Heart Disease Prediction Using Late Fusion Over Convolutional Neural Networks

Chronic heart failure (CHF) has evolved as not only a prevalent but also an increasing health problem as the heart continuously fails to perform an adequate task of delivering anticipated perfusion of blood to tissue and organ systems (physiologic demands) [1, 2]. The occurrence of CHF is rapidly rising to epidemic levels at 3% per year. CHF has a prevalence in developed nations, affecting about 2–3% of the general population and having a particularly high incidence of 12% in people over the age of 65 [3, 4]. The consequences of CHF are not limited to the field of health, as nowadays, the diagnosis and treatment of this disease account for roughly 3% of the annual costs of health care [5]. This number is an indication of the large economic burden that CHF imposes [6].

Based on these significant health questions, our research is interested in cardiac disease prediction with a special focus on late fusion techniques instead of convolutional neural network (CNN) [7]. Utilizing the insights derived from machine learning (ML), specifically CNNs, our research aims to increase the accuracy and efficiency of prediction models for cardiac diseases by analyzing the heart's sound [8, 9]. By applying the methods of late fusion, this means to join the heterogeneous sources of information in order to identify possibly interacting patterns and increase the accuracy and credibility of forecasts [10, 11].

In electrocardiogram (ECG) signal processing, diverse forms of noise within the acquired signals often undermine the precision of diagnostic data [12, 13]. ECG equipment, renowned for its high sensitivity, frequently encounters a range of interferences, including contact noise, electromagnetic interference (EMI) noise, baseline drift, power line interference, and motion artifacts. Numerous denoising techniques have been extensively investigated on a global scale in response to the urgent requirement for noise reduction [14, 15]

Moreover, implementing a dual-band filter bank alongside a product filter and a lowpass filter to mitigate ECG data noise reflects a notable scholarly achievement [16]. Another approach involves using a deep neural network (DNN) in conjunction with a wavelet scale-adaptive threshold technique and a denoising autoencoder (DAE), resulting in an improved signal-to-noise ratio and root mean square error [17, 18].

An alternative approach uses multi-scale principal component analysis (MSPCA) for efficient removal of noise from the ECG signals prior to processing [19, 20]. While not very well established in the categorization of ECG, the application of MSPCA has shown potential. To eliminate noise from ECG signals, a technique combining empirical mode decomposition and genetic algorithm threshold technology is used [21]. However, there is awareness of the fact that within the optimization procedure, the genetic algorithm may encounter local optima, which can be a challenge [22, 23]. The utilization of CNNs has been prominent due to their remarkable efficacy in processing multidimensional inputs, such as ECG time series data (1D) and medical images (2D and 3D), demonstrating substantial promise [24, 25]. Within the domain of ECG signal classification, CNNs have shown notable efficacy, exhibiting enhanced accuracy and the ability to extract relevant features from input heartbeats [26] directly.

In addition to ECG applications, deep learning (DL) approaches, particularly CNNs, exhibit exceptional performance in analyzing radiological images, thereby demonstrating their utility and versatility across other medical domains [27, 28]. The combination of multiple disciplines creates an interesting intersection between cardiac disease prediction, ECG signal denoising, and the deep impact of ML and DL in healthcare diagnostics. The primary objective is to improve the accuracy of forecasts and tackle the significant health and economic issues associated with CHF [29].

Recent advances in DL have shown the potential of end-to-end learning, a revolutionary method, in which ML models are able to directly learn from raw data without having pre-identified characteristics [30, 31]. This methodology has proven to be exceptionally efficacious in a wide variety of domains such as pattern recognition, image manipulation, natural language comprehension, speech and audio processing, and sensor data analysis [32, 33]. In the context of detection of CHF, applying a hybrid approach with both traditional ML methods and end-to-end DL methods exhibits better performance than either method separately.

Whereas the traditional methods of ML heavily rely on a huge number of features that are defined by experts. In contrast, end-to-end DL approaches use learning approaches based on the raw phonocardiogram (PCG) signal in the time domain, and a temporal-domain representation (e.g., the spectrogram [34, 35]). Utilizing a hybrid technique that has been effectively applied in previous research on human activity recognition using smartphone sensor data proves to be a very promising approach to improve the performance and accuracy in CHF and other healthcare domains. The combination of traditional ML, which provides a myriad of ways to learn features, and end-to-end DL, providing ways to see directly from the raw data, is a promising framework to break new ground in the field of healthcare diagnostics.

In the schematic illustration (Figure 1), the way of detecting heart disease using late fusion over CNN is described. The first dataset, the PhysioNet Challenge 2020 dataset, is extracted from the INCART Database, PTB Database, CPSC Database, and the Georgia 12-lead ECG Challenge (G12EC) Database, as well as an unidentified database. Then the elaborate pre-processing phase begins. These include key procedures such as batch generation, class weight calculation, handling of undefined classes, informative processing, one-hot encoding, and preparation of validation data. The Late Fusion over CNN model is then applied to the processed images, which show symptoms of heart disease. Finally, the results undergo several evaluations in model procedures, enabling one to confidently determine whether heart disease is present or not.

image003.png

Figure 1. Framework of heart disease prediction using late fusion over a convolutional neural network (CNN)

The Contributions of the research paper are elaborated below:

1. This study adds to the field because it succeeds at integrating the demographic information and the ECG signals so as to predict heart disease. The holistic strategy proposed in this study aims to help manage the various problems encountered in multi-modal data fusion.

2. The study also presents impressive work on optimizing CNNs for maximum performance using the PhysioNet 2020 Challenge dataset. As a result, the introduced process ensures that the model is well-calibrated to the unique characteristics of the provided dataset, which increases its predictive accuracy.

3. The design element of the fusion mechanism involves the introduction of modality-specific streams and the integration of an attention mechanism for fusion. The novelty of this methodology enhances the model's ability to capture relevant characteristics from diverse data sources, thereby improving the overall predictive efficiency of the model.

4. The study focuses on the rigorous optimization of the dataset with proper consideration of the need to tailor the model to the complexity of the PhysioNet 2020 Challenge dataset. Careful tuning of the model improves the efficiency of discovering relevant clinically related patterns related to heart disease.

5. The introduction of the late fusion design is a major improvement, as it allows the model to use data from many ECG leads. It is helpful to perform a well-rounded scan, enabling the model to detect complicated patterns in the context of heart disease across all dimensions of the input data.

The sections below have been presented systematically. Having discussed in detail the existing work done on the subject, Section Two examines that subject matter. Section 3 goes further to explain the various stages that are involved in the data pre-processing and gives a thorough discussion of the datasets used. The study includes the examination of the methods used in Section 4, which gives insight into some of the techniques used for selection. Section 5 gives an extensive account of the experimental setup and the results. After an in-depth discussion of the results in Section 6, Section 7 is dedicated to a coherent conclusion of the paper by briefly summarizing the results and proposing potential directions for further investigation.

This study is included in the PhysioNet/Computation in Cardiology (CinC) Challenge 2020 [7, 16, 31]. It aims to classify 27 cardiac abnormalities by utilizing a dataset consisting of 43,101 12-lead ECG recordings that have been provided. The dataset was classified using a late fusion technique that we developed.

The numbers for this challenge have been sourced from a diverse range of references, such as:

• CPSC Database and CPSC-Extra Database
• INCART Database
• PTB and PTB-XL Database
• The Georgia 12-lead ECG Challenge (G12EC) Database
• Undisclosed Database

The China Physiological Signal Challenge (CPSC 2018) was held in Nanjing, China, at the 7th International Conference on Biomedical Engineering and Biotechnology. The file includes public CPSC Database data and CPSC-Extra Database data. The unused data does not contain CPSC 2018 test data. The final, private repository contains CPSC-2018 exam data. The training set comprises 6,877 individuals (3,699 males and 3,178 females) and 3,453 individuals (1,843 males and 1,610 females). These individuals underwent 6- to 60-second 12-lead ECGs. The recordings are sampled at 500 Hz.

The public dataset from St. Petersburg, the INCART 12-lead Arrhythmia Database, is supplemental. The collection includes 74 annotated recordings from 32 Holter albums. Each 30-minute recording has 12 257 Hz-sampled standard leads.

Third, the Physikalisch-Technische Bundesanstalt (PTB) provides two public databases: one is the PTB Diagnostic ECG Database, and the other is the PTB-XL, which is a huge electrocardiography dataset. The first PTB database involves personalities who have 516 records, 377 males and 139 females. Each recording was sampled at 1000 Hz. PTB-XL contains a set of 21,837 clinical 12-lead ECGs. The ECGs were collected at a frequency of 500 Hz and for a duration of 10 s. There are a total of 11,379 male and 10,458 female ECGs in the collection.

The dataset of 43,101 ECG records was neatly divided into training, validation, and test phases, ensuring balance, while also being subject-independent to avoid data leakage. Specifically, 70 percent of the data was used for training, 15 percent for validation, and 15 percent for testing, with stratification employed to maintain the class distribution across all subsets. Additionally, to enhance the reliability and robustness of the model, a 5-fold cross-validation was employed, allowing for consistent evaluation of performance across various data splits. Every fold entailed repacking of the data without altering class balances to avoid bias from overrepresented cardiac conditions.

This approach ensured that the model's performance measures were based on generalizable learning, rather than memorization of the dataset. To address the heterogeneity of data combinations across multiple ECG databases, including CPSC, INCART, PTB, PTB-XL, and Chapman Shaoxing, uniform pre-processing pipelines were employed. All the signals were resampled to a fixed frequency of 500 Hz, z-score normalized, and segmented into equal time windows to ensure consistency in signal length. Band-pass filtering (0.5–50 Hz) was used to remove baseline wander and powerline interference, and reduce noise. The harmonization of labels was also performed to align the diagnostic classes across datasets with the PhysioNet 2020 Challenge label taxonomy, ensuring compatibility and avoiding semantic overlap. This method of systematic pre-processing and data processing reduced the impact of domain variance, allowing the model to generalize well.

The fourth source is a Georgia database with Southeastern demographic data. The training set comprises 10,344 10-second, 500-Hz, 12-lead ECGs. The number of male ECGs is 5,551, while the number of female ECGs is 4,793.

The fifth source is an unidentified American database that differs geographically from the state of Georgia. The 10,000 ECGs in this source are test data only.

For complete data transmission, WFDB is used. The diagnosis and other patient information are stored in a WFDB header format text file, with each ECG recording included. Also provided is a binary MATLAB v4 file with ECG signal data. The MATLAB load function and the Python scipy.io. load mat function may read binary files. Refer to our baseline models for examples of data loading. The first line of the header displays the aggregate lead count and the number of samples or data points per lead. The following paragraphs describe how to conserve each lead, and the last sections cover diagnostics and demographics.

3.1 Data imbalance handling

To address the class imbalance issue inherent in the PhysioNet 2020 dataset, we employed a multi-pronged approach, supplementing the adjustment of class weights. First, the skew in the data was measured by determining the distribution of the 27 cardiac conditions. According to this analysis, dynamic class weighting was employed in the model's training, with underrepresented classes assigned higher weights in the loss function. This ensured that the model was more aggressive towards misclassification of minority classes, promoting the network to learn meaningful representations of rare cardiac abnormalities.

Other than weighted loss, minorities were over-sampled, and dominated classes were under-sampled during batch generation. In particular, time-warping, random cropping, and lead-specific noise injection synthetic augmentation methods were employed to undersample overrepresented classes, thereby providing a diverse range of examples without overfitting. Such augmentation techniques not only equalized the number of training samples for each class but also increased the model's resilience to changes in real-world ECG signals.

As a result, the network became more generally mapped to input signals and cardiac condition labels. Validation was significantly enhanced by the joint action of dynamic weighting and data augmentation at the target level. The model was also in a better position to identify minority-class events by reducing bias towards dominant classes, as indicated by the higher recall and AUC values of the validation set. This specific management of the imbalance between classes resulted in an increase in the precision, recall, and accuracy of the validation set to approximately 98%, while the training performance remained high.

The separation of the 12-lead ECG into two streams of six leads has a basis in both physiological and technical arguments. Physiologically, the 12-lead ECG can be classified into limb leads (I, II, III, aVR, aVL, aVF) and precordial leads (V1-V6), which record different electrophysiological views of cardiac activity. The limb leads primarily represent the frontal plane, whereas the precordial leads provide information about the horizontal plane. The model can learn modality-specific features independently on each plane by splitting the leads into two streams, while still maintaining the important spatial and temporal patterns that would have been washed out had all the leads been learned together.

This specific processing enables the network to be more proficient in identifying minor irregularities in specific sets of leads, thereby enhancing classification accuracy. The attention mechanism focuses on how salient properties are prioritized in these two streams to provide a physiologically understandable weighting of lead information. Attention weights enable the model to actively pay attention to leads that provide the most diagnostic information about a particular cardiac condition. Technically, this mechanism reduces the chances of overfitting on redundant features and increases the cross-lead integration of features. It has a twofold theoretical basis, as it is based on the principles of cardiac electrophysiology and the best practices of DL, which make the multi-lead representation more efficient. This enables the network to capture both local and global patterns relevant to predicting heart diseases.

This comparative analysis of the two datasets will demonstrate the effectiveness of our model and its ability to accurately predict both known and unknown data. This section will examine the performance of the training and validation sets using multiple evaluation measures, and then present the results obtained from the training set and validation set.

5.1 Model evaluation metrics

i. Loss

During training, the loss function measures the discrepancy between the actual labels and the model's predicted values. Reduce the loss as much as possible to increase the model's predictive accuracy.

$Binary Crossentropy =-\left[y * \log \left(y^{\prime}\right)+(1-y) * \log \left(1-y^{\prime}\right)\right]$   (2)

While many different loss functions exist, Binary cross-entropy is frequently employed in binary classification applications. Where y' is the predicted probability and y is the true label (0 or 1).

ii. Accuracy

The ratio of correctly predicted instances to the total number of instances is the accuracy metric. To guarantee a high percentage of accurate forecasts, maximize accuracy.

$Accuracy =(T P+T N) /(T P+T N+F P+F N)$   (3)

True positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) are represented in this way.

iii. Recall (or Sensitivity)

The model's recall measures its capacity to locate all pertinent instances, especially those that are favourable. To reduce FNs, maximize recall.

$Recall =T P /(T P+F N)$   (4)

where, FN stands for false negatives and TP for true positives.

iv. Precision

Precision gauges how well the model predicts positive outcomes. To reduce false positives, maximize precision.

$Precision =T P /(T P+F P)$   (5)

where, FP stands for false positives and TP for true positives.

v. AUC (Area under ROC curve)

The AUC assesses how well the model can differentiate between positive and negative occurrences across various probability thresholds.

To guarantee strong discriminating power, maximize AUC.

The ROC curve, which plots the TP rate against the FP rate at various thresholds, serves as the foundation for calculating the AUC.

vi. Confusion Matrix

The confusion matrix is a tabular representation that effectively demonstrates the performance of a classification model by comparing the predicted and observed classes of a dataset. The provided information pertains to the model's ability to accurately categorize cases as either successful or unsuccessful classifications.

The metric that quantifies the number of instances in which the model correctly predicts the positive class is referred to as the TP measure. In the cases above, the occurrence of heart disease aligns with the model's projected outcomes. However, the diagnostic assessment yields a favourable result.

The metric denoting the frequency at which the model correctly predicts the negative class is referred to as the TN. In these cases, the model suggested the absence of heart disease; yet, the empirical evidence demonstrates its existence.

The occurrence of the positive class being mistakenly predicted by the model is also known as FP or Type I Error. In these particular cases, the model made inaccurate predictions by falsely indicating the presence of heart disease, while in reality, it was absent.

The frequency of instances in which the model mistakenly predicts the negative class is also known as FN or Type II Error. In the cases mentioned earlier, the model exhibited erroneous predictions by incorrectly indicating the absence of heart disease, despite a confirmed positive diagnosis.

Collectively, these evaluation measures provide a comprehensive understanding of the model's efficacy in predicting heart disease. While precision, recall, and accuracy focus on specific aspects of classification performance, the AUC metric evaluates the overall discriminative ability of the model. Within the framework of predicting cardiac disease, the amalgamation of various indicators facilitates a comprehensive assessment of the model's strengths and limitations.

The overfitting observed is primarily due to the intrinsic imbalance and noisy nature of the multi-modal biomedical dataset, where specific classes are underrepresented and inter-subject variability is high. The high accuracy of its training enables the model to recall the defining features of the dominant classes. However, the validation recall is rather low, indicating that the model is less able to generalize to unseen data, particularly for minority classes. The modified framework addresses this by incorporating sophisticated regularization methods, such as dropout layers, L2 weight decay, and early stopping, along with data augmentation methods (e.g., random signal stretching, Gaussian noise injection, and spectral masking).

These interventions are designed to enhance the robustness of a given model by applying it to a broader range of signal perturbations, thereby improving memorization and generalization across modalities. Additionally, a scheduling mechanism for the adaptive learning rate and a cross-validation protocol have been proposed to maintain the same performance on the validation set and avoid overfitting to particular folds. Additionally, the fusion architecture is enhanced to incorporate attention-based gating, enabling the model to dynamically repress noisy or redundant features during training. These adjustments stabilize the learning process, provide equal contributions to modality, and enhance recall performance. The results of post-implementation analysis showed better validation recall and a smaller training-validation gap, which validated that the added regularization and optimization measures helped overcome the issue of overfitting that the reviewers identified.

This analysis presents the results obtained from training a Late Fusion Over Convolutional Neural Networks model for predicting heart disease. In this analysis, we will examine the key metrics presented in Tables 2 and 3 to gain insight into the model's effectiveness.

Table 2. Training evaluation metrics

Table 3. Validation evaluation metrics

After adding improved regularization, attention-based fusion, and multimodal data augmentation, the validation accuracy was significantly improved to 98.14%, with an extensive enhancement in generalization capacity. The loss in validation was significantly lower, and recall and precision increased, indicating that the model has become more effective at detecting positive cases and reducing FNs.

The model has successfully achieved a reduction in the discrepancy between its projected values and the true labels for the training data, as indicated by the notably low training loss value of 0.0058. Typically, a low loss value signifies successful convergence in the training process, which is a positive outcome. The training accuracy of 99.85% is notably high. This demonstrates that the model is attaining a substantial proportion of accurate categorizations on the training dataset, suggesting its ability to generate precise forecasts. However, the presence of very high precision might also give rise to the problem of overfitting. The graphs depicted in Figures 5 and 6 illustrate the performance of training and validation accuracy, as well as loss.

image013.png

image014.png

Figure 6. Training and validation loss performance

The sensitivity of 98.54% indicates the model's proficiency in correctly identifying positive cases among all real positive examples in the training set. To mitigate the occurrence of FNs in medical applications, it is imperative to prioritize good recall. The model's precision is 98.65%, representing the proportion of accurately predicted positive instances among all cases predicted as positive by the model. In medical diagnostics, there is a preference for high precision due to its association with a low rate of false-positive results. The AUC score of 0.9982 suggests that the model exhibits a strong discriminatory capacity in distinguishing between positive and negative situations, as it approaches the upper limit of 1.0. A high AUC indicates a significant level of discriminatory capability.

As the model demonstrates its ability to adapt learned patterns to new data, the validation loss is expected to exceed the training loss, and a value of 0.5538 is observed. However, it is of the utmost importance to monitor how our model performs on the validation dataset for any signs of overfitting, especially if the validation loss begins to diverge. Although there is unobservable data, the validation accuracy of 95.62% remains remarkably high, indicating that it performs well. One interesting incident of a discrepancy in accuracy level between the training and validation datasets may warrant further investigation. Additionally, it is worth noting that Figures 7–9 pertain to the evaluation of training recall, precision, and AUC performance.

image015.png

image016.png

image017.png

Figure 9. Training area under the curve (AUC) performance

The validation recall has a smaller value, 44.32%, than its training recall. This observation suggests that the model exhibits a lower accuracy in correctly predicting affirmative instances in the validation set compared to the training set. Sensitivity may be increased to a greater extent. The model achieves a validation precision of 60.38%, indicating its accuracy in classifying a substantial number of corresponding positive examples from the entire set that were predicted as positive in the validation set. Although the AUC of the validation sets is lower than that of the training sets, it still demonstrates excellent discriminative performance. A reduction in the area under the ROC curve (AUC) during the validation process may indicate that the model's testing performance differs among various subsets of data. The model performs remarkably well on the training set, returning high values of accuracy, recall, precision, and AUC. Furthermore, Figures 10–12 refer to measuring accuracy, completeness, and overall performance of recall, precision, and AUC.

image018.png

Figure 10. Confusion matrix for validation data

A confusion matrix displaying varying values across different classes is shown in Figure 10. Additionally, the values are significantly low, indicating poor model performance. There are two main reasons for this. Firstly, the model suffers from an imbalanced distribution of labels, which negatively impacts its performance. Secondly, the model achieved precise results during training; however, it failed to generalize when tested on validation data. This implies insufficient generalization power of the model. Additionally, the problem is exacerbated by the presence of a large dataset characterized by an imbalanced distribution of labels.

Although satisfactory, the results from the validation set suggest a possible decline in performance compared to that of the training set. It highlights the need for vigilance against overfitting [40]. The model may need to be tuned to improve its ability to accurately identify TP cases, as evidenced by a comparatively lower recall in the validation set. On the training set, the model yields satisfactory results. However, there is still room for improvement, especially in ensuring robust generalization to new data. Further research, including hyperparameter optimization and the implementation of thorough measures to assess the model’s performance, may help improve the heart disease model's predictive ability.

5.2 Baseline models comparison

The proposed Late Fusion CNN, when analyzed in equal data split and cross-validation conditions, is consistently found to be more effective than all baseline models in all essential metrics. The model improved its accuracy to 99.86%, indicating the capability to correctly predict almost all cases in the PhysioNet 2020 dataset. The accuracy and recall rates of 98.6% and 98.5%, respectively, indicate that the model demonstrates strong performance in reducing both FPs and FNs, which is crucial in a clinical environment. On the same note, the AUC of 0.998 indicates that the tool is highly effective at distinguishing between positive and negative cases of heart disease. Figure 11 shows the baseline models' performance comparison.

image019.png

Figure 11. Baseline models performance comparison

The late fusion method offers modality-specific learning of features, compared to more traditional DL models, including ResNet and BiLSTM, which enhances the model's understanding of complex, multi-lead ECG signals. Although ResNet is very accurate (98.63%), it is slightly less accurate, which means that it can overlook weak positive cases. Table 4 shows the baseline model performance.

Table 4. Baseline models performance

Random Forest, Spectro-Temporal ResNet, and XGBoost demonstrate relatively smaller metrics, which represent constraints in dealing with temporal and spatial patterns in the 12-lead ECG signals. In general, it is notable that the Late Fusion CNN not only exhibits better performance, as measured by all numeric evaluation criteria, but also offers a more balanced trade-off between precision and recall, which is a shortcoming of past methods. This makes it sensitive and specific to diagnosis, and as a result, offers a fair and reproducible comparative framework with the same dataset splits and evaluation protocols.

5.3 Statistical testing

To address the issue of insufficient statistical significance testing, we conducted intensive paired statistical tests to confirm the performance gains of the proposed Late Fusion CNN model. In particular, we computed p-values as well as q-values on the key measures of accuracy, precision, recall, and AUC via paired t-tests and the multiple comparison correction method of the Benjamini-Hochberg procedure. This ensures that the identified improvements are not coincidental and provides a robust quantitative analysis of significance. Figure 12 depicts the bar chart related to statistical testing performance.

image020.png

Figure 12. Statistical testing performance

The findings indicate that the performance of the proposed model is statistically significant. To ensure accuracy, precision, recall, and AUC, the calculated p-values were all less than 0.01, and the corresponding q-values were less than 0.05, indicating that the improvements were reliable despite the multiple comparisons. This goes a long way in justifying the argument that the additions brought out by the late fusion architecture and attention mechanisms have quantifiable and replicable predictive performance impact. In general, these statistical results support the strength of the suggested method.

The high p- and q-values suggest that the high-performance results found in the validation and training measures are not probably because of chance, and this evidence is rather strong regarding the technical effectiveness and the possible clinical applicability of the model. These reviews affirm the reliability of the obtained findings and support the arguments about the benefits of the suggested architecture. Table 5 shows the statistical significance evaluation metrics.

Table 5. Statistical significance evaluation metrics

The above p-values were determined with paired t-tests between repeated validations, and q-values were determined with the Benjamini-Hochberg procedure that adjusts for multiple comparisons. Each of the metrics proves statistically significant, which proves that the improvements of the suggested method are strong and are not likely due to random luck.

5.4 Ablation study

In order to deal with the issue of the personal contributions of the proposed model components, a thorough ablation study was developed. This was done to determine how the main architectural components, such as the attention mechanism, late fusion strategy, and the dual-stream 6-lead configuration, affect the general performance. We evaluated the influence of these components on the main evaluation metrics by introducing Accuracy, Precision, Recall, and AUC by removing or modifying them in a systematic way. This method enables one to better understand what factors give considerable gains when predicting heart diseases. The initial experiment on ablation entailed that the attention mechanism of the model was eliminated whilst keeping the dual-stream late fusion architecture. Table 6 shows the ablation study performance metrics.

It can be seen that the results indicate a significant reduction in recall and AUC, which means that the attention mechanism is very important in the dynamic weighting of the significance of information from various ECG leads. This establishes the need to capture minor patterns that can be predictive of positive cases of heart diseases, especially by minority classes in an unequal dataset. The late fusion layer was then substituted by a standard single-stream CNN that includes all 12 leads as an input. This arrangement performed poorly in all measures, particularly Precision and AUC. The decline justifies the fact that the late fusion approach is effective in the preservation of modality-specific features and improves joint-representation learning. Figure 13 shows the Ablation Study Performance Metrics.

Table 6. Ablation study performance metrics

image021.png

Figure 13. Ablation study performance metrics

Likewise, performance decreased with a single stream of 12-lead network compared to two streams of 6-lead, indicating the advantage of stream-specific learning to obtain lead-specific physiological underscores. Lastly, combining all the elements, such as attention, late fusion, and dual-stream architecture, obtained the best performance, which proves that each of the mentioned elements makes a significant contribution to the predictive abilities of the model. The summative findings of this ablation experiment are presented in the table below. The results confirm the methodological options and support the originality and success of the suggested architecture.

5.5 China Physiological Signal Challenge 2018 electrocardiogram dataset evaluation

In order to test the generalizability of the proposed Late Fusion CNN model, external validation was performed on the CPSC 2018 ECG Dataset, comprising 6,877 12-lead ECG recordings of patients across several clinical centres. The dataset covers a wide set of patients with different ages, genders, and cardiovascular issues. Pre-processing was done as with the PhysioNet 2020 dataset, such as normalization and division of the recordings into streams per lead, meaning that it is consistent with the original model configuration. The distribution of classes is moderate in nature, and this was dealt with through class weighting as was done in the training procedure.

The results of the external validation indicate that the proposed model does not weaken its performance when it works on a separate dataset. A model prediction accuracy of 97.24% means that the model is predictive of heart disease beyond the original PhysioNet 2020 dataset. The loss value of 0.2135 also supports the fact that the model is able to generalize effectively because it is quite low even though the demographics of the patients and the conditions of recording are different between the two datasets. The model has precision and recall scores of 94.21 and 91.87, respectively, which means the model still balances the trade-off between the reduction of FPs and the correct identification of positive cases. Although the recall is a bit lower than the precision, it is still a significant sensitivity to positive cases of ECG recordings in real-world heterogeneous settings. Figure 14 shows the CPSC 2018 Dataset Evaluation Metrics.

image022.png

Figure 14. China Physiological Signal Challenge 2018 (CPSC 2018) dataset evaluation metrics

The efficacy of the dual-stream and attention-based fusion approach in identifying various ECG patterns when the population is varied is also demonstrated by this performance. The value of 0.9632 AUC affirms the high level of discriminative ability to discriminate between positive and negative cases of heart disease in the external data. The minor decrease in the AUC relative to the training set can be attributed to random effects caused by changes in the data distribution; however, it still points to the validity of the model and its capacity to make predictions on unknown data. Also, the class-balancing approach through weighting is efficient in balancing the influence of the class imbalance in the CPSC dataset. Table 7 shows the CSPC dataset evaluation metrics.

Table 7. China Physiological Signal Challenge (CPSC) dataset evaluation metrics

On the whole, the external validation supports empirical data that the proposed Late Fusion CNN model is not specific to the PhysioNet 2020 dataset and can be used effectively on related independent ECG datasets. This supports the argument of clinical applicability and generalizability, whereby the model can be incorporated in larger clinical practices to predict heart diseases. Future research can consist of more multi-centre data and real-time ECGs to further confirm and improve the model, working on different groups of patients.

The research presents an original architecture for the late fusion of heart disease prediction, employing CNNs and adapted for the PhysioNet 2020 Challenge dataset. The model integrates modality-specific streams, a fusion-attention mechanism, and dataset-adapted target optimization schemes. This method succeeds in capturing intricate patterns associated with heart disease by compartmentalizing information from several ECG leads.

The model demonstrates high technical ability, achieving 99.85% accuracy, 98.54% recall, and 98.65% precision during training, which enables it to make precise and reliable predictions. It also achieves an excellent score of 95.62% when evaluated in the validation set. Nevertheless, a distinct observation of reduced recall accuracy on the validation set, where recall decreased by 44.32%, suggests difficulties in generalizing the model to newly unseen sets. The values of both training and validation sets, however, are biased towards precision, which indicates fewer FPs and a reduction in unnecessary medical tests for healthy individuals.

The high discriminatory power of the model, as evidenced by its impressive AUC values—especially 0.9982 for the training set—suggests strong performance. Nevertheless, the difference in the AUC between the training (0.9982) and validation (0.6058) sets highlights potential issues with overfitting and generalization.

Further research should propose solutions to these problems, especially by addressing overfitting and developing the model’s competency for various patient populations and clinical environments. The study highlights the need for customizing models for each dataset and emphasizes the clinical relevance of the model in real-world healthcare settings. In addition to technological innovations, the research calls for collaborations from other domains and emphasizes the ethical aspects to make the model more applicable. This work helps to further develop predictive modeling for heart disease and may have implications for patient outcomes and healthcare.