Machine Learning based Fall Detection Algorithms for Toddlers using Public and Experimental Datasets

Falling and its adverse impacts, especially for children under the age of six, have life long consequences. Fall detection, in such cases can be paramount for administering care quickly and efficiently. This paper addresses the concern of a dearth of fall datasets, with the falls being performed by actual toddlers, and presents a novel experimental dataset for the same. It also provides a comparison between multiple publicly available datasets, which are recorded by simulating falls by adults, across deep learning and machine learning models. Using a binary classification scheme to detect falls, the machine learning models are tested for both the experimental dataset as well as the publicly available ones. In the comparison of models, the kNN model showed favorable results, having the highest accuracy to our experimental dataset, that of 96 percent, with a sensitivity of 96.5 percent and specificity of 95.8 percent. Similar tests of models with the publicly available datasets highlight the significant difference in results when utilizing fall data simulated by adults and children. The research also sheds light on the application of such models in various wearables as well as the metrics to be deliberated over when choosing a suitable algorithm for this particular implementation.