[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Next Article in Journal
Sensor Network Attack Synthesis against Fault Diagnosis of Discrete Event Systems
Next Article in Special Issue
Predicting the Healing of Lower Extremity Fractures Using Wearable Ground Reaction Force Sensors and Machine Learning
Previous Article in Journal
Genetic Algorithm-Based Cooperative Coding and Caching Data Dissemination Scheme in Multi-UAV-Enabled Internet of Vehicles
Previous Article in Special Issue
Using Computer Vision to Annotate Video-Recoded Direct Observation of Physical Behavior
You seem to have javascript disabled. Please note that many of the page functionalities won't work as expected without javascript enabled.
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Sensors
Volume 24
Issue 14
10.3390/s24144444
sensors-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

On-Device Semi-Supervised Activity Detection: A New Privacy-Aware Personalized Health Monitoring Approach

by
Avirup Roy
,
Hrishikesh Dutta
,
Amit Kumar Bhuyan
and
Subir Biswas
*
Department of Electrical and Computer Engineering, Michigan State University, East Lansing, MI 48824, USA
*
Author to whom correspondence should be addressed.
Sensors 2024, 24(14), 4444; https://doi.org/10.3390/s24144444
Submission received: 5 June 2024 / Revised: 3 July 2024 / Accepted: 5 July 2024 / Published: 9 July 2024
(This article belongs to the Special Issue IMU and Innovative Sensors for Healthcare)
Browse Figures
Figure 1
<p>A wearable on-device human activity detection system.</p> "> Figure 2
<p>Feature distribution of the three classes.</p> "> Figure 3
<p>Pre-processing and classification pipeline.</p> "> Figure 4
<p>Iterative semi-supervised learning framework.</p> "> Figure 5
<p>Example cluster model evolution with incoming episodes self-trained using iterative semi-supervised learning paradigm.</p> "> Figure 6
<p>(<b>a</b>–<b>h</b>) Impacts of PCA on accuracy parameters for population-based SSL; (<b>a</b>–<b>c</b>): true positive for the three classes; (<b>d</b>–<b>f</b>) false positive for the three classes; (<b>g</b>) overall accuracy; (<b>h</b>) classification rate.</p> "> Figure 7
<p>(<b>a</b>–<b>h</b>) Impacts of PCA on accuracy parameters for distance-based SSL; (<b>a</b>–<b>c</b>) true positive for the three classes; (<b>d</b>–<b>f</b>) false positive for the three classes; (<b>g</b>) overall accuracy; (<b>h</b>) classification rate.</p> "> Figure 8
<p>Impact of PCA on complexity for population-based SSL; (<b>a</b>) computational time; (<b>b</b>) CPU memory usage.</p> "> Figure 9
<p>Impact of PCA on complexity for distance-based SSL; (<b>a</b>) computational time; (<b>b</b>) CPU memory usage.</p> "> Figure 10
<p>Post-convergence accuracy parameters’ results for (<b>a</b>–<b>c</b>) population-based SSL and (<b>d</b>–<b>f</b>) distance-based SSL, with varying <math display="inline"><semantics> <mrow> <mi>α</mi> </mrow> </semantics></math>.</p> "> Figure 11
<p>Total learning time and CPU memory usage for all the episodes using (<b>a</b>) population-based SSL and (<b>b</b>) distance-based SSL, with varying <math display="inline"><semantics> <mrow> <mi>α</mi> </mrow> </semantics></math>.</p> "> Figure 12
<p>Post-convergence accuracy parameters’ results for (<b>a</b>–<b>c</b>) population-based SSL and (<b>d</b>–<b>f</b>) distance-based SSL, with varying <span class="html-italic">Nc</span>.</p> "> Figure 13
<p>Total learning time and CPU memory usage for all the episodes using (<b>a</b>) population-based SSL and (<b>b</b>) distance-based SSL, with varying <math display="inline"><semantics> <mrow> <msub> <mrow> <mi>N</mi> </mrow> <mrow> <mi>c</mi> </mrow> </msub> </mrow> </semantics></math>.</p> "> Figure 14
<p>Post-convergence accuracy parameters’ results for (<b>a</b>–<b>c</b>) population-based SSL and (<b>d</b>–<b>f</b>) distance-based SSL for different clustering methods.</p> "> Figure 15
<p>Total learning time and CPU memory usage for 8262 episodes using (<b>a</b>) population-based SSL and (<b>b</b>) distance-based SSL for different clustering methods.</p> ">
Versions Notes

Abstract

:
This paper presents an on-device semi-supervised human activity detection system that can learn and predict human activity patterns in real time. The clinical objective is to monitor and detect the unhealthy sedentary lifestyle of a user. The proposed semi-supervised learning (SSL) framework uses sparsely labelled user activity events acquired from Inertial Measurement Unit sensors installed as wearable devices. The proposed cluster-based learning model in this approach is trained with data from the same target user, thus preserving data privacy while providing personalized activity detection services. Two different cluster labelling strategies, namely, population-based and distance-based strategies, are employed to achieve the desired classification performance. The proposed system is shown to be highly accurate and computationally efficient for different algorithmic parameters, which is relevant in the context of limited computing resources on typical wearable devices. Extensive experimentation and simulation study have been conducted on multi-user human activity data from the public domain in order to analyze the trade-off between classification accuracy and computation complexity of the proposed learning paradigm with different algorithmic hyper-parameters. With 4.17 h of training time for 8000 activity episodes, the proposed SSL approach consumes at most 20 KB of CPU memory space, while providing a maximum accuracy of 90% and 100% classification rates.

1. Introduction

An increasingly sedentary lifestyle in modern societies has compounded many health problems such as diabetes, obesity, and high blood pressure. The high-resolution monitoring of daily activities and providing personalized feedback are found to be effective in reducing sedentary lifestyle. There are many commercially available wearable products such as the Apple Watch [1] and Fitbit watch [2] that can monitor and classify human activities using supervised machine learning methodologies. Such learning typically requires model training using pre-labelled data collected from many users. Once centrally trained, the model is loaded in those wearable devices before consumers can use them.
While this approach of model training works, the design cycle can be improved in the following two fronts. First, the activity data for supervised learning model training can potentially raise privacy concerns. This is because the activity data from a large population needs to be recorded for training. Inadvertent access to such personal fitness- and health-oriented data can be abused by insurance companies and organizations during job placement and hiring. Second, the model used in commercially available wearable is trained using data collected from a generalized population. This contrasts with model training in a subject-specific manner, which is known to provide better classification accuracies for activity monitoring applications.
This paper sets out to address these two design issues by employing a self-supervised activity classification approach in which activity data are recorded from the same subject for whom a classification model is trained iteratively. By not relying on data from others, it addresses the privacy concerns, and by training a semi-supervised model with the same person’s activity data, thus making it personalized. In the proposed approach, the entire learning process happens on-device, thus eliminating the needs for (i) transferring data over network links and (ii) storing and processing it on an out-of-device system. These further add to privacy preservation. The tradeoff is that the mechanism needs to be computationally and storage-wise light enough to be able to run on low-power embedded devices.
The on-device semi-supervised learning (SSL) [3] approach is specifically designed for learning with sparsely labelled datapoints. This light-weight mechanism is fully implemented in embedded wearable devices. The targeted application scenario is as follows: When a user starts using such a device, the semi-supervised learning algorithm in it starts “training itself” with the activity data collected from the device-integrated inertial measurement sensors (IMU) such as accelerometers. Based on pre-stored templates, if the algorithm can classify a current activity, it labels these data. As the device is used, this mechanism is used to label a small subset of all collected activity data. The self-supervised training and classification are performed simultaneously in an iterative manner. As the device is worn by the user for longer durations, the objective is to improve the classification performance over time. Special care is taken to reduce the computational and storage complexity of the iterative mechanism to keep it suitable for low-power wearable embedded devices. Furthermore, since training and data collection both happen on the same device, the target application setting ensures full privacy by not allowing the user data to ever leave the device. Personalization is achieved since the algorithm that classifies a user’s activities and is trained mainly using the data from the same user.
The main contributions of the paper are as follows First, it introduces a privacy-preserving and personalized semi-supervised learning mechanism that can perform human activity classification with accuracies that are comparable to neural network-based supervised learning. It also explores how this can be achieved with fewer labelled data and with low computational resources represented by computational and memory complexities. Second, two SSL variants, namely, population-based, and distance-based, are explored and evaluated for the application considered in the paper. Finally, all the above mechanisms are experimentally analyzed and evaluated from the standpoints of classifiability, classification accuracy, and computational/storage complexity. Accuracy figures are compared with those from benchmark supervised learning models.

2. Related Work

Human activity detection (HAD) in commercially available devices such as smartphones and smart watches is often performed with embedded sensors such as accelerometers and gyroscopes [4]. Different combinations of sensing modalities and their positions on the human body have been explored in the literature [5,6,7,8]. On the machine learning front, different forms of HAD systems have been realized using a variety of classifiers based on supervised learning [7,9]. The above works explore different elements of HAD systems like intraclass variability and interclass similarity. The challenges of disparity in relevant and irrelevant data have also been discussed.
To deal with person-specific activity types, which are often affected by individual physical attributes [10,11] and such, personalized human activity detection has been explored in [12,13,14,15,16,17,18,19,20,21]. While ref. [12] deals with it using biometry, refs. [13,14,15,16,17,18,19,20,21] deal with it at the classifier level. The works in [13,16] train supervised learning models using personalized data from the target user. Both of these works involve the transfer of the data collected from the target user to be transferred to an external server for supervised training. This method affects user privacy since the data are shared to an external device. Some relevant works in [17,18,19] consider transfer learning in a way of domain adaptation methods. Transfer learning has the convenience where explicit training data are not required from the target user device but has the disadvantage of the increased likelihood of selecting incorrect class data for training, thus degrading the overall classification performance. Refs. [20,21] present similar works which consider personalized learning models for classification tasks.
On-device learning is an effective solution for applications related to personalized applications. A relevant work in [22] uses similar pre-trained supervised learning models for on-device hydration-tracking. This hydration-tracking system also preserves the privacy of the target user data, but the usage of the supervised learning model requires a huge amount of training data. At times enough labelled data may not be available from the target user to train a supervised learning model. In this way, either privacy or personalization or both for a classifier model may be sacrificed, which becomes a point of major concern for these kinds of applications which require both privacy and personalization.
Semi-supervised learning (SSL) [3] is a possible solution for on-device learning in which classifiers can be trained on the device from a specific target user’s data, thus preserving the personalization feature of the model as discussed in [23]. SSL is also useful in this context because unlike in fully supervised learning, learning with SSL does not rely on the availability of extensive labelled data. It was shown in [24] that SSL models can potentially learn with sparsely labelled datapoints. This mechanism has proven to be effective for on-device applications involving limited computational power. A relevant work in [25] uses an SSL method based on the combination of deep learning and transfer learning. The usage of a deep learning model involves training using data from different users other than the target user. This affects the personalization feature of our premise in this paper. The work in [26] develops a human activity detection system using a bi-view semi-supervised learning to detect semantic human activities like having dinner, shopping, etc. This method also uses a windowed datapoint extraction technique and clustering mechanism as the basis of the classifier model, but not on the same device where the sensors are present. This method also involves a two-layered framework for the classification task, which becomes computationally expensive and is thus not very suitable for on-device self-training and classification.
The above works in the literature provide many definitive ideas for classifying human activities with or without using personalized classifiers. While semi-supervised learning has been proposed as an effective tool to train personalized classifiers of HAD, none of these works considers the feasibility of training on devices with computational constraints. Also, the aspect of computational complexity is absent in those approaches. The work in this paper addresses both these issue by developing a low-computation self-training mechanism using a semi-supervised learning algorithm. Both classification accuracy and computational complexity are considered as evaluation parameters. The different hyper-parameters and system parameters are analyzed based on their relevance for implementation on wearable devices with computational constraints.

3. System and Data Model

This section outlines the adopted system and data model used by the proposed semi-supervised algorithms. Figure 1 depicts the entire system where the wearable device on the wrist of the user reads the human motion using embedded IMU sensors. The different classes of human activities are classified using the semi-supervised learning model self-trained using the motion data from the same user on the same wearable device.

3.1. Sensing Modality and Dataset

Activity classification is performed on data collected from a wearable embedded device containing Inertial Measurement Unit (IMU) sensors, storage, and a micro-controller. The IMU sensors capture user movements in terms of the acceleration of the wearable unit along three orthogonal axes. We have used a public-domain human activity dataset (i.e., Wireless Sensor Data Mining Lab-WISDM lab dataset from the University of California, Irvine [27]) generated from such a wearable unit. The dataset contains accelerometer time-series data collected from a smartwatch as 36 subjects perform six activities.

3.2. Segmentation and Feature Extraction

As the first processing step, the raw accelerometer data are pre-processed as follows: First, l 2 norm [28] is applied to standardize the data across all axes in order to ensure that the data are bounded within the range of −1 to +1. Subsequently, a window-based approach is used for segmenting data into 2 s long episodes. For each episode, two features are defined for each of the accelerometer axes. The first feature is the coefficient of variation, which reflects the acceleration variability across 40 samples spanning the 2 s long episode. The second feature is the number of mean crossing points within an episode. Physically, this one indicates the frequency of movements of the subject from its mean position. There may be a possibility of an exception occurring when there is no significant movement across the mean position in either of the three orthogonal directions. In order to handle this, ‘0’ has been set as the feature value for these special cases. Based on their physical implications, these two features are selected for their discriminatory abilities towards activity classification. Windowing and feature extraction from the raw time-series data helps in reducing the impact of noises in the raw data on the model. Windowing the time series into a fixed number of samples prevents the accumulation of errors which may happen in a continuous time series when the raw samples are provided to the model as inputs. Extracting features from consecutive time series’ samples means representing the set of samples by some feature values, which considers only the magnitude of significant movement by the subject in a particular direction during the 0.5 s episode duration. These feature values are fed to the learning model, which in turn helps in avoiding errors (if any) and any kind of model bias.

3.3. Class Definition

Sedentary lifestyle has been attributed [29] as a main contributor to widespread obesity in modern urban societies in recent years. Monitoring and detecting sedentary lifestyles have been projected by the health research community as a useful tool to mitigate obesity and its many health effects. This has led researchers in recent years to actively investigate technology for classifying sedentary and non-sedentary activities [30,31]. Motivated by these works and the research trend, this paper sets out to deal with the problem of detecting sedentary activities but in a more personalized and computationally efficient manner. User activities are categorized into three distinct classes, namely, sedentary, moderately active, and active. The sedentary class encompasses stationary activities such as standing and sitting. Moderately active activities involve ambulatory motion, primarily represented by walking. Finally, the active class pertains to more vigorous physical activities, such as jogging. The correlation among all the features over the entire activity dataset is depicted in Figure 2.

3.4. Processing Pipeline

Figure 3 depicts the entire processing dataflow for the proposed semi-supervised learning mechanism. It starts with acceleration sensing, with a sampling rate of 20 Hz. The resulting time series is then segmented into 2 s long episodes. The data in each window are subsequently normalized before extracting features as described above. The dimensions of the extracted features are then reduced for managing the computational complexity of processing. Such dimensionality reductions are achieved by choosing the top three principal components after applying principal component analysis (PCA) [32]. The proposed iterative semi-supervised learning (ISSL) is then applied in the presence of a small pool of pre-labelled data. Details of the ISSL process including adaptive data clustering and dynamic cluster-labelling are presented in the next section. This processing pipeline enables continuous improvement in activity prediction accuracy over time, thus providing a real-time mechanism to classify human activity with two-second temporal granularity.

4. Semi-Supervised Learning with Sparsely Labelled Datapoints

Figure 4 depicts the iterative semi-supervised learning framework which forms the basis of the classification model for human activities. The pre-processed accelerometer data arrives in the form of episodes of 2 s worth of data. The episodes are represented by the extracted features followed by a dimensionality reduction step. Reducing the dimensionality helps to reduce the computational load of the learning system, as described in Section 3. Each unlabelled incoming episode is added to a data pool ( S p o o l ). A very small percentage ( α ) of incoming episodes which have extreme feature values beyond a threshold are pre-labelled with a designated activity class and added to a pre-labelled data pool ( L S p o o l ). These are considered pre-labelled because they can be easily identified to be the parent of a class based on their extreme feature values matching to the corresponding class. These pre-labelled episodes are used in the SSL training phase for the cluster labelling step. The SSL training and activity prediction is performed on the unlabelled data pool ( S p o o l ). The training is performed after every N e incoming episode. In the training step, first, all the existing episodes in the S p o o l are grouped into N c number of clusters based on their feature similarity in the clustering step. This clustering is performed by some popular clustering methods like k-means [33] or the Gaussian mixture model (GMM) [34]. There is the possibility of an exception when the number of datapoints in S p o o l (count) is less than the number of clusters ( N c ). In those cases, there might be a run-time error during the clustering process. In order to handle this, a condition is placed where SSL training will only happen when the ‘count’ is greater than or equal to N c . The N c clusters are labelled using the pre-labelled episodes ( L S p o o l ), using any of the two cluster labelling methods: population-based labelling and distance-based labelling.
In the population-based labelling, the cluster is labelled on the basis of the highest number of pre-labelled episodes present in that cluster. For example, if cluster i has M 1 pre-labelled episodes from class A and M 2 pre-labelled episodes from class B , and if M 1 > M 2 , then cluster i gets labelled as class A . If cluster i does not have the highest number of pre-labelled episodes class of pre-labelled episodes or does not have a pre-labelled episode at all, it remains unlabeled. Thus, all the episodes in that cluster remain unclassified. This cluster labelling mechanism is shown in Algorithm 1.
Algorithm 1: Population-based cluster labelling
1. Input:  L S p o o l , Cluster representatives (CR) of N c clusters.
2. Output:  Labels of N c clusters.
3. For i = 1   t o   N c
4.    Calculate the number of pre-labelled samples in i -th cluster M 1 , M 2 ,…, M k , for k classes in i -th cluster.
5.    If max ( M ) exists://max ( M ) is the unique class with the maximum number of pre-labelled episodes
6.         Label cluster i with label of m a x ( M )
7.      Else:
8.            Cluster i remains unlabeled.
9.        End If-Else
10.   End For
In distance-based labelling, the cluster is labelled based on label of the closest pre-labelled episode. For example, if cluster i has the minimum distance from the j -th pre-labelled episode which belongs to class A, then i is labelled class A . No clusters remain unlabeled in this variant of cluster labelling, until there are pre-labelled datapoints in the pre-labelled data pool ( L S p o o l ). The clusters will remain unlabeled when there are no pre-labelled datapoints. This cluster labelling mechanism is shown in Algorithm 2.
Algorithm 2: Distance-based cluster labelling.
1.  Input:  L S p o o l , Cluster representatives (CR) of N c clusters.
2. Output:  Labels of N c clusters.
3. For i = 1   t o   N c
4.    Initialize the distance of cluster representative- i   ( C R i ) from pre-labelled samples as D ;
5.    For j = 1   t o   L S p o o l // L S p o o l consists of the pre-labelled datapoints
6.        Calculate distance of pre-labelled sample j from the C R i as D j
7.    End For
8.     D ( i ) ← m i n ⁡ ( D j ) ;//pre-labelled episode number with the minimum distance from ‘i’-th CR.
9.    Label cluster ‘ i ’ with the label of ‘ j ’-th pre-labelled episode
10. End For
After the clusters are labelled using any one of the above two cluster labelling algorithms, the labelled clusters act as the trained classifier model. When a newly arrived and unseen episode comes in between training iterations, the episode is classified based on the most recently trained model, which are the labelled clusters. The new episode obtains the class label of the nearest cluster in the feature space. A few representative cluster models trained with incoming movement data from a user are shown in Figure 5. It can be seen how the model with 20 clusters expands and the classification accuracy improves with more incoming episodes. The quantitative performance of the classifier is formally presented in the next section.
Personalization and Privacy: The 2 seconds’ episodes represented by the extracted features are classified by the semi-supervised model embedded in a wearable device. The model is iteratively trained as more labelled episodes, which are very infrequent, come in. It also classifies unlabeled episodes in between those infrequent training episodes. Since both training and classification happen for the same user, the semi-supervised model acts in a very personalized manner. Since in this process no data need to be transferred outside the embedded wearable device, which performs both model training and classification, privacy of the user’s movement data is automatically preserved. This makes the approach personalized and privacy preserving at the same time.

5. Experiments and Results

An open-source human activity dataset (i.e., Wireless Sensor Data Mining Lab-WISDM lab dataset from the University of California, Irvine [27]) has been used for training and validation of the iterative semi-supervised learning model. Three broad classes of activities viz. sedentary, moderately active and active have been used from testing the proposed model as discussed in Section 3. The time-series data have been pre-processed and windowed into two-minute episodes. Six features, as described in Section 3, are extracted from each episode. Moreover, 8262 episodes (2754 from each class of activities) are used for the model evaluation experiments of the proposed iterative semi-supervised learning framework explained in Section 4.
The SSL model is trained using three principal components (i.e., using (PCA) [32]) extracted from the six features discussed earlier. Experiments involving different clustering algorithms, the hyper-parameter number of clusters ( N c ), and system parameter α (i.e., % of pre-labelled datapoints) have been performed to evaluate the accuracy and complexity of the proposed SSL model. The accuracy metrics are true positive (TP), false positive (FP), overall accuracy (Acc), and the classification rate, which is defined as the percent of classified datapoints among the actual number of datapoints. The computational complexity of the model is represented by the computational time and the CPU memory usage for each learning cycle. The accuracy and complexity are analyzed based on the performance of the SSL algorithm in order to train 8262 episodes (datapoints) for 100 runs. All the results involving the SSL algorithm are means of the results collected from 100 independent runs.

5.1. Pre-Trained Supervised Learning Model as a Benchmark

A pre-trained supervised model evaluated with a 10-fold cross-validation using the same dataset has been used as a benchmark for the proposed semi-supervised learning framework. The six extracted features mentioned in Section 3 have been directly used as the input to the NN model. The NN model is trained using the hyper-parameters mentioned in Table 1.
The mean performance (as mentioned in Table 2) of each accuracy parameter for all 10-fold cross-validation tests is used as a performance benchmark for the classification accuracy of the proposed model. It is observed that the NN model has mean true positives of 98.47%, 97.41%, and 98.19% for the three classes (i.e., sedentary, moderately active, and active), respectively. The mean overall accuracy is 98.03%.

5.2. Impacts of Feature Dimensionality Reduction on SSL

The primary motivation of the proposed self-training algorithm is its suitability for the wearable system and its privacy-preserved personalized applications. Ideally, the semi-supervised learning-based self-training model needs to be computationally light due to the resource-constrained nature of the embedded wearable devices. To that end, PCA has been used to reduce the number of features from six to three. Figure 6 and Figure 7 depict the differences in accuracy parameters (TP, FP, Acc, and λ ) for training the SSL model with and without using PCA, using the population-based and distance-based SSL algorithms, respectively. The results are presented for the two presented algorithm viz. population-based and distance-based. Moreover, 20 clusters formed by the k-means clustering algorithm are used for these experiments.
For both SSL algorithms, true positives (TP) and false positives (FP) for all the three classes are almost same for both the scenarios with and without using PCA. The overall accuracy (Acc) and the classification rate ( λ ) are also the same for both algorithms. The only observable difference is in the stability in TP, FP, and Acc performance with more incoming episodes. The performances with PCA have less oscillatory behavior after convergence compared to the case not using PCA. This makes it evident that a feature space with less dimensions makes the cluster models more definite in less time.
Figure 8 and Figure 9 present the impacts of dimensionality reduction from the six original features to three principal components. It can be observed that more dimensions (features) of a data sample results in higher computational complexity. Thus, for both, the types of SSL algorithms, both computational time, and CPU memory usage go up when PCA is not used. Thus, the experiments analyzing the different hyper-parameters and system parameter of the SSL model are carried out using the three principal components obtained by using PCA on the original six features.

5.3. Impacts of Pre-Labelled Data Volume on SSL Performance

Figure 10 depicts the TP, FP, overall accuracy, and classification rate for SSL operating with different amounts of pre-labelled data, which is represented as α (in percentage). These experiments are performed with 20 clusters formed by the k-means clustering method in the SSL model. It can be observed that all the accuracy parameters converge faster and reach better performance values with increasing values of α . For the population-based SSL, lower α   (i.e., 0.05%, 0.1%, 0.5%, and 1%) values are unable to achieve a 100% classification rate. The reasons are as follows: First, many clusters may the lack predominance of prelabelled datapoints from a particular class. Second, a cluster may not have any pre-labelled episodes altogether. All the datapoints in those clusters remain unlabeled and thus unclassified. Figure 10f shows that using the distance-based SSL, all the clusters and their datapoints are always labelled since this variant of SSL does not depend on the presence of any pre-labelled datapoints in a cluster. Rather, a cluster is labelled according to the label of the nearest pre-labelled datapoint in the distance-based SSL. The labelling in this mechanism is not impacted by whether the nearest pre-labelled datapoint is present inside or outside the cluster. Also, it can be observed that convergence is achieved faster in the distance-based SSL than in the population-based SSL.
Figure 11a,b depicts the total semi-supervised learning computational complexity for both variants of SSL. Complexity is represented by the total computational time in seconds and CPU memory usage in Bytes. It can be observed that the computational time and CPU memory usage are not dependent on the amount of pre-labelled datapoints for the population-based SSL since the number of pre-labelled datapoints vary in different clusters for different runs of the learning algorithm. On the other hand, it can be observed that more computational time is required for the distance-based SSL to learn 8262 episodes since the greater the number of pre-labelled datapoints, the greater the number of distance calculations required.

5.4. Impacts of the Number of Clusters ( N c ) on SSL Performance

The classification in the proposed model occurs based on the clusters that actually contain datapoints based on their similarities. The number of clusters in the learning model is an important hyper-parameter which can determine the precision of classification as well as the computational complexity to train the cluster model. Figure 12 depicts the accuracy parameters after learning convergence. These results represent the accuracy parameters for both population-based and distance-based SSL algorithms with varying number of clusters in the model. These above results are obtained from experiments performed by setting the system parameter α to 10% and k-means as the clustering algorithm.
It can be observed that true positive (TP) values improve and false positive (FP) values decrease with higher N c for all three classes and for both the variants of SSL. As a result, the overall accuracy improves with a higher number of clusters ( N c ). The overall accuracy performance does not improve much beyond 10 clusters for both population-based and distance-based SSL. Also, population-based SSL has marginally better accuracy (90%) compared to the distance-based SSL (88%). Both the variants of SSL reach a 100% classification rate, but distance-based SSL reaches this faster as all the clusters are labelled from the first cycle of learning as opposed to the population-based SSL.
In the Figure 13a,b, it can be observed that both variants of SSL become computationally expensive with a higher number of clusters. Total CPU memory usage required to train all the episodes of activities is the same for both population-based and distance-based SSL for each value of N c . The total computational time required for iterative learning is marginally higher for the distance-based SSL compared to the population-based SSL. This is mainly because population-based SSL only counts the pre-labelled episodes in a cluster which is computationally less expensive than calculating the distance from each pre-labelled datapoint from a cluster center.

5.5. Impacts of Different Clustering Algorithms

Among the popular clustering methods, k-means (KM) and the Gaussian mixture model (GMM) are the best methods for the current approach. GMM has three useful variants based on the relevant covariance types: spherical (GMM_s), full (GMM_f), and diagonal (GMM_d). Based on the accuracy and complexity performances using the different clustering methods mentioned above, the best clustering method suited for the ideal SSL performance is analyzed.
Figure 14a,b,d,e depict the TPs and FPs for all the three classes of activities along with the overall accuracy and classification rate for different clustering methods. The experiments are performed to provide 20 clusters ( N c = 20 ) using the different clustering methods, with α = 10 % . It turns out that GMM_f has the highest TPs and the least FPs for all the three classes. Subsequently, GMM_f has the best overall accuracy (Acc) for both the population-based and the distance-based SSL variants as shown in Figure 14c,f. This is because the full covariance type has each component with their own general covariance matrix, which makes the grouping of datapoints belonging to the same class of activities more precise. Population-based SSL has marginally better overall accuracy compared to distance-based SSL.
Figure 15a,b show that GMM_f is computationally expensive with a very high total computational time and CPU memory usage for learning the entire dataset (8262 episodes). GMM_s has the second best overall accuracy after GMM_f with less computational time. But, the CPU memory usage is still very high. K-means requires much less memory usage compared to the GMM clustering types. Thus, when SSL is implemented on the wearable device with limited CPU memory, k-means should be preferred over GMM.

6. Discussion

This paper presents a fully on-device approach for human activity detection using a semi-supervised learning paradigm. The proposed SSL mechanisms self-trains in a fully personalized and privacy-preserving manner. The performance of the mechanism is compared with that of a pre-trained supervised learning model (NN) which forms a comparison benchmark. The post-convergence performance numbers for the benchmark and the different variants of the proposed SSL algorithm with the best algorithmic parameters (post-convergence) are summarized in Table 3. Since the NN model was pre-trained outside the wearable device, the computational time and CPU memory usage are not considered for the summarization purpose. Advantages of using an unsupervised dimensionality reduction technique prior to the SSL algorithm are also discussed in Section 5.2.
The proposed SSL model is analyzed and validated based on different learning hyper-parameters and system parameters. The results suggest that 5% or higher pre-labelled activity datapoints add precision to the model, therefore improving its classification performance. It has been also observed that using 20 clusters in the model with GMM full-type clustering yields an overall accuracy of 90%, with a 100% classification rate after learning convergence. However, since GMM_full is a computationally expensive method, a k-means clustering method is also explored for resource-limited scenarios. With 4.17 h of training time for 8000 activity episodes the k-means clustering-based SSL approach consumes at most 20 KB of CPU memory space while providing a maximum accuracy of 90% with a 100% classification rate. It is true that the SSL model cannot reach an accuracy (98%) as high as the pre-trained supervised model (NN) where the entire training dataset is pre-labelled. However, the accuracy of the SSL model (90%) with only 10% of pre-labelled datapoints is still comparable and within a practically acceptable range. These results indicate the feasibility of the proposed SSL paradigm for practical human activity detection using a personalized and privacy-preserving health monitoring system.
The proposed semi-supervised learning framework has the following limitations. It can be observed in Figure 8 and Figure 9 that the computational time and CPU memory usage keep increasing monotonically as the learning progresses with the accumulation of incoming episodes (datapoints) in the data pool. As the size of the data pool keeps increasing, more computational resources (time and CPU memory) are required. Thus, a necessary upgrade of the proposed framework is required such that the learning process should stop after the classification model has converged. Following the interest of the sedentary research community [29,30,31], this paper deals with three major relevant classes, namely, sedentary, moderately active, and active. The presented SSL framework can be scaled up to handle more activities beyond the sedentary-targeted clinical applications.

7. Conclusions

This paper presents an on-device semi-supervised human activity detection system that can learn and predict human activity patterns in real time. The proposed semi-supervised learning (SSL) framework uses sparsely labelled user activity events acquired from Inertial Measurement Unit (IMU) sensors installed as wearable devices. The objective is to learn classification in real-time. The proposed cluster-based learning model in this approach is trained with data from the same target user, thus preserving data privacy while providing personalized activity detection. Two different cluster labelling strategies, namely, population-based and distance-based, are employed to achieve the desired classification performance.
A comparative study of the proposed strategy has been presented in terms of classifiability and classification accuracies. Different accuracy parameters for each of the classes of human activities (true positive and false positive) as well as the overall accuracy of classifying all the classes of activities have been considered to evaluate the learning model formed by the proposed SSL framework. The proposed system is shown to be computationally efficient, which is relevant in the context of limited computing resources on typical wearable devices. The trade-off between classification accuracy and computation complexity is analyzed for different algorithmic hyper-parameters (number of clusters) and other system parameters (% of pre-labelled datapoints). Different clustering algorithms like k-means and the Gaussian mixture model (GMM) have also been experimented with as part of the framework evaluation. Computational time and CPU memory usage have been used as parameters to measure the time and space complexity of the proposed learning model for the different algorithmic parameters mentioned above. Extensive experimentation and simulation studies have been conducted on multi-user human activity data from the public domain to validate the proposed learning paradigm. It has been observed that the proposed semi-supervised online learning framework can achieve a maximum classification accuracy of around 90% with 10% pre-labelled as opposed to the pre-trained supervised learning model which achieves 98% classification accuracy with 100% pre-labelled datapoints.

Author Contributions

Conceptualization, A.R., H.D., A.K.B. and S.B.; methodology, A.R.; software, A.R.; validation, H.D., A.K.B. and S.B.; formal analysis, A.R.; investigation, A.R.; resources, A.R. and S.B.; data curation, A.R.; writing—original draft preparation, A.R.; writing—review and editing, H.D., A.K.B. and S.B.; visualization, A.R.; supervision, S.B.; project administration, S.B.; funding acquisition, S.B. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

An open-source human activity dataset (i.e., Wireless Sensor Data Mining Lab-WISDM lab dataset from the University of California, Irvine) has been used for the experiments of this research. The WISDM dataset can be found at https://archive.ics.uci.edu/dataset/507/wisdm+smartphone+and+smartwatch+activity+and+biometrics+dataset (accessed on 4 September 2023).

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Lui, G.Y.; Loughnane, D.; Polley, C.; Jayarathna, T.; Breen, P.P. The Apple Watch for Monitoring Mental Health–Related Physiological Symptoms: Literature Review. JMIR Ment. Health 2022, 9, e37354. [Google Scholar] [CrossRef] [PubMed]
  2. Bunn, J.A.; Navalta, J.W.; Fountaine, C.J.; Reece, J.D. Current State of Commercial Wearable Technology in Physical Activity Monitoring 2015–2017. Int. J. Exerc. Sci. 2018, 11, 503–515. [Google Scholar] [PubMed]
  3. Hady, M.F.A.; Schwenker, F. Semi-supervised Learning. In Handbook on Neural Information Processing; Bianchini, M., Maggini, M., Jain, L.C., Eds.; Intelligent Systems Reference Library; Springer: Berlin/Heidelberg, Germany, 2013; pp. 215–239. [Google Scholar] [CrossRef]
  4. Rasekh, A.; Chen, C.-A.; Lu, Y. Human Activity Recognition using Smartphone. arXiv 2014, arXiv:1401.8212. [Google Scholar] [CrossRef]
  5. Atallah, L.; Lo, B.; King, R.; Yang, G.-Z. Sensor Positioning for Activity Recognition Using Wearable Accelerometers. IEEE Trans. Biomed. Circuits Syst. 2011, 5, 320–329. [Google Scholar] [CrossRef] [PubMed]
  6. Bulling, A.; Blanke, U.; Schiele, B. A tutorial on human activity recognition using body-worn inertial sensors. ACM Comput. Surv. 2014, 46, 1–33. [Google Scholar] [CrossRef]
  7. Wang, Y.; Cang, S.; Yu, H. A survey on wearable sensor modality centred human activity recognition in health care. Expert Syst. Appl. 2019, 137, 167–190. [Google Scholar] [CrossRef]
  8. Twomey, N.; Diethe, T.; Fafoutis, X.; Elsts, A.; McConville, R.; Flach, P.; Craddock, I. A Comprehensive Study of Activity Recognition Using Accelerometers. Informatics 2018, 5, 27. [Google Scholar] [CrossRef]
  9. Nweke, H.F.; Teh, Y.W.; Al-Garadi, M.A.; Alo, U.R. Deep learning algorithms for human activity recognition using mobile and wearable sensor networks: State of the art and research challenges. Expert Syst. Appl. 2018, 105, 233–261. [Google Scholar] [CrossRef]
  10. Del Rosario, M.B.; Wang, K.; Wang, J.; Liu, Y.; Brodie, M.; Delbaere, K.; Lovell, N.H.; Lord, S.R.; Redmond, S.J. A comparison of activity classification in younger and older cohorts using a smartphone. Physiol. Meas. 2014, 35, 2269–2286. [Google Scholar] [CrossRef] [PubMed]
  11. Albert, M.V.; Toledo, S.; Shapiro, M.; Kording, K. Using Mobile Phones for Activity Recognition in Parkinson’s Patients. Front. Neurol. 2012, 3, 158. [Google Scholar] [CrossRef] [PubMed]
  12. Weiss, G.M.; Yoneda, K.; Hayajneh, T. Smartphone and Smartwatch-Based Biometrics Using Activities of Daily Living. IEEE Access 2019, 7, 133190–133202. [Google Scholar] [CrossRef]
  13. Weiss, G.M.; Lockhart, J.W. The Impact of Personalization on Smartphone-Based Activity Recognition. In Proceedings of the Workshops at the Twenty-Sixth AAAI Conference on Artificial Intelligence, Toronto, ON, Canada, 22–26 July 2012. [Google Scholar]
  14. Ferrari, A.; Micucci, D.; Mobilio, M.; Napoletano, P. On the Personalization of Classification Models for Human Activity Recognition. IEEE Access 2020, 8, 32066–32079. [Google Scholar] [CrossRef]
  15. Weiss, G.M.; Timko, J.L.; Gallagher, C.M.; Yoneda, K.; Schreiber, A.J. Smartwatch-based activity recognition: A machine learning approach. In Proceedings of the 2016 IEEE-EMBS International Conference on Biomedical and Health Informatics (BHI), Las Vegas, NV, USA, 24–27 February 2016; pp. 426–429. [Google Scholar] [CrossRef]
  16. Berchtold, M.; Budde, M.; Gordon, D.; Schmidtke, H.R.; Beigl, M. ActiServ: Activity Recognition Service for mobile phones. In Proceedings of the International Symposium on Wearable Computers (ISWC) 2010, Seoul, Republic of Korea, 10–13 October 2010; pp. 1–8. [Google Scholar] [CrossRef]
  17. Siirtola, P.; Koskimäki, H.; Röning, J. From User-Independent to Personal Human Activity Recognition Models Using Smartphone Sensors. Available online: https://oulurepo.oulu.fi/handle/10024/22078 (accessed on 9 May 2024).
  18. Roggen, D.; Förster, K.; Calatroni, A.; Tröster, G. The adARC pattern analysis architecture for adaptive human activity recognition systems. J. Ambient. Intell. Humaniz. Comput. 2013, 4, 169–186. [Google Scholar] [CrossRef]
  19. Cook, D.; Feuz, K.D.; Krishnan, N.C. Transfer learning for activity recognition: A survey. Knowl. Inf. Syst. 2013, 36, 537–556. [Google Scholar] [CrossRef]
  20. Horiguchi, S.; Amano, S.; Ogawa, M.; Aizawa, K. Personalized Classifier for Food Image Recognition. IEEE Trans. Multimed. 2018, 20, 2836–2848. [Google Scholar] [CrossRef]
  21. Cho, Y.; Lee, A.; Park, J.; Ko, B.; Kim, N. Enhancement of gesture recognition for contactless interface using a personalized classifier in the operating room. Comput. Methods Programs Biomed. 2018, 161, 39–44. [Google Scholar] [CrossRef]
  22. Roy, A.; Dutta, H.; Griffith, H.; Biswas, S. An On-Device Learning System for Estimating Liquid Consumption from Consumer-Grade Water Bottles and Its Evaluation. Sensors 2022, 22, 2514. [Google Scholar] [CrossRef] [PubMed]
  23. Singh, G.; Chowdhary, M.; Kumar, A.; Bahl, R. A Personalized Classifier for Human Motion Activities With Semi-Supervised Learning. IEEE Trans. Consum. Electron. 2020, 66, 346–355. [Google Scholar] [CrossRef]
  24. Roy, A.; Dutta, H.; Bhuyan, A.K.; Biswas, S.K. Semi-Supervised Learning Using Sparsely Labelled Sip Events for Online Hydration Tracking Systems. In Proceedings of the 2023 International Conference on Machine Learning and Applications (ICMLA), Jacksonville, FL, USA, 15–17 December 2023; pp. 1799–1804. [Google Scholar] [CrossRef]
  25. Oh, S.; Ashiquzzaman, A.; Lee, D.; Kim, Y.; Kim, J. Study on Human Activity Recognition Using Semi-Supervised Active Transfer Learning. Sensors 2021, 21, 2760. [Google Scholar] [CrossRef] [PubMed]
  26. Lv, M.; Chen, L.; Chen, T.; Chen, G. Bi-View Semi-Supervised Learning Based Semantic Human Activity Recognition Using Accelerometers. IEEE Trans. Mob. Comput. 2018, 17, 1991–2001. [Google Scholar] [CrossRef]
  27. Weiss, G. WISDM Smartphone and Smartwatch Activity and Biometrics Dataset. 2019. Available online: https://archive.ics.uci.edu/dataset/507/wisdm+smartphone+and+smartwatch+activity+and+biometrics+dataset (accessed on 4 September 2023). [CrossRef]
  28. Yang, M.; Meng, Z.; King, I. FeatureNorm: L2 Feature Normalization for Dynamic Graph Embedding. In Proceedings of the 2020 IEEE International Conference on Data Mining (ICDM), Sorrento, Italy, 17–20 November 2020; pp. 731–740. [Google Scholar] [CrossRef]
  29. Martínez-González, M.; Martínez, J.A.; Hu, F.; Gibney, M.; Kearney, J. Physical inactivity, sedentary lifestyle and obesity in the European Union. Int. J. Obes. 1999, 23, 1192–1201. [Google Scholar] [CrossRef]
  30. Kańtoch, E. Recognition of Sedentary Behavior by Machine Learning Analysis of Wearable Sensors during Activities of Daily Living for Telemedical Assessment of Cardiovascular Risk. Sensors 2018, 18, 3219. [Google Scholar] [CrossRef] [PubMed]
  31. Montoye, A.H.; Pivarnik, J.M.; Mudd, L.M.; Biswas, S.; Pfeiffer, K.A. Validation and Comparison of Accelerometers Worn on the Hip, Thigh, and Wrists for Measuring Physical Activity and Sedentary Behavior. AIMS Public Health 2016, 3, 298–312. [Google Scholar] [CrossRef] [PubMed] [PubMed Central]
  32. Greenacre, M.; Groenen, P.J.F.; Hastie, T.; D’Enza, A.I.; Markos, A.; Tuzhilina, E. Principal component analysis. Nat. Rev. Methods Primer 2022, 2, 100. [Google Scholar] [CrossRef]
  33. Sinaga, K.P.; Yang, M.-S. Unsupervised K-Means Clustering Algorithm. IEEE Access 2020, 8, 80716–80727. [Google Scholar] [CrossRef]
  34. Zhang, Y.; Li, M.; Wang, S.; Dai, S.; Luo, L.; Zhu, E.; Xu, H.; Zhu, X.; Yao, C.; Zhou, H. Gaussian Mixture Model Clustering with Incomplete Data. ACM Trans. Multimed. Comput. Commun. Appl. 2021, 17, 1–14. [Google Scholar] [CrossRef]
Sensors 24 04444 g001
Figure 1. A wearable on-device human activity detection system.
Figure 1. A wearable on-device human activity detection system.
Sensors 24 04444 g001
Sensors 24 04444 g002
Figure 2. Feature distribution of the three classes.
Figure 2. Feature distribution of the three classes.
Sensors 24 04444 g002
Sensors 24 04444 g003
Figure 3. Pre-processing and classification pipeline.
Figure 3. Pre-processing and classification pipeline.
Sensors 24 04444 g003
Sensors 24 04444 g004
Figure 4. Iterative semi-supervised learning framework.
Figure 4. Iterative semi-supervised learning framework.
Sensors 24 04444 g004
Sensors 24 04444 g005
Figure 5. Example cluster model evolution with incoming episodes self-trained using iterative semi-supervised learning paradigm.
Figure 5. Example cluster model evolution with incoming episodes self-trained using iterative semi-supervised learning paradigm.
Sensors 24 04444 g005
Sensors 24 04444 g006
Figure 6. (a–h) Impacts of PCA on accuracy parameters for population-based SSL; (a–c): true positive for the three classes; (d–f) false positive for the three classes; (g) overall accuracy; (h) classification rate.
Figure 6. (a–h) Impacts of PCA on accuracy parameters for population-based SSL; (a–c): true positive for the three classes; (d–f) false positive for the three classes; (g) overall accuracy; (h) classification rate.
Sensors 24 04444 g006
Sensors 24 04444 g007
Figure 7. (a–h) Impacts of PCA on accuracy parameters for distance-based SSL; (a–c) true positive for the three classes; (d–f) false positive for the three classes; (g) overall accuracy; (h) classification rate.
Figure 7. (a–h) Impacts of PCA on accuracy parameters for distance-based SSL; (a–c) true positive for the three classes; (d–f) false positive for the three classes; (g) overall accuracy; (h) classification rate.
Sensors 24 04444 g007
Sensors 24 04444 g008
Figure 8. Impact of PCA on complexity for population-based SSL; (a) computational time; (b) CPU memory usage.
Figure 8. Impact of PCA on complexity for population-based SSL; (a) computational time; (b) CPU memory usage.
Sensors 24 04444 g008
Sensors 24 04444 g009
Figure 9. Impact of PCA on complexity for distance-based SSL; (a) computational time; (b) CPU memory usage.
Figure 9. Impact of PCA on complexity for distance-based SSL; (a) computational time; (b) CPU memory usage.
Sensors 24 04444 g009
Sensors 24 04444 g010
Figure 10. Post-convergence accuracy parameters’ results for (a–c) population-based SSL and (d–f) distance-based SSL, with varying α .
Figure 10. Post-convergence accuracy parameters’ results for (a–c) population-based SSL and (d–f) distance-based SSL, with varying α .
Sensors 24 04444 g010
Sensors 24 04444 g011
Figure 11. Total learning time and CPU memory usage for all the episodes using (a) population-based SSL and (b) distance-based SSL, with varying α .
Figure 11. Total learning time and CPU memory usage for all the episodes using (a) population-based SSL and (b) distance-based SSL, with varying α .
Sensors 24 04444 g011
Sensors 24 04444 g012
Figure 12. Post-convergence accuracy parameters’ results for (a–c) population-based SSL and (d–f) distance-based SSL, with varying Nc.
Figure 12. Post-convergence accuracy parameters’ results for (a–c) population-based SSL and (d–f) distance-based SSL, with varying Nc.
Sensors 24 04444 g012
Sensors 24 04444 g013
Figure 13. Total learning time and CPU memory usage for all the episodes using (a) population-based SSL and (b) distance-based SSL, with varying N c .
Figure 13. Total learning time and CPU memory usage for all the episodes using (a) population-based SSL and (b) distance-based SSL, with varying N c .
Sensors 24 04444 g013
Sensors 24 04444 g014
Figure 14. Post-convergence accuracy parameters’ results for (a–c) population-based SSL and (d–f) distance-based SSL for different clustering methods.
Figure 14. Post-convergence accuracy parameters’ results for (a–c) population-based SSL and (d–f) distance-based SSL for different clustering methods.
Sensors 24 04444 g014
Sensors 24 04444 g015
Figure 15. Total learning time and CPU memory usage for 8262 episodes using (a) population-based SSL and (b) distance-based SSL for different clustering methods.
Figure 15. Total learning time and CPU memory usage for 8262 episodes using (a) population-based SSL and (b) distance-based SSL for different clustering methods.
Sensors 24 04444 g015
Table 1. Hyper-parameters for pre-trained neural network (supervised learning) model.
Table 1. Hyper-parameters for pre-trained neural network (supervised learning) model.
# of input features6
# of hidden layers1 (128 neurons)
Activation functiontanh (hidden layers), soft-max (output layer)
OptimizerAdam
Loss functionCategorical cross entropy
Table 2. 10-fold validation results for a pre-trained neural network model.
Table 2. 10-fold validation results for a pre-trained neural network model.
FoldsClass 1: SedentaryClass 2: Moderately ActiveClass 3: ActiveOverall Accuracy (%)
True Positive (%)False Positive (%)True Positive (%)False Positive (%)True Positive (%)False Positive (%)
198.5824220.704792697.1740011.492537398.4331060.707642598.06321453
298.2071710.487309697.2458491.369311399.0228010.903795998.15736382
398.5898470.746871296.5931861.295284498.7799920.984528897.98278644
498.9077670.825946897.412051.491033698.0730630.485338798.13071544
598.2456140.561685198.4362472.347227897.069450.222446997.91554599
698.3713360.783132597.61.782820197.9838710.464083997.98278644
798.4793920.789953497.4068071.630434898.2179020.523454898.03657881
898.9549841.232821396.8472111.854091197.2918350.36275797.70037655
998.2668280.585267497.2436161.549607698.6334410.788197398.0500269
1098.1033090.52440598.2179021.711294598.3929290.404285498.23830016
Mean98.4708670.724218597.4176871.652364398.1898390.584653198.02576951
Table 3. Summary of post-convergence accuracy results and total computational complexity numbers for the pre-trained NN model (benchmark) and the different variants of semi-supervised learning model.
Table 3. Summary of post-convergence accuracy results and total computational complexity numbers for the pre-trained NN model (benchmark) and the different variants of semi-supervised learning model.
Learning MechanismClass 1: SedentaryClass 2: Moderately ActiveClass 3: ActiveOverall Accuracy (%)Classification Rate (%)Total Computational Time to Train 8262 Episodes (s)Total CPU Memory Usage to Train 8262 Episodes (Bytes)
True Positive (%)False Positive (%)True Positive (%)False Positive (%)True Positive (%)False Positive (%)
Neural Network (Benchmark)98.470.7297.411.6598.180.5898.03100N/AN/A
Population-based SSL
(K-means, N C = 20, α = 10%)		94.191.3381.216.8192.178.0389.1899.1714,308.927455.104
Distance-based SSL
(K-means, N C = 20, α = 10%)		95.322.5876.365.992.89.2788.1610015,293.887559.6
Population-based SSL
(GMM_d, N C = 20, α   = 10%)		95.11.678.254.8295.19.3489.4899.9913,454.9122,122.97
Distance-based SSL
(GMM_d, N C = 20, α   = 10%)		95.572.0976.095.2893.699.9488.4510014,281.2722,273.25
Population-based SSL
(GMM_f, N C = 20, α   = 10%)		97.041.5382.45.5891.727.390.3899.9915,003.9122,930.06
Distance-based SSL
(GMM_f, N C = 20, α   = 10%)		96.931.5575.73.8695.1410.6989.25510016,059.3623,080.42
Population-based SSL
(GMM_s, N C = 20, α   = 10%)		94.451.3482.136.6892.157.689.5799.9913,060.7885822,117.19
Distance-based SSL
(GMM_s, N C = 20, α   = 10%)		95.161.7977.65.9692.6229.5588.4610014,382.6612422,267.82
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Roy, A.; Dutta, H.; Bhuyan, A.K.; Biswas, S. On-Device Semi-Supervised Activity Detection: A New Privacy-Aware Personalized Health Monitoring Approach. Sensors 2024, 24, 4444. https://doi.org/10.3390/s24144444

AMA Style

Roy A, Dutta H, Bhuyan AK, Biswas S. On-Device Semi-Supervised Activity Detection: A New Privacy-Aware Personalized Health Monitoring Approach. Sensors. 2024; 24(14):4444. https://doi.org/10.3390/s24144444

Chicago/Turabian Style

Roy, Avirup, Hrishikesh Dutta, Amit Kumar Bhuyan, and Subir Biswas. 2024. "On-Device Semi-Supervised Activity Detection: A New Privacy-Aware Personalized Health Monitoring Approach" Sensors 24, no. 14: 4444. https://doi.org/10.3390/s24144444

APA Style

Roy, A., Dutta, H., Bhuyan, A. K., & Biswas, S. (2024). On-Device Semi-Supervised Activity Detection: A New Privacy-Aware Personalized Health Monitoring Approach. Sensors, 24(14), 4444. https://doi.org/10.3390/s24144444

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop