Page 70 - AIH-1-4

Artificial Intelligence in Health ML models for heartbeat classification

Table 1. Summary of mappings between beat annotations highlights several challenges, including (I) handling
and Association for the Advancement of Medical EC57 missing and imbalanced data, (II) selecting a robust
categories 18 classification algorithm, (III) managing ECG signal
complexity, (IV) addressing computational demands,
Label Category Annotation and (V) recognizing methodological limitations. To
0 N • Normal achieve the objectives of this study, seven classification
• Left/right bundle branch block
• Atrial escape algorithms were used to classify heartbeat categories, and
• Nodal escape their performance was evaluated. Herein, ML models-
1 S • Atrial premature nearest neighbors (KNN), naive Bayes (NB) classifier,
• Aberrant atrial premature random forest (RF) classifier, logistic regression (LR),
• Nodal premature eXtreme gradient boosting (XGBoost) classifier, support
• Supraventricular premature vector machines (SVMs), and decision trees (DTs) were
2 V • Premature ventricular contraction employed, along with the incorporation of the FT and
• Ventricular escape Gaussian noise injection techniques. The overall design of
3 F • Fusion of ventricular and normal the implementation method is presented in Figure 5. In this
study, we utilized the advantages of Pearson correlation and
the associated p-values, which serve as statistical tools, to
unveil meaningful relationships and dependencies within
our data. In addition, we introduced controlled noise to
enhance the robustness of our models, allowing them to
better adapt to real-world variations. Furthermore, the FT
technique was leveraged to extract essential frequency-
domain features from our data. This combination enabled
our models to make more accurate predictions and better
understand complex data patterns, offering valuable
insights for various applications, comparable to state-of-
art algorithms.
The ML models enhanced through the proposed
approach can exhibit improvements in terms of key metrics
Figure 2. Distribution of the training set such as accuracy and F1 score. KNN and NB are favored
for their simplicity and real-time efficiency, while RF and
XGBoost excel in handling complex interactions and large
datasets, offering robustness and feature importance. LR is
valued for its interpretability in binary classification, SVMs
are known for managing high-dimensional data and noise,
and DTs are favored for their clear and interpretable results.
These algorithms were selected for their effectiveness in
handling complex ECG data, computational efficiency,
adaptability, and noise robustness, ensuring reliable
performance and scalability. Notably, the ability of ML
models to handle motion artifacts in ECG signals varies.
RF and XGBoost are particularly robust under noisy
environments due to their ensemble nature, while SVMs
19
effectively maintain accuracy using kernel methods. KNN
20
and NB face challenges with respect to noise sensitivity;
Figure 3. Distribution of the test set however, preprocessing techniques such as FT can help in
this regard. LR and DT require feature engineering and
2.3. Methods used noise reduction for exhibiting better performance, 21,22 with

Enhancing the classification prediction accuracy can aid ensemble methods further boosting their robustness.
in the early diagnosis of cardiovascular diseases. However, As detailed in a previous study, Pearson’s correlation
23
a review of current state-of-the-art research indicates that test is a statistical method used to assess the relationship
metrics and prediction rates often fall short. Literature between two continuous variables. It yields a coefficient

Volume 1 Issue 4 (2024) 64 doi: 10.36922/aih.3543