1. Random Forest-Based Classification of Heart Rate Variability Signals by Using Combinations of Linear and Nonlinear Features
Alan Jovic, Nikola Bogunovic
Faculty of Electrical Engineering and Computing,
University of Zagreb, Croatia

2. Contents Problem description Methods HRV records Results Discussion
Feature extraction
Classification and evaluation
HRV records
Results
Discussion
Conclusion

3. Problem description
Heart rate variability (HRV) analysis examines fluctuations in the sequence of cardiac interbeat (RR) intervals
Each cardiac rhythm has a pattern (regular or irregular) in these RR interval fluctuations
HRV is a strong predictor of arrhythmic mortality

4. Problem description
Some rhythms have very similar patterns of HRV, e.g. normal and (accelerated) junctional rhythm
Some patterns cannot be efficiently detected by HRV analysis alone, e.g. bundle branch blocks, differentiating atrial disorders (atrial fibrillation vs. wandering atrial pacemaker)
Many rhythms and anomalies can be automatically detected and classified using HRV analysis alone
The main question are:
How accurately can a rhythm be classified?
What should the optimal length of the analyzed segment be?
Which features should be used for which type of rhythm?

5. Motivation Nonlinear phenomena are involved in genesis of HRV:
Lots of nonlinear features described in
literature, very few comparisons of
different features’ combinations on
the same dataset
The aim of this work is to
evaluate a number of different
combinations of (sometimes
interrelated) linear and nonlinear HRV
features in classification of several
types of cardiac rhythms

6. Feature extraction
We consider that a feature is linear if it is unable to take into account the nonlinear dynamics of the HRV signal
Examples of linear features include:
Time domain statistical and geometric measures
Frequency domain spectral features
Nonlinear features try to encompass and quantify the observed complexity of the HRV signal changes
Most of the employed nonlinear features make no assumptions on whether the changes are deterministic or stochastic in origin
Some of the features are specifically designed for HRV analysis, others have more broad areas of application

7. Feature extraction 1 SDNN, pNN20, pNN50, RMSSD, HTI 5
Scheme
number
Features in scheme
Number of features
Description
Comment 1
SDNN, pNN20, pNN50, RMSSD, HTI 5
Linear, time domain 2
PSD, VLF, LF, HF, LF/HF
Linear,
frequency domain 3
Linear (time domain), linear (frequency domain) 10
Linear 4
Linear, SD1/SD2 ratio, Fano factor, Allan factor 13
Linear +
nonlinear
Spatial filling index (SFI), Correlation dimension (D2), Central tendency measure (CTM)
Nonlinear chaos attractor features
Time interval (lag) , T={1,2,5,
10, and 20}, reconstruction
dimension d=2 6
Approximate entropy (ApEn1-ApEn4), Maximum approximate entropy (MaxApEn), r for MaxApEn, Multiscale sample entropy (SampEn1-SampEn20), Multiscale Carnap 1D entropy(Carnap1-Carnap20) 46
Entropies
Dimension m=2 for ApEn and SampEn 7
Advanced sequential trend analysis (ASTA): ASTA1-ASTA11 11
ASTA 8
Detrended fluctuation analysis (DFA): DFA 5, DFA 7, DFA 10, DFA 15, DFA 20
DFA 9
SFI, D2, CTM, ApEn1-ApEn4, SDNN, pNN20, RMSSD, HTI
Features combination
T=1, d=2, m=2
All features 78
Advanced linear + nonlinear chaos attractor features (T=1) + entropies + ASTA + DFA

8. Feature extraction
Most of the linear and nonlinear features were implemented in our own framework for HRV called ECG Chaos Extractor
The only exceptions were frequency domain features, which were extracted in Matlab using the autoregressive function
Some newly proposed nonlinear features
ASTA, Carnap 1D (tessellation) entropy – both methods require more elaborate further research
Not all of the nonlinear HRV features covered in literature were inspected (e.g. Lyapunov exponents, spectral entropy...)

9. Classification and evaluation
For the best results on a large number of features, a strong classification algorithm is required
We opted for Random Forests (RF), an ensemble of random decision trees developed by Breiman in 2001
Internal mechanism for feature selection makes it a valuable tool in the case of a large number of potentially insignificant features
We have also tried other classifiers: ANN, SVM, and C4.5 decision tree, however none of the algorithms gives better results in terms of accuracy and speed

10. Classification and evaluation
RF was constructed with 40 trees for each feature scheme
Stratified 10x10-fold cross-validation evaluation procedure was executed
Analysis overview:
HRV annotations records
ECG Chaos Extractor
Matlab
Weka
Classification accuracy
DATABASES
FEATURE EXTRACTION
FEATURE SELECTION AND CLASSIFICATION
RESULTS

11. Heart rhythm (total no. of feature vectors) ECG annotations records
HRV records
Heart rhythm (total no. of feature vectors)
PhysioBank database
ECG annotations records
RR intervals analyzed
Normal heart rhythm (665)
MIT-BIH Normal Sinus Rhythm Database,
Normal Sinus Rhythm RR Interval Database
MIT-BIH:
NSR: nsr001-nsr054
1-500, , , , , , , ,
Any arrhythmia
(492)
MIT-BIH Arrhythmia Database,
CAST RR Interval Sub-Study Database
MIT-BIH:
CAST: e001a-e130a, f001a-f130a
1-500,
Supraventricular arrhythmia (312)
MIT-BIH Supraventricular
Arrhythmia Database
1-500, , ,
Congestive heart failure (747)
BIDMC Congestive Heart Failure Database,
Congestive Heart Failure
RR Interval Database
BIDMC: chf01-chf15
CHF RR: chf201-chf219
1-500, , , ... , ,
Four types of cardiac rhythms (seven databases)
500 RR intervals analyzed, with overlapping
A total of 2216 feature vectors

12. Nonlinear chaos attractor features
Results
Scheme
number
Description 1
Linear, time domain 2
Linear,
frequency domain 3
Linear 4
Linear +
nonlinear 5
Nonlinear chaos attractor features 6
Entropies 7
ASTA 8
DFA 9
Features combination 10
All features

13. Nonlinear chaos attractor features
Results
Scheme
number
Description 1
Linear, time domain 2
Linear,
frequency domain 3
Linear 4
Linear +
nonlinear 5
Nonlinear chaos attractor features 6
Entropies 7
ASTA 8
DFA 9
Features combination 10
All features

14. Nonlinear chaos attractor features
Results
Scheme
number
Description 1
Linear, time domain 2
Linear,
frequency domain 3
Linear 4
Linear +
nonlinear 5
Nonlinear chaos attractor features 6
Entropies 7
ASTA 8
DFA 9
Features combination 10
All features

15. Nonlinear chaos attractor features
Results
Scheme
number
Description 1
Linear, time domain 2
Linear,
frequency domain 3
Linear 4
Linear +
nonlinear 5
Nonlinear chaos attractor features 6
Entropies 7
ASTA 8
DFA 9
Features combination 10
All features

16. Discussion Good performance was achieved with schemes: 4, 10, 3, and 9
The most promising combination is the one in scheme 4 consisting of the following features:
SDNN, pNN20, pNN50, RMSSD, HTI, PSD, VLF, LF, HF, LF/HF, SD1/SD2 ratio, Fano factor, Allan factor
High increase in the number of nonlinear features did not significantly improve classification accuracy
For further research, each segment should be labeled based on the beats or rhythm it contains, and not based on database it originated from
Improvement in accuracy is to be expected
Drawback is the time needed for careful labeling of the rhythms
Additional research is required for useful applicability of certain methods (ASTA, Carnap entropy)

17. Conclusion
Results suggest high efficiency of linear features in the classification problems
Some of the nonlinear features contribute to greater accuracy of the models
Random forest proved valuable for:
Finding the most relevant subset of features
Efficient classification of different cardiac rhythms
The authors recommend a combination of several time domain, frequency domain and nonlinear features for the best results on medium-sized HRV segments

18. Thank you!
Questions?