Using machine learning to predict older adults’ movement behaviours
The number of older adults continues to grow at an unprecedented rate globally, with individuals who are 60 years or older accounting for 8.5% (617 million) of the populace in 2016; and projected to rise to 17% (1.6 billion) by 2050. Such an increase in the ageing population presents several public health challenges, thus positive lifestyle is often encouraged among older adults to maintain good health, functionality and independent living.
Body-worn accelerometers are the most popular method for objectively assessing physical activity (PA) in older adults. On that note, the results presented are motivated by a recent study by Sanders et al. which remains the only study to have used a post-data collection analytical process to estimate generic cut-points for sedentary behaviour (SB) and moderate to vigorous physical activity (MVPA) in older adults. Sanders et al. used the GENEActive (wrist worn) and ActiGraph (hip worn) accelerometers to produce outputs from a heterogeneous sample of 34 older adults (mean age = 69.6, SD = 8.0) to determine raw acceleration cut-points for SB and MVPA. The results were promising but left some areas to improve upon, such as:
- The use of a single feature for PA characterisation, i.e., ENMO values derived from the accelerometer.
- The generic cut-point approach assumes that ‘one size fits all’ which is rarely the case in real life.
- The potential bias in their validation approach, because they calibrated and tested with data from the same group of individuals.
The current study expands the work conducted by Sanders et al. and proposes a machine learning method for personalising activity intensity cut-points for older adults according to their individual characteristics (e.g. age, gender, weight, height, and BMI).
The current study answers the following research questions (RQ) and makes the following contributions:
- Is post data collection analysis using machine learning to predict personalised PA acceleration cut-points for older adults more accurate than the evaluation of raw accelerometry approach conducted by Sanders et al?
To the best of our knowledge, this is the first study to personalise activity classification thresholds among this age group using a standardised approach. Our results produced lower accuracy values than those previously reported. In other words, the ‘one size fits all’ assumption is clearly impractical in real world scenarios.
- Is our approach comparable to the state-of-the-art?
We predicted cut-points for SB and MVPA based on the general characteristics of a person such as age, gender, weight, height, BMI etc. This contradicts the state-of-the-art, where accelerometry data is known and used to calibrate cut-points. Our results showed only minor improvements over the state-of-art model. It is still an improvement, nonetheless..
- Are personalised accelerometer cut-points feasible and superior?
For the ActiGraph device, the Standard Error of estimation from the machine learning approach was lower by 0.33 (Youden optimised SB), 9.50 (Sensitivity optimised SB), 0.64 (Youden optimised MVPA), and 22.11 (Specificity optimised MVPA). Likewise, the Standard Error of estimation from the machine learning approach was lower for the GENEActiv device by 2.29 (Youden optimised SB), 41.65 (Sensitivity optimised SB), 4.31 (Youden optimised MVPA) and 347.15 (Specificity optimised MVPA).
An advantage of the machine learning approach is that it can be easily replicated, thereby providing greater methodological transparency and improved comparability between different studies and accelerometer devices.
Overall, results indicate that personalised PA acceleration cut-points obtained through machine learning are a feasible and superior alternative to generic, evaluation cut-points. The personalisation approach presented prove absolute superiority over the state-of-the-art. We believe that a larger training dataset would lead to further improvement and thus, a result that can be generalised.
The results are very promising, especially when we consider that the machine learning methodology predicts cut-points without prior knowledge of accelerometry data, unlike the state-of-the-art, raw accelerometry methodology conducted by Sanders et al. More data is required to expand the scope of the experiments presented in this paper.
George Sanders is a Research Fellow in Pubic Health and Obesity in the Carnegie School of Sport at Leeds Beckett University. After completing undergraduate and masters degrees at Durham and Loughborough Universities, respectively, his academic career started at Edge Hill University where he completed a Sport England funded PhD with Professor Stuart Fairclough.