PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Md Shahriar Sajid

arXiv:2608.24771v1cs.CVcs.AI

TL;DR

Stroke is a major cause of mortality and disability, motivating earlier prediction. This study evaluates supervised learning models for stroke classification, with Stacking Classifier achieving 99.52% accuracy and ensemble learning emerging as the standout approach.

  • Problem

    Stroke is a leading cause of mortality and disability, underscoring the importance of early prediction.

  • Method

    The study evaluates supervised machine learning models for stroke prediction, including TabNet's sequential attention mechanism.

  • Results

    99.52% accuracy was achieved by the Stacking Classifier.

  • Takeaways & Limitations

    Ensemble learning emerged as the standout factor distinguishing this research from previous works.

  • Takeaways & Limitations

    The numerical dataset limits prediction compared with using brain images such as MRI or CT scans.

Abstract

from arXiv · show

Brain stroke, known for its high mortality and incidence rates, poses significant health risks and requires rapid intervention for survival. Early diagnosis and preventive measures can greatly reduce life loss and disabilities. Recent advancements in deep learning have led to novel computer-aided diagnostic techniques for early stroke detection. This study proposes an intelligent system that predicts potential strokes using eleven features, evaluated through seven supervised machine learning algorithms. The process includes a literature review, dataset visualization, data preprocessing, and model evaluation. Ensemble methods like Random Forest, Stacking Classifier, and Bagging Classifier achieved high accuracies of 99.52%, while Decision Tree reached 98.24%. Other models, including KNN and TabNet, demonstrated reliable performance, achieving accuracies of 96.73% and 96.49%, respectively. The custom feedforward model achieved 94.91%, while SVC and logistic regression had lower accuracies at 88.06% and 77.03%. The results highlight the effectiveness of ensemble methods in stroke classification.

CERTIFICATE … 1.3 Literature review

The paper motivates early stroke prediction through the substantial mortality and disability burden of stroke and reviews machine-learning and CNN-based approaches for classification. Prior studies report performance from 82% to 99.99%, while noting limitations involving clinical adequacy, dataset bias, and generalizability.

  • CERTIFICATE: The thesis is presented as an original work conducted under direct supervision and corrected according to external guidance, with satisfactory quality.
  • 1.1 Overview: The chapter covers background and motivation, prior authors’ work, the research gap, objectives, thesis planning, estimation, and outline, with keywords including machine learning and ROC Curve.
  • 1.2 Background and Motivation: Strokes are the second leading cause of mortality and third leading cause of disability worldwide, while early prediction and identification are presented as requiring an expert system.
  • 1.2 Background and Motivation: 75% of stroke-related deaths and 81% of stroke-related disability worldwide occur in developing countries, underscoring the motivation for machine-learning systems that uncover patterns in patient datasets.
  • 1.3.2 Analyzing E. Dritsas et al. Work: 82% accuracy was achieved by Naive Bayes in E. Dritsas et al.’s study, but the reported accuracy might not be adequate for real-world clinical decision-making.
  • 1.3.3 Analyzing Bhagyashree Rajendra Gaidhani et al. Work: CNN-based studies reported classification accuracy of 96-97% and segmentation accuracy of 85-87%, while another Random Forest study achieved 96% accuracy on the Stroke Prediction dataset.

1.4 Research Gap · 1.5 Objectives of the Research · 1.6 Vision

The paper identifies dataset limitations affecting complex-model training, generalization, and stroke prediction, then proposes using public data and CNN-based methods for early prediction and comparative evaluation. It also envisions CNNs continuing to advance computer-vision applications through sophisticated feature extraction and analysis.

  • 1.4 Research Gap: The Kaggle stroke-prediction dataset is presented as a starting point, but further research is needed to address its limitations.
  • 1.4 Research Gap: Dataset power limitations may hinder training exceptionally complex deep-learning models and their ability to capture strong patterns.
  • 1.4 Research Gap: Models may perform well on accurate datasets yet generalize poorly to real-world scenarios with many affected populations.
  • 1.4 Research Gap: A higher proportion of healthy records than trauma-related data may create imbalance and limit the paradigm’s analytical power.
  • 1.5 Objectives of the Research: The proposed CNN framework uses open data to predict stroke events, motivated by the importance of early injury classification and treatment.
  • 1.5 Objectives of the Research: The objectives include collecting valid public data and developing a sustainable algorithm to predict early occurrences of brain strokes.
  • 1.5 Objectives of the Research: The study plans to analyze data with proposed methods, feed models, calculate accuracy, and compare prediction performance using different evaluation matrices.
  • 1.6 Vision: CNNs extract significant features from photographs and motion pictures and support tasks including image segmentation, object recognition, and facial applications.

1.7 Application Scenarios … 1.10 Thesis Plan

The paper outlines CNN applications, performance challenges, thesis contributions, and a staged plan covering data collection, processing, modelling, and thesis completion. It emphasizes overfitting mitigation, feature reduction, algorithm comparison, and use of a public stroke-prediction dataset.

  • 1.7 Application Scenarios: CNNs support healthcare diagnosis and treatment planning, mobile facial recognition and image segmentation, and financial fraud detection.These applications reflect CNNs’ ability to extract features and learn patterns from visual data.
  • 1.8 Challenges: CNNs require large labeled datasets, while limited datasets increase susceptibility to overfitting.Data collection and labeling can be labor-intensive, time-consuming, and costly.
  • 1.8 Challenges: Data augmentation, regularization, and dropout are identified as techniques for mitigating CNN overfitting.Examples include rotating, scaling, or cropping images, penalizing complex models, and randomly ignoring features during training.
  • 1.9 Contributions: The study proposes an algorithm and compares its performance with various machine learning algorithms.This comparison forms one of the thesis’s stated fundamental contributions.
  • 1.9 Contributions: The thesis uses the public “Stroke-prediction-dataset” and reduces features using feature importance scores.The dataset contains information collected from hospitals worldwide.
  • 1.10.1 Work plan: The early stroke-detection workflow comprises data collection, data processing, and modelling.Data comes primarily from public resources, processing identifies stroke-related patterns, and modelling produces and tests predictive models.
  • 1.10.2 Timeline Diagram: The thesis timeline progresses from dataset and programming study through literature review, dataset selection, coding, preprocessing, supervisor review, drafting, correction, and final submission.The literature review covers numerical and image stroke-detection datasets before focusing on the selected stroke prediction dataset.

1.11 Thesis Estimation … 3.2 Implementation

The thesis demonstrates a Kaggle-based stroke-prediction workflow spanning dataset selection, preprocessing, imbalance handling, model design, and implementation. It also situates the study against prior work and outlines subsequent evaluation, impact, engineering, and future-work chapters.

  • 1.11 Thesis Estimation: The research used personal computing resources and Kaggle’s free tier, demonstrating valuable results with readily available tools despite limited access to more powerful resources.Kaggle provided public data used for training and evaluating the explored machine learning models.
  • 1.12 Thesis Outline: The thesis outline covers methodology and literature review, design and implementation, results comparison, social and environmental influence, complex engineering activities, and conclusion with future work.The outlined implementation chapter addresses dataset selection, preprocessing, model selection, and model implementation.
  • 2.1 Overview / 2.2 Case study: The study selected T. Tazin et al.’s research as its case study because it used the same dataset and similar machine learning models.The reference study used voting classifiers, decision trees, random forests, and logistic regression on the publicly available Stroke Prediction dataset.
  • 2.2 Case study / 2.3 Detailed Methodology / 2.4 Justification of the study: 96% accuracy was achieved by the reference Random Forest model, which outperformed the other models and motivated the present study’s use of ensemble techniques.The present study aimed to preprocess data more efficiently and apply bagging and stacking classifiers with voting-based base-model selection.
  • 3.2.7 Machine learning architecture: CNN / 3.2.8 Model Selection / 3.2.10 Logistic Regression / 3.2.11 Support Vector Machine / 3.2.12 Decision Tree Classifier / 3.2.13 Random Forest / 3.2.14 Bagging Classifier / 3.2.16 TabNet Classifier: The implementation introduces CNN architecture concepts alongside logistic regression, SVM, decision trees, random forests, bagging, and TabNet-based model selection.CNNs extract hierarchical patterns through convolution, activation, pooling, and fully connected layers, while TabNet uses a sequential attention mechanism.

3.3 Conclusion

The conclusion summarizes the dataset characteristics, visualization, analysis, preprocessing, sampling techniques, and model overviews presented in the chapter.

  • The chapter reviewed the dataset characteristics and their description.
  • It covered data visualization, analysis, and preprocessing.
  • It also described sampling techniques and the models’ characteristics.

4.1 Overview

This section evaluates the thesis work’s results and successes by analyzing the findings with supporting explanations, tables, and graphics.

  • 4.1 Overview: The chapter evaluates the thesis work’s results and successes.The passage identifies evaluation as crucial.
  • 4.1 Overview: The analysis presents findings with full explanations.
  • 4.1 Overview: Tables and graphics support the chapter’s analysis.

4.2 Evaluation Metrics

The section defines accuracy, precision, recall, F1-score, and confusion matrices as evaluation criteria for the proposed model. It also describes ROC curves and AUC-ROC for assessing binary classification performance and class discrimination.

  • Accuracy: Accuracy measures the proportion of correctly classified samples relative to the total number of samples.
  • Precision: Precision measures how many samples predicted as positive are truly positive, helping assess false-alarm risk.
  • Recall: Recall, also called sensitivity or true positive rate, measures how many actual positive cases the model correctly identifies.Recall is particularly important when false negatives are a major concern, such as in medical diagnosis.
  • F1-score: F1-score combines precision and recall through their harmonic mean, balancing performance when class distributions are uneven.It does not account for true negatives.
  • Confusion Matrix and ROC Curve: A confusion matrix records true positives, true negatives, false positives, and false negatives, while ROC curves plot true positive rate against false positive rate.AUC-ROC summarizes the model’s general ability to distinguish positive and negative classes.

4.3 Outcome of methodology

The methodology evaluated seven machine learning classifiers using twelve selected features after preprocessing and cleaning. The dataset was split into training and testing subsets, with 5-fold cross-validation used to determine accuracy.

  • Model training: Seven classifiers—Logistic Regression, KNN, SVM, Decision Tree, Tab-Net, Stacking, and Bagging—were trained.The models used twelve selected features from the feature-selection task.
  • Data preparation: The original data were preprocessed and cleaned before being separated into training and testing sets.
  • Data preparation: 80% of the data were used for training and 20% for testing.
  • Model evaluation: 5-fold cross-validation was performed to determine model accuracy.The resulting model metrics were summarized in the employed-models table.

4.4 Details on each model’s result

Section 4.4 presents visual details of the employed models’ results through metric tables, charts, training-versus-testing curves, and confusion matrices.

  • Visual metric analysis: The tables and chart types display different metrics of the employed models for better insight.The section uses visual presentations to summarize model metrics.
  • Training vs Testing graphs: Training-versus-testing graphs show the employed models’ training and testing curves.These curves are presented in Figure 16.
  • Confusion matrices: Figure 17 presents confusion matrices for the employed models.The figure provides a visual confusion-matrix view of model results.

4.5 Comparison

The comparison shows ensemble methods achieving the strongest classification performance, led by Stacking Classifier at 99.81% accuracy. Decision Tree also performed robustly, while lower-accuracy classifiers may remain useful when interpretability or computational efficiency is prioritized.

  • Ensemble methods: 99.81% accuracy was achieved by the Stacking Classifier, making it the top performer among the compared models.The results attribute this performance to ensemble methods’ collective decision-making across multiple models.
  • Ensemble methods: 99.52% accuracy was achieved by Random Forest, reinforcing the effectiveness of ensemble methods for classification.Random Forest was identified as following closely behind the Stacking Classifier.
  • Other classifiers: 98.30% accuracy was achieved by the Decision Tree model, demonstrating robust performance despite its simple classification approach.The passage characterizes Decision Tree as simple yet powerful.
  • Other classifiers: KNN, TabNet Classifier, and the Custom Feedforward Model achieved slightly lower accuracies but still demonstrated respectable performance.Their results may remain useful in certain contexts or datasets.
  • Classifier selection: Logistic Regression and SVC showed relatively lower accuracies, suggesting limitations in capturing complex dataset relationships, though they may suit interpretability or computational-efficiency priorities.The comparison presents classifier selection as dependent on dataset and task characteristics, including the problem domain’s requirements.

4.6 State of art · 4.7 Conclusion: · 5.1 Introduction

The study’s ensemble methodology achieved 99.52% accuracy on the stroke-prediction dataset, while the conclusion emphasizes ensemble learning’s potential to improve robustness, generalizability, and reliability. The introduction broadens the discussion to the societal and environmental impacts of CNN-based stroke prediction.

  • 4.6 State of art: 99.52% accuracy was achieved by the ensemble methodology, tying stacking classifier and random forest classifier.This result is presented as a comparison with prior work using similar methodologies and the same dataset.
  • 4.7 Conclusion:: Ensemble learning combines multiple CNN models to address overfitting and sensitivity to training-data variations.The conclusion identifies these as advantages over relying on a single CNN model.
  • 4.7 Conclusion:: Ensemble approaches can improve accuracy, generalizability, and robustness in stroke prediction compared with a single CNN model.The conclusion contrasts ensemble learning with reliance on one CNN model.
  • 4.7 Conclusion:: Ensemble methods can potentially mitigate the impact of biases present in training datasets.The passage frames bias mitigation as a potential benefit of ensemble methods.
  • 4.7 Conclusion:: Ensemble learning approaches are presented as promising for developing more reliable and trustworthy models that can benefit patient care.The conclusion describes this promise as significant as stroke prediction continues to evolve.
  • 5.1 Introduction: CNN-based stroke prediction requires consideration of broader societal and environmental impacts beyond medical image analysis.The introduction frames these impacts as crucial aspects of a comprehensive approach.
  • 5.1 Introduction: The chapter examines potential benefits and challenges associated with CNN-based stroke prediction models.This scope connects the models’ technical use with their broader impacts.

5.2 Societal Impact … 6.6 Conclusion

The paper presents CNN-based stroke prediction as potentially beneficial for healthcare while emphasizing equity, safety, transparency, cultural acceptance, and environmental sustainability. It also frames dataset use and engineering practices as essential to developing reliable, generalizable models for practical stroke detection.

  • 5.2.1 Financial and Health Influences: Early and accurate prediction may accelerate intervention, reduce long-term disabilities, improve prognoses, lessen healthcare burdens, and support targeted preventive measures.The paper also identifies affordability, accessibility, and dataset bias as concerns requiring attention.
  • 5.2.2 Safety, Legal, and Cultural Issues: Misdiagnosis risks, limited explainability, and cultural differences may affect the safety, trust, and acceptance of CNN-based stroke prediction tools.False positives may cause unnecessary procedures, whereas false negatives may delay critical interventions; XAI can improve transparency.
  • 5.3 Impact on Environment: CNN development and deployment can increase energy consumption through intensive training and create environmental costs through large-scale medical-image storage.The paper recommends energy-efficient training and sustainable storage solutions.
  • 5.4 Sustainability Issues: Sustainable deployment requires optimizing computational resources, using smaller and diverse datasets, and maintaining models through efficient retraining and updates.These practices aim to reduce storage and redeployment costs while preserving effectiveness and accuracy.
  • 5.5 Conclusion: A holistic approach addressing bias, transparency, and sustainable practices is needed to make CNN-based stroke prediction a beneficial and responsible healthcare tool.The conclusion connects the technology’s healthcare potential with its societal and environmental responsibilities.
  • 6.1 Introduction: The chapter examines complex engineering problems and related activities involving multiple variables, uncertainties, interdependencies, and constraints.It positions engineering managers as needing appropriate abilities and resources to develop practical solutions.
  • 6.5 Leveraging the stroke-prediction-dataset: The stroke-prediction-dataset supports CNN development through data exploration, bias analysis, class-distribution assessment, stratified splitting, and balanced evaluation.Sharing results, limitations, and insights can benefit the broader stroke-detection research community.
  • 6.6 Conclusion: Preprocessing, transfer learning, appropriate evaluation metrics, explainable AI, continuous exploration, and effective dataset use can produce reliable, generalizable CNN models for practical stroke detection.The paper identifies robust model development as challenging but potentially capable of significantly improving stroke detection in practice.

7.1 Introduction … Footnote

The thesis explored methodologies for knowledge transfer, finding ensemble learning to be a distinguishing factor while identifying data limitations and future directions. Its strongest models included Random Forest and Stacking Classifier, each achieving 99.52% accuracy.

  • 7.1 Introduction: The chapter introduces its structure: a synopsis of findings and implications, limitations, and future projects.It follows a preceding chapter reviewing the experiments.
  • 7.2 Synopsis of the Thesis: The thesis examined innovative methodologies for facilitating knowledge transfer across diverse domains, with most models producing satisfactory results despite some data-handling difficulties.The study concludes that ensemble learning was the key factor distinguishing the thesis from previous work.
  • 7.2 Synopsis of the Thesis: Ensemble learning distinguished the thesis by aggregating predictions from multiple base classifiers and leveraging their complementary strengths.Random Forest, Stacking Classifier, and Bagging Classifier were identified as pivotal ensemble methods.
  • 7.3 Discussion of Key Discoveries: 99.52% accuracy was achieved by both the Random Forest and Stacking Classifier models, while Bagging Classifier reached 99.27%.These models delivered the highest reported accuracies in the discussion of key discoveries.
  • 7.3 Discussion of Key Discoveries: 98.24% accuracy was reported for Decision Tree, followed by KNN at 96.73% and TabNet Classifier at 96.49%.Decision Tree was described as simple yet effective, KNN as proximity-based, and TabNet as using sequential decision making.
  • 7.4 Limitations: The numerical dataset limited stroke prediction compared with image data from MRI or CT scans, which the thesis listed for future progress.The limitation specifically concerns predicting stroke occurrence from numerical rather than brain-imaging data.
  • 7.5 Future Endeavors: Future endeavors include a Bangladesh stroke-patient repository, a mobile/web-based stroke prediction and suggestions application, and an image-based dataset.These directions are presented as two areas of concentration with three listed developments.
  • 7.6 Conclusions: The conclusion identifies ensemble learning as the key factor in the thesis’s success, enhancing predictive accuracy and robustness through collective model strengths.It also states that most models yielded satisfactory results while positioning ensemble learning as promising for future interdisciplinary research.
Loading 2608.24771v1…