TY - JOUR
T1 - Activity detection and classification from wristband accelerometer data collected on people with type 1 diabetes in free-living conditions
AU - Cescon, Marzia
AU - Choudhary, Divya
AU - Pinsker, Jordan E.
AU - Dadlani, Vikash
AU - Church, Mei Mei
AU - Kudva, Yogish C.
AU - Doyle, Francis J.
AU - Dassau, Eyal
N1 - Funding Information:
The determination of appropriate nutritional and insulin dosing strategies in daily life depends on the sound understanding of what, or a combination of, energy systems and substrates are used to support the activities carried out. In the first place this entails being able to recognize the specific activity and exercise performed, in addition to their intensity level and duration. Prompted by this, we used pattern recognition methods [29,30] of diverse nature to classify subject behavior. The methods we considered were linear models such as Logistic Regression (LR) [31], sparse kernel machines such as Support Vector Machine (SVM) [32], tree-based methods such as the RF [27], probabilistic classifiers as the Naive Bayes (NB) [30] and instance-based learning such as k-nearest neighbor (k-NN) [33]. We trained the models on the training dataset of each subject, with our curated best-fit features, and compared their relative performances with a 10-fold cross-validation method. In addition to the above mentioned methods, we have implemented 2 approaches based on deep learning (DL): the Multilayer Perceptron (MLP) classifier, and an artificial recurrent neural networks (RNN) model with long short-term memory (LSTM) [34]. For the first DL-based method, we used the MLP classifier from the scikit-learn library in Python. MLP does not take into account the time-correlations inherent in the dataset and hence, we took the approach of describing each time point by the features that were crafted for the machine learning algorithms described hereinabove. We evaluated MLP classifiers with different sizes of hidden layers, and noticed an increase in accuracy of classification when increasing the number of layers. The optimized configuration consisted in 4 hidden layers activated by Rectified Linear Units (ReLu) activations. As for the second DL approach, we preprocessed the data structure to make it fit to the LSTM model such that a rolling window with 50% overlap was applied all along the data for each participant. This allowed the creation of a sample-wise dataset in which each sample is a window of data with length 128 timestamps (4 s) and width equal to the number of features we have already extracted. Using one-hot encoding, we created an output dataset corresponding to the input and adjusted to the number of activity classes for each patient, such that each input sample is assigned an output sample based on the corresponding activity the patient had at that window of time. We used Keras API for training a model containing one LSTM layer with 128 units followed by a hidden dense layer of 128 unit. In the output, a dense layer containing neurons equal to the number of activity classes for each patient is followed by a SoftMax activation function along with a categorical cross entropy loss function which for each input, outputs the probability of belonging to each class. For both MLP and LSTM we considered an 80/20% split ratio between train and test sets, respectively. For the training process, the train-set was further divided into subsets of training and validation with ratio of 80% and 20%, respectively. Details on the classifiers and parameters used are reported in Appendix A. Fig. 3 outlines the step-wise procedure implemented for behavior classification.Funding for this project was made possible through collaboration between the Juvenile Diabetes Reasearch Foundation and The Leona M. and Harry B. Helmsley Charitable Trust (Grant 2-SRA-2017-503-M-B), and the National Institutes of Health (Grant DP3DK113511). D.C. received support from the Khorana Program for Scholars. J.P. was supported in part by a grant from the William K. Bowes Jr. Foundation (WKB-2017-22754).Dr. Pinsker reports receiving grant support, provided to his institution, and consulting fees and speaker fees from Tandem Diabetes Care; grant support, provided to his institution, and advisory board fees from Medtronic; grant support, provided to his institution, and consulting fees from Eli Lilly; grant support and supplies, provided to his institution, from Insulet; and supplies, provided to his institution, from Dexcom.Dr. Kudva reports product support from Dexcom, Roche Diabetes and Tandem.Dr. Dassau reports receiving grants from JDRF, NIH, and Helmsley Charitable Trust, personal fees from Roche and Eli Lilly, patents on artificial pancreas technology, and product support from Dexcom, Insulet, Tandem, and Roche. Dr. Dassau is currently an employee and shareholder of Eli Lilly and Company. The work presented in this manuscript was performed as part of his academic appointment and is independent of his employment with Eli Lilly and Company.
Funding Information:
Funding for this project was made possible through collaboration between the Juvenile Diabetes Reasearch Foundation and The Leona M. and Harry B. Helmsley Charitable Trust (Grant 2-SRA-2017-503-M-B ), and the National Institutes of Health (Grant DP3DK113511 ). D.C. received support from the Khorana Program for Scholars. J.P. was supported in part by a grant from the William K. Bowes Jr. Foundation ( WKB-2017-22754 ).
Publisher Copyright:
© 2021 Elsevier Ltd
PY - 2021/8
Y1 - 2021/8
N2 - This paper introduces methods to estimate aspects of physical activity and sedentary behavior from three-axis accelerometer data collected with a wrist-worn device at a sampling rate of 32 [Hz] on adults with type 1 diabetes (T1D) in free-living conditions. In particular, we present two methods able to detect and grade activity based on its intensity and individual fitness as sedentary, mild, moderate or vigorous, and a method that performs activity classification in a supervised learning framework to predict specific user behaviors. Population results for activity level grading show multi-class average accuracy of 99.99%, precision of 98.0 ± 2.2%, recall of 97.9 ± 3.5% and F1 score of 0.9 ± 0.0. As for the specific behavior prediction, our best performing classifier, gave population multi-class average accuracy of 92.43 ± 10.32%, precision of 92.94 ± 9.80%, recall of 92.20 ± 10.16% and F1 score of 92.56 ± 9.94%. Our investigation showed that physical activity and sedentary behavior can be detected, graded and classified with good accuracy and precision from three-axial accelerometer data collected in free-living conditions on people with T1D. This is particularly significant in the context of automated glucose control systems for diabetes, in that the methods we propose have the potential to inform changes in treatment parameters in response to the intensity of physical activity, allowing patients to meet their glycemic targets.
AB - This paper introduces methods to estimate aspects of physical activity and sedentary behavior from three-axis accelerometer data collected with a wrist-worn device at a sampling rate of 32 [Hz] on adults with type 1 diabetes (T1D) in free-living conditions. In particular, we present two methods able to detect and grade activity based on its intensity and individual fitness as sedentary, mild, moderate or vigorous, and a method that performs activity classification in a supervised learning framework to predict specific user behaviors. Population results for activity level grading show multi-class average accuracy of 99.99%, precision of 98.0 ± 2.2%, recall of 97.9 ± 3.5% and F1 score of 0.9 ± 0.0. As for the specific behavior prediction, our best performing classifier, gave population multi-class average accuracy of 92.43 ± 10.32%, precision of 92.94 ± 9.80%, recall of 92.20 ± 10.16% and F1 score of 92.56 ± 9.94%. Our investigation showed that physical activity and sedentary behavior can be detected, graded and classified with good accuracy and precision from three-axial accelerometer data collected in free-living conditions on people with T1D. This is particularly significant in the context of automated glucose control systems for diabetes, in that the methods we propose have the potential to inform changes in treatment parameters in response to the intensity of physical activity, allowing patients to meet their glycemic targets.
KW - Artificial pancreas
KW - Automated insulin delivery
KW - Free-living conditions
KW - Physical activity
KW - Supervised learning
KW - Type 1 diabetes mellitus
KW - Wearable devices
KW - Wrist-worn accelerometer
UR - http://www.scopus.com/inward/record.url?scp=85111531677&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85111531677&partnerID=8YFLogxK
U2 - 10.1016/j.compbiomed.2021.104633
DO - 10.1016/j.compbiomed.2021.104633
M3 - Article
C2 - 34346318
AN - SCOPUS:85111531677
SN - 0010-4825
VL - 135
JO - Computers in Biology and Medicine
JF - Computers in Biology and Medicine
M1 - 104633
ER -