Optimized Attention Enhanced Temporal Graph Convolutional Network-based Cloud Resource Allocation Supported IoT for Students' Health Monitoring System

Summary

This protocol describes an optimized attention-enhanced temporal graph convolutional network for cloud-based Internet of Things (IoT) student health monitoring.

Abstract

Sensor technology progressions have paved the way for the rapid expansion of the Internet of Things (IoT) applications to construct behavioral and physiological monitoring systems, like an IoT-based student healthcare monitoring system. The status of student health observation is necessary because the number of students who survive loneliness is increasing in large geographical areas. This research article presents an approach named optimized attention enhanced temporal graph convolutional network-based cloud resource allocation supported Internet of Things for students' health monitoring system (HMS-AETGCN-NGOA-IoT). The proposed HMS-AETGCN-NGOA-IoT is implemented using MATLAB. To detect students' health status, performance metrics like precision, accuracy, F1-score, Recall (Sensitivity), Specificity, Error rate, Computation time, and ROC are considered. The HMS-AETGCN-NGOA-IoT approach achieves 19.11%, 24.12%, and 28.13% higher specificity; 24.93%, 23.04%, and 9.51% lower computation time; 15.2%, 25.45%, and 13.91% higher ROC values; and 8.45%, 20.98%, and 27.55% higher accuracy compared with the existing Health Monitoring System based on Message Passing Neural Network for Internet of Things(HMS-MPNN-IoT), Health Monitoring System based on Support Vector Machine for Internet of Things(HMS-SVM-IoT) and Health Monitoring System based on Deep Neural Network for Internet of Things(HMS-DNN-IoT) methods, respectively.

Introduction

One of the most valuable and exciting research areas is cloud computing¹. This computing technique purchases structure and software services, and user-requested services from Internet². A number of clients, together with cloud computing requests, are rising day by day. As a result, enhancing the speed and precision of cloud computing is critical³. Cloud computing improves patient monitoring⁴. The cloud offers a stable foundation for hard and massive computing tasks, such as data storing and processing, device services, and other information processing activities⁵. Many businesses and people favor cloud computing due to the requirement to archive, analyze, and comprehend massive amounts of data⁶. The financial market and the scholarly community have given the healthcare services sector priority recently. Due in large part to its potential for research, this industry has attracted a large amount of financing⁷. The healthcare industry is dealing with a number of problems, such as a lack of resources, high healthcare expenditures, and inefficient hospital resource management⁸. Additional logistical tasks include the limited beds, equipment, healthcare specialists, as well as the volume of medical care that patients required⁹.

Patient monitoring is typically done passively, which means that medical teams control medicine based on changes in the patient's health status¹⁰. The patient's condition could deteriorate in emergency situations; hence, the response time is crucial. IoT offers a wide range of opportunities in the e-health sector¹¹. This technology has the ability to progress various fields and enhance health care services¹². IoT is crucial to healthcare services and serves as an industry accelerator. Body sensor networks (BSNs) analyze physiological characteristics and symptoms while transmitting critical signals for a physician to examine in order to spot unusual and unexpected events¹³. As a result, emergency medication might be given to avoid life-threatening situations¹⁴.

Using smart technologies under IoT, the medical care progress has made it easy and enhanced the health services¹⁵. IoT offers a platform for achieving objectives like enhancing the quality of life for those who are ill or disabled and making their daily lives easier¹⁶. It has made it simpler for those who are unable to physically be with the patients to provide care¹⁷. If properly used, the branch of study known as e-health presents a number of advantages for society¹⁸. Furthermore, it permits remote data storage on cloud-based systems and allows authorized users, including external specialists and the inside team, to collaborate¹⁹. It is also an ideal platform for mobile/tablet/PDA-based health applications²⁰.

The use of smartphones opens up several prospects for the development of effective healthcare services and solutions²¹. Furthermore, most existing systems still have significant limitations, such as offline data collecting, processing, and analysis, as well as focusing solely on diagnosing chronic conditions such as cardiovascular disease and diabetes without employing predictive methodologies²². Due to these constraints, the developed student health observing method depends on the health function report in IoT trails, along with biological and behavioral markers at the same time²³. The suggested platform includes biological sensors and health nursing tools for collecting and storing health data from students²⁴. The acquired data is then used by deep learning algorithms to track the student's overall health function status²⁵. The major contributions of the research work are summarized below.

In this research, HMS-AETGCN-NGOA-IoT is proposed (Figure 1). Monitoring a student's health state is done by assessing two types of health condition parameters: behavioral and biological change parameters. To achieve optimal prediction, the Attention Enhanced Temporal Graph Convolutional Network (AETGCN) technique was used to detect students' health function status. HMS-AETGCN-NGOA-IoT method is implemented in MATLAB, and the effectiveness is assessed using several performance metrics (Figure 2). The efficiency of the suggested method is analyzed with the existing HMS-MPNN-IoT, HMS-SVM-IoT, and HMS-DNN-IoT methods.

Protocol

All required materials, software, and equipment used in this study are mentioned in the Table of Materials.

Data acquisition and preparation

The Data Set used in this work is the Student Mental Health Dataset, which was obtained from the publicly available Kaggle repository²⁶. The dataset contains self-reported questions and responses collected from university students, covering demographic information, academic stress, sleep patterns, and mental health-related indicators. The data includes both sensitive and non-sensitive health conditions with a balanced distribution, which is suitable for supervised learning and performance evaluation. This dataset is divided into two categories: sensitive and non-sensitive events. Sensitive events are the conditions that threaten the health of students, like extreme academic stress, prolonged lack of sleep, high levels of anxiety or depression symptoms, and unhealthy lifestyle habits like excessive smoking and drinking, while non-sensitive events are associated with normal or stable mental health. This data set is split into training and testing sets, allocating 70% of the samples for training and 30% for testing the model. Implement tenfold cross-validation for model evaluation.

Data preprocessing using two-way recursive filtering (TWRF)

Apply TWRF to remove noise from the input dataset²⁷. Implement the filtering process using the mathematical formulation: For an input row X with value x_i at node i, apply the recursive filter.

y_i = a₁y_i-1 + a₂y_i-2 + ... + a_ky_i-k + b₀x_i + b₁x_i-1 + ... + b_kx_i-k

where a represents negative coefficients and b represents current data values. Calculate the value using the node function:

f(i) = min(255, max(0, round(y_i)))

where parameter 255 represents the maximum filtering value. Compute the node product vector and determine the desired filtering value. Execute the complete TWRF process to generate clean, preprocessed data using the final filtering equation.

Feature extraction using adaptive synchroextracting transform (AST)

Apply AST to extract features from the preprocessed data²⁸. Extract 12 numerical features: age, weight, heartbeat, sleep hours, fasting blood sugar, Cholesterol, HDL, LDL, Triglyceride, TSH, T4, and T3. The above features are selected because of their strong clinical relevance in assessing physical and mental health conditions²². For instance, Heartbeat and sleep hours represent stress and fatigue, while blood sugar and lipid values (cholesterol, HDL, LDL, triglycerides) represent metabolic health, and thyroid hormones (TSH, T3, T4) represent mood and mental health. Extract 5 nominal features such as gender, body temperature, body pressure, cigarette consumption, and alcohol consumption. These features are extracted because they define important demographic, physiological, and lifestyle variables that affect health status. Utilize the ridge curve identification method within AST using the time-frequency representation:

AST_x(t,ξ) = SET_x(t,ξ)δ(ξ - ξ_x(t))

Apply the adaptive ridge extraction algorithm to reduce interference items and extract optimal features.

Model implementation and optimization

Configure the attention-enhanced temporal graph convolutional network (AETGCN) for health status detection²⁹using graph convolution

h_i^(I+1) = σ(∑_j∈N(i) 1/c_ij W^(I) h_j^(I))

where h_i is the feature of node i, W is the weight function, and c_ij is the normalizing term.

Implement the purpose of AT-GCN units using the gate mechanisms and output states. Incorporate spatial and temporal attention mechanisms:

Spatial: e_ij = LeakyReLU(a^T[Wh_i∥Wh_j])

Temporal: β_i = q^T tanh(W_hh_i + W_xx_i + b_h)

Calculate normalized significance scores and implement temporal attention for frame importance weighting.

Northern Goshawk optimization algorithm (NGOA) implementation

Initialize NGOA for optimizing AETGCN parameters³⁰ with population matrix:

X = [X₁, X₂, ..., X_n]^T

Evaluate the objective function for each population member. Implement the exploration phase for prey identification:

P_i = X_k, i = 1,2,...,N, k = 1,2,...,i-1,i+1,...,N

X_i^new, P¹ = {X_i + r(P_i - IX_i), FP_i < F_iX_i + r(X_i - P_i), FP_i ≥ F_i

Execute the exploitation phase for parameter optimization:

X_i^new, P² = {X_i + R(0,1) × (P_i - I×X_i), FP_i < F_iX_i + R(0,1) × (X_i - P_i), FP_i ≥ F_i

Iteratively optimize the weight parameters until the termination criteria are met.

Model training and evaluation

Train the optimized AETGCN model using the processed features. The performance of the proposed model was measured using a confusion matrix by calculating the true positive (TP), true negative (TN), false positive (FP), and false negative (FN) values for sensitive and non-sensitive health events. Using these values, performance metrics such as accuracy, precision, recall, F1-score, specificity, error rate, and ROC were calculated to measure the classification ability of the proposed model. The results were compared with HMS-MPNN-IoT, HMS-SVM-IoT, and HMS-DNN-IoT models using the same experimental environment, and the computational complexity was measured using the execution time taken for processing.

Results

The representative results validate the effectiveness of the proposed HMS-AETGCN-NGOA-IoT approach in identifying sensitive and non-sensitive health issues of students. Enhanced accuracy and F1-score values prove the effectiveness of the attention mechanism in the temporal graph convolutional network in identifying the temporal patterns and relationships of the health features. The high specificity and ROC curves ensure accurate identification with fewer false alarms, and the lower computation time proves the efficiency ...

Discussion

The proposed HMS-AETGCN-NGOA-IoT model demonstrates a significant advancement in automated student health monitoring by effectively integrating IoT data acquisition with a sophisticated deep learning framework. The core of this model's success lies in the Attention Enhanced Temporal Graph Convolutional Network (AETGCN), which is specifically designed to handle the complex, relational, and time-dependent nature of health data. By modeling students and their physiological/behavioral parameters as a dynamic graph, the AETGC...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
MATLAB	MathWorks		R2023a or later
Operating system	windows 10
Personal Computer	N/A		Memory 8 GB RAM
Processor	Intel, Core i5
Student Mental Health Dataset	Kaggle		https://www.kaggle.com/datasets/shariful07/student-mental-health