Automated detection of wandering patterns in people with dementia

Abstract

Purpose: This study focuses on travel patterns of people with dementia (PWD), which can be classified as direct, pacing, lapping, or random based on the Martino-Saltzman (MS) model. Method: We take the movement data of five nursing home residents with dementia, comprising 220 travel episodes of room-to-room movements, and manually applied MS model to classify the travel patterns in each episode. Next, we propose two approaches to automatically classify the travel patterns: a machine learning approach and a deterministic predefined tree-based algorithm. In the machine learning approach, eight classical algorithms including Naïve Bayes (NB), Multilayer Perceptron (MLP), Pruned decision trees (C4.5), Random Forests (RF), Logiboost (LB) and Bagging (BAG) with pruned C4.5 trees as base classifiers, k-Nearest Neighbor (k-NN) with one neighbor, and Support Vector Machine (SVM) are employed. Results: RF yields the best classification results. The sensitivity, specificity, precision, recall, F1-measure of the RF are 92.3%, 92.3%, 92.2%, 92.3%, 92.2% respectively. The best classification latency, which is 0.01s, is achieved by NB, C4.5, BAG, and k-NN. In the deterministic approach, we have developed a set of predefined tree-based algorithms to rectify the shortcomings of classical machine learning algorithms. Experimental results indicate that the deterministic algorithm is able to classify direct and various models of indirect travel with 98.2% sensitivity, 98.1% specificity, 98.2% precision, 98.2% recall, 98.2% F1-measure, and 0.0003s classification latency. Conclusion: The deterministic algorithm is simple to implement and highly suitable for real time applications aiming to monitor wandering behavior of PWD in long term care settings.