Ensemble of Convolutional Neural Networks for StrokePrediction: Towards Improved Diagnostic Accuracy

1.3.1 Analyzing Oznur Ozaltin et al. Work

• OzNet extracted features for machine-learning classification of brain CT scans; its mRMR-Naïve Bayes combination achieved 98.42% accuracy and 0.99 AUC.The review notes that the paper did not explicitly describe how it addressed potential training-data bias.Source: §1.3.1 · p. 15

1.3.3 Analyzing Bhagyashree Rajendra Gaidhani et al. Work

• LeNet classified MRI images as normal or abnormal, while SegNet performed semantic segmentation of abnormal regions; classification models reached 96-97% accuracy and segmentation models achieved 85-87%.Source: §1.3.3 · pp. 15–16

1.3.4 Analyzing T. Tazin et al. Work

• Random Forest outperformed the other models with 96% accuracy in Tazin et al.’s study using the open-access Stroke Prediction dataset.The study also trained Logistic Regression, Decision Tree, and Voting Classifier models.Source: §1.3.4 · p. 16

1.3.6 Analyzing N. Biswas et al. Work

• In the reported classifier results, optimization yielded the highest accuracy for SVM at 99.99%, while Random Forest ranked second with 99.87% accuracy.Before data balancing, classifiers showed over 90% accuracy; four exceeded 96% after oversampling, and the most accurate model led to user-friendly mobile and web applications.Source: §1.3.6 · p. 16

1.3.7 Analyzing B. Akter et al. Work

• Random Forest recorded the highest accuracy at 95.30% in Akter et al.’s comparison of RF, SVM, and Decision Tree classifiers.Their process included data collection, preprocessing, transformation, and standardization during training and testing.Source: §1.3.7 · pp. 16–17

1.3.8 Analyzing M. U. Emon et al. Work

• Emon et al.’s weighted-voting classifier achieved 97% accuracy and outperformed its base classifiers.The study used ten classifiers on Kaggle public data and combined their outputs using weighted voting.Source: §1.3.8 · p. 17

1.3.9 Analyzing A. Srinivas et al. Work

• A soft-voting ensemble of Random Forest, Extremely Randomized Trees, and Histogram-Based Gradient Boosting reached up to 96.88% accuracy on a UCI stroke dataset.The ensemble combines classifier predictions through an average probability-weighted sum.Source: §1.3.9 · p. 17
• The subsection calls for incorporating swarming optimization with intelligence.Source: §1.3.9 · p. 17

1.3.10 Analyzing T. I. Shoily et al. Work

• Limited information about the dataset and its biases restricts generalizability, despite the reported strong algorithm performance on that data.The review says Shoily et al. examined Random Forest, J48 decision trees, k-Nearest Neighbors, and Naive Bayes.Source: §1.3.10 · pp. 17–18
• 99.8% accuracy was achieved by Random Forest, the best of four classifiers in Shoily et al..The review also describes their patient dataset as unidentified and their classification procedure as using a publicly available Kaggle dataset.Source: §1.3.10 · pp. 17–18

1.4 Research Gap

• The research gap includes dataset limitations that may hinder complex-model training, introduce bias in sequential or linear estimates, and weaken real-world generalization.Source: §1.4 · p. 18
• The dataset has a higher proportion of healthy than trauma-related records, which may limit analytical power and motivates investigating mitigation strategies.Source: §1.4 · p. 18

1.5 Objectives of the Research

• The objectives include collecting valid public data and using an open-data CNN framework to predict stroke events.Source: §1.5 · p. 18
• The study aims to develop a sustainable algorithm for predicting early brain-stroke occurrences and analyze the data with proposed methods and models.Source: §1.5 · p. 19
• The objectives call for calculating accuracy and comparing prediction performance across methods using different evaluation metrics.Source: §1.5 · p. 19

1.6 Vision

• The vision describes CNNs extracting significant features from photographs and motion pictures for tasks such as segmentation and object recognition.Source: §1.6 · p. 19

1.7 Application Scenarios

• CNNs are described as tools for analyzing medical scans to assist diagnosis and treatment planning, and for smartphone facial recognition and image segmentation [6].The healthcare examples include detecting mammogram anomalies for early breast cancer detection.Source: §1.7 · p. 19

1.8 Challenges

• CNNs require large labeled datasets, and limited datasets make overfitting more pronounced as models learn noise or unrepresentative details.Collecting and labeling enough examples can be labor-intensive, time-consuming, and costly.Source: §1.8 · p. 20
• Data augmentation using operations such as rotation, scaling, and cropping, along with regularization and dropout, is presented as a way to mitigate overfitting.Source: §1.8 · p. 20

1.9 Contributions

• The thesis uses a public “Stroke-prediction-dataset” collected from hospitals worldwide and reduces its features using feature-importance scores.Source: §1.9 · p. 21
• The thesis summarizes related machine-learning studies and proposes an algorithm whose performance is compared with other machine-learning algorithms.Source: §1.9 · p. 21

1.10.1 Work plan

• The early stroke-detection work plan has three stages: gathering data from public resources, processing it to identify patterns, and developing predictive models.Source: §1.10.1 · p. 21

1.10.2 Timeline Diagram

• The thesis timeline moves from learning about datasets and libraries and reviewing literature to selecting the stroke-prediction dataset, planning the work, and building code.Source: §1.10.2 · p. 22
• The timeline continues with data cleaning and preprocessing, code review by the supervisor, thesis drafting, revision, and submission of the finalized thesis.Source: §1.10.2 · pp. 22–23

1.11.1 Overall Estimation

• The thesis uses personal computing resources and Kaggle’s free tier, which provides public data for training and evaluation despite limiting access to more powerful computing resources.The authors present this setup as demonstrating the feasibility of obtaining valuable results with readily available tools.Source: §1.11.1 · p. 23

1.12 Thesis Outline

• The outline assigns Chapter 2 to methodology and literature review, Chapter 3 to design and implementation, Chapter 4 to results and overview, and Chapter 8 to conclusion and future work.Source: §1.12 · pp. 23–24

2.2 Case study

• 96% accuracy was reported for Random Forest, which outperformed the other models in the selected prior study.The study used the publicly available Stroke Prediction dataset and compared voting classifiers, decision trees, random forests, and logistic regression.Source: §2.2 · p. 25

2.3 Detailed Methodology

• The prior study selected a Kaggle stroke-detection dataset of 5,110 instances with 12 features, handling missing values and label encoding, then used SMOTE to balance the dataset.Among the four employed models, Random Forest outperformed the other machine learning models.Source: §2.3 · pp. 25–26

2.4 Justification of the study

• The thesis proposes more efficient preprocessing and extends four-model prior work with basic classifiers and bagging and stacking ensembles, selecting base models by voting.The stated aim is a sustainable accuracy that outperforms existing studies; the prior paper’s highest reported accuracy was 96%.Source: §2.4 · pp. 26–27

3.1 Design overview

• The design and implementation chapter covers hardware, dataset selection and description, preprocessing, model selection, and model implementation; its hardware section provides remote and local configuration tables.Source: §3.1 · p. 28, Table 1 · p. 28, Table 2 · p. 28

3.2.1 Dataset selection

• The study begins with Kaggle’s public stroke-prediction-dataset, a numerical dataset with 5,110 rows and 12 columns, including physiological features and a binary stroke output (1 or 0).Source: §3.2.1 · p. 29, §3.2.2 · p. 29

3.2.2 Description of the Dataset

• The numerical dataset has 5,110 rows and 12 columns; its stroke output records 249 rows with value 1, indicating stroke risk, and 4,861 rows with value 0, indicating no identified stroke risk.Source: §3.2.1 · p. 29, §3.2.2 · p. 29

3.2.3 Description of the attributes

• The dataset’s attributes include heart disease, marital status, work and residence categories, glucose level, and BMI.The text also identifies stroke status as the output column and smoking status as an attribute.Source: §3.2.3 · p. 31
• The attribute description also identifies smoking status and stroke status, with stroke serving as the output column.Source: §3.2.3 · p. 31

3.2.4 Dataset Splitting

• The dataset is split 80:20, with 80% used for training and 20% for testing to evaluate performance on unseen data.Source: §3.2.4 · pp. 32–33

3.2.5 Dataset Preprocessing

• Preprocessing drops the non-informative id column and fills missing BMI values with the column mean.Source: §3.2.5 · p. 34

3.2.6 Handling Imbalanced Data

• The 5,110-row dataset has 249 stroke-positive and 4,861 stroke-negative records; undersampling reduces the majority class to 249 records.The resulting dataset has 249 records in each class.Source: §3.2.6 · pp. 34–35

3.2.7 Machine learning architecture: CNN

• The described CNN applies convolutional filters to produce feature maps, then uses ReLU, pooling, and a SoftMax output to classify inputs.Source: §3.2.7 · pp. 36–38

3.2.8 Model Selection

• The model-selection section states that preprocessed data is fed into various machine-learning models and lists a custom feedforward model.Source: §3.2.8 · p. 38

3.2.10 Logistic Regression

• Logistic regression models a binary outcome from one or more independent variables, using coefficients to estimate the log-transformed likelihood of the outcome.Source: §3.2.10 · pp. 38–39
• Logistic regression estimates class probabilities by mapping real-valued inputs through a sigmoid function to values between 0 and 1.The passage identifies the output as the probability of the positive class.Source: §3.2.10 · p. 39

3.2.11 Support Vector Machine

• SVM seeks a separating hyperplane that maximizes the margin between classes, with the nearest data points serving as support vectors.Those support vectors influence the hyperplane’s position and orientation.Source: §3.2.11 · pp. 40–41

Kernel Trick and Nonlinear Separation:

• For data that is not linearly separable, SVM uses kernel functions to implicitly transform features into a higher-dimensional space.The described kernels include polynomial and radial basis function (RBF) kernels.Source: Kernel Trick and Nonlinear Separation · p. 41

3.2.12 Decision Tree Classifier

• Decision trees classify data through hierarchical feature nodes and class-label leaves, with root-to-leaf paths encoding decision rules.Their hierarchical rules support interpretation and visualization of decision boundaries.Source: §3.2.12 · pp. 41–42

3.2.13 Random Forest

• Random forests combine votes from multiple decision trees trained with distinct characteristics, reducing overfitting risk and improving generalization.The method injects diversity into individual trees through the training data.Source: §3.2.13 · p. 43

3.2.14 Bagging Classifier

• Bagging Classifier trains base classifiers independently on bootstrapped subsets and combines their predictions by polling for classification or averaging for regression.The passage describes bagging as an ensemble method intended to improve model stability and accuracy.Source: §3.2.14 · p. 44
• Bagging trains base classifiers on subsets and predicts the class receiving the most votes.The task distinguishes Non-stroke and Stroke classes.Source: §3.2.14 · pp. 45–46

3.2.15 Stacking Classifier

• Stacking combines predictions from different base classifiers as inputs to a meta-classifier that produces the final prediction.For multi-classification, outputs from binary base models can be fed to the meta-classifier.Source: §3.2.15 · pp. 46–47

3.2.16 TabNet Classifier

• TabNet uses sequential attention to select and weight features during training, supporting interpretable tabular-data modeling.Its feature focus is described as providing insight into influential features while combining tree-like mapping with deep-learning performance.Source: §3.2.16 · p. 48

3.2.17 Feed Forward model

• The feed-forward model passes information forward through layers, multiplying inputs by weights, summing them, and applying a threshold for classification.The described output is 1 when the sum exceeds a zero-based threshold and -1 otherwise.Source: §3.2.17 · pp. 49–50

3.2.18 Model Implementation

• Implementation trains initialized models on preprocessed data, evaluates accuracy, precision, recall, and F1 score, and identifies the best performer.The section also describes integrating models into an application for predictions on unseen data.Source: §3.2.18 · p. 50

3.2.19 Proposed Workflow

• The proposed workflow collects patient medical histories, preprocesses data to extract relevant features, and addresses missing and imbalanced data.The workflow is presented as a process for preparing and comparing machine-learning models.Source: §3.2.19 · p. 51
• The proposed workflow proceeds from preprocessing and encoding to feeding data into machine-learning models and comparing them to select the best-performing model.Figure 14 presents the proposed research flow model.Source: §3.2.19 · p. 51, Figure 14 · p. 52
• Models are validated and tested using separate datasets to evaluate their performance.The algorithms are trained to identify patterns indicative of a stroke.Source: §3.2.19 · p. 52

3.3 Conclusion

• The chapter summarizes the dataset’s characteristics.Source: §3.3 · p. 52
• The overview covers data visualization, analysis, preprocessing, and sampling techniques.Source: §3.3 · p. 52
• The chapter also reviews the models, their characteristics, and their descriptions.Source: §3.3 · p. 52

• The chapter analyzes the thesis work’s findings and evaluates its results and successes.Source: §4.1 · p. 53
• The findings are presented with full explanations, tables, and graphics.Source: §4.1 · p. 53

4.2 Evaluation Metrics

• The proposed model is evaluated using accuracy, precision, recall, F1-score, and a confusion matrix.The section also explains ROC curves and AUC-ROC as measures of binary-classification performance.Source: §4.2 · p. 53

Accuracy

• Accuracy measures the proportion of correctly classified samples, but can conceal failures to identify a minority class in an imbalanced dataset.A model that predicts only the majority class may still appear accurate even when the minority class is more important.Source: Accuracy · p. 53
• Precision measures how many predicted positives are truly positive, making it useful when false positives are costly.The section notes that high precision is associated with fewer false alarms.Source: Accuracy · pp. 53–54
• Recall, also called sensitivity or true positive rate, measures how many actual positive cases the model identifies and matters when false negatives are consequential.The section highlights medical diagnosis, where a missed disease can have serious effects.Source: Accuracy · p. 54
• F1-score is the harmonic mean of precision and recall, balancing them for uneven class distributions, but it does not account for true negatives.It is described as useful when class counts are uneven or false positives and false negatives have different costs.Source: Accuracy · p. 54
• A confusion matrix shows true positives, true negatives, false positives, and false negatives, with an N × N structure for N classes.The section presents it as a more complete view of classification results and applicable to binary or multiclass problems.Source: Accuracy · p. 55
• A ROC curve plots true positive rate against false positive rate across classification levels, while AUC-ROC summarizes the model’s ability to distinguish positive and negative classes.The section describes perfect classification as reaching the upper-left ROC point, (0, 1).Source: Accuracy · p. 56

• Seven classifiers, including LR, KNN, SVM, Decision Tree, Tab-Net, Stacking, and Bagging, were trained using twelve selected features.Source: §4.3 · p. 56
• The cleaned and preprocessed dataset was divided into 80% for training and 20% for testing.Source: §4.3 · p. 56
• Five-fold cross-validation was used to determine accuracy, and Table 4 summarizes the employed models’ metrics.Source: §4.3 · p. 56, Table 4 · p. 57

• The section displays different metrics of the employed models using charts for better insight.Source: §4.4 · p. 58
• Figure 16 presents training-versus-testing curves for the employed models.Source: §4.4 · p. 58, Figure 16 · p. 59
• Figure 17 presents confusion matrices for the employed models.Source: Figure 17 · p. 61

4.5.1 Model performance

• 99.81% accuracy made Stacking Classifier the top performer, followed by Random Forest at 99.52%.The passage attributes ensemble performance to the collective decision-making power of multiple models.Source: §4.5.1 · p. 61
• 98.30% accuracy showed robust performance for the Decision Tree model.Source: §4.5.1 · pp. 61–62
• KNN, TabNet Classifier, and the Custom Feedforward Model achieved lower but respectable accuracies than the top-performing models.The passage notes that these classifiers may remain useful in certain contexts or datasets.Source: §4.5.1 · p. 62
• Logistic Regression and SVC had relatively lower accuracies, but may suit scenarios prioritizing interpretability or computational efficiency.The comparison suggests these models may be limited in capturing complex dataset relationships.Source: §4.5.1 · p. 62
• Random Forest, Stacking Classifier, and Bagging Classifier are described as robust choices for classification, particularly with complex datasets.The passage advises evaluating each classifier's strengths and weaknesses against the problem domain's requirements.Source: §4.5.1 · p. 62

4.6 State of art

• [11] achieved 96% accuracy with CNN ensemble learning models, while B. Akter et al. reached 95.90% using basic machine learning CNN models on the same dataset.Source: §4.6 · p. 62
• G. Sailasya et al. achieved 82% accuracy with Naïve Bayes, while I. G. Ivanov et al. reported 98.776% as their highest accuracy on the same dataset.Source: §4.6 · p. 62
• Ashrafuzzaman et al. employed logistic regression, random forest, decision tree, naïve Bayes, and support vector machine, alongside a custom proposed model.Source: §4.6 · pp. 62–63
• 99.52% accuracy was achieved by the ensemble methodology, tied between the stacking classifier and random forest classifier.Source: Table 5 · p. 63, §4.6 · p. 63

4.7 Conclusion:

• By combining multiple CNN models, ensemble techniques can address overfitting and sensitivity to training-data variation, potentially improving accuracy, generalizability, and robustness over a single CNN.Source: §4.7 · p. 63
• Ensemble methods may mitigate training-data biases and support more reliable, trustworthy stroke-prediction models that benefit patient care.Source: §4.7 · p. 63

5.1 Introduction

• The chapter considers the broader societal and environmental impacts of CNN-based stroke prediction, including its potential benefits and challenges.Source: §5.1 · p. 64

5.2.1 Financial and Health Influences

• Early, accurate stroke prediction may speed intervention, improve prognoses, reduce long-term disability and care burdens, and enable targeted prevention.The paper also flags equitable access and affordability, and warns that training-data biases could disproportionately benefit some demographics.Source: §5.2.1 · p. 64

5.2.2 Safety, Legal, and Cultural Issues

• False positives may prompt unnecessary or harmful procedures, false negatives may delay intervention, and transparency, explainability, and cultural acceptance affect trust and adoption.The paper identifies XAI techniques as a way to support transparency.Source: §5.2.2 · p. 65

5.3 Impact on Environment; 5.4 Sustainability Issues

• Training CNNs can consume substantial energy, while storing large medical-image datasets also carries environmental costs; resource optimization and smaller, diverse datasets are proposed mitigations.The sustainability discussion also recommends maintaining models with efficient updates and retraining.Source: §5.3 · p. 65, §5.4 · p. 65

5.5 Conclusion

• The paper concludes that addressing bias, transparency, and sustainable practices is necessary for CNN-based stroke prediction to be a responsible healthcare tool.It presents the approach as having the potential to benefit healthcare, subject to considering societal and environmental impacts.Source: §5.5 · p. 66

6.1 Introduction; 6.2 Definition

• The chapter introduces complex engineering problems as difficult to define, analyze, and solve because of multiple variables, uncertainties, interdependencies, and constraints.It introduces the chapter as exploring engineering problems and related activities.Source: §6.1 · p. 67, §6.2 · p. 67

6.3 Addressing Complex Engineering Problems; 6.4 Addressing Complex Engineering Activities

• Tables 6 and 7 present ranges of complex engineering problem-solving cases and engineering activities, respectively.The text introduces them as brief tabulations.Source: §6.2 · p. 67, Table 6 · p. 67, Table 7 · p. 68

6.5 Leveraging the stroke-prediction-dataset

• Effective dataset use includes exploring distributions and potential biases, stratifying balanced training, validation, and test sets, and sharing findings with the research community.The passage presents these practices as ways to maximize the stroke-prediction dataset's utility for stroke detection.Source: §6.5 · p. 69

6.6 Conclusion

• Preprocessing, transfer learning, appropriate evaluation metrics, and explainable AI can help address challenges in building robust, generalizable CNN stroke-detection models.The conclusion says that continued exploration and effective use of the stroke-prediction dataset could improve CNN-based stroke detection in practice.Source: §6.6 · p. 70

7.1 Introduction

• The chapter introduces a summary of the main findings and implications, then addresses limitations and future projects.Source: §7.1 · p. 71

7.2 Synopsis of the Thesis; 7.6 Conclusions

• The thesis emphasizes ensemble learning as a key factor, aggregating classifiers to use their complementary strengths and enhance predictive accuracy and robustness.Source: §7.2 · p. 71, §7.6 · p. 73

7.3 Discussion of Key Discoveries

• 99.52% accuracy was achieved by both Random Forest and Stacking Classifier, followed by Bagging Classifier at 99.27%.Source: §7.3 · p. 71
• 98.24% for Decision Tree, 96.73% for KNN, and 96.49% for TabNet were also reported.Source: §7.3 · p. 72
• 94.91% for the Custom Feedforward Model, 88.06% for SVC, and 77.03% for Logistic Regression completed the reported accuracy results.Source: §7.3 · p. 72

7.4 Limitations

• The work used numerical data, and identifies brain MRI, CT, or other image data as a better approach for predicting stroke occurrence.Source: §7.4 · p. 72

7.5 Future Endeavors

• Future plans include a Bangladesh stroke-patient repository, a mobile or web-based prediction and suggestions application, and an image-based dataset.Source: §7.5 · p. 72

Footnote

• The footnote thanks thesis supervisor Oishi Jyoti, teacher Md. Nahiduzzaman, and friends for their support and guidance.Source: Footnote · p. 81