Long-term Structural Health Monitoring (SHM) is crucial for the study of ageing structures and infrastructures. Unlike fast-varying damage, which manifests abruptly, slow-varying damage, such as corrosion, evolves progressively, presenting different challenges for detection and analysis. These challenges are particularly pronounced in existing structures, where slow-varying damage often progresses gradually and can be overlooked by traditional methods designed to detect sudden changes. This study focuses on characterizing slow-varying damage using vibration data recorded over years of monitoring, operating a comparative analysis of damage-sensitive features to assess their effectiveness in identifying long-term trends. These features are extracted from acceleration data from statistical, time and frequency domains. A selection of the possible features was operated to ensure that the monitoring process captures gradual changes while managing large volumes of data. A Long Short-Term Memory (LSTM) network was trained on each feature timeseries, to predict the long-term trends, and their prediction errors calculated on the scaled data were used to compare their performance. The selected features were tested using a new benchmark that simulates long-term ageing scenarios in a controlled numerical environment. The benchmark data is generated from a Single Degree of Freedom (SDOF) system subjected to dynamic loading, including the effects of environmental and operational variabilities (EOVs), providing a testing dataset to later further expand the results to a wider range of structural scenarios.