INTRODUCTION
A stroke, or cerebrovascular accident, results from an interruption in blood flow to the brain caused by a blockage [ischemic stroke (IS)] or bleeding (hemorrhagic stroke) ( Karthik et al., 2020). This interruption causes harm to brain cells and may lead to different neurological impairments. Rapid medical intervention is necessary to lessen the severity of brain damage and the likelihood of permanent impairment following a stroke. IS is caused by the obstruction of an artery providing blood to the brain due to a blood clot or plaque accumulation. ISs are the predominant form, comprising about 85% of all strokes ( Sirsat et al., 2020). A hemorrhagic stroke occurs as a blood artery in the brain bursts, causing bleeding into the brain tissue around it. Conditions including excessive blood pressure, aneurysms, or arteriovenous malformations may lead to hemorrhagic strokes.
There are several facets of brain stroke (BS) patients’ disability and quality of life, including physical, cognitive, emotional, and social aspects ( Rahman et al., 2023). Rehabilitation, social support systems, and healthcare access are crucial in aiding stroke survivors in managing their disability and enhancing their quality of life. Depending on its location and severity, individuals may encounter weakness or paralysis ( Krishna et al., 2021). Strokes may influence coordination and balance, resulting in challenges with walking or daily activities. Aspiration pneumonia and other problems may develop from dysphagia, which is a typical symptom experienced by stroke survivors ( Abbasi et al., 2023). Stroke survivors frequently express being exhausted, which might hinder their capacity to perform routine tasks. Strokes may result in sensory problems, including numbness or tingling in certain regions ( Tursynova et al., 2023).
An individual’s ability to acquire new information or remember recent events may be impaired due to problems with either short-term or long-term memory following a BS. Stroke survivors may have difficulties sustaining focus and concentration, which can hinder their ability to engage in discussions or activities ( Phong et al., 2017). Deficits in executive function, including planning, problem-solving, and decision-making, might impede autonomy and everyday activities. Aphasia, a language issue, often occurs after a stroke and affects speech, comprehension, reading, and writing skills ( Li et al., 2020). Patients who have had a BS may develop mood problems, including sadness and anxiety, due to physical restrictions, lifestyle modifications, or neurological causes. Some individuals show emotional lability, which involves abrupt and unpredictable changes in mood, such as laughing or sobbing, that are not connected to the circumstances ( Li et al., 2020). Dealing with disability and reliance may cause emotions of reduced self-esteem and sense of self, especially if the stroke leads to significant lifestyle changes. Disability and communication challenges may cause stroke survivors to withdraw socially and become isolated, making it difficult for them to participate in social activities or sustain relationships.
Stroke survivors may face stigma or prejudice because of their apparent disability or cognitive limitations, impacting their social relationships and inclusion in society. Physical and cognitive limitations may restrict independence and autonomy, resulting in heightened dependency on caretakers and decreased capacity to participate in everyday activities ( Zhang et al., 2021). Individuals who have had a BS may struggle to engage in hobbies, leisure activities, or employment, resulting in a diminished sense of pleasure and satisfaction. Medical bills, rehabilitation costs, and loss of income from disability may lead to financial pressure and strain on resources. These challenges affect stroke survivors’ quality of life.
Artificial intelligence (AI) algorithms are trained on extensive datasets of medical imaging scans, including computed tomography (CT) and magnetic resonance imaging (MRI) images, to identify patterns linked to stroke, such as ischemic regions or hemorrhages ( Pereira et al., 2018). Radiologists and physicians may benefit from these algorithms’ ability to automatically evaluate new imaging scans for possible stroke symptoms, leading to a faster and more precise diagnosis ( Diker et al., 2023). Conventional approaches to manually examining medical photographs may be laborious and influenced by personal opinions. AI systems can rapidly analyze vast amounts of imaging data, offering physicians prompt diagnostic information. This effectiveness is of the utmost importance in instances of stroke, when prompt action is essential to lessen the extent of brain damage and enhance recovery.
AI algorithms embedded in electronic health records may examine patient information such as clinical notes, test findings, imaging reports, and prescription histories to provide decision-making assistance to healthcare professionals ( Diker et al., 2023). The algorithms may discover critical clinical data, detect possible stroke instances, provide suitable diagnostic procedures or therapies, and notify doctors of relevant recommendations or protocols. AI-driven clinical decision support systems seek to boost diagnostic precision, minimize medical mistakes, and optimize clinical processes. AI-driven risk prediction models use machine learning methods to categorize stroke patients according to their unique risk profiles ( Alon and Dehkharghani, 2021). The models use several patient characteristics, including age, sex, comorbidities, imaging results, and biomarkers, to predict the probability of adverse events or recurrence. Healthcare professionals may enhance patient outcomes and save healthcare costs by identifying high-risk people, implementing focused interventions, optimizing treatment procedures, and allocating resources efficiently.
AI algorithms are essential for improving clinical trial design, patient recruitment, and data analysis in stroke research ( Akter et al., 2022). Machine learning models examine various data sources such as electronic health records, medical imaging, genomics, and wearable sensor data to identify suitable patients, forecast trial results, and enhance trial procedures. AI-based methods optimize trial procedures, improve patient recruitment and retention, and enable real-time monitoring of trial progress and safety. AI technologies help speed up the creation and assessment of new treatments, which aids in the progress of stroke research and enhances patient care.
One of the leading causes of death and disability globally is a cerebrovascular accident, which is commonly referred to as a BS. The expenditures related to long-term care, rehabilitation, and medical treatments substantially load healthcare systems ( Akter et al., 2022). Timely identification and management are vital in minimizing the consequences of cerebral strokes. Early detection of stroke risk factors and timely diagnosis may expedite medical action, thereby mitigating the severity of stroke effects and enhancing patient outcomes. MRI is a commonly used diagnostic instrument for identifying brain irregularities like stroke. It offers precise visual representations of brain regions and can detect indications of ischemia or hemorrhagic strokes ( Cheon et al., 2019). This assists medical professionals in accurately diagnosing and strategizing treatment plans. Deep learning (DL) approaches have shown significant potential in applications involving the processing of medical images. DL models can autonomously acquire complex patterns and characteristics from medical pictures, facilitating precise and rapid identification of diverse illnesses like strokes.
MRI is safe for repeated imaging investigations. It may be used for susceptible groups, including pregnant women, children, and patients needing long-term follow-up since it does not utilize ionizing radiation ( Cheon et al., 2019). This benefit decreases the likelihood of radiation exposure linked to other imaging techniques, including CT or positron emission tomography. MRI offers a thorough examination of the whole brain, enabling the assessment of both supratentorial and infratentorial structures in a single imaging session ( Yalçın and Vural, 2022). This feature allows radiologists to comprehensively evaluate the size and spread of stroke damage and related issues, including bleeding, swelling, and pressure effects.
Stroke MRIs can demonstrate a wide range of lesion sizes, shapes, intensities, and locations. Stroke detection may rely on various factors in MRI scans, such as voxel intensities, texturing, and spatial correlations ( Shoily et al., 2019). Ensemble learning (EL) methods can identify and rank the most valuable features for classification, resulting in more efficient and discriminative models. EL approaches use a variety of classifiers to capture various elements of variability, leading to more robust and generalizable stroke detection models. Ensemble approaches can employ random subspace sampling or feature bagging techniques to decrease the dimensionality of MRI images. Ensemble classifiers may enhance computing efficiency and maintain prediction accuracy by choosing specific subsets of features or input channels to address the curse of dimensionality. These features motivated the authors to employ EL approaches and MRI images to detect BS. The authors intend to build an EL-based BS detection model using the MRI images in this study. The study novelties are:
LITERATURE REVIEW
DL methods like convolutional neural networks (CNNs) can analyze medical images, including CT scans, MRIs, and diffusion-weighted imaging (DWI) ( Alon and Dehkharghani, 2021). Using these techniques, imaging data of stroke lesions may be automatically detected and segmented. In addition to DL, conventional machine learning techniques such as RFs, support vector machines, and gradient boosting models, including LightGBM and XGBoost, are used for stroke detection applications ( Yu et al., 2020). In order to detect strokes, these algorithms may employ a variety of characteristics retrieved from clinical and imaging data. Computer-aided design systems are being created to assist radiologists and physicians in analyzing medical pictures to diagnose strokes. These systems usually use AI algorithms to aid in analyzing and interpreting imaging data, offering insights and identifying possible areas of BS ( Meijs et al., 2020). Improvements in imaging technologies, including perfusion imaging, spectroscopy, and functional MRI, allow for a thorough evaluation of brain tissue perfusion, metabolism, and functional connectivity. This may help in the early identification and description of stroke lesions. Mobile applications and wearable devices with sensors may monitor physiological parameters and identify stroke-related anomalies. These methods might facilitate prompt identification and timely action, especially in distant or neglected regions.
MRI may detect ischemic and hemorrhagic strokes due to its tissue sensitivity. Early identification and intervention are made possible by the great contrast and resolution seen in even the most minor lesions ( Lee et al., 2022). MRI provides several imaging sequences that give additional information on brain tissue properties, blood circulation, and structure. The sequences include T1-weighted, T2-weighted, DWI, perfusion-weighted imaging (PWI), and gradient echo (GRE). MRI can thoroughly evaluate stroke lesions and adjacent tissue by integrating data from several sequences. MRI can distinguish between ischemic and hemorrhagic strokes, guiding specific treatment strategies and predicting outcomes ( Babutain et al., 2021). DWI is very sensitive to acute IS, whereas GRE sequences are effective in identifying hemorrhagic strokes such as intracerebral hemorrhage and subarachnoid hemorrhage.
MRI perfusion methods, like PWI, may detect areas of reduced blood flow around the central infarct in IS ( Qiu et al., 2020). This information is vital for determining treatment decisions since it identifies brain tissue that may be saved and benefit from reperfusion therapy, including thrombolytic or endovascular intervention. MRI may provide information on tissue viability and functional outcomes after a stroke. Diffusion tensor imaging and MRI can evaluate white matter integrity, neuronal connections, and brain activation patterns, aiding in forecasting recovery paths and directing rehabilitation approaches ( Nielsen et al., 2018). MRI enables consecutive imaging examinations to track stroke advancement, treatment response, and the emergence of issues such as edema, hemorrhagic transformation, and mass effect ( Ali Aljarallah et al., 2023). This longitudinal evaluation is essential for improving patient care and modifying treatment plans as necessary. In contrast to invasive CT imaging, which exposes patients to dangerous radiation levels, MRI is free of ionizing radiation ( Gaidhani et al., 2019; Al-Mekhlafi et al., 2022). This permits repeated imaging and makes it safer for sensitive groups, including pregnant women and children.
The development of models that are both clinically interpretable and generally applicable is of equal importance to attaining high accuracy in stroke detection. Several current studies may lack rigorous validation procedures or be unable to provide insights into the characteristics acquired by DL models, which poses a challenge for physicians to have confidence in and incorporate these models into clinical practice. Subsequent investigations have to focus on creating interpretable EL models and verifying their efficacy using a range of real clinical datasets. Although DL and EL models have the potential for stroke diagnosis, their real-time deployment and incorporation into clinical processes have not been thoroughly investigated. The practical difficulties of implementing these models in actual healthcare settings, such as those involving computational effectiveness, scalability, legal compliance, and physician acceptability, require additional study.
MATERIALS AND METHODS
The authors build an EL-based BS detection model using MobileNet V3, LightGBM, CatBoost, and RF models. The MobileNet V3 model is used to extract the intricate patterns of BS. The architecture of MobileNet V3 supports the proposed model in generating features with limited computational resources. LightGBM and CatBoost models are used as base models to predict BS using the MRI features. Finally, the RF model is used as a meta-model to identify BS using the outcomes of the base models. The proposed research methodology is outlined in Figure 1.
The architecture of MobileNet V3 is specifically designed to efficiently extract hierarchical characteristics from medical pictures, especially those related to BS diagnosis, making it very useful for mobile and embedded devices. MobileNet V3 models pre-trained on extensive datasets such as ImageNet may be further trained on medical imaging datasets with fewer annotated samples. This process, known as transfer learning, allows the models to adapt specifically to detecting BSs. LightGBM, CatBoost, and RF are widely used EL algorithms renowned for their efficacy in managing structured data and identifying intricate associations among attributes. These models can exceptionally capture complex linkages and interactions between variables, which is essential for identifying subtle patterns that indicate BS in MRI images. Moreover, these models exhibit lower susceptibility to overfitting compared to deep neural networks, particularly when confronted with high-dimensional datasets with restricted sample sizes, which is a common occurrence in medical imaging datasets.
The authors obtained the MRI images from the public database ( Liu et al., 2022). The dataset covers 2888 clinical MRI images of individuals admitted at a National Stroke Center in Baltimore, MD, USA. The images are classified into IS, hemorrhagic, and not visible. The authors employed a data augmentation technique to generate substantial MRI images to train the proposed model.
MobileNet V3 is a state-of-the-art CNN architecture built for efficient and accurate inference on mobile and embedded devices. MobileNet V3 balances computational economy and predictive performance by using inverted residual blocks, efficient feature fusion methods, and sophisticated training methodologies. This has made it a popular option for several computer vision tasks. MobileNet V3 is often used for image classification tasks such as object identification, scene categorization, and fine-grained classification. Its small design and effective deduction process make it suitable for use in mobile applications, online services, and devices at the edge of a network. MobileNet V3 utilizes sophisticated regularization methods such as stochastic depth regularization and label smoothing to enhance generalization performance and increase resilience against overfitting. The authors enhanced the feature extraction by integrating automated neural architecture search methods that analyze a vast range of potential network structures to find the best configurations based on pre-determined performance standards. The inverted residual blocks in MobileNet V3 are enhanced with squeeze and excitation (SE) blocks, which improve feature representation and adaptability. SE blocks adjust channel-wise feature responses by learning channel-wise attention weights, enabling the model to emphasize key features of BS and reduce the effect of irrelevant features. MobileNet V3 uses efficient feature fusion methods like concatenation and element-wise addition to merging features from various network levels. This allows the network to efficiently collect geographical and semantic information at several scales while reducing computing burden.
The need for scalable methods to accurately diagnose strokes in massive datasets is growing as a result of the proliferation of electronic health records and medical imaging technology. In several machine learning contests and practical applications, LightGBM has shown to be the most effective and extensively utilized approach. Leveraging a successful BS detection model may boost reliability. LightGBM is a gradient boosting EL approach. Gradient boosting constructs a robust predictive model by iteratively integrating the forecasts of several weak learners (decision trees). Like other gradient boosting methods, LightGBM trains a sequence of decision trees, with each tree aiming to rectify the mistakes of the preceding trees. The process involves minimizing the loss function by sequentially incorporating new trees into the ensemble, emphasizing areas of the feature space where the model exhibits suboptimal performance. LightGBM is a robust gradient boosting framework known for its rapid training, excellent accuracy, and built-in capability to handle categorical variables. The novel characteristics of this model, including leaf-wise tree development, histogram-based computation, and improved category splitting, make it suitable for many machine learning applications, especially when working with extensive datasets and intricate feature spaces. Moreover, its emphasis on interpretability and scalability improves its usefulness in practical applications where performance, speed, and model transparency are crucial. LightGBM and CatBoost aim to provide fast and precise gradient boosting frameworks but vary in their methods for dealing with categorical variables, regularization, missing data, and model interpretability. LightGBM natively supports handling categorical variables without requiring encoding or pre-processing. The system automatically transforms category information into numerical representations using one-hot or binary encoding methods to enhance accuracy and efficiency. LightGBM utilizes a proficient technique to manage category factors when constructing trees. The method employs histogram-based splitting, categorizing comparable target values into bins, and determining the optimal split points using histogram statistics. The authors enable parallel training of LightGBM models on several CPU cores or graphics processing units (GPUs) to speed up the training process. LightGBM provides parallelization via multi-threading and GPU acceleration, utilizing computing resources efficiently and accelerating convergence.
CatBoost is a robust gradient boosting technique created to effectively manage categorical variables, control model complexity, handle missing values cleanly, and provide scalable performance. The particular characteristics, regularization methods, and inherent backing for categorical data make it suitable for many predictive modeling tasks such as classification, regression, and ranking. Moreover, its emphasis on interpretability and explainability increases its usefulness in practical scenarios where model transparency is crucial. CatBoost is intended to effortlessly manage categorical data without requiring encoding or pre-processing beforehand. It internally transforms categorical information into numerical representations during training to enhance accuracy and efficiency. CatBoost utilizes categorical embedding, an advanced method, to convert categorical data into numerical representations. This embedding method captures the intrinsic links and interactions across many categories, improving the model’s capacity to learn from categorical input successfully. CatBoost incorporates the management of missing values within its internal optimization process, guaranteeing uniform handling of missing data across all model training and prediction phases. To improve the performance of the CatBoost model, the authors employ Optuna hyperparameter optimization.
To increase the accuracy and resilience of predictions, RF uses a sophisticated EL approach that combines the predictions of several decision trees. It may serve as a meta-model to combine the forecasts of base models trained on various data subsets, resulting in more precise predictions of BS incidence or risk. Meta-modeling using RF combines the advantages of many base models to improve forecast accuracy by merging valuable insights from different origins. RF reduces overfitting by combining the forecasts of several trees that are trained on various subsets of the dataset. Ensemble averaging mitigates individual tree biases and decreases volatility, resulting in more reliable and broadly applicable predictions. RF is adept at managing high-dimensional data with many characteristics, making it ideal for evaluating intricate medical datasets like neuroimaging data for predicting BSs. The RF model offers metrics of feature relevance, showing the relative impact of each feature on the model’s predicted accuracy. These data may assist in determining the most significant predictors of BS and selecting aspects for further research. The RF meta-model combines the predictions of base models by aggregating the individual decision trees. Aggregating the forecasts enhances the target variable’s precision by using several base models’ unique advantages. Utilizing RF for meta-modeling may enhance forecast accuracy compared to stand-alone base models. The meta-model integrates many predictions from different models to identify intricate patterns in the data, reduce overfitting, and provide more dependable forecasts of BS likelihood.
Assessing the efficacy of a BS detection model by MRI requires several phases to evaluate its precision, dependability, and practicality in a clinical setting. Sensitivity and specificity assess the BS model’s accuracy in identifying stroke-positive and stroke-negative instances. Accuracy offers a thorough evaluation of the model’s performance across all categories. Precision (positive predictive value) and recall (true positive rate) assess the model’s accuracy in recognizing stroke cases and its capability to recover all pertinent stroke instances. The F1 score offers a fair evaluation of the model’s performance by taking into account both false positives and false negatives. The area under the receiver operating characteristic (AU-ROC) score is used to evaluate the model’s discriminating performance at various threshold levels. The area under the precision–recall curve (AU-PRC) measures the trade-off between accuracy and recall at different categorization thresholds.
RESULTS AND DISCUSSION
The experimental validation is conducted on Windows 10, Intel i7, 16 GB RAM, and NVIDIA GeForce RTX 3050 Ti. The authors used Python 3.8.1 programming environment, Keras, and TensorFlow libraries to implement the proposed BS detection model. The pre-trained CNN models were developed in Python 3.8.0. Fivefold cross-validation was followed to evaluate the performance of the BS detection models. The cross-validation results are presented in Table 1. In each fold, the suggested model obtained a considerable outcome. The presented feature extraction produced the crucial patterns of BS from the MRI images.
Results of performance validation.
| Folds | Accuracy | Precision | Recall | F1 score | Specificity | Sensitivity |
|---|---|---|---|---|---|---|
| 1 | 98.3 | 97.1 | 97.2 | 97.1 | 96.7 | 97.1 |
| 2 | 98.5 | 96.6 | 96.3 | 96.4 | 97.5 | 96.6 |
| 3 | 97.9 | 97.3 | 97.1 | 97.2 | 96.8 | 97.2 |
| 4 | 98.5 | 98.1 | 98.3 | 98.2 | 95.4 | 96.6 |
| 5 | 97.8 | 97.4 | 97.5 | 97.4 | 96.2 | 97.1 |
| Average | 98.2 | 97.3 | 97.2 | 97.2 | 96.6 | 96.9 |
The ability to find the specific type of BS is highlighted in Table 2. The presented model showed a significant performance in detecting discern patterns of dyslexia. LightGBM and CatBoost models supported the RF model to present an exceptional performance. The outcomes of the performance validation are presented in Figure 2.
Classes’ identification performance evaluation.
| Folds | Precision | Recall | F1 score | Accuracy | Specificity | Sensitivity |
|---|---|---|---|---|---|---|
| IS | 98.5 | 97.5 | 98.0 | 98.7 | 95.8 | 96.8 |
| Hemorrhagic | 98.6 | 97.7 | 98.1 | 98.8 | 96.7 | 96.9 |
| Not visible | 98.2 | 96.6 | 97.3 | 98.9 | 96.9 | 97.5 |
Abbreviation: IS, ischemic stroke.
Table 3 outlines the results of the comparative analysis. The dataset was highly imbalanced. The author faced challenges in optimizing the MobileNet V3-based feature extraction. The recommended feature extraction improved the capability of the proposed model. The findings outlined the exceptional performance of the suggested model on the MRI dataset. The proposed BS detection model produced optimal results compared to the existing models. Figure 3 reveals the performance of the BS detection models on the MRI dataset.
Findings of comparative analysis.
| Classes | Precision | Recall | F1 score | Accuracy | Specificity | Sensitivity |
|---|---|---|---|---|---|---|
| Proposed model | 97.3 | 97.2 | 95.3 | 98.2 | 96.6 | 96.9 |
| DenseNet-121 | 95.4 | 95.3 | 96.0 | 96.1 | 94.4 | 95.4 |
| MobileNet V3 | 96.3 | 95.8 | 95.4 | 95.2 | 95.3 | 96.3 |
| SqueezeNet V1.1 | 95.8 | 95.1 | 96.0 | 94.9 | 95.5 | 95.7 |
| EfficientNet B7 | 96.1 | 95.9 | 95.9 | 95.6 | 94.7 | 94.9 |
| An et al. (2023) | 95.7 | 96.2 | 96.3 | 95.5 | 95.2 | 95.6 |
| Soleimani and Farezi (2023) | 96.6 | 96.1 | 96.5 | 96.6 | 95.9 | 96.1 |
| Liu et al. (2021) | 96.7 | 96.3 | 95.3 | 95.9 | 96.1 | 95.5 |
Table 4 presents the exceptional performance, including high AU-ROC and AU-PRC scores, and minimal computational loss. The model has high discriminative ability, precision–recall trade-off, and efficient optimization during training. The AU-ROC score of 0.95 suggests that the model can differentiate between stroke and non-stroke patients. This indicates that the model’s predictions are accurate, with a low level of uncertainty between positive and negative samples in the anticipated probability. Considering the balance between accuracy and recall, the AU-PRC achieved an impressive score of 0.92, demonstrating the model’s ability to sustain high precision at elevated recall levels. Accuracy is vital in medical applications such as BS diagnosis since incorrect results may lead to severe outcomes. The strong performance of our model has important clinical implications. With its excellent AU-ROC and AU-PRC scores, our model may be a dependable tool for radiologists and clinicians to prioritize and manage stroke patients effectively, enabling prompt intervention and treatment. The minimal computational loss also guarantees that our model effectively uses the provided training data, resulting in improved convergence and parameter estimates during training.
AU-ROC, AU-PRC, and computational loss outcomes.
| Classes | AU-ROC score | AU-PRC | Loss | Parameters (in millions) | Flops (in giga) |
|---|---|---|---|---|---|
| Proposed model | 0.95 | 0.90 | 0.45 | 48 | 53 |
| DenseNet-121 | 0.87 | 0.81 | 1.12 | 41 | 77 |
| MobileNet V3 | 0.91 | 0.83 | 1.36 | 54 | 59 |
| SqueezeNet V1.1 | 0.89 | 0.81 | 0.89 | 51 | 66 |
| EfficientNet B7 | 0.91 | 0.85 | 1.05 | 50 | 61 |
| An et al. (2023) | 0.83 | 0.89 | 0.61 | 52 | 70 |
| Soleimani and Farezi (2023) | 0.95 | 0.90 | 0.75 | 53 | 63 |
| Liu et al. (2021) | 0.92 | 0.88 | 0.61 | 57 | 59 |
Abbreviations: AU-PRC, the area under the precision–recall curve; AU-ROC, the area under the receiver operating characteristic score.
The proposed BS prediction model improves stroke diagnosis, treatment results, and resource allocation for healthcare providers and patients. EL methods can identify minute patterns and signs of stroke in MRI images. Early identification allows physicians to initiate thrombolytic treatment or mechanical thrombectomy to reduce brain damage and improve patient outcomes. EL models use MRI features to determine patient risk. Ensemble models personalize stroke risk projections by combining unique patient characteristics, enabling tailored preventative and lifestyle treatments. Healthcare practitioners use the proposed model to diagnose, plan, and manage stroke patients. The proposed model enables clinicians to make decisions by combining the outcomes of the LightGBM and XGBoost models. Ensemble models empower patients to self-manage stroke risk factors by teaching them about their risk and promoting lifestyle changes, including smoking cessation, cardiovascular control, and regular physical activity. In order to better diagnose and manage strokes, researchers are turning to EL models, which pave the way for new algorithms, biomarkers, and diagnostic tools. The proposed model can streamline scientific discoveries, enhance diagnostic accuracy, and improve stroke treatment by integrating diverse knowledge and collaborative research.
AI-powered advancements promise to transform the way stroke is detected, diagnosed, treated, and studied. AI technologies enable healthcare practitioners to offer prompt, tailored, and effective stroke treatment by utilizing machine learning, image analysis, telemedicine, decision support systems, and predictive analytics. Integrating AI into clinical practice requires cooperation across interdisciplinary teams, rigorous validation, regulatory monitoring, and continuous development of algorithms to guarantee accuracy, safety, and ethical use in patient care.
EL-based BS detection models outperform individual models in prediction performance, resilience, and generalization. However, they have limits. EL generally combines numerous base models, which increases system complexity and computational cost, particularly for massive datasets or complex algorithms. An ensemble model’s efficacy is directly proportional to the variety and quality of its base learners. Understanding an ensemble model’s decision-making process might be challenging, limiting its application in interpretability-critical fields. As the dataset or base learners expand, ensemble model building and maintenance might get complicated. EL improves BS detection models and other machine learning applications despite these drawbacks. Addressing these restrictions and making suitable design decisions may help ensemble approaches reach their full potential.
The study’s findings may be limited by the size and diversity of the dataset used for model training and evaluation. Medical imaging datasets, especially those specific to BS, can be relatively small and may not adequately represent the full spectrum of stroke cases and imaging variations encountered in clinical practice. The generalizability of the developed model to different patient populations, imaging protocols, and healthcare settings may be limited. The model’s performance in real-world scenarios remains uncertain without validating external datasets from diverse sources. While MobileNet V3 is employed for feature extraction, the choice of features and their representation may not fully capture the nuanced characteristics of BS in MRI images. More informative features or alternative feature extraction approaches might exist that were not explored in this study. EL techniques, while effective, can introduce additional complexity to the model training and interpretation process. The optimal combination of base models, hyperparameters, and fusion strategies may not be exhaustively explored, potentially limiting the model’s performance and scalability.
CONCLUSION
The authors have addressed the existing challenges in detecting BS using MRI images. A feature extraction technique using the MobileNet V3 backbone was suggested to extract the key patterns of BS and non-stroke images. The early stopping strategy supported the authors in overcoming the challenges of feature extraction. LightGBM and CatBoost models were applied to generate outcomes. These models were fine-tuned in order to find BS with limited resources. Lastly, the authors used the RF classifier to differentiate BS and non-stroke MRI images. The suggested model was generalized on the MRI dataset. The generalization findings highlighted the importance of the proposed model in BS detection. The model has outperformed the existing models by obtaining exceptional results. The reliability of the proposed BS detection model can help radiologists and clinicians in finding BS. Fine-tuning the recommended model can improve the reliability and interpretability of this study’s findings.