In the medical field, disease diagnosis through cloud computing is a promising technology to focus more on patient care. It offers benefits for both patients and doctors. We aimed to develop a cloud based classification method for patients with brain haemorrhage in rural (or) remote areas. The detection of brain haemorrhage is critical for neurological diagnosis and treatment. Although deep learning models have shown promise in medical image analysis, predicting brain haemorrhage remains challenging due to its complexity and variability. To address this, we propose an ensemble approach for detection of Intracranial Haemorrhage diagnosis using cloud computing. Cloud services enable data to be stored on remote servers and then be accessed via the Internet. The user does not need to be in a specific location to access it, which allows the user to work remotely. The ensemble comprises diverse deep learning models (ResNet50, VGG16, and DenseNet121) with Global Average Pooling 2D layers, benefiting from the cloud's computational resources for efficient training. Each model brings unique strengths, enabling us to capture a wider range of data patterns and improve prediction accuracy. Data pre-processing involves dividing the dataset into training and validation sets. The base models are trained on the training data, and hyper parameters are fine-tuned via cross-validation on the validation set. The stacking and blending techniques combine base model predictions. Stacking uses validation predictions to train a meta-model that intelligently combines individual predictions, while blending directly averages predictions for a simpler ensemble. To overcome the Challenges of privacy preserving, data protection, improve the performance and fast diagnosis, the proposed work is also deployed in a cloud based framework. Performance evaluation employs standard metrics on a separate test set. Results show that the ensemble approach significantly enhances Intracranial Haemorrhage Prediction, improving clinical decision-making for neurological emergencies and fast diagnosis. Our approach incorporates strategies to handle class imbalance, such as using suitable loss functions and data augmentation during training.