<p>Human Activity Recognition (HAR) plays a crucial role in the development of intelligent systems for smart healthcare, sustainable living environments, and enhanced human well-being. However, conventional machine learning (ML) and deep learning (DL) approaches often face challenges in handling noisy real-world sensor data or require high computational resources. To address these limitations, this study proposes a novel two-stage hybrid ensemble framework that integrates probabilistic outputs from multiple ML classifiers with advanced DL models for robust HAR. In the first stage, diverse ML classifiers, including K-Nearest Neighbors (KNN), Decision Tree (DT), Naive Bayes (NB), Logistic Regression (LR), and Support Vector Machines (SVM) are independently trained to generate class probability vectors. These outputs are then fused to construct a meta-feature representation. In the second stage, deep learning models such as Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Bi-directional Long Short-Term Memory (BiLSTM), and Autoencoder are trained on the meta-feature space to perform final classification. The proposed framework is evaluated on the benchmark Human Activity Recognition Using Smartphones dataset and achieves superior performance with an average accuracy of 99.21%, precision of 99.26%, recall of 99.24%, and F1-score of 99.25%. Statistical validation confirms the robustness and generalization capability of the model compared to state-of-the-art approaches. By enabling accurate, efficient, and scalable activity recognition, the proposed system contributes to the advancement of intelligent healthcare solutions, remote patient monitoring, and inclusive digital ecosystems. This work supports the development of sustainable and human-centric intelligent systems that enhance quality of life and promote well-being in real-world environments.</p>