhythmia absoluta and BPM are performances on their dataset.
two of the most common features cardiologists are looking
for in real-life clinical practice (along with the absence of Our proposed algorithm provides visual reasoning and feed-
P-wave). Thus, she acknowledged that our model appeared back for decision making that can significantly boost effi-
to learne some medically relevant feature without explicitly ciency of AF detection in collaboration with experts. For
programming to do so. deeper analysis of the performance, see appendix B. The
website of our project is available at http://physionet.
4.4 Most relevant segments of recordings itk.ppke.hu/
We divided the signal to 50ms long segments and calculated Lastly, to help doctors to analyze long ECG recordings easily
the output of the neural network for each of them. As we did and quickly, we designed a tool that colors the background
previously when we measured the ”confidence” of the deci- of the ECG plot highlighting the segments according to the
sion of the model, we took again the output of the three neu- prediction of the model.
rons of the soft-max layer, responsible for Normal, AF and
Other classes, respectively. For each 50 ms long segment, 6. REFERENCES
the neuron that produced the highest value determined the
color of the background behind the current section. When [1] C. A. Morillo, A. Banerjee, P. Perel, D. Wood, and
the neuron responsible for the normal class produced the X. Jouven, “Atrial fibrillation: the current epidemic,”
highest value, then the background was colored to green. Journal of geriatric cardiology: JGC, vol. 14, no. 3, p.
Similarly, the blue background indicated that the neuron of 195, 2017.
the other class had the highest output value, and red in-
dicated atrial fibrillation. Additionally, higher values are [2] B. B. Kelly, V. Fuster et al., Promoting cardiovascular
translated to darker colors, so the darkness of background health in the developing world: a critical challenge to
indicates the ”confidence” of the prediction at a certain seg- achieve global health. National Academies Press,
ment. An example of these graphs can be seen in Figure 3. 2010.
By implementing that algorithm, we aimed to help doctors
by highlighting the most relevant regions of ECG recordings. [3] G. H. Tison, J. M. Sanchez, B. Ballinger, A. Singh,
While our algorithm cannot substitute doctors, it might be J. E. Olgin, M. J. Pletcher, E. Vittinghoff, E. S. Lee,
a good tool to speed up the evaluation of long ECG record- S. M. Fan, and R. A. e. a. Gladstone, “Passive
ings while unburdening physicians drawing their attention detection of atrial fibrillation using a commercially
to the most important parts of the signal. available smartwatch,” JAMA Cardiology, vol. 3, no. 5,
p. 409, 2018.
4.5 Computational Complexity
[4] S. P. Shashikumar, A. J. Shah, Q. Li, G. D. Clifford,
From the perspective of practical applicability in real-life and S. Nemati, “A deep learning approach to
medicine, our method is not just designed for classification,
but performs well as a real-time detector by the nature of
Fully Convolutional Networks: after the initial warm-up de-
lay of 1.2 sec, we can generate new responses in less than
2 msec taking the last 20 second history into consideration.
If the evaluation is centralized and we allow to compute re-
sponses in batches, the time required per sample is less than
0.5 msec.
StuCoSReC Proceedings of the 2019 6th Student Computer Science Research Conference 78
Koper, Slovenia, 10 October
sifying recordings classified with the lowest confidence was Sections with green background are detected as nor-
the high amplitude noise and the irregular baseline changes mal rhythm sections, red indicates AF, and other
as it made P-wave detection very difficult. In case of hardly arrhythmias are highlighted with blue.
detectable P-waves, they both would rather look at RR in-
tervals to inspect whether they are regular or not. 5. CONCLUSION
In addition, Expert-2 noted that the model appeared to rec- While it is clear that current conditions demand expertise
ognize the regularity of RR-intervals and used it as a strong in both domains of cardiology and machine learning, the
evidence of normal class (doctors also consider it as a sign emergence of cheap hand-held devices creates a niche for
of healthy heart rhythm). Unfortunately, it was misleading approaches capable of utilizing larger amount of data, and
sometimes because some recordings exhibited both unusu- that gives rise to adaptive and scalable algorithms, such as
ally regular RR-intervals and also some clear signs of AF, Deep Neural Networks.
and thus the model predicted normal class (with low confi-
dence) instead of AF. Besides, Expert-1 noticed that all of We carried out an extensive architecture and hyperparame-
the confidently classified AF recordings have arrhythmia ab- ter search, and reported our findings on the Computing in
soluta (i.e., the RR intervals are always changing), and most Cardiology 2017 Challenge. To contribute to the field, we
of these recordings have high BPM value, while recordings have open-sourced our project providing a general training
of low confidence prediction have much lower BPM on aver- environment for practitioners to quickly evaluate baseline
age. She mentioned that arrhythmia absoluta and BPM are performances on their dataset.
two of the most common features cardiologists are looking
for in real-life clinical practice (along with the absence of Our proposed algorithm provides visual reasoning and feed-
P-wave). Thus, she acknowledged that our model appeared back for decision making that can significantly boost effi-
to learne some medically relevant feature without explicitly ciency of AF detection in collaboration with experts. For
programming to do so. deeper analysis of the performance, see appendix B. The
website of our project is available at http://physionet.
4.4 Most relevant segments of recordings itk.ppke.hu/
We divided the signal to 50ms long segments and calculated Lastly, to help doctors to analyze long ECG recordings easily
the output of the neural network for each of them. As we did and quickly, we designed a tool that colors the background
previously when we measured the ”confidence” of the deci- of the ECG plot highlighting the segments according to the
sion of the model, we took again the output of the three neu- prediction of the model.
rons of the soft-max layer, responsible for Normal, AF and
Other classes, respectively. For each 50 ms long segment, 6. REFERENCES
the neuron that produced the highest value determined the
color of the background behind the current section. When [1] C. A. Morillo, A. Banerjee, P. Perel, D. Wood, and
the neuron responsible for the normal class produced the X. Jouven, “Atrial fibrillation: the current epidemic,”
highest value, then the background was colored to green. Journal of geriatric cardiology: JGC, vol. 14, no. 3, p.
Similarly, the blue background indicated that the neuron of 195, 2017.
the other class had the highest output value, and red in-
dicated atrial fibrillation. Additionally, higher values are [2] B. B. Kelly, V. Fuster et al., Promoting cardiovascular
translated to darker colors, so the darkness of background health in the developing world: a critical challenge to
indicates the ”confidence” of the prediction at a certain seg- achieve global health. National Academies Press,
ment. An example of these graphs can be seen in Figure 3. 2010.
By implementing that algorithm, we aimed to help doctors
by highlighting the most relevant regions of ECG recordings. [3] G. H. Tison, J. M. Sanchez, B. Ballinger, A. Singh,
While our algorithm cannot substitute doctors, it might be J. E. Olgin, M. J. Pletcher, E. Vittinghoff, E. S. Lee,
a good tool to speed up the evaluation of long ECG record- S. M. Fan, and R. A. e. a. Gladstone, “Passive
ings while unburdening physicians drawing their attention detection of atrial fibrillation using a commercially
to the most important parts of the signal. available smartwatch,” JAMA Cardiology, vol. 3, no. 5,
p. 409, 2018.
4.5 Computational Complexity
[4] S. P. Shashikumar, A. J. Shah, Q. Li, G. D. Clifford,
From the perspective of practical applicability in real-life and S. Nemati, “A deep learning approach to
medicine, our method is not just designed for classification,
but performs well as a real-time detector by the nature of
Fully Convolutional Networks: after the initial warm-up de-
lay of 1.2 sec, we can generate new responses in less than
2 msec taking the last 20 second history into consideration.
If the evaluation is centralized and we allow to compute re-
sponses in batches, the time required per sample is less than
0.5 msec.
StuCoSReC Proceedings of the 2019 6th Student Computer Science Research Conference 78
Koper, Slovenia, 10 October
