Monitoring the health of Hajj pilgrims is challenging due to the intense physical and emotional demands of the pilgrimage. Existing solutions often fail to accurately integrate diverse physiological data and adapt to dynamic conditions. We propose a two-step LSTM-TabNet model that first captures temporal dependencies with an LSTM and then uses TabNet for robust feature selection and classification. This approach effectively addresses these challenges, providing real-time, context-aware health monitoring for pilgrims using wearable devices. Data collected from 19 participants demonstrated that our model achieves high accuracies of 97.0% for physical tiredness level, 95.0% for emotional mood level, and 95.0% for rukun (Hajj ritual) activity classification, outperforming existing frameworks. To promote reproducibility, the code and anonymized dataset are publicly available.