Pneumonia is a major global health issue, with prompt diagnosis critical for effective treatment. Chest X-rays (CXRs) are an important diagnostic tool, but manual interpretation can be difficult and error-prone. This paper offers a feature-level ensemble learning strategy for automating pneumonia diagnosis in CXRs, which employs convolutional neural networks (CNNs) to improve diagnostic accuracy and assist healthcare practitioners. The model combines features from three pre-trained CNN architectures—VGG19, EfficientNet-B0, and DenseNet121—taking advantage of each architecture’s specific strengths. By adding trainable weights to the features collected by multiple models, instead of only concatenating features, this method enables the model to dynamically prioritize the most relevant feature representations while capturing a greater variety of features. In order to improve interpretability and assist physicians in making decisions, gradient-weighted class activation mapping, or Grad-CAM, is also utilized to produce heatmaps that highlight important regions in CXRs. The proposed model obtains 93.1% classification accuracy on the test dataset, demonstrating its ability to aid in the precise and efficient detection of pneumonia. This study deals with the need for dependable AI tools in medical diagnostics, with the goal of reducing radiologist burden and improving decision-making in resource-constrained environments.