Developing and validating an accelerometer-based algorithm with machine learning to classify physical activity after acquired brain injury

ABSTRACT

Purpose: To develop and validate an accelerometer-based algorithm classifying physical activity in people with acquired brain injury (ABI) in a laboratory setting resembling a real home environment.

Materials and methods: A development and validation study was performed. Eleven healthy participants and 25 patients with ABI performed a protocol of transfers and ambulating activities. Activity measurements were performed with accelerometers and with thermal video camera as gold standard reference. A machine learning-based algorithm classifying specific physical activities from the accelerometer data was developed and cross-validated in a training sample of 11 healthy participants. Criterion validity of the algorithm was established in 3 models classifying the same protocol of activities in people with ABI.

Results: Modeled on data from 11 healthy and 15 participants with ABI, the algorithm had a good precision for classifying transfers and ambulating activities in data from 10 participants with ABI. The weighted sensitivity for all activities was 89.3% (88.3–90.4%) and the weighted positive predictive value was 89.7% (88.7–90.7%). The algorithm differentiated between lying and sitting activities.

Conclusion: An algorithm to classify physical activities in populations with ABI was developed and its criterion validity established. Further testing of precision in home settings with continuous activity monitoring is warranted.

KEYWORDS:

• Algorithms [G17.035]
• Brain Injuries [C10.228.140.199]
• Monitoring, Ambulatory [E01.370.520.500)
• Neurological Rehabilitation [E02.760.169.063.500.477]
• Validation Study [V03.950]

Acknowledgments

The authors wish to thank all participants; clinical staff at HNURC who helped facilitate the data collection; MSE Niels Estrup Andersen, who readily offered technical assistance, and Professor Morten Pilegaard for linguistic revision.

Declaration of interest

The authors declare no conflict of interest and no external funding. The study was conducted as part of a PhD-project.

Notes

1. Description of the Random Forest Algorithm:1. If the number of cases in the training set is N, sample N cases at random but with replacement from the original data. This sample will be the training set for growing the tree.2. If there are M input variables/features, a number m ≪ M is specified such that at each node, m variables are selected at random out of the M and the best split on these m is used to split the node. The value of m is held constant during the forest growing.3. Each tree is grown to the largest extent possible. There is no pruning.4. Predict new data by aggregating the predictions of the n trees (majority voting)

2. Sensitivity was the proportion of events classified as the specific physical activity by the algorithm (“positive” events) among events classified as the same activity by the gold standard reference (“true positive” events).Specificity was the proportion of events not classified as the specific physical activity by either method (“negative” events) among events not classified as the same activity by the gold standard reference (“true negative” events).Positive predictive value was the proportion of events correctly classified as the specific physical activity by the algorithm (“positive” events) among events classified as the same activity by the algorithm (“predicted” events).Negative predictive value was the proportion of events not classified as the specific physical activity by either method (“negative” events) among events not classified as the same activity by the algorithm (“not predicted” events).