This paper presents a real-time early warning system designed to monitor and respond to anomaly points in vital signs and activity patterns using an unsupervised transformer approach. The system is capable of detecting the symptom onset of infectious diseases using data from wearable sensor devices. To address the challenges posed by missing values and sparse data, we propose an unsupervised anomaly detection transformer model. The model includes an encoder that learns individuals’ vital signs and activity patterns and a classifier decoder that computes anomaly scores. It detects anomalies by analyzing the discrepancy between local and global similarity, identifying points that resemble nearby points but differ from those further away. Leveraging the transformer architecture, the model can effectively capture complex patterns in the data, providing a robust framework for early anomaly detection. Experimental results demonstrate the effectiveness of our approach in accurately detecting symptom onset. By continuously monitoring vital signs and activity patterns, the system can provide early warning alerts for potential health issues, enabling timely intervention and improving public health monitoring and response strategies through its innovative data-driven approach.