This paper presents a study on wrist-based fall detection systems, focusing on improving generalization across different datasets. The authors propose a new dataset combining three existing datasets, containing over 1300 fall samples and 28,000 negative samples. The dataset includes a standardized method for adding samples, allowing future data sources to be incorporated. The authors evaluated their dataset using cost-sensitive machine learning methods to handle class imbalance. They extracted temporal and frequency features from raw data of accelerometers and gyroscopes using a sliding window of 2 seconds with 50% overlap. They tested the generalization properties of each dataset by evaluating them against each other and the performance of their new dataset. The model achieved a recall of 90.57%, specificity of 96.91%, and an AUC-ROC of 98.85% against the combination of three datasets. The study highlights the challenges of fall detection due to the imbalance in data, with falls being a rare event. Traditional machine learning models often fail to generalize well due to this imbalance. The authors propose a solution by combining multiple datasets to improve generalization. They also discuss the importance of using wrist-based sensors for fall detection, as they are practical, sensitive to body movements, and have low energy consumption. The study compares different fall detection systems, including camera-based, context/ambient, and wearable systems. It also discusses the limitations of existing datasets and the need for more diverse and representative data. The authors evaluated their model using three datasets: UP Fall, WEDA Fall, and UMA Fall. They performed preprocessing to standardize the data, including normalization and labeling. They extracted features from the data using a sliding window approach and selected the most important features using a random forest model. They tested the model's performance using various metrics, including recall, specificity, and AUC-ROC. The results showed that the model achieved high accuracy in distinguishing falls from daily activities, with the best performance on the UMA Fall dataset. The study concludes that combining multiple datasets improves the generalization of fall detection models and that using wrist-based sensors is a promising approach for fall detection. Future work includes expanding the dataset with more diverse data and using more advanced machine learning models.This paper presents a study on wrist-based fall detection systems, focusing on improving generalization across different datasets. The authors propose a new dataset combining three existing datasets, containing over 1300 fall samples and 28,000 negative samples. The dataset includes a standardized method for adding samples, allowing future data sources to be incorporated. The authors evaluated their dataset using cost-sensitive machine learning methods to handle class imbalance. They extracted temporal and frequency features from raw data of accelerometers and gyroscopes using a sliding window of 2 seconds with 50% overlap. They tested the generalization properties of each dataset by evaluating them against each other and the performance of their new dataset. The model achieved a recall of 90.57%, specificity of 96.91%, and an AUC-ROC of 98.85% against the combination of three datasets. The study highlights the challenges of fall detection due to the imbalance in data, with falls being a rare event. Traditional machine learning models often fail to generalize well due to this imbalance. The authors propose a solution by combining multiple datasets to improve generalization. They also discuss the importance of using wrist-based sensors for fall detection, as they are practical, sensitive to body movements, and have low energy consumption. The study compares different fall detection systems, including camera-based, context/ambient, and wearable systems. It also discusses the limitations of existing datasets and the need for more diverse and representative data. The authors evaluated their model using three datasets: UP Fall, WEDA Fall, and UMA Fall. They performed preprocessing to standardize the data, including normalization and labeling. They extracted features from the data using a sliding window approach and selected the most important features using a random forest model. They tested the model's performance using various metrics, including recall, specificity, and AUC-ROC. The results showed that the model achieved high accuracy in distinguishing falls from daily activities, with the best performance on the UMA Fall dataset. The study concludes that combining multiple datasets improves the generalization of fall detection models and that using wrist-based sensors is a promising approach for fall detection. Future work includes expanding the dataset with more diverse data and using more advanced machine learning models.