A patient-adaptable ECG beat classifier using a mixture of experts approach
Summary (3 min read)
Introduction
- A large in-house ECG database is developed and maintained to test each ECG processing algorithm to be incorporated into the product.
- The result is a complicated classifier which is costly to develop, maintain, and update.
- The authors may include the training algorithm and the database used to develop the classifier to be delivered to the users, so that the classification algorithm can be finetuned to each patient.
A. ECG Beat Classification Techniques
- Automated ECG beat classification was traditionally performed using a decision-tree-like approach, based on various features extracted from an ECG beat [1], [4], [5], [13], [20], [22].
- The features used include the width and height of QRS complex, RR interval, QRS complex area, etc.
- Their abilities to learn from examples and extract the statistical properties of the examples presented during the training sessions, make them an ideal choice for an automated process that imitates human logic.
- Several efforts have been made to apply ANN’s for the purpose of ECG beat detection and classification.
- They have achieved an average recognition accuracy of 90% in classifying the beats into two groups; normal and abnormal.
B. Self-Organization Map (SOM) and Learning Vector Quantization (LVQ)
- SOM and LVQ are both clustering based algorithms proposed by Kohonen [14], [15].
- In LVQ1, for a given input vector , a code word is found such that (2) The code word is then updated as follows: (3) where if the classification is correct [i.e., and have the same class label] and , otherwise.
- As such, the development of the code book and eventually decision boundary can be made completely transparent to the user.
- The resulting code book then will be submitted to the LVQ PAK to facilitate fine tuning and classification.
III. MIXTURE OF EXPERTS (MOE)
- This user adaptation problem bears certain resemblance to the incremental learning problem in that new data are to be incorporated to improve existing classifier’s performance.
- The LE represents a specialized ECG beat classifier, trained on a small segment of annotated ECG beats taken from the specific patient.
- In the MOE method, the combined th output vector of both the experts is given by (6) where is the input feature vector, , are the weighting vectors for each expert from a gating network and are defined by (7) where ’s are the weight vectors of the gating network.
- Define , and , , to be the subregion in the feature space where the classifier makes correct classification of and let be defined the same way, also known as Theorem 1.
- The authors further partition this training data set into two subsets: one for the training of the user-specific classifier , and the other for estimating and .
IV. EXPERIMENT
- The purpose of this experiment is to demonstrate the usefulness of the proposed user-adaptation procedure.
- In particular, the authors will show that an ECG beat classifier trained on general patient records does not perform well when presented with patient records which contain rare beat types.
- Moreover, the authors show that the performance of the MOE classifier is able to gain significant performance enhancement with a small amount of annotated patient specific training data.
A. Data Preparation
- The authors concentrate on the classification of ventricular ectopic beats (VEB’s).
- According to the AAMIrecommended practice, records containing the paced beats (four records) can be excluded from the reporting requirements.
- The first group is intended to serve as a representative sample of a variety of waveforms and artifacts which an arrhythmia detector might encounter in routine clinical use.
- If this GE classifier were a commercial device, it will be deemed not-applicable (due to low performance) to many of these 20 test records.
- Each of the four categories included beats of several types as shown in Table III.
B. Training and Testing Procedure
- A GE classifier was developed with SOM and LVQ algorithms using the data from the records of the first group (100–124).
- The objective of this paper is to classify the QRS beats into one of the four different categories.
- This is a reasonable assumption since each code word is obtained using the SOM clustering algorithm based on the L norm distance measure.
- The LE classifier is developed in exactly the same manner as the global classifier, except that it uses only the first two and half minutes in the tape, and is constructed separately for each particular “patient tape” (tape #200–234) in the MIT/BIH database.
- The output of the classifier is calculated as given by (6).
C. Results
- The classifier was tested with the selected 20 records of the second group of the MIT database.
- The GE was left intact and is used as is for testing the 25 min of data from each 30-min testing record with first 5 min excluded as they are used to develop the LE and the “gating network.”.
- All detection statistics are founded on the mutually exclusive categories of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN).
- Sensitivity, specificity, and positive predictivity are used to compare the results, also known as Three statistics.
- These three statistics, together with the percentage classification rates, are reported for each individual testing file as required by the AAMI-recommended practice [18].
D. Discussion
- 1) From Tables V and VI, the authors observe that the MOE approach is capable of significantly enhancing the performance of an ECG beat classifier over the global classifier.
- 2) Comparing the LE and ME, the authors found that LE outperformed ME in terms of classification rate, mainly due to higher specificity (ability to correctly classify normal beats), but with lower sensitivity (ability to correctly classify PVC beats as PVC).
- Hence, although a LE classifier performs well, the availability of a global classifier does help to further enhance its performance.
- 4) A potential drawback of this proposed method is the need to develop a LE classifier for each individual patient, even with only 5 min of patient’s ECG record.
- Since this must be performed by a physician or a ECG specialist, potentially it would be very costly.
V. CONCLUSION
- The authors developed a novel approach to demonstrate the feasibility of having a patient-adaptable ECG beat classification algorithm.
- The authors outlined the basic requirements of such a system, namely accuracy, cost-effectiveness and protection of the device manufactures intellectual property rights.
- The authors presented a SOM/LVQ-based approach to illustrate that these requirements can be met.
- The potential benefit of patient adaptation is immense and is worth pursuing further.
- The authors believe it can be easily adapted to other automated patientmonitoring algorithms and eventually support decentralized remote patient-monitoring systems.
Did you find this useful? Give us your feedback
Citations
2,933 citations
1,449 citations
Cites background or methods or result from "A patient-adaptable ECG beat classi..."
...For comparison, the classification performance of a published system [11] is shown in Table VII(b)....
[...]
...Classifiers methods employed include linear discriminants [7], back propagation neural networks [8]–[10], self-organizing maps with learning vector quantization [11], and self-organizing networks [12]....
[...]
...[11], as this study looked at the problem of distinguishing VEB from non-VEB heartbeats on the MIT-BIH arrhythmia database....
[...]
...TABLE VII (A) CLASSIFICATION PERFORMANCE OF CONFIGURATION IX ON EACH RECORDING OF DS2 USING THE AAMI RECOMMENDED PERFORMANCE MEASURES, (B) AGGREGATE CLASSIFICATION PERFORMANCE FOR THE SYSTEM IN [11], AND (C) THE CLUSTERING PERFORMANCE OF CLASSIFIER IX AND SYSTEM FROM [12]...
[...]
...The AAMI standards are adopted in this study and our results have been compared to those of [11] and [12]....
[...]
1,300 citations
Cites background or methods from "A patient-adaptable ECG beat classi..."
...There have been several methods for generic and fully automatic ECG classification based on signal processing techniques, such as frequency analysis [2], wavelet transform [3] and filter banks [4], statistical [5] and heuristic approaches [6], hidden Markov models [7], support vector machines [8], artificial neural networks (ANNs) [9], and mixture-of-experts method [10]....
[...]
...With the limited training data as proposed in [10] and [14]–[17], we shall demonstrate that simple CNNs will suffice to achieve a superior classification performance rather than the complex ones that are commonly used for deep learning tasks....
[...]
...Among all, only few of them with a patient-specific design [10], [12], [15]–[18] have, in particular, demonstrated significant performance improvements over the automatic and generic ECG classification methods thanks to their ability to adapt or optimize the classifier body according to each patient’s ECG signal....
[...]
...However, among many methods in the literature, only few [10], [14]–[18] have in fact used the AAMI standards along with the complete data from the benchmark MIT-BIH arrhythmia database [22]....
[...]
...To perform a more extensive and accurate comparative performance evaluation, the base performance of the proposed system is compared with the three existing algorithms, [10], [15], and [16], all of which comply with the AAMI standards....
[...]
586 citations
555 citations
References
8,197 citations
"A patient-adaptable ECG beat classi..." refers methods in this paper
...According to Kohonen, there are three different LVQ algorithms, called LVQ1, LVQ2, and LVQ3 developed at subsequent stages to handle classification problems with different natures....
[...]
...SOM and LVQ are both clustering based algorithms proposed by Kohonen [14], [15]....
[...]
...[14] T. Kohonen, Self-Organization and Associative Memory....
[...]
7,883 citations
6,686 citations
Additional excerpts
...algorithm [21]....
[...]
4,338 citations
2,933 citations
Related Papers (5)
Frequently Asked Questions (12)
Q2. What is the classification of ECG beats?
Automated ECG beat classification was traditionally performed using a decision-tree-like approach, based on various features extracted from an ECG beat [1], [4], [5], [13], [20], [22].
Q3. How many records were excluded from the study?
Since this study is to evaluate the performance of a classifier that can identify a premature ventricular contraction (PVC), certain records in the database with no PVC’s (11 records) were excluded from the study, leaving 33 records of interest.
Q4. What is the drawback of this proposed method?
4) A potential drawback of this proposed method is the need to develop a LE classifier for each individual patient, even with only 5 min of patient’s ECG record.
Q5. How many beats are in the MIT/BIH database?
Since each record in the MIT/BIH database is of length 30 min, the 2.5 min segment account for 1/12th of total available patient specific data and contains approximately 150 ECG beats.
Q6. How will the GE classifier be adapted to the MOE approach?
with the MOE approach, the authors will adapt this GE classifier with a LE classifier to gain significant performance enhancement at low cost.
Q7. What is the morphological template of the GE?
The information of each beat is stored as a 13-element vector, with the first nine elements representing the transformed morphological template, and the next three elements representing the temporal parameters.
Q8. How can the LE be used to improve the performance of the patient?
the LE is able to pick up those patient-specific beats, and therefore, provide significantly enhanced performance (from 3.65% to 98.4%).
Q9. What is the cost of annotating a brief segment of patient-specific ECG?
In practice, the attending cardiologist or any expert in ECG beat annotation will have to annotate a brief segment of patient-specific ECG in order to take advantage of the MOE approach.
Q10. How many points were picked up to form the template?
The position of annotation labels is used to identify the peak of the QRS waveform and 14 points on either side of the peak were picked up to form the template.
Q11. What is the a posterior probability of the classifier output?
To enable the “soft combination” of the classifier output, it is desired that the outputs of each classifier be an estimate of the a posterior probability of the feature vector belonging to that class.
Q12. How many records are used in this study?
The second group of 20 records is used to simulate the ECG records of 20 patients, which are to be classified by the GE classifier.