Smart home technology has revolutionized daily living by enabling automated activity recognition and behavior monitoring, particularly in multi-resident environments. This study presents an advanced framework for analyzing sensordriven activity data to identify routines, shared tasks, and behavioral deviations overtime. By leveraging machine learning models such as Bi-directional Long Short-Term Memory (BiLSTM), Random Forest, Tuned Random Forest, and XGBoost, the research explores the complexities of overlapping tasks, personal schedules, and routine shifts. Among the models evaluated, XGBoost achieved the highest test accuracy of 91.39%, effectively handling imbalanced datasets and prioritizing key features. Random Forestand its tuned variant demonstrated strong reliability with test accuracies of 88.31% and 88.90%, respectively, offering interpretability and robustness. While BiLSTM performed well in capturing sequential dependencies, it struggled with overlapping activities, achieving a test accuracy of 73.66%. Beyond activity classification, the study emphasizes behavioral deviation analysis to detect variations in daily routines. These deviations, identified through changes in task frequencies and durations, serve as indicators of potential health concerns, lifestyle modifications, or shifts in social engagement. The findings highlight the potential of smart home systems not only to enhance daily convenience but also to facilitate proactive health monitoring and personalized care interventions.