Enhancing Brain Stroke Detection: A Novel Deep Neural Network with Weighted Binary Cross Entropy Training

This article introduces and evaluates a novel Deep Neural Network (DNN) designed specifically for brain stroke detection.Highlighting the innovation of using a weighted Binary Cross Entropy (BCE) loss function to address dataset imbalances, this work also aims to propose a less computationally complex model that makes predictions using very basic demographic and health-related information.In the realm of stroke detection using brain imaging, various methods, including Convolutional Neural Networks (CNNs) and Hybrid Algorithms (OzNet-mRMR-NB and Linearized Scattering Operator with ML Algorithms), have shown potential but suffer from computational complexity and time consumption.Analyzing factors such as Gender, Age, Hypertension, Heart Disease, Work Type, Residence Type, Average Glucose Level, Body Mass Index (BMI), and Smoking Status allows the proposed model to avoid the computational burdens of image-based feature extraction, focusing instead on accessible data to improve prediction accuracy.The selected model architecture features six hidden layers with dimensions (31, 64, 128, 64, 32,16) to effectively process these inputs and accurately identify stroke risk.By using weighted BCE as the loss function, the model counteracts dataset bias towards more common nonstroke cases, emphasizing the identification of critical stroke instances during training.Furthermore, the proposed model demonstrated significant improvements in detecting stroke instances, particularly in the minority class, by employing weighted BCE.While accuracy slightly decreased from 95.36% with standard BCE to 75.36% with weighted BCE, this reduction is outweighed by substantial gains in Recall (from 0 to 34.91), Precision (from 0 to 100), and F1-Score (from 0 to 51.75).In addition to making a significant methodological contribution, this study has the potential to improve stroke prediction greatly, allowing for more efficient and rapid treatment.