The paper presents NEAPOLIS, an automatic real-time system for detecting arrhythmia conditions from electrocardiogram (ECG) readings. NEAPOLIS extracts features from ECG signals like wavelet transforms and interval descriptors. Random forest feature selection identified the top 5 important features. NEAPOLIS was tested on the MIT-BIH database, classifying beats with over 97% average sensitivity, specificity, precision and F1 score. Results for individual arrhythmia classes like left bundle branch block and premature ventricular contraction also exceeded 94%. NEAPOLIS is part of the larger ATTICUS ambient-intelligent telemonitoring system for improving human sustainability.

1. Automatic Real-time Beat-to-beat
Detection of Arrhythmia Conditions
Giovanni Rosa, Gennaro Laudato, Angela Rita Colavita, Simone Scalabrino, and Rocco Oliveto

2. IoT for Healthcare

3. Internet of Medical Things
Preventive care
Long-term care and
chronic diseases

4. Internet of Medical Things
AI and Smart Wearables
for
early anomaly detection

5. Internet of Medical Things
Preventive medical support
thanks to
early notified anomalies

6. Zhao et al. (2005)

7. Zhao et al. (2005) Haque et al. (2009)

8. Zhao et al. (2005) Leonarduzzi et al. (2010)
Haque et al. (2009)

9. Zhao et al. (2005) Leonarduzzi et al. (2010) Li et al. (2016)
Haque et al. (2009)

10. A need for automatic systems having real-time
anomaly detection with high accuracy

11. NovEl APproach for the autOmatic
reaL-time beat-to-beat detectIon
of arrhythmia conditions
(NEAPOLIS)

12. Arrythmia conditions
Left and Right Bundle Branch
Block (LBBB and RBBB)
Premature Ventricular
Contraction (PVC)
Atrial Premature Beats (APB)

13. NEAPOLIS in a nutshell

14. NEAPOLIS in a nutshell

15. NEAPOLIS in a nutshell

16. NEAPOLIS in a nutshell

17. NEAPOLIS in a nutshell

18. pre-RR interval
post-RR interval
local-RR interval
Single beat
Requires at least 10 heart beats
Current and post heart beats
Previous and current heart beats
Pandey and Janghel (2020)
Selected features

19. Selected features
Maximum Overlap Discrete
Wavelet Transform (MODWT)
Autoregressive Model (AR)
Multifractal Wavelet Leader
Fast Fourier Transform (FFT)
R-R interval descriptors
ECG Signal
Extracted features

20. Experiment

21. MIT-BIH Database
48
ECG Recordings
~110,000
Labelled heart beats
30
Minutes of recording
Goldberger et al. (2000); Moody and Mark (2001)

22. Data extraction
Extraction of R peaks
annotations
Beat segmentation Baseline removal
2-step median filter
200 ms
600 ms
Filtered ECG signal
180 samples
(360hz signal)

23. Data extraction
~99,000
heart beats
Normal LBBB RBBB PVC APB
70,000
35,000
0

24. What are the most important features for
the beat-to-beat classification of
arrhythmia conditions?
RQ1

25. Selected features
Maximum Overlap Discrete
Wavelet Transform (MODWT)
Autoregressive Model (AR)
Multifractal Wavelet Leader
Fast Fourier Transform (FFT)
R-R interval descriptors
ECG Signal
Feature selection
using
Random Forest
Remove
co-correlated features
(Pearson > 0.9)
Extracted features
Feature selection

26. Selected features - Top 5
dw_4 fft_1 fft_3 cfr_1 pre_rr
Importance
level
dw = Discrete Wavelet
fft = Fast Fourier Transform
pre_rr = pre-RR interval
cfr = AR model reflection
coefficient

27. What is the accuracy of NEAPOLIS?
RQ2

28. Heart Beats classification
via LSTM Model
Pandey and Janghel (2020)
Selected baseline

29. Classification
Train/Test
Random Split
Repeated 1,000 times, due to split randomness
DS1 (training set)
and
DS2 (test set)
50% - 50%

30. Classification
Class balancing
and
Feature scaling
Random Forest
Classifier
Train/Test
Random Split
Repeated 1,000 times, due to split randomness
DS1
DS1 (training set)
and
DS2 (test set)
50% - 50%

31. Classification
Class balancing
and
Feature scaling
Random Forest
Classifier
Validation
Train/Test
Random Split
Repeated 1,000 times, due to split randomness
DS1 (training set)
and
DS2 (test set)
DS1
DS2
50% - 50%

32. Overall Results
Average metrics NEAPOLIS (Pandey et al., 2020)
Sensitivity
Specificity
Precision
F1 score
97.16 (+ 2.27)
99.53 (+ 0.39)
97.22 (+ 0.49)
97.18 (+ 1.41)
94.89
99.14
96.73
95.77

33. Results (for each class)
Metric Normal
Sensitivity
Specificity
Precision
F1 score
LBBB RBBB PVC APB
99.18 (+ 0.21)
99.97 (+ 0.04)
99.68 (+ 0.63)
99.43 (+ 0.42)
98.28 (+ 3.10)
99.61 (- 0.02)
95.02 (- 0.05)
96.62 (+ 1.49)
90.48 (+ 7.00)
99.81 (+ 0.02)
92.49 (+ 0.85)
91.47 (+ 4.10)
98.53 (+ 1.01)
99.96 (+ 0.04)
99.50 (+ 0.45)
99.01 (+ 0.73)
99.34 (+ 0.03)
98.29 (+ 1.84)
99.43 (+ 0.59)
99.39 (+ 0.32)

34. NEAPOLIS is part of a real IoMT system
ATTICUS
Ambient-intelligent Tele-monitoring System

35. Summary
The paper has been funded by the
PON project ARS01_00860 “Telemonitoraggio e telemetria in ambienti intelligenti per migliorare la sostenibilità umana”
ATTICUS - RNA/COR 576347