Purpose: In this study, we attempted to construct an improved anomaly detection system. In our previous study, we developed an anomaly detection system for the elderly. However, the detection performance may not be sufficient for relatively mild abnormalities such as fever or low peripheral oxygen saturation (SpO₂), which frequently occur in elderly people. Methods: The purposed method used 3 types of models: natural language processing model, data generation model, and anomaly detection model. Among the residents of a long-term care facility, 79 (16 males and 63 females, 88.54 ± 7.21 years of age) elderly people were selected as participants. Results: The success rate of predicting physical anomaly (mean ± standard deviation) using our previous method was 24.82 ± 32.55% and that using the method proposed in this study was 42.12 ± 39.20%, with a significant difference between the two methods. Conclusions: In this study, we succeeded to improve the performance of an anomaly detection system for the elderly. Difference in detection performance between subject was observed. Further research is needed to investigate the relationship between subject condition and detection performance.