A Trustworthy Healthcare Management Framework Using Amalgamation of AI and Blockchain Network

Abstract

Over the last few decades, the healthcare industry has continuously grown, with hundreds of thousands of patients obtaining treatment remotely using smart devices. Data security becomes a prime concern with such a massive increase in the number of patients. Numerous attacks on healthcare data have recently been identified that can put the patient’s identity at stake. For example, the private data of millions of patients have been published online, posing a severe risk to patients’ data privacy. However, with the advent of Industry 4.0, medical practitioners can digitally assess the patient’s condition and administer prompt prescriptions. However, wearable devices are also vulnerable to numerous security threats, such as session hijacking, data manipulation, and spoofing attacks. Attackers can tamper with the patient’s wearable device and relays the tampered data to the concerned doctor. This can put the patient’s life at high risk. Since blockchain is a transparent and immutable decentralized system, it can be utilized for securely storing patient’s wearable data. Artificial Intelligence (AI), on the other hand, utilizes different machine learning techniques to classify malicious data from an oncoming stream of patient’s wearable data. An amalgamation of these two technologies would make the possibility of tampering the patient’s data extremely difficult. To mitigate the aforementioned issues, this paper proposes a blockchain and AI-envisioned secure and trusted framework (HEART). Here, Long-Short Term Model (LSTM) is used to classify wearable devices as malicious or non-malicious. Then, we design a smart contract that allows only of those patients’ data having a wearable device to be classified as non-malicious to the public blockchain network. This information is then accessible to all involved in the patient’s care. We then evaluate the HEART’s performance considering various evaluation metrics such as accuracy, recall, precision, scalability, and network latency. On the training and testing sets, the model achieves accuracies of 93% and 92.92%, respectively.

1. Introduction

Modernization in the Internet of Things (IoT) technology makes a paradigm shift from human-centric to machine-centric interactions. It inspires especially the healthcare sector, which has been revolutionized from 1.0 to 4.0 by incorporating the promising characteristics of IoT technology. The essential benefits of IoT, such as data collection capability, deployment easiness, customer centricity, and less operational cost, ushered the healthcare sector to offer intelligent medical services, such as remote diagnosis, tracking of chronic diseases, and on-demand medical report analysis. Recently, a well-designed application of IoT, i.e., wearable devices, has been preferred in the healthcare sector, which measures the blood oxygen, temperature, blood pressure, breathing rate, and glucose state for the remote diagnosis of the patient. It utilizes Artificial intelligence (AI) algorithms to analyze the above-mentioned measures and provide significant inferences about the patient [1,2]. Furthermore, information communication tools and technology advancements make wearable devices reliable and efficient. However, it also makes it vulnerable to security threats where an attacker can exploit and manipulate the patient’s healthcare data [3]. Furthermore, from the literature, we can analyze that the collected healthcare data are processed and analyzed at the centralized system, which is also susceptible to various security attacks, such as session hijacking, data manipulation, and spoofing attacks [4,5]. Additionally, attackers can find vulnerabilities in the centralized system and resist healthcare centres from data access. Consequently, it can put an individual’s life at risk.

Various researchers from academia/industry have given indispensable solutions to resist security attacks associated with wearable devices. For instance, researchers have employed cryptographic-based solutions to protect healthcare data. The authors of [6] presented a dynamic cryptographic cipher structure and data availability strategy to guarantee healthcare data security in terms of confidentiality, availability, source authentication, and integrity. Their proposed solution is lightweight and needs a minimum number of iterations. Ref. [7] presented a recoverable encrypted and watermarked image approach for medical image security in healthcare information systems. They utilize symmetric key encryption in their approach and employ a client authentication scheme that grants access to medical images. However, the aforementioned cryptographic solutions cannot withstand modern computing capabilities that can easily break off the encryption ciphers of wearable healthcare data. If this happens, it is effortless for an attacker to attempt unauthorized access and fetch the crucial healthcare data of the patients. Therefore, there is a stringent requirement for a technology that can seamlessly differentiate malicious and non-malicious intents of the exploited wearables.

Authors in [8] proposed a PDiag system based on Bayesian classification for efficient and privacy-ensured diagnosis. The online medical primary diagnostic service processes sensitive personal health information without compromising privacy. In particular, a lightweight polynomial aggregation approach is utilized to offer an efficient and privacy-preserving classification strategy using an improved naive Bayes classifier. The encrypted user query is sent straight to the service provider and the user can only decrypt the assessment result. Furthermore, to ensure healthcare data security, ref. [9] suggested a robust watermarking methodology for healthcare data using key and transform coefficients of the source image. This obfuscates the entire process of sending and receiving medical data and ensures protection against intruders. While AI is advantageous in the healthcare industry, it has certain downsides. One significant disadvantage of AI-based systems is that they are relatively easy to copy once constructed. This implies that an attacker can perform adversarial attacks, i.e., poisoning the training and learning of the AI model.

The advent of blockchain technology attracted the attention of the entire scientific community. It is a shared public ledger that securely stores wearable healthcare data in an immutable manner to improve data privacy and establish trust between various entities of the healthcare systems [10,11]. Ref. [12] introduced a medical care information preservation system that considered the complete data storage process in wearables and medical centre servers. The procedure is secure and follows the Health Insurance Portability and Accountability (HIPAA) privacy and security requirements. Their suggested approach employed the advanced chaotic map technique key negotiation, reducing the amount of processing required, and achieving lightweight quantification. It also takes advantage of blockchain technology’s characteristics that secure healthcare data from data manipulation attacks and boost data security. Then, the authors of [13] offered a blockchain and smart contract-based trustworthy and transparent medicine supply chain to prevent doctors’ prescription information from being manipulated by attackers.

However, blockchain technology solely cannot confront the security threats associated with wearable devices. Thus, there is a need for a blockchain-based secure and trusted monitoring framework that takes the benefits of AI algorithms to strengthen the security of wearable device data. The proposed framework, i.e., HEART utilizes the LSTM model trained on the WUSTL EHMS 2020 dataset, which comprises wearable healthcare and attack data. Then, the learning of the pre-trained LSTM model is integrated into wearable devices to classify it as malicious or non-malicious by analyzing its feature space. Once the devices are classified, only legitimate wearable devices’ healthcare data are stored in the blockchain network to ensure data security. A blockchain-based smart contract is designed to verify the authenticity of different healthcare sector entities, such as doctors, wearable devices, patients, and healthcare data. Only the verified members can access the wearable healthcare data, and unauthenticated members are discarded from the HEART. Furthermore, the HEART operates on the Sixth-Generation (6G) communication technology, which offers ultra-low latency, extremely high reliability, and high availability compared to conventional 3G, 4G, and 5G communication technologies. Incorporating a 6G network interface is crucial for critical services, such as healthcare, which can exchange patients’ healthcare data efficiently with healthcare centers that potentially save a person’s life. The performance of the HEART is evaluated using different performance metrics, such as accuracy, precision, recall, loss curve, scalability, and network latency. The LSTM model shows better accuracy in terms of training and testing accuracies, i.e., 93% and 92.92%.

1.1. Motivation

There are millions of people participating in the healthcare sector, whether they are directly or indirectly affiliated. Hospitals and health facilities gather information such as a person’s name, social security number, address, and other personal information. The most difficult difficulty that the healthcare business faces is securely storing data so that unauthorized users cannot access it. There have been multiple cyberattacks on the healthcare business, and millions of people’s data have been released online. Upon analysis of different research [14,15] regarding cybersecurity in the healthcare sector, it is observed that those approaches only utilize a single technology. This results in a dilemma of whether to store the patient data securely or to analyze the data for tampering with the patient’s health data. As a result of this, we present a blockchain and AI-based solution for securely storing patient data and analyzing it for any tampering that could have taken place by an attacker.

1.2. Background

Healthcare is a very important sector as it deals with very sensitive information about patients, which, if compromised, can put the lives of millions of patients at risk. To combat this, the research community has spent countless hours integrating Artificial Intelligence (AI) with healthcare in order to combat the rising cybersecurity attacks on healthcare systems. AI as well as machine learning applications are revolutionizing healthcare delivery. Health organizations have amassed massive amounts of data in the form of health records and photographs, demographic data, claims data, and clinical trial data. AI technologies are perfectly adapted to evaluating this data and uncovering patterns and insights that humans would not be able to discover on their own. Deep learning from AI may assist healthcare companies to make better financial and clinical choices and enhance the quality of experience they give.

AI alone cannot be suitable enough to resist cyberattacks. There needs to be a place to securely store the patient’s data which can not be modified once it is stored there. This is where blockchain comes into the picture, as it provides security as well as the immutability of patient’s health data. The use of blockchain technology to store patient data will reduce the number of cyberattacks as the security aspect of blockchain will deter attackers from trying and infiltrate the blockchain. The actual benefit of interoperability might be unlocked via a blockchain-powered health information exchange. Current middlemen’s friction and expenses might be reduced or eliminated by blockchain-based solutions. Blockchain’s promise has far-reaching ramifications for stakeholders in the healthcare ecosystem. This technology offers immutability and security of data. The decentralized nature of this technology indicates that, in order to change the patient’s data, the attacker must gather control of 51% of nodes in the blockchain and on a public blockchain; this task is next to impossible. This makes it unfeasible to change a patient’s record in the blockchain. Making use of this technology offers the ability to integrate disparate systems in order to produce insights and better determine the worth of care. A statewide blockchain network for electronic medical information can increase efficiency and promote improved health outcomes for patients in the long run.

1.3. Research Contributions

The following are the salient contributions of the paper:

• We propose a blockchain and AI-based secure and trusted framework, i.e., HEART that efficiently detects the security attacks from the healthcare wearable devices;
• An AI-based LSTM model is adopted to detect the presence of malicious wearable devices by utilizing its pre-training on the standard dataset, i.e., the WUSTL EHMS 2020 dataset;
• A blockchain-based smart contract is deployed, which allows only legitimate wearable devices to store their healthcare data inside the blockchain’s immutable ledger;
• The HEART is evaluated by considering several assessment metrics, such as training accuracy, loss, blockchain scalability, and latency. As a result, the Adam-based LSTM model surpasses other optimizers in terms of accuracy, i.e., 93%.

1.4. Paper Organization

The flow of the remaining paper is described as follows: Section 2 discusses the motivation behind the proposed work by comparing it with the existing baseline works. Section 3 presents the system model and the problem formulated, i.e., to enhance the security and privacy concerns of the smart healthcare system. Section 4 shows the proposed framework-HEART. Section 5 illustrates the results and discussion. Lastly, Section 6 provides concluding remarks and shows the future scope of this work.

2. Related Works

This section discusses the existing work that calls attention to cybersecurity difficulties in the healthcare industry.

2.1. IoT Security in Healthcare

Ref. [16] has suggested a solution to secure smart gadgets physically while simultaneously maintaining data secrecy. The disadvantage of this suggested system is that it has only been tested on an IoT testbed. Therefore, the system’s performance in the real world might vary, and its behavior is uncertain. Ref. [17] has presented a lightweight authentication technique for wireless medical sensor networks to secure data exchange from sensors to doctors [18]. Next, the authors of [19] presented an authentication scheme for the internet of medical things that takes advantage of physically unclonable features. Their suggested method prevents attackers from accessing servers and harmful devices from entering. The aforementioned solutions are highly efficient; however, the authentication schemes do not resist modern decryption tools that easily break off the encryption keys.

2.2. AI Security in Healthcare

Ref. [20] has proposed an online medical prediagnosis system that uses a nonlinear support vector machine to protect the privacy of medical information. Their suggested approach is highly efficient in terms of prediction. Then, Ref. [21] has presented a machine learning (ML)-based solution for identifying harmful activity in smart healthcare systems. Their proposed system employs four machine learning techniques: artificial neural networks, decision trees, random forests, and k-nearest neighbors. Their result outperforms in terms of accuracy, i.e., 90%. Next, the authors of [22] introduced a trust-based strategy for detecting rogue devices in the healthcare industry using Bayesian inference. Their technology is capable of recognizing rogue medical equipment. However, the proposed solutions have not discussed how they confront data integrity and manipulation attacks.

2.3. Blockchain in Healthcare

Ref. [23] proposed an automated health record system that utilizes blockchain technology. The proposed framework is designed to exchange health information on the blockchain in order to create a smart e-health system. The system uses a modified Merkle tree data structure for faster and more secure access to patient logs [24]. The disadvantage, however, is that the proposed system did not consider the possibility of malicious data being uploaded into the blockchain. Attackers can manipulate the patient’s data and upload it onto the blockchain, which can cause many problems in the system. Ref. [25] has proposed a bribery self-mining system in the blockchain that works on Ethereum. The proposed system describes a new way of essentially gaining higher rewards and spending less Ethereum. The system is tested on a simulated healthcare system, and it is seen that harmful attacks can be made on a healthcare system due to higher rewards and lower spending costs. The disadvantage of this system is that it considers the attacks being made in the form of mining and not the malicious data sent to the blockchain. Ref. [26] proposed a blockchain-orchestrated deep learning approach for secure data transmission of data in IoT-enabled healthcare systems. The proposed system has been developed to detect intrusion in healthcare systems using deep learning and blockchain. The demerit of this system is that it does not consider the patient’s health data, which can threaten the patient’s life if incorrect data are fed into the blockchain. Ref. [27] has proposed a 5G blockchain architecture for healthcare systems. The proposed system aims to provide a comparable database for each user. The disadvantage of this system is that it is not tested in the real world; therefore, its performance is unknown. Ref. [28] has presented a framework for detecting anomalies in patients’ health data. In this system, blockchain is utilized to securely store the patient’s data. However, the proposed solution has a higher computational complexity as it has processed all the healthcare data. Ref. [29] presented a blockchain-based electronic medical record and patients, practitioners, laboratories, and admin to form a secured environment in the blockchain. Still, the system is not functioning as a full-fledged healthcare management system. The aforementioned analysis shows that none of the existing work has amalgamated blockchain and AI technology to alleviate the security attacks from wearable devices. Moreover, most solutions are centric on one security attack, wherein HEART uses different security threats, such as man-in-the-middle attacks, spoofing, data injection, and data integrity attacks to analyze the security of wearable devices. Therefore, there is a requirement for a blockchain and AI-based secure and trusted framework that analyzes different security attacks from wearable devices and improves the security and privacy of the healthcare industry.

Table 1. A comparative table between the existing state-of-the-art works and the proposed work.

Author	Year	Issues Targeted	Solution Proposed	Pros	Cons
Chelladurai et al. [23]	2022	Automated health record system using blockchain	The proposed system exchanges health information on the blockchain to create a smart e-health system	The system utilizes a modified Merkle tree data structure which provides secure and fast access of patient logs.	The system does not consider the possible manipulation of patient data by attackers.
Wang et al. [25]	2022	Bribery self mining in blockchain	Proposed a new selfish mining scheme and ways to tackle it	The system proposes a unique way of selfish mining in Ethereum which is more efficient than others	The proposed scheme has lower ethereum costs and higher rewards, leading to more harmful attacks on healthcare systems.
Khujamatov et al. [27]	2020	Blockchain for healthcare	Blockchain-based 5G healthcare architecture	The solution utilizes the 5G technology alongside blockchain, which provides ultra-low latency and high data rate, which are crucial for healthcare systems.	The proposed framework does not check whether the data received by the blockchain is tampered with or not.
Sharma et al. [29]	2020	Privacy of patient’s data	Blockchain-enabled electronic health records privacy preserving	Transactions are secured via blockchain technology which improves transparency in the healthcare industry.	The system is not functioning as a full-fledged healthcare management system.
Newaz et al. [21]	2019	Security of smart healthcare devices	AI-based secure smart healthcare framework	Detected malicious intents in smart healthcare systems.	The data collected from smart healthcare systems are considered not compromised.
Yanambaka et al. [19]	2019	Security of Internet of Medical Things	Physical unclonable function-based lightweight authentication scheme	A device authentication mechanism for the internet of medical things that take advantage of physically unclonable features.	The proposed system does not use technologies such as 5G/6G, which provide even faster authentication time for each device.
Meng et al. [22]	2018	Trust management in healthcare	Trust management in healthcare systems	Capable of recognizing rogue medical equipment.	Large traffic, security policy enforcement, and the deployment of extra security measures are some of the system’s downsides.
Yeh et al. [16]	2016	Security of healthcare systems	Securing healthcare system with body sensor networks	Secure smart gadgets physically while simultaneously maintaining data secrecy and resistance against forgery attacks and replay attacks.	Overall performance and computation cost are not measured
HEART	2022	Security of wearable devices and patient’s healthcare data	Blockchain and AI-based framework	HEART authenticates the integrity of wearable devices	-

shows the comparative analysis of the existing approaches and the proposed work.

3. System Model and Problem Formulation

3.1. System Model

The proposed system model is represented in Figure 1. Every patient has a wearable gadget that actively monitors the patient’s health data. The primary goal of the proposed system is to securely store health data such that there is no data breach by an unauthorized user. To do this, the wearable device’s integrity is first evaluated. If the wearable gadget is attacked, the data it transmits are not recorded on the blockchain and is therefore deleted. Once the device’s integrity has been certified, the patient’s data are transmitted to the blockchain for storage. The blockchain’s smart contract allows just the most critical data to be stored on the network. It is also in charge of data access control. Only those participating in the patient’s therapy have access to the data. The forthcoming section provides a detailed description of the system model.

3.2. Problem Formulation

Consider wearable devices W and patients P in the following way:

$$\begin{array}{c}\hfill W=\{W_1,W_2,W_3,\cdots ,W_k,\cdots ,W_N\}\end{array}$$

(1)

$$\begin{array}{c}\hfill P=\{P_1,P_2,P_3,\cdots ,P_j,\cdots ,P_N\}\end{array}$$

(2)

where k represents a single wearable device and N represents the total number of wearable devices such that $1\le k\le N$. j represents a single individual, and each patient $P_j$ is equipped with a wearable device $W_k$, so that

$$\begin{array}{c}\hfill P_j\leftarrow W_k\end{array}$$

(3)

$$\begin{array}{c}\hfill \forall 1\le j,k\le N\end{array}$$

(4)

Each $P_j$’s health data is gathered by $W_k$ and kept on a centralized system C. An attacker A can enter the system and gain access to patient information, such as

$$\begin{array}{c}D=\{D_1,D_2,D_3,\cdots ,D_j,\cdots D_N\}\end{array}$$

(5)

$$\begin{array}{c}C\leftarrow D\end{array}$$

(6)

$$\begin{array}{c}A\stackrel{\mathrm{Attacks}}{\to }C\end{array}$$

(7)

$$\begin{array}{c}A\leftarrow C_{comp}\left(D_j\right)\end{array}$$

(8)

$$\begin{array}{c}\forall 1\le j\le N\end{array}$$

(9)

where $D_j$ represents the patient’s medical records, and $C_{comp}$ represents the hacked centralized system. The aforementioned equations show that, without sufficient security measures, the system is readily infiltrated, and A may take patient data for malicious intent. To address the above-mentioned issues, the HEART employs AI and blockchain to create a more robust and secure method of storing patient data. Rather than storing $D_j$ straight into C, it is first examined by the AI-based system S to confirm its integrity. Once verified to be saved, $D_j$ is recorded in blockchain B, which is thereafter viewable only to those participating in the patient’s care. Mathematically, it is defined as follows:

$$\begin{array}{c}S\stackrel{\mathrm{Analyzes}}{\to }{D}_j\end{array}$$

(10)

$$\begin{array}{c}ifsafe,B\stackrel{\mathrm{Store}}{\leftarrow }{D}_j\end{array}$$

(11)

$$\begin{array}{c}IvP\stackrel{\mathrm{Access}}{\to }{D}_j\end{array}$$

(12)

$$\begin{array}{c}\forall 1\le j\le N\end{array}$$

(13)

where $IvP$ refers to the persons participating in the treatment. The usage of blockchain makes it next to impossible for an attacker to acquire access to $P_j^{\prime }s$ data. In order to identify that data have been manipulated, the AI model will be continuously monitoring the incoming patient health data. Since these data are time series data, the trained LSTM model is best suited for detecting the manipulation of patient data. The entire system is connected to the public blockchain, which indicates that, if malicious data are found, all the participants will be informed and the data will be discarded. The forthcoming section contains detailed information on the HEART.

4. HEART: The Proposed Framework

The HEART is presented in Figure 2. The system is dissected into three layers, i.e., user, analytic, and blockchain layers. Algorithm 1 depicts the entire system flow. The following is a detailed description of all the layers.

Algorithm 1: Flow of the system.

Input: Wearable device data j = 1 for$j\le N$do     $P_j\leftarrow W_j$     $D_j=\{HeartRate,SpO2,\cdots \}$     $P_j\stackrel{\mathrm{has}}{\to }{D}_j$     $W_j\leftarrow D_j$     $S\stackrel{\mathrm{examines}}{\to }{W}_j\left(D_j\right)$     if $W_j==A$ then         Discard data of $W_j$     else         Smart Contract $\stackrel{\mathrm{sent}}{\leftarrow }{D}_j$         Smart Contract $\stackrel{\mathrm{stores}\mathrm{in}}{\to }B$         $IvP\stackrel{\mathrm{accesses}}{\to }{D}_j\stackrel{\mathrm{from}}{\to }B$         If necessary, Doctor $\underset{\mathrm{treatment}}{\overset{\mathrm{medicine}}{\to }}{P}_j$     end if end for

4.1. User Layer

The patients $P_1,P_2,P_3,\cdots P_N\in P$ each have wearable devices $W_1,W_2,W_3,\cdots W_N\in W$, which measure the patient’s health data. The wearable device collects the data from the users, such that

$$\begin{array}{c}{W}_j\leftarrow \{P_j\left(HR\right),P_j\left(O2\right),P_j\left(T\right),\cdots \}\end{array}$$

(14)

$$\begin{array}{c}\forall 1\le j\le N\end{array}$$

(15)

where $P_j\left(HR\right)$, $P_j\left(O2\right)$, and $P_j\left(T\right)$ are the user’s heart rate, blood oxygen level, and temperature. In this layer, apart from regular users, there may be a possibility that an attacker is also among them. The attacker can grab the data from the communication channel and manipulate the data such that

$$\begin{array}{c}A\stackrel{\mathrm{Changes}}{\to }\{P_j\left(HR\right),P_j\left(O2\right),P_j\left(T\right),\cdots \}\end{array}$$

(16)

$$\begin{array}{c}{W}_j\leftarrow \{P_j^C\left(HR\right),P_j^C\left(O2\right),P_j^C\left(T\right),\cdots \}\end{array}$$

(17)

$$\begin{array}{c}\forall 1\le j\le N\end{array}$$

(18)

where $P_j^C$ denotes the changed data of the user. This can lead to severe problems; for instance, a patient who requires a critical medical diagnosis cannot acquire it because the attackers have manipulated the patient’s data and forwarded it to the medical practitioners. Hence, such situations can be life-threatening for the user. To detect an attacker in the system, the forthcoming security layer deploys a DL model for the same.

4.2. Analytic Layer

The analytic layer collects all data from wearable devices and saves it in a database. Then, the obtained data are cleaned, and preprocessing is performed on it. The following section provides a detailed description of the dataset and data preparation.

4.2.1. Dataset Description

A dataset named WUSTL EHMS 2020 dataset [30] is utilized for training the LSTM-based sequence model. For the purpose of continuously detecting whether the patient’s wearable data has tampered or not, there are no proper time-series datasets other than WUSTL EHMS 2020. This dataset is ideal for our use case as the patient’s data would be continuously communicated via an open channel to the blockchain. Thus, the dataset consists of different network characteristics which can further aid in the detection of incoming malicious data. The dataset also features information regarding the patient’s health data obtained from a wearable device. Using this data will allow the trained machine learning model to learn to analyze the health data and check if it has been tampered with. The dataset contains 44 characteristics, 35 of which are network traffic measures, eight biometric features from patients, and one feature for the class label that specifies the attack data or not. The dataset comprises a total of 16,318 samples, of which 14,272 are normal samples, and 2046 are attack samples.

Table 2. Parameters considered for patient’s data.

Patient Data Type	Description
Temp	The temperature of the patient
SpO2	The SpO2 reading of the patient
Pulse Rate	The measurement of heart rate
Sys	The systolic pressure of the patient
Dia	The diastolic pressure of the patient
Heart Rate	The rate at which the heart beats per minute
Resp Rate	The rate at which the patient is breathing
ST	Electrically neutral area between ventricular depolarization (QRS complex) and repolarization (T wave) in millivolts

represents the patient data considered in the system. Since there are a lot of components for the purpose of obtaining the security information of a network, the security information is instead described by the following equation:

$$\mathrm{Security}\mathrm{information}\in \mathrm{Different}\mathrm{networking}\mathrm{components}$$

(19)

4.2.2. Data Preprocessing and Preparation

The collected dataset requires preprocessing before it is forwarded to training and prediction. Figure 3 depicts the data preprocessing and preparation stages in detail. Only 26 of the 44 qualities truly contribute to forecasting the result. Flags, IP addresses, MAC addresses, and other trivial information are removed from the dataset as they are not competent features compared to the others. After obtaining the relevant features, their values are normalized using the min-max scaler from the sklearn package. Because the values of characteristics have varying ranges, normalization is required. Normalizing the data puts the mean closer to 0 and speeds up convergence. The values are normalized to be between 0 and 1, inclusive. The min-max scaler formula is as follows:

$$\begin{array}{c}\hfill {\tau }_{std}=\frac{\tau -{\tau }_{min}}{{\tau }_{max}-{\tau }_{min}}\end{array}$$

(20)

$$\begin{array}{c}\hfill {\tau }_{scaled}={\tau }_{std}\cdot (max-min)+min\end{array}$$

(21)

where $\tau$ represents the value that needs to be normalized, ${\tau }_{min}$ and ${\tau }_{max}$ denotes the feature’s minimum and maximum value. ${\tau }_{scaled}$ is the normalized value (using min-max scalar), and $max$ and $min$ express the maximum and minimum value of the new range for normalization. The preprocessing procedure is completed after the data have been normalized. To use the data for training, it must be reshaped into three-dimensional data, as LSTM requires three-dimensional input. There are m features and n samples for each $k^{th}$ device. As a result, the data are reshaped as $m\times n\times 1$ for training purposes for each device. Thus, the input shape $\chi$ for every device is ${\chi }_{m\times n\times 1}$.

4.2.3. LSTM

After the data have been processed, an LSTM is trained on it. LSTMs are Recurrent Neural Networks (RNNs) that can learn order dependency in sequence prediction challenges [31,32]. RNNs are a form of neural network that predicts sequential or time-series data. RNNs are differentiated by their “memory”, which allows them to alter the current input and output by using information from previous inputs [33]. Unlike standard feedforward networks, the output is determined by the preceding items in the sequence.

The LSTM cell with $C_{t-1}^s$, $C_t^s$, $O_{t-1}$, $O_t$, $\varpi$, $\sigma$, and $tanh$ representing the previous cell state, cell state, previous output, output, input data, sigmoid, and tangent activation functions [31,34]. The LSTM has three gates: a forget gate, an input gate, and an output gate. As the name implies, the forget gate eliminates unneeded information before merging into the cell state. This gate takes as an input $x_t$ and $h_{t-1}$. To reduce extraneous data, the inputs are passed via a sigmoid function. The input gate is used to update the cell state with fresh information. Like the forget gate, the input gate accepts $O_{t-1}$ and $\varpi$ as input and runs it via a sigmoid and a $tanh$ function. The outcomes are then multiplied, and the usable data are added to the current cell state. The output gate stores relevant information based on $C_t^s$, $O_{t-1}$, and $\varpi$. After the input and forget gates are combined, the $C_t^s$ is sent via a $tanh$ function. $O_{t-1}$ and $\varpi$ are sent through a sigmoid function to determine which values need to be updated. Finally, the result of both operations is multiplied and produces the cell’s final output. The equations for all the gates are shown below [35]:

$$\begin{array}{c}\hfill in{p}_t=sig(z_i\cdot [O_{t-1},\varpi ]+q_i)\end{array}$$

(22)

$$\begin{array}{c}\hfill for{g}_t=sig(z_f\cdot [O_{t-1},\varpi ]+q_f)\end{array}$$

(23)

$$\begin{array}{c}\hfill ou{t}_t=sig(z_o\cdot [O_{t-1},\varpi ]+q_o)\end{array}$$

(24)

where $in{p}_t$, $for{g}_t$ and $ou{t}_t$ are the equations for input, forget and output gate. $sig$ is the sigmoid function. $z_i$, $z_f$ and $z_o$ are the weight for each of the respective gates. The equations for candidate cell, $C_t^s$ and final output are as follows:

$$\begin{array}{c}\delta =tanh(z_c\cdot [O_{t-1},\varpi ]+q_c)\end{array}$$

(25)

$$\begin{array}{c}{C}_t^s=for{g}_t\cdot O_{t-1}+in{p}_t\cdot \delta \end{array}$$

(26)

$$\begin{array}{c}{O}_t=ou{t}_t\cdot tanh\left(C_t^s\right)\end{array}$$

(27)

where $\delta$ is the candidate cell at timestamp t. Figure 4 shows the overall architecture of the LSTM model which is used for training and prediction. Algorithm 2 represents the prediction flow of the LSTM-based sequence model. Since LSTM is a type of recurrent neural network, the computational complexity of these models is dependent on the length of the input sequence. This indicates that, if the length of the input sequence is l, then the computational complexity of LSTM will be $O\left(l\right)$. The complexity is $O\left(l\right)$ because, in order to obtain the output, the LSTM model must traverse the entire sequence of input.

Algorithm 2: Flow of the LSTM-based prediction model.

Input: ${\chi }_{m\times n\times 1}$   Output: Attacker_Type model = Sequential(),         ▹ ml = model ml.add(InputLayer(input_shape = (m, 1))) ml.add(LSTM(150, return_sequences = True)) ml.add(BatchNormalization()) ml.add(Dropout(0.2)) ml.add(LSTM(150)) ml.add(Dense(10, activation= relu)) ml.add(Dense(1, activation = sigmoid)) ml.compile(loss=log_loss,optimizer= adam) ml.load_weights() Attacker_Type = ml.predict($\chi$) if Attacker_Type == 0 then     The device is not an attacker. else if Attacker_Type == 1 then     The device is an attacker. end if

When the dataset is completely processed, the training of the LSTM model begins. The processed time-series data are fed into the LSTM model for the prediction of attacker type. The weights and biases at the beginning of the training are randomly initialized. As the training proceeds, the weights and biases start to converge toward the optimum point. The optimum point is achieved when the model’s accuracy becomes stagnant, i.e., the model no longer converges. This optimum point differs for every use case and, in our case, the optimum point is achieved when the model achieves 90% accuracy. At this point, it can be said that the LSTM model has converged. The specific time to converge for a model varies from model to model and, in our case, the LSTM model took around 1300 s to converge. Once the LSTM model is trained on the standard dataset, i.e., the WUSTL EHMS 2020 dataset, it is then deployed in the wearable devices to validate its pre-trained learning on the unknown wearable healthcare data. Finally, suppose the LSTM model successfully bifurcates legitimate wearable devices based on its learning from the WUSTL EHMS 2020 dataset. In that case, its healthcare data are forwarded to the blockchain network to protect it from data manipulation attacks.

4.3. Blockchain Layer

The blockchain layer is in charge of securing data stored in the blockchain to tackle data injection attacks [36,37,38]. To check the validity of the data, a smart contract is employed. Smart contracts are programs executed based on the satisfaction of certain conditions and are stored in the blockchain network. The data are initially routed through a smart contract, determining who has access to the information.

Blockchain is particularly important in the healthcare industry because the system protects sensitive data, which is very valuable in the field of healthcare. With the growth of electronic medical records, the healthcare industry is continuously attempting to secure patient, hospital, insurance, and billing records. In addition, with so much data available, keeping track of it all while maintaining privacy is difficult. Blockchain’s essential properties, such as transparency, immutability, distributed network, and reliability, strengthen security and offers trust in healthcare systems. In healthcare systems, it is used to monitor the manufacturing of prescription medications and their safety. It provides a platform for storing patient medical records that practitioners can quickly access to give remote medical aid. It is utilized to lessen the risk of fraud and inaccuracy in the medical industry by expediting qualification verification [39].

In the HEART, the patient’s health records are only sent by the smart contract to be stored on the blockchain. The patient and the doctor can only access the data. If required, the doctor can prescribe medication and also suggest necessary treatment to the patient.

Figure 5 shows the implementation of the aforementioned smart contract for the HEART. The smart contract has four major functionality: aUser, healthData, prescription and viewPatientData. The aUser function is responsible for checking the type of user, i.e., whether they are a patient or a doctor. If the user is a patient, the wearable device automatically enters the patient’s health data in the healthData function. This data are then stored onto the blockchain and can be viewed either by the patient or the doctor using the viewPatientData function. Figure 5a,b show the above mentioned functionalities. The user must enter a unique patient ID to view the patient’s data. If the user is a doctor, they can use the prescription function to enter the medication, if necessary, to be given to the patient and also provide insights into what kind of treatment is necessary for the said patient. These functionalities are shown in Figure 5c and Figure 5d, respectively. If any unauthorized user tries to view a patient’s data, a warning message is shown, and the patient’s data will not be visible. The warning message is shown in Figure 5e. Furthermore, the healthcare data are stored inside the interplanetary file system (IPFS)-based public blockchain. A Merkle Tree or Merkle Directed Acyclic Graph, similar to the one used in the Git Version Control system, is employed inside IPFS to efficiently track changes to files on a network in a decentralized manner. Each data element is identified by the cryptographic hash of its contents, known as content-addressing to improve the response time and scalability of the blockchain network. Each block contains a header, which includes metadata such as the timestamp, the hash of the previous block, and the nonce (a value used in the mining process). The block also contains a list of validated transactions, and a Merkle Tree root, which is a single hash that represents the entire collection of transactions in the block. Furthermore, each block also has a unique digital signature, called a “block hash”, which is generated using a cryptographic hashing algorithm. The block hash is based on the contents of the block, including the transaction data and the previous block’s hash.

The AI-based model was developed using Python 3.7.12 on Google Colaboratory (Colab) with 12.69 GiB of RAM and Tensorflow 2.6.0. To speed up the training, the Tesla K80 GPU is enabled. The smart contract is written in solidity language (version 0.8.7) and deployed over the remix ide (version 0.21.1).

5. Results and Discussion

This section assessed the proposed LSTM-based system on the test data and also describe the experimental setup and tools which were utilized for conducting the experiment. To obtain the best working model, many variable parameters are tuned.

Experimental Setup and Tools

This section describes the experimental setup for the proposed HEART framework. The LSTM model in the analytics layer is developed with Python v3.7.12 as the base environment. Tensorflow v2.9.2 and Keras v2.9.0 were utilized to develop the LSTM model for detecting an attacker. To perform mathematical calculations and processing of the dataset, Numpy v1.21.6 and Pandas v1.3.5 are used respectively. Matplotlib v3.2.2 is used to obtain the performance evaluation graphs for the LSTM model. The LSTM model is simulated on a cloud computing platform named Google Colab, which utilizes an Intel(R) Xenon(R) CPU @ 2.20 GHz, 12.69 Gb of RAM, and a Tesla K80 GPU, where this simulation is run for 150 epochs. Furthermore, once the LSTM model is trained, the non-malicious data detected by the model are sent to a smart contract. The smart contract is developed for blockchain using Remix IDE v0.29.2 and solidity v0.8.7 is the programming language with which the smart contract is written. The smart contract has four major functionalities which are as follows; $aUser$, $healhData$, $prescription$, and $viewPatientData$. A detailed description of each function is provided in the aforementioned section. Since the patient’s health data are a piece of very critical information, the data are stored in a blockchain, which ensures that the data cannot be tampered with by anyone. A block in a blockchain has the following attributes; magic number, which is used to identify a block in a blockchain, block size which defines the size of data that can be written in the block, block header which contains critical block information such as previous block hash, merkle root, timestamp, nonce and more, transactions is a list of all the transactions in that particular block, which in our case is the patient’s health data. The blockchain is also available to every medical personnel associated with a particular patient which ensures transparency and also prevents tampering with data. Since the data are stored on a public blockchain, a hash is calculated using all the patient’s health data and is appended to the block. This ensures that, if the patient’s health data have been tampered with, the new hash generated from the patient’s data by medical personnel and the hash stored in the block will not match, thus indicating that the data are manipulated by a malicious user. This malicious user can also be identified by medical personnel because the user must be registered onto the blockchain to manipulate the patient’s health data.

Table 3. Simulation parameters for the proposed HEART framework.

Parameters	Values
Analytics Layer Parameters
Epochs	150
Optimizer	Adam
Activation Function	tanh
Learning Rate	0.001
Loss Function	Binary Crossentropy
Blockchain Layer Parameters
Solidity Compiler	v0.8.7
Remix IDE	0.29.2
Gas Limit	3,000,000

shows the configuration set-up for entire experiment.

Multiple evaluations and validation techniques are used to accurately. For that, we used a few benchmarked performance parameters, such as precision, recall, and the F1 score. The trained model’s accuracy, recall, and F1-score are displayed in

Table 4. Model scores’ comparison on training and testing set.

Statistical Measure	Training Set	Testing Set
Precision value (%)	0.87	0.88
Recall value (%)	1.00	1.00
F1-Score (%)	0.93	0.93

. Both the training and testing sets are used to compute the scores. The recall value specifies the proportion of expected positive values that outnumber the predicted ones. Precision score, on the other hand, is defined as the proportion of actually expected positive values that exceed the actual ones. Lastly, the F1 score balances the accuracy and recall values to strengthen the prediction performance for the proposed work [40]. The trained LSTM model is also compared with a Recurrent Neural Network (RNN), another type of model that works on time series data. The comparison between the two models is provided in

Table 5. Comparison of LSTM and RNN.

Statistical Measures	LSTM	RNN
Precision Value (%)	0.88	0.84
Recall Value (%)	1.00	0.92
F1-Score (%)	0.93	0.88
Training Time (seconds)	1500	2250

. The table shows that the LSTM model outperforms the RNN model in every evaluation measure and also takes significantly less time to train. This is because RNNs suffer from vanishing gradient problems for longer input sequences. A vanishing gradient problem is one where the gradient of the loss function of the model reaches closer to zero, making it harder to train the model. Due to the vanishing gradient problem, the ability of RNNs to retain information for longer sequences is severely limited. LSTM, on the other hand, utilizes input and forget gates which helps the model retain longer dependencies and overall better flow control of gradients. The gates in an LSTM unit regulate the addition or removal of information in the unit. This allows the LSTM model to maintain long-term dependencies. Thus, the LSTM model is taken as the model of choice for the detection of a malicious user.

For training purposes, all optimizers are given a constant learning rate of 0.001. Figure 6a,b represent the model’s loss. A loss is a penalty for making an incorrect forecast. The loss is 0 if the model’s forecast is flawless; otherwise, the loss is significant. Figure 6a measures the loss over each iteration, whereas Figure 6c measures the loss over each epoch. An epoch is a single pass of the whole dataset, whereas an iteration is the total number of passes with a defined number of instances in each pass [41]. This number is referred to as the batch size, which in our instance is 256. Figure 6a shows that the model is trained for three different optimizers where the Stochastic Gradient Descent (SGD) performs marginally worse than the rest. The model’s training and testing loss are observed in Figure 6c. It is seen that there are spikes in the graph, and the overall loss is decreasing, indicating that the model is learning.

The model’s accuracy is another criterion for evaluating it. Accuracy gives a decent indication of the model’s prediction ability. A higher accuracy gives a good insight that the model is performing well on the given dataset. The model’s accuracy is determined at each iteration and epoch in the same way as the loss measure. Figure 6b represents the model’s correctness over all iterations. This graph shows that all of the optimizers produce quite comparable accuracies. Figure 6d displays the model’s accuracy estimated over all epochs for both training and testing sets. From the above-mentioned graphs ( Figure 6), it is clear that the accuracy parameter is performing well, i.e., achieving 93% and 92.92% on the training and testing set, respectively. We want to mention that the above outcome is obtained without modifying any parameters (e.g., number of epochs, batch size) in the LSTM model. Furthermore, various optimizers are compared for the LSTM model to showcase the convergence speed of each optimizer in terms of training loss and accuracy.

Figure 6e illustrates the scalability comparison between a conventional blockchain and the 6G-based proposed blockchain network. The persistent characteristic of a 6G network, such as ubiquitous data rate (1 Tbps), drives the blockchain network to operate quickly, allowing it to handle more transactions. From the graph, we can depict that, as the number of transactions increases, the scalability of a 6G-based proposed blockchain network increases. On the contrary, the conventional blockchain cannot process more transactions due to the absence of 6G network properties.

Figure 6f displays the latency comparison among traditional cellular networks, such as 4G and 5G and 6G. Latency is an essential parameter that needs to be measured in a critical application such as wearable technology. It is evident that, due to the enriched properties of a 6G network, such as ubiquitous high data rates (1 Tbps), high scalability (${10}^9$ devices/sqm), and high reliability (99.99999%), the latency of the HEART decreases, i.e., low latency (<1 ms). Furthermore, from the graph, we can observe that, as the number of transactions increases, the latency of a 6G network decreases compared to a 4G and 5G network interfaces. This is because the inclusion of a 6G network reduces the packet drop ratio and increases the packet delivery ratio compared to conventional network interfaces. Consequently, the 6G network improves the latency, i.e., reduces the delay in the HEART.

6. Conclusions

In this study, we proposed a blockchain and AI-assisted secure and trusted framework, i.e., HEART to protect wearable healthcare data from numerous security vulnerabilities, such as man-in-the-middle, session hijacking, data manipulation, and spoofing attacks. First, we trained the LSTM model on a standard dataset, i.e., the WUSTL EHMS 2020 dataset and integrated it into wearable devices, which detects whether the device is malicious or not. Furthermore, the healthcare data of the legitimate wearable device are forwarded to the blockchain network, which securely stores the data inside the immutable ledger. The entire HEART is communicating via a 6G network interface, improving latency and scalability. Lastly, the HEART is evaluated against numerous performance metrics, such as accuracy, latency, scalability, and loss curve. The results show that the Adam-based LSTM model achieves 93% and 92.92% training and testing accuracies compared to other optimizers.

In the future, we will improve the AI model by using LSTM and Gated Recurrent Unit (GRU) in an ensemble model which will enhance the capabilities of the model further and will become much more robust. We will also improvise the security aspects of the HEART by including modern-day attacks, such as ransomware, Sybil attack, and zero-day attacks.