Towards a smart fall detection system using wearable sensors

A fall-detection system is employed in order to monitor an older person or infirm patient and alert their carer when a fall occurs. Some studies use wearable-sensor technologies to detect falls, as those technologies are getting smaller and cheaper. To date, wearable-sensor-based fall-detection approaches are categorised into threshold and machine-learning-based approaches. A high number of false alarms and a high computational cost are issues that are faced by the threshold- and machine-learning basedapproaches, respectively. The goal of this thesis is to address those issues by developing a novel low-computational-cost machine-learning-based approach for fall detection using accelerometer sensors. Toward this goal, existing fall-detection approaches (both threshold- and machine-learning-based) are explored and evaluated using publicly accessible datasets: Cogent, SisFall, and FARSEEING. Four machine-learning algorithms are implemented in this study: Classification and Regression Tree (CART), k-Nearest Neighbour (k-NN), Logistic Regression (LR), and Support Vector Machine (SVM). The experimental results show that using the correct size and type for the sliding window to segment the data stream can give the machine-learning-based approach a better detection rate than the threshold-based approach, though the difference between the threshold- and machine-learning-based approaches is not significant in some cases. To further improve the performance of the machine-learning-based approaches, fall stages (pre-impact, impact, and post-impact) are used as a basis for the feature extraction process. A novel approach called an event-triggered machine-learning approach for fall detection (EvenT-ML) is proposed, which can correctly align fall stages into a data segment and extract features based on those stages. Correctly aligning the stages to a data segment is difficult because of multiple high peaks, where a high peak usually indicates the impact stage, often occurring during the pre-impact stage. EvenT-ML significantly improves the detection rate and reduces the computational cost of existing machine-learning-based approaches, with an up to 97.6% F-score and a reduction in computational cost by a factor of up to 80 during feature extraction. Also, this technique can significantly outperform the threshold-based approach in all cases. Finally, to reduce the computational cost of EvenT-ML even further, the number of features needs to be reduced through a feature-selection process. A novel genetic-algorithm-based feature-selection technique (GA-Fade) is proposed, which uses multiple criteria to select features. GA-Fade considers the detection rate, the computational cost, and the number of sensors used as the selection criteria. GAFade is able to reduce the number of features by 60% on average, while achieving an F-score of up to 97.7%. The selected features also can give a significantly lower total computational cost than features that are selected by two single-criterion-based feature-selection techniques: SelectKBest and Recursive Feature Elimination. In summary, the techniques presented in this thesis significantly increase the detection rate of the machine-learning-based approach, so that a more reliable fall detection system can be achieved. Furthermore, as an additional advantage, these techniques can significantly reduce the computational cost of the machine-learning approach. This advantage indicates that the proposed machine-learning-based approach is more applicable to a small wearable device with limited resources (e.g., computing power and battery capacity) than the existing machine-learning-based approaches.