This paper addresses the challenge of fall detection in older adults, focusing on the imbalance between fall and daily activity (ADL) samples and the scarcity of data. The authors propose a new dataset that combines data from three existing datasets (UP Fall, UMA Fall, and WEDA Fall), resulting in over 1300 fall samples and 28,000 ADL samples. They evaluate the dataset using classic cost-sensitive machine learning methods to handle class imbalance. The model is trained and validated using temporal and frequency features extracted from raw data collected by accelerometers and gyroscopes on the wrist. The study assesses the generalization properties of each dataset and the performance of the combined dataset. The results show that the model achieves a recall of 90.57%, specificity of 96.91%, and AUC-ROC of 98.85%. The WEDA Fall dataset is noted for its strong performance in ADL identification, while the UP Fall dataset demonstrates robustness in detecting falls. The UMA Fall dataset shows balanced performance in both fall and ADL detection. The research highlights the importance of adapting normalization techniques to the specific nature of each dataset and model, and provides valuable insights for future advancements in fall detection systems.This paper addresses the challenge of fall detection in older adults, focusing on the imbalance between fall and daily activity (ADL) samples and the scarcity of data. The authors propose a new dataset that combines data from three existing datasets (UP Fall, UMA Fall, and WEDA Fall), resulting in over 1300 fall samples and 28,000 ADL samples. They evaluate the dataset using classic cost-sensitive machine learning methods to handle class imbalance. The model is trained and validated using temporal and frequency features extracted from raw data collected by accelerometers and gyroscopes on the wrist. The study assesses the generalization properties of each dataset and the performance of the combined dataset. The results show that the model achieves a recall of 90.57%, specificity of 96.91%, and AUC-ROC of 98.85%. The WEDA Fall dataset is noted for its strong performance in ADL identification, while the UP Fall dataset demonstrates robustness in detecting falls. The UMA Fall dataset shows balanced performance in both fall and ADL detection. The research highlights the importance of adapting normalization techniques to the specific nature of each dataset and model, and provides valuable insights for future advancements in fall detection systems.