This project consists of a wearable device for the elderly. Its aim is two-fold: (1) preventing independent people from becoming frail people, and (2) detecting the degradation of one’s health and bring assistance.
New technologies currently allow to track patients from home, but also act as motivational coaches. First, patients become aware of their own health state. Second, they actually improve it to be able to share data with a doctor or even with friends.
Prevention
Age, diseases, environment and social behaviour are among the factors that can result in a vulnerability of the elderly. Whenever one of these factors starts to deteriorate, a sensitivity to stress can be depicted, which then leads to an autonomy loss. Fried’s criteria have been widely used to determine whether a patient is still independent, becoming to lose his autonomy or even frail:
- Slow Walking Speed ( < 1.0 m/s, see below)
- Decrease in Physical Activity
- Unintentional Loss of Weight
- Fatigue, Low Cardio, Lack of Energy
- Decrease in Caloric Expenditure
- Low Grasp Force
No criterion depicts a Robust person; One or two criteria depict a Pre-Frail person; Three criteria depict a Frail person.
In 2015, in France, 50% of the > 65 year-old were considered as pre-frail, whilst 15% were frail.
Frailty is defined as a state of pre-dependence, however it is reversible with a walking-based fitness program.
The RESPECT project is hence born from a need to translate patients’ behaviour data into readable data for doctors to analyse.
On the users’ side, it aims to act as a motivational coach. On the doctors’ side, it aims at detecting the activity-related Fried’s criteria. On the designers’ side, this is translated into a need to characterise the users’ walking speed, their daily number of steps and the distance they travelled, as well as their weight variation, with a cheap, robust and reliable disappearable* system.
*A disappearable is a wearable that is discreet enough to fully disappear into an everyday clothing item, while achieving its tasks.
Detection
My work on this project was regarding the detection of weight loss and the daily step count. An alert must be sent out to the patient’s doctor when an important weight loss is noticed (5 to 10% of the global person’s weight). We defined that the minimum requirement for the system was to detect at least a 3kg weight loss.
But when to measure weight to actually be able to notice a weight loss?
This measure needs to be taken in the same conditions, at a normal speed walk, on flat surfaces. I hence developed a classification method for detecting this condition against walking at high/low speeds, in staircases, or when descending/ascending slopes.
First, the classification consisted in analysing which data was the most promising to get our results. I compared multiple Machine Learning algorithms (LDA, Naive Bayes, SVM, K-neighbours, Trees) and found out that using the best methods (LDA linear, Naive Bayes, Weighted Distances, Classification tree) were the best using a single accelerometer (80% prediction accuracy), an accelerometer coupled with two pressure sensors (under the heel and the foot metatarsals) – 95% prediction accuracy or with only pressure data (97% prediction accuracy).
After noticing the redundancy of standard deviations and means in the previous algorithms, I computed these indicators for different walking stages. I then proposed a classification tree based on the acceleration amplitude’s standard deviation to define the types of walks and their speed.
I then refined the algorithm for counting the users’ steps. The previous algorithm relied on numerical thresholds and delays between strides, which could not be scaled to any user or type of walk. I hence defined a stride detection algorithm, efficient for all types of walks (even in staircases or while running). The precision of this algorithm was above 95% accuracy.
More details are available here or in my internship report (in French).