Project W13: Telehealth - Fall Detection Solution

The World Health Organization (WHO) de nes falls as involuntary events resulting in an individual coming to rest at a lower level after initial impact. Falls pose a public health threat across proad populations including the elderly, children, disabled individuals and patients with orthopedic or neurological disorders. While external factors that raise potential for falls might be controlled, it is often impossible to eliminate the intrinsic physiological factors. As such, there is a pressing need for technology that reliably detects falls in real time to minimize occurence of related injuries and healthcare costs. Wearable sensors o er a promising avenue in this regard by allowing unrestricted wireless monitoring of subject movements.

In this assignment, you will be analyzing wearable-sensor data from a group of subjects performing either a fall action (F) or non-fall action (NF; i.e. basic motor activities of daily living such as walking, standing, ascending, descending). Wearable-sensor data are provided in the le falldetection_dataset.csv, which contains 566 samples of motor actions along with respective action labels (’F’ or ’NF’) presented as rows. The 306 features from the sensors that re ect various properties including velocity, acceleration, temperature and pressure are presented as columns.


Part A) [40 pts]

Exploratory data analysis via clustering can serve as an a priori step prior to targeted model ts. In this step, we would like to infer whether samples of sensor data form segregated clus-ters in the space spanned by the sensor features. Since the number of features are relatively large to visualize, rst perform principal components analysis (PCA) on the 566 306 fea-ture matrix to extract the top two PCs, noting how much variance they capture. Using the projections of 566 data samples onto the rst and second PCs, run k-means clustering (or another method of choice) to separate data into clusters. Try di erent numbers of clusters (N), and state your assessment on a proper number of clusters. Taking the clusters obtained when N=2, check the degree of percentage overlap/consistency between the cluster mem-berships and the action labels originally provided. Comment on whether fall detection is possible based on these measurements.

Part B) [60 points]

Once the exploration is complete, we can proceed onto the supervised learning stage. In this step, the goal is to build a classi er that detects the action label (fall or non-fall) with high accuracy. We will be testing two separate models for this purpose: a support-vector-machine (SVM) classi er and a multi-layer perceptron (MLP) classi er. Freely experiment with various hyperparameters included in these models to maximize classi cation accuracy. To do this, however, you need to implement a cross-validation procedure with a three-way split of data into non-overlapping training/validation/testing sets (e.g., 70%, 15%,15%). The models with various di erent hyperparameters are t on the training set, the parameter selection is performed based on the validation set, and the nal classi cation accuracy is reported on the testing set. Show/report all intermediate results, and comparatively evaluate SVM and MLP models. Comment on the success of fall detection based on wearable sensors.