The Performance of Post-Fall Detection Using the Cross-Dataset: Feature Vectors, Classifiers and Processing Conditions

Abstract

In this study, algorithms to detect post-falls were evaluated using the cross-dataset according to feature vectors (time-series and discrete data), classifiers (ANN and SVM), and four different processing conditions (normalization, equalization, increase in the number of training data, and additional training with external data). Three-axis acceleration and angular velocity data were obtained from 30 healthy male subjects by attaching an IMU to the middle of the left and right anterior superior iliac spines (ASIS). Internal and external tests were performed using our lab dataset and SisFall public dataset, respectively. The results showed that ANN and SVM were suitable for the time-series and discrete data, respectively. The classification performance generally decreased, and thus, specific feature vectors from the raw data were necessary when untrained motions were tested using a public dataset. Normalization made SVM and ANN more and less effective, respectively. Equalization increased the sensitivity, even though it did not improve the overall performance. The increase in the number of training data also improved the classification performance. Machine learning was vulnerable to untrained motions, and data of various movements were needed for the training.

Support Vector Machine Classifiers Show High Generalizability in Automatic Fall Detection in Older Adults

“…For SVM, results have been mixed: Igual and colleagues [13] found that SVM or neural network classifiers worked either well or only moderately depending on the dataset on which they were trained, but no accidental falls were investigated. Similarly, Koo et al [31] found that SVM and ANN classifiers may work well in internal testing but during external testing only when training data were subjected to a number of preprocessing steps. Of note, they used their own laboratory data for training and validated on the SisFall dataset, but not on accidental falls.…”

“…One reason for better performance might have been that training data were impact aligned using the SampEn algorithm. Impact alignment was recently shown to improve classifier performance by Koo and colleagues [31] using a different technique for event detection. Some previously published algorithms reached extremely high performance values up to 100% sensitivity or accuracy (e.g., [60,61]) while others are of similar value as in our study (e.g., [29,42,44,62]).…”

“…Igual et al [13] found that SVM had good generalizability during cross-dataset validation. SVM and ANN were used by Koo et al [31] in a cross-validation study and showed mixed results. Both studies used fall repositories containing only simulated falls.…”

“…Segments were identified as not containing a fall impact and discarded if their SampEn value was lower than a prespecified threshold and kept otherwise, labeled as "fall". The datasets were thus temporally aligned to the fall impact [31]. The threshold used was empirically determined for all subjects such that it was only surpassed by the SampEn value yielded from a fall impact.…”

A Hierarchical Approach to Activity Recognition and Fall Detection Using Wavelets and Adaptive Pooling

“…Post fall intelligence is an important research area in the field of fall detection as it can be useful in determining various post fall injuries [ 18 ] and serve as an intelligence parameter [ 19 ] for doctors. Koo et al [ 20 ] present experiments for post fall detection from a combination of self collected data and the SisFall dataset. They conduct tests using sliding windows as well as discrete windows from these signals and compute statistical features from them.…”