Data-driven Structural Health Monitoring based on modal properties as damage sensitive features typically relies on statistical pattern recognition approaches to detect anomalies. Considering the strong influence of environmental and operational variables on the vibration response, considerable research efforts have been made to develop data normalization techniques aimed at compensating the normal variability of data to limit the occurrence of false or missed alarms. In this framework, a further fundamental step is the setting of an appropriate alarm threshold to distinguish regular structural conditions from anomalous states. The alarm threshold is set in the training phase based on the statistical distribution of the damage sensitive feature in this period. Since anomaly detection methods focus on finding outliers, assuming a Gaussian distribution of data involves a course hypothesis about the tails of the distribution. The present paper analyzes the use of Extreme Value Statistics to set an alarm level for anomaly detection, focusing on a procedure based on the Block Maxima method to collect extreme values and on the Generalized Extreme Value distribution. A comparison with other conventional approaches is presented to demonstrate the significant detection accuracy that can be achieved by using the extreme value theory. The natural frequency time histories of the Z24 bridge are assumed as benchmark data to show the effectiveness of the method in enhancing the reliability of anomaly detection.