Systematic Review on AI-Blockchain Based E-Healthcare Records Management Systems

Electronic health records (EHRs) are digitally saved health records that provide information about a person’s health. EHRs are generally shared among healthcare stakeholders, and thus are susceptible to power failures, data misuse, a lack of privacy, security, and an audit trail, among other problems. Blockchain, on the other hand, is a groundbreaking technology that provides a distributed and decentralized environment in which nodes in a list of networks can connect to each other without the need for a central authority. It has the potential to overcome the limits of EHR management and create a more secure, decentralized, and safer environment for exchanging EHR data. Further, blockchain is a distributed ledger on which data can be stored and shared in a cryptographically secure, validated, and mutually agreed-upon manner across all mining nodes. The blockchain stores data with a high level of integrity and robustness, and it cannot be altered. When smart contracts are used to make decisions and conduct analytics with machine-learning algorithms, the results may be trusted and unquestioned. However, Blockchain is not always indestructible and suffers from scalability and complexity issues that might render it inefficient. Combining AI and blockchain technology can handled some of the drawbacks of these two technical ecosystems effectively. AI algorithms rely on data or information to learn, analyze, and reach conclusions. The performance of AI algorithms is enhanced through the data obtained from a data repository or a reliable, secure, trustworthy, and credible platform. Researchers have identified three categories of blockchain-based potential solutions for the management of electronic health records: conceptual, prototype, and implemented. The purpose of this research work is to conduct a Systematic Literature Review (SLR) to identify and assess research articles that were either conceptual or implemented to manage EHRs using blockchain technology. The study conducts a comprehensive evaluation of the literature on blockchain technology and enhanced health record management systems utilizing artificial intelligence technologies. The study examined 189 research papers collected from various publication categories. The in-depth analysis focuses on the privacy, security, accessibility, and scalability of publications. The SLR has illustrated that blockchain technology has the potential to deliver decentralization, security, and privacy that are frequently lacking in traditional EHRs. Additionally, the outcomes of the extensive analysis inform future researchers about the type of blockchain to use in their research. Additionally, methods used in healthcare are summarized per application area while their pros and cons are highlighted. Finally, the emphasized taxonomy combines blockchain and artificial intelligence, which enables us to analyze possible blockchain and artificial intelligence applications in health records management systems. The article ends with a discussion on open issues for research and future directions.

for their implementation in the healthcare industry sector. 89 In addition, the advent of modular IT systems has been 90 observed since the implementation of healthcare provisions 91 in the 1970s. 92 Healthcare 1.0 is the name given to this period. Because 93 of a lack of funding, healthcare services were limited and not 94 coordinated with digital systems during this period. On the 95 other hand, bio-medical machines had not yet been built 96 and did not integrate with networked electronic devices. 97 Paper-based medications and reports were commonly used 98 in healthcare institutions during this period, resulting in 99 increased costs and time.
overcome the ''Information and Resource Island'' (infor-197 mation silo), for example, due to privacy concerns and 198 restrictions. In addition, the information silo contributes to 199 unnecessary data redundancy and bureaucracy. 200 In this situation, the United States Congress passed and 201 signed the Health Insurance Portability and Accountability 202 Act (HIPAA) in 1996 [190]. It established standards to pro-203 tect the privacy and security of personal health information, 204 as well as many programs to combat fraud and abuse in the 205 healthcare system, including the following five rules:

219
• The Rule of Enforcement. For breaking HIPAA rules, 220 there will be an investigation and fines.

221
ISO 27789 [9] is another typical audit trail for EHRs that 222 keeps personal health information auditable across systems 223 and domains. A secure audit record must be created every 224 time an operation is triggered by a system that complies with 225 ISO 27789. As a result, a collaborative and open data-sharing 226 system is essential, as it simplifies auditing and post-incident 227 inquiry or forensics in the event of alleged misbehavior (e.g., 228 data leakage). Forensic scholars also do highlight this concept 229 (forensic-by-design) [9], [10]. 230 When the next generation of secure EHR systems has been 231 generated, we should follow the next requirements based on 232 the relevant standards listed above: 233 • Data accuracy and integrity: e.g., unauthorized data 234 modification is not allowed and can be detected.

235
• Data security and privacy.

237
• The patient control mechanism allows the patient control 238 mechanism of EHRs (e.g., the patients will have control 239 over their records and can get a notification if there is 240 unauthorized access or loss of their data).

246
• Safety and security. The blockchain-based decentralized 247 system is resistant to a single point of failure and insider 248 attacks.

249
• The use of a pseudonym. Each node is assigned a 250 pseudonymous public address to safeguard its true 251 identity.
252 VOLUME 10,2022 responding to the coronavirus pandemic [7] is based on these 307 prospective advantages. 308 Disintermediation is defined as the absence of a centralized 309 authority that collects, processes, and validates data & models 310 designed and shared. It enables a reduction in the time, error, 311 and cost of process performance aimed at building and updat-312 ing a predictive model that supports clinical practice and risk 313 management. Transactions certified by the blockchain, and 314 the data included within them are irreversible, in the sense that 315 they cannot be changed or erased, ensuring their legitimacy 316 while also strengthening the security of the system in which 317 the activities take place [9]. Furthermore, the cryptographic 318 system, the immutability of the data communicated across the 319 network, and the lack of a centralized authority foster greater 320 trust in the system, as the need to maintain this confidence 321 among the parties involved in the process fades [10].

323
In the wake of the COVID-19 pandemic, current healthcare 324 systems have come under scrutiny. Currently some existing 325 healthcare systems may be overburdened by the COVID-19 326 outbreak. As of right now, there is no trustworthy data 327 monitoring system in place [18] to give key healthcare orga-328 nizations the information they require about potential epi-329 demics in real-time. In fact, most of the current coronavirus 330 information comes from separate sources such as the public, 331 hospitals, clinical labs with a large amount of inaccurate data 332 without being monitored thoroughly. The use of unreliable 333 information makes it challenging for potential outbreak iden-334 tification and quarantine. Another limitation is the current 335 time-consuming and in-accuracy coronavirus detection pro-336 cedure that often takes several hours to complete the virus 337 tests. This is unacceptable in light of the rapid spread of the 338 coronavirus. It is critical to learn how to swiftly and accu-339 rately identify coronaviruses. Coronavirus data processing 340 utilizing human-dependent medicinal software is exceedingly 341 tough, especially when dealing with complex patterns and 342 enormous volumes. Blockchain technology offers promis-343 ing security solutions to aid in the fight against pandemics. 344 Indeed, the blockchain creates immutable transaction ledgers 345 for medical data sharing systems. More importantly, the com-346 bination of blockchain and smart contract technology elimi-347 nates the need for central servers to ensure fairness among 348 transaction parties. Traceability and decentralization are two 349 key characteristics of blockchain that are not found in other 350 traditional security techniques. Furthermore, blockchain can 351 provide reliable data analytics. Data collection is an important 352 step in disease analytics. How to ensure the reliability of 353 collected data during data collection is important for ensuring 354 the high quality of disease data analytics [66]. The use of 355 incorrect data or untrustworthy database sources can lead to 356 biased analytical results, which can have fatal consequences, 357 such as incorrect diagnosis. Furthermore, in an emergency 358 epidemic situation, many sources of contagious disease data 359 are collected without protection from hospitals, the public, 360 or the media, which can result in data modifications. These 361  [4]. However, human 375 rights and privacy advocates have objected to the plan since 376 it might potentially disclose citizens' private information, 377 which could lead to major civil liberties abuses. To combat 378 the spread of the coronavirus, real-time monitoring systems 379 that protect user privacy are needed. As privacy become more 380 of a concern, secret blockchain networks, that uses Privacy 381 by Blockchain Design (PbBD) technologies to customize the 382 level of privacy, are now gaining attention. 385 We briefly outline blockchain technology to assist readers 386 in comprehending the remainder of the article. In the fol-387 lowing subsections, we will cover the fundamental structure  and P2P distributed consensus to guarantee ledger consis-398 tency and user security. Hence, these time stamped blocks are 399 linked together by a cryptographic hash [11]. Typically, each 400 block contains transaction records that have been verified by 401 peers, often known as miners. The chain is continually length-402 ened, with each new block being added to the end. Each new 403 block, on the other hand, contains a reference to the preceding 404 block's header, which is essentially a cryptographic hash 405 (e.g., SHA-256). the creation of each block has been with 406 pseudonymity, transparency, and immutability [12], [13].

407
A block is made up of the block header and the block body, 408 defined below, as seen in Figure 1.  To obtain a consensus protocol among all the distributed 486 nodes before a new block can be attached to the blockchain, 487 different protocols have been developed [15].

488
• PoW (Proof of Work): PoW is the name of Bitcoin's 489 consensus algorithm (Proof of Work). Before receiving any 490 rewards, a miner node with a certain level of computing 491 (hashing) power must perform laborious task of mining to 492 prove that he is not malicious [99], [100]. To find an eligible 493 nonce value that is smaller than (or equal to) the target hash 494 value, the node must continually perform hash calculations. 495 It is difficult to generate a nonce, yet it is trivial for other 496 nodes to check its validity. The task is costly as a result of the 497 numerous computations required (computational resources). 498 If the blockchain network were to be attacked by a 51 percent 499 attack [191], this would be an extreme case. A miner or a 500 group of miners having more than 51% of the processing 501 power can delay the generation of new blocks and create 502 fraudulent records of transactions that benefit the attackers. 503 • Proof of Stake (PoS) Compared to PoW, PoS uses less 504 power. It is widely believed that nodes with the highest stakes 505 (such as cash) are less likely to attack the network [105]. 506 It's unfair to decide based on account balance because the 507 wealthiest node is more likely to take over the network, 508 making it a centralized one.

509
• Delegated Proof of Stake (DPoS) Similar to PoS, DPoS 510 can also be used. The key distinction between DPoS and 511 PoS is that the DPoS is democratically representative [192], 512 whereas the PoS selection is based on all nodes. Stakeholders 513 can elect delegates to decide who generates and validates new 514  also known as decentralized applications, that allow com-545 munication between patients and doctors to take place with-546 out the need for a third-party intermediary other than the 547 Ethereum network. This will allow patients to have greater 548 control over their records [137]. Pilot programs around the world have begun to study the 554 application of blockchain technology in hospitals, and some 555 of these projects are now underway. After developing and 556 launching a blockchain-based pilot platform in the United 557 States last year, Booz Allen Hamilton Consulting was tasked 558 with advising the Food and Drug Administration's Office of 559 Translational Sciences on how to apply the technology in 560 healthcare data management ( Figure 3). The pilot project, 561 which is presently being implemented at four large hospitals 562 and which makes use of Ethereum to regulate data access 563 via virtual private networks, is being run by the Ethereum 564 Foundation. As a result of its use of IPFS, the project can 565 employ encryption and decrease data replication by utilizing 566 off-chain cloud components and cryptographic techniques to 567 facilitate user sharing [19]. Internet of Things devices. A variety of stakeholders, includ-610 ing patients, doctors, pharmaceutical specialists, and payers 611 will benefit from real-time AI-powered healthcare analyt-612 ics [24], [25]. A key data source for blockchain service 613 providers is the complete PHR service trajectory, which is 614 becoming increasingly important. (See Figure 4).

615
Blockchain technology is also a feasible solution for 616 managing personal electronic health records. Patients may 617 be reimbursed with tokens for providing health data with 618 physicians and research collaborators through the use of 619 so-called ''smart contracts,'' which are electronic contracts 620 that exchange data between parties. Using blockchains to 621 tokenize data, Health Wizz, for example, is experiment-622 ing with a blockchain-and Fast Healthcare Interoperability 623 Resources (FHIR)-enabled EHR aggregator mobile app that 624 will allow patient groups to aggregate and organize their 625 medical records in a safe manner, as well as exchange, donate 626 and/or swap their medical records [24]. To facilitate improved 627 coordination between healthcare institutions and caregivers 628 for a higher level of care, the goal is to make it as simple 629 as managing online bank accounts to manage one's health 630 information.

631
In the context of an EHR blockchain business [24], medical 632 chain allows a variety of healthcare agents to apply for and 633 obtain authorization to view and communicate with patients' 634 medical records. These agents include physicians, hospitals, 635 laboratories, pharmacies, and insurers. In the medical chain's 636 distributed ledger, each interaction is recorded as a transac-637 tion, and the ledger is auditable, open, and stable at all times. 638     In this section, we discuss the issues surrounding the use 708 of AI for health record systems such as the portability and 709 transparency of these algorithms, as well as the requirement 710 of the training sizes necessary for satisfactory productivity. The issue of inadequate transparency related to elaborate 713 algorithms of machine learning such as deep learning creates 714 hurdles to their application in phenotyping tasks, especially 715 when the stakes are significant and end-user confidence is 716 essential. Clinicians, for example, would prefer the algo-717 rithms to supplement or enhance their expertise as opposed 718 to merely dictating their decision-making process [140]. Reg-719 ulatory authorities, on the other hand, require algorithms 720 to be decipherable for transparency reasons because their 721 classification system may have substantial legal or financial 722 ramifications [32]. As a result, improving the interpretability 723 of such ''black box'' models is crucial.

724
The results from previous research that employed a 725 recently established approach in interpreting deep learning 726 model predictions were outstanding [32], [33]. Researchers 727 in [32] applied a modified variant of saliency, which is dubbed 728 'saliency' [34], to classify the most appropriate terms from 729 clinical material and were subsequently utilized for predic-730 tion purposes by convolutional neural networks. Based on 731 the authors, clinicians would assess these terminologies as 732 more descriptive and applicable to the desired trait than the 733 most crucial characteristics determined using a more stan-734 dard definition of extraction-based NLP method.
[33] created 735 heatmaps by using mappings of class activation in agreement 736 with radiologists' assessments, representing the most signif-737 icant portions of chest X-ray images applied by their deep 738 neural network for the prediction of chest diseases [35]. Such 739 initiatives to improve the transparency and interpretability 740 of complicated machine learning models strengthen the trust 741 and confidence of physicians and other end users in these 742 technologies, hence encouraging the number of uses.   This article brings machine learning and data mining together 803 for a joint discussion because both disciplines are based 804 on data science and frequently cross [44]. However, there 805 are a few fundamental differences between data mining and 806 machine learning. The study of methods that can extract 807 information automatically is known as machine learning [44]. 808 Forecasting future events requires two sets of data (training 809 data and test data). On the other hand, data mining is an 810 iterative process of uncovering various types of novel and 811 useful patterns in data.

812
Data mining can employ machine learning, but it can 813 also use other techniques besides or in addition to machine 814 learning to identify new patterns. Machine learning and data 815 mining technologies are employed mainly in the healthcare 816 industry to extract knowledge from vast amounts of electronic 817 health data. Machine learning and data mining approaches 818 were included in the analysis in [45] because they use similar 819 mechanisms for disease prediction and are frequently dis-820 cussed together in the literature. Artificial neural networks (ANNs) were first proposed by 824 McCulloch and Pitts [46] and popularized in the 1980s 825 by [47]. They can handle a range of categorization issues. 826 The word ''neutral'' in their name implies brain-inspired 827 systems designed to mimic how human brains learn cate-828 gories. ANNs were created to mimic the way the human brain 829 works, in which a vast number of neurons are coupled to one 830 another via many axon junctions. Neuron connections can 831 be strengthened or decreased by reinforcing labeled training 832 data, just as they can be in biological learning. A weighted 833 matrix can be used to represent these neuronal connections. 834 This matrix is referred to as a layer, similar to the cortical 835 layers in the brain. The training data used in ANNs serves 836 as a form of 'biological learning' for people. In an ANN 837 framework, there can be one or more hidden layers in addition 838 to the input and output layers. ANNs are taught to generate 839 an output from a set of input variables.

840
Several ANN research focused on the survival prediction 841 problem were found in the literature. However, a few research 842 relying on electronic health data were found.

843
Deep learning is a subfield of machine learning that deals 844 with ANN-inspired algorithms [48]. These algorithms have 845 been utilized to model illness symptoms and hazards in recent 846 years. Liu et al. [49] created a deep learning-based tech-847 nique for early identification of Alzheimer's disease and mild 848 cognitive impairment in 2014. Neuroimaging data from the 849 Alzheimer's Disease Neuroimaging Initiative database was 850 used.

851
To get around the bottleneck, they used stacked auto-852 encoders. Cheng et al. [50] suggested a method for pheno-853 typing patient electronic health records (EHRs) using deep 854 learning. Each patient's EHR was initially represented as a 855 temporal matrix, with time on one axis and events on the 856 other. The researchers built a four-layer convolution neural 857 network (CNN) to extract phenotypes and forecast risk.  Association analysis has been frequently utilized in data min-918 ing and machine learning literature for prediction because 919 it can extract hidden and relevant information from huge 920 datasets [52]. This function generates a collection of dataset 921 item association rules [53]. Power of association is an impli-922 cation statement with X → Y, where X and Y are disjoint 923 item sets (i.e., X ∩ Y = Ø). It means that the existence of 924 X things in current transactions may result in one or more Y 925 items appearing in future transactions. As a result, association 926 analysis has been widely utilized with market basket data to 927 forecast retail sales behavior, where each object reflects a 928 customer's purchase [52].

929
If an item is related to disease and the item set is specified 930 as the patient's set of conditions until now, this method can be 931 applied to the medical context to predict future disease risk. 932 The Hierarchical Association Rule Model (HARM) was 933 introduced in [55] to predict illness risk from medical data 934 using association analysis and a Bayesian estimate. First, a set 935 of association rules is developed utilizing association analysis 936 methods in this modeling technique. Then, using Bayesian 937 estimation, these association rules are ranked. HARM can 938 anticipate a patient's likely future medical issues based on her 939 previous and present history of reported ailments, assuming 940 that each patient regularly consults healthcare professional. 941

942
A network can be represented as a graph, which is made up 943 of nodes (also known as vertices or actors) and edges (also 944 known as ties or links). Edges represent the relationships 945 between things, while nodes represent the entities themselves. 946 Many scientific problems can be represented as graphs and 947 modeled as networks. Many graphs theory approaches and 948 algorithms for analyzing various problems, including dis-949 ease prediction in the healthcare area, can be found in the 950 literature.

951
Many statistical and data mining methods for predicting 952 disease risk from healthcare data do not explicitly take into 953 account the link between diseases and symptoms. Chronic 954 and non-communicable diseases, on the other hand, do not 955 arise in isolation [52]. They frequently share a risk factor, 956 which might be genetic, environmental, or behavioral in 957 nature.

958
These risk factors have a synergistic influence on health 959 outcomes, which makes it difficult to forecast if they are stud-960 ied separately. A network method may be more applicable in 961 this scenario. Statistical methods are also used in a network-962 based approach. Another comparable approach is Social Net-963 work Analysis (SNA), which is built on a solid theoretical 964 foundation drawn from network and graph theories. SNA is 965 the study of the pattern of relationships among network enti-966 ties, such as a group of people, departments, or organizations, 967 as the name suggests. If the elements in the dataset have a lot 968 of relationships between them, SNA can be especially useful. 969 need to confer among themselves about a patient's illness 971 diagnosis. Patients are additionally cared for by pharmacists, 972 nurses, and medical technicians. As a result, the recordings 973 of these dialogues are bound by a network structure.

974
Each sort of entity participating in the healthcare data is 975 represented as a node to describe the health care infrastruc-976 ture as a social network. Edges linking the corresponding 977 node pairs represent relationships between entities. SNA has 978 been utilized to better analyze physician-patient partnerships 979 as well as collaborations throughout a hospital network. 980 Uddin et al. [56] suggested an SNA framework to analyze the

1027
Federated learning is a machine learning technique that is 1028 carried out over numerous computing nodes with the confi-1029 dentiality and privacy of sensitive data protected throughout 1030 data sharing as a precondition. By exchanging encrypted 1031 datasets, different medical organizations can collaborate to 1032 create high-accuracy prediction models. To establish account-1033 ability and reliable cooperation, blockchain as a regulator can 1034 record associated training transactions in an immutable and 1035 transparent manner. Medical organizations and researchers 1036 will be more ready to share encrypted datasets to advance 1037 medical treatment and public health in this circumstance.

1038
The security of data input is ensured by blockchain as a 1039 dependable backbone for machine learning algorithms. The 1040 first challenge raised by [72] is the sharing of huge datasets 1041 across different applications and domains. In reality, however, 1042 homo-morphic encryption has a substantial computational 1043 expense. Perhaps sensitive data can be encrypted in the future 1044 without affecting machine learning for intelligent services. 1045 If the rate of erroneous predictions is high, blockchain 1046 can also be used to store rollback models. The pointers to 1047 essential data of retrained models are stored in a safe and 1048 immutable manner on the blockchain. In the context of erratic 1049 arrhythmia alarm rate, [73] argued that retraining models 1050 indexed by pointers in the blockchain can improve accuracies 1051 for continuous remote systems.

1052
AI can also be used to automate the production of 1053 smart contacts, making processes more secure and adaptable. Whereas they assert ownership of the first attempt to 1058 use blockchain for an EHR management system in their 1059 paper [81], they present a system for exchanging EHRs in 1060 their report, with a focus on security and ease-of-transfer. 1061 The system, however, is still a concept and has yet to be 1062 deployed and tested for its stated areas of progress.
[74] used 1063 blockchain technology to create MedRec, a decentralized 1064 EHR management framework. For simplicity and adaptabil-1065 ity, their modular architecture was combined with an in-place 1066 data storage system. They enticed the medical community 1067 and EHR stakeholders to participate as miners in the net-1068 work's Proof of Work [20] verification. Permission to view 1069 aggregated and anonymized data will be granted in exchange. 1070 In collaboration with the Harvard Medical School Teaching 1071 Hospital, they developed and tested the first working pro-1072 totype. They suggested that future research focus on areas 1073 where miners can rank their preferences for data attributes 1074 (demographic, gender, age group, and so on) to allow preci-1075 sion medicine and targeted research.

1076
In [78], researchers built a prototype that differed greatly 1077 from the MedRec framework's permissionless mining. From 1078 a medical standpoint, they decided to create a closed, access-1079 controlled blockchain EHR system. As a result, cloud storage 1080   2) What standards are used to store EHRs in the 1120 blockchain? After reviewing papers from various categories, selected 1126 papers are presented in this portion. As indicated in 1127 Section VII-B, the article selection query was intentionally 1128 extensive to evaluate as many research issues as feasible.

1129
Selected papers are presented in this segment after screen-1130 ing from various categories. The selection query for the arti-1131 cles was purposely long enough to consider as many research 1132 questions as possible, as described in SectionVII-B. Using the 1133 searching mechanism, we were able to retrieve 1280 research 1134 articles from the scientific repositories, as shown in Figure. 6. 1135 After the first screening step, we removed duplicates and 1136 retrieved 159 papers. Using the second and third screening 1137 methods (here, exclusion was based on title and abstract), 1138 a total of 32 articles were deleted accordingly, leaving 1139 127 papers for further processing. We uploaded the remain-1140 ing papers to the Mendeley software for thorough reading. 1141 Finally, all articles that did not serve the purpose of the SLR 1142 were deducted, and a total of 113 articles were there.

1143
The second analysis we ran, as part of our systematic inves-1144 tigation, was to determine the purpose or field of blockchain 1145 application in the healthcare industry. As indicated in Table 3, 1146

1161
The concept of privacy refers to a person's right to select 1162 when, how, and to what extent they can access, change, and 1163 share their own EHRs. [125]. A healthcare provider may pur-1164 posefully or unintentionally misuse electronic health records 1165 (EHRs) to violate patients' privacy, for example [126]. Many 1166 patients are concerned about their electronic health records 1167 (EHRs), according to the study article [127].

1168
About half of those polled [128] thought that sharing health 1169 data would make it more difficult to protect their personal 1170 information. As a result, when comparing blockchain-based 1171 solutions that claim to protect EHRs' privacy, privacy is an 1172 important consideration.

2) SECURITY
an individual's electronic health records (EHRs) are confined to authorized individuals. According to [129], about half 1177 of patients are concerned about the security of their EHRs 1178 because they must travel via the Internet.

1179
EHR security is more important to a doctor than to patients, 1180 according to [130], and the majority of doctors prefer paper 1181 records over EHRs because they believe the former are safer.
Because doctors use digital health records, they are more vul-1183 nerable to security breaches than paper-based records [131].
be thoroughly studied first. These aspects indicate that the 1186 security of EHRs should be seriously considered.  We see varying degrees of privacy and anonymity [141] 1233 depending on the implementation type of the blockchain: 1234 public, private, or licensed. According to [142], 1235 CORDA [143] protects the transaction's privacy by requiring 1236 validation to be performed only by the persons participating 1237 in the transaction. In the field of Industry 4.0, we discover 1238 the blockchain-based Secure Mutual Authentication System 1239 (BSeIn) [144], which aims to provide privacy and security 1240 assurances such as anonymous authentication, audit capa-1241 bilities, and secrecy. It demonstrates the scalability enabled 1242 by Smart Contracts. They enable privacy via the various 1243 consensus methods employed in blockchain [145]. In other 1244 instances, anonymity is used in [146]. While the work in [147] 1245 emphasizes conditional privacy, it considers traceability of 1246 operations important in the event of a public audit by all 1247 entities participating in the blockchain.

1248
The first references we found to anonymization were 1249 through pseudonymization [141], which is the process of 1250 obliterating some of the information required to identify 1251 an entity. Although they assert in [ 1261 A mechanism in which the identity of the sender is frequently 1262 concealed behind a public key, but other transaction charac-1263 teristics are made public. This presents a difficulty for health 1264 data. One option to limit public exposure is to utilize approved 1265 blockchain technology. One way to safeguard sensitive data 1266 is to implement an out-of-chain solution [141], [153]. The 1267 approach entails locating sensitive data in a system other than 1268 the blockchain and anchoring it to the blockchain's link. This 1269 technique is advantageous for systems that manage enormous 1270 amounts of data, and it would be impractical to incorporate 1271 these data into the blockchain structure. Additionally, it is 1272 recommended for systems that handle highly sensitive data 1273 and require greater access control, such as health data.

1274
The requirement for confidence and privacy necessitates 1275 the development of a mechanism that safeguards cars against 1276 forgeries while safeguarding privacy from monitoring threats. 1277 The work in [154] proposes a Blockchain-Based Anony-1278 mous Reputation System (BARS) to construct a trust model 1279 VOLUME 10, 2022

1338
-Design patterns: Within the confines of the ''hide strat-1339 egy'', design patterns take on a variety of forms. One such 1340 pattern is data encryption (in transit or at rest, anonymization 1341 or pseudonymization), which refers to strategies that disen-1342 tangle certain related events. Data encryption is a type of 1343 security that encrypts data so that it may be accessed only 1344 with the correct encryption key. It converts data to another 1345 format and hence requires a decryption key to retrieve the 1346 data [164]. . This technique necessitates a distributed pro-1353 cessing solution rather than a centralized one. Data from mul-1354 tiple sources should be stored independently and separately. 1355 -Design patterns: No specific design pattern for this strat-1356 egy has been identified to date [143]. 1357 4. Aggregate: According to this technique, personal data 1358 should be managed with the fewest feasible details and at 1359 the highest level of aggregate possible. As a result, this data 1360 becomes less sensitive. When the data is sufficiently uneven, 1361 the group over which it is aggregated is large, and only a 1362 small quantity of data can be ascribed to a single individual, 1363 resulting in privacy protection [143].   [173]. This feature enables the auditing 1426 of the entire blockchain if necessary and that, in the case 1427 of sensitive information, such as EHR, may result in the 1428 exposure of information that enables the transaction's identity 1429 to be determined [141]. Depending on how blockchain is implemented, various pri-1432 vacy concerns may occur, making it easy to track an entity's 1433 transactions. A notable example is given in [141], where 1434 an entity's public key corresponds to its identity in the 1435 blockchain system, allowing for the discovery of all trans-1436 actions linked with that public key. This scenario would be 1437 catastrophic in a public blockchain and might also present an 1438 issue in a private blockchain, as not all members may require 1439 access to transaction data. The case in [141] refers to specific 1440 blockchain implementations that enable selective publication 1441 of private information and rely on zero-knowledge cryptog-1442 raphy for verification. How to apply the GDPR-mandated 1443 right to be forgotten for a patient's data is one of the disad-1444 vantages demonstrated when implementing blockchain in the 1445 health area. Among the downsides of blockchain technology 1446 are the costs involved with authenticating connected data, 1447 auditing different entities and transactions, and the cost of 1448 interoperability provided to the network of participants. The 1449 pseudonym does not ensure transaction privacy, and it is even 1450 feasible to de-anonymize a user's identity through analysis of 1451 incoming and outgoing transactions. Privacy by Blockchain Design develops on data privacy 1454 solutions for the disruptive and rapidly growing new tech 1455 ecosystem. Blockchains can not only be GDPR-compliant, 1456 but they can also help raise data protection levels and truly 1457 give back data ownership to individual patients or their legal 1458 guardians (e.g., family members or the state), by establish-1459 ing general principles and methods for handling personal 1460 data in blockchain ecosystems. PbBD specifies technical and 1461 organizational measures for data protection while taking into 1462 account the principle of ''privacy by design'' as well as 1463 specifications that are inspired by legal frameworks, such 1464 as GDPR. As such, the Blockchain as a great tool for pri-1465 vacy and want to encourage the industry to take the lead in 1466 this area.    [86]. These strategies facilitate the discovery of the most 1507 suitable solutions in identifying the pertinent data sources in 1508 pervasive environments, the best cloud or edge servers for 1509 processing the data and application, as well as in allowing 1510 resource-efficient information management in extensive dis-1511 tribution of computing settings.

1512
The optimization process at present is implemented by 1513 centralizing control and taking into account system-wide and 1514 application-wide enhancement objectives, causing unneces-1515 sary and irrelevant management of data and poor perfor-1516 mance of the system or the application itself [87]. The 1517 application of blockchain enables decentralized optimization 1518 methodologies to bring up new research and development 1519 possibilities. By analyzing highly applicable data, the decen-1520 tralized optimization techniques are advantageous in terms of 1521 improving system performance, particularly when numerous 1522 techniques are executed concurrently across the systems and 1523 applications.

1525
AI apps and systems use planning approaches to collab-1526 orate with other systems and applications, as well as to 1527 solve complex problems in new situations. Planning strate-1528 gies improve the operational efficiency and resilience of AI 1529 systems by gathering current input conditions and performing 1530 different logic and rule-based algorithms to achieve preset 1531 goals [88]. Currently, centralized planning is a laborious 1532 and time-consuming activity. Consequently, decentralized AI 1533 planning techniques based on blockchain are required to 1534 provide a higher degree of robustness with provenance his-1535 tory and continuous monitoring. It is worth noting that the 1536 blockchain ecosystem can also be used to create immutable 1537 and critical blueprints for task-essential systems and relevant 1538 applications. 1539 VOLUME 10, 2022 ble, and deep learning, remain to be the heart of AI systems in 1543 facilitating knowledge discovery and autonomous processes. AI applications must manage data in such a way that is 1594 highly applicable and precise, with full datasets obtained 1595 from credible data sources, along with effective decentralized 1596 storage. In the underlying network, AI applications tradi-1597 tionally have used centralized data management techniques 1598 operated across all nodes [97]. These strategies include but 1599 are not limited to, data segmentation, filtration, context-aware 1600 storage systems and transmission in underlying architecture, 1601 as well as temporal and intelligent management of data sys-1602 tems. When considering decentralized storage networks and 1603 blockchain immutability requirements, inefficient centralized 1604 data management may arise, resulting not only in data redun-1605 dancy in terms of small modifications but also in the transfer 1606 of comparable information several times. In the event of large 1607 datasets is being utilized, the massive size of data transfer 1608 would cause bandwidth to overload quickly and raise the 1609 issue of backhaul network traffic, thus, necessitating decen-1610 tralized data processing for AI systems based on blockchain 1611 structure. By taking into account the data's temporal and 1612 spatial features, decentralized data infra-structure strategies 1613 are intended for application at the network node level. Fur-1614 thermore, decentralized data management systems may place 1615 metadata on the blockchain network to assure data security 1616 and provenance while the conventional large-capacity storage 1617 solutions, including cloud clusters and data centers, might be 1618 utilized to store actual data. For client-centric small datasets, 1619 the metadata and real data are maintained on the blockchain, 1620 with the management of data being done through the network 1621 via token-based incentives for nodes carrying various shards 1622 or participants in swarms.

1624
A trained model's true performance is evaluated after the 1625 distribution in production settings. Model deployment, on the 1626 other hand, is a regular and repetitive process as the devel-1627 opers must constantly improve the models and rectify bias 1628 by generating a certain set of findings while disregarding the 1629 rest of the options to provide particularly useful and educated 1630 judgments. Model deployment is considered a simple itera-1631 tive process in centralized systems. In decentralized systems, 1632 however, poses quite a challenge [98]. Intelligent contract-1633 based model deployment overcomes these difficulties by con-1634 stantly logging changes and preserving unchangeable model 1635 versioning. Furthermore, a model collaboration between var-1636 ious AI applications would be safer and more reliable since 1637 developers can monitor the origin and traces of a specific 1638 model version.    [98]. 1699 The emergence of BaaS is projected to benefit both con-1700 sortium and private blockchain firms by allowing them to 1701 concentrate on creating value through apps development, 1702 validation, and implementation rather than worrying about 1703 the infrastructures associated with the storage, underlying 1704 network, and computation. Besides the fact that the installa-1705 tion of BaaS facilitates the formation of new cross-industry 1706 private-public partnerships, it also helps in the develop-1707 ment of new opportunities and company-customer interaction 1708 models. To construct smart contracts, developers have access 1709 to a single-click setup of BaaS services. On that note, the 1710 incorporation of BaaS with AI services opens up a new world 1711 of possibilities for apps developers, considering that the main 1712 cloud providers currently are offering a plethora of cloud 1713 services for AI applications.

1715
Traditional blockchain systems built a linear infrastructure 1716 using a mixture of a connected list of data frameworks 1717 and hashing algorithms. Nonlinear infrastructure, built upon 1718 graph theory and buffering data modeling, on the other hand, 1719 is growing to meet the needs of instantaneous applications 1720 and to manage massive volumes of data.

1721
• Linear: Blockchain system based on a single chain that 1722 expands linearly, with new blocks inserted at the chain's 1723 end. The early adoption phase of a decentralized system 1724 usually uses single chains despite several flaws asso-1725 ciated with it. For example, single chains would scale 1726 sluggishly, affecting the real-time performance of decen-1727 tralized applications [99], [100]. Furthermore, because 1728 each business situation has its single chain, information, 1729 asset, and value exchange in different chains would be 1730 a challenging task. Single-chain blockchains instead, 1731 may be used for single-task AI systems that conduct 1732 search, refinement, and training, as well as autonomous 1733 AI applications that function in homogenous environ-1734 ments. PoS has shown to be more energy-efficient than PoW, and 1816 it also solves the vulnerability issue by eliminating pseudony-1817 mous validators and allowing only those who possess the 1818 blockchain's native currency to participate. Validators, on the 1819 other hand, have little to risk if they do not authenticate 1820 the transactions on the blockchain, which may delay the 1821 development of new blocks. Although PoS is useful for the 1822 lag-tolerant AI apps, it is not ideal for AI systems, espe-1823 cially in the management of flowing data, changing the iden-1824 tification, and making intelligent decisions on a real-time 1825 basis. PoAc is a mixture of PoW and PoS protocols. Such protocol 1828 aims to address the 51 percent attack problem by implement-1829 ing the PoW algorithm on blank blockchains [105]. This 1830 is done by PoAc protocol solves complicated mathematical 1831 problems first and validators begin to receive incentives, 1832 increasing their holding on the ledger. This allows for the 1833 validators with a sufficient stake in the blockchain to use the 1834 PoS algorithm. Additionally, PoAc is effective in terms of 1835 security, memory, and network connectivity.

1836
As a result, it may be advantageous for AI programs that 1837 require less data accessibility and higher security. According to the PoB protocol, validators can only spend 1840 their coins if they send them to a public, valid, unusable, and 1841 faulty address. After burning their money, users are instantly 1842 authorized to develop new blocks and collect incentives [99]. 1843 Users could benefit from PoB since it allows them to con-1844 tribute in advance and earn interest on the chain while also 1845 becoming approved validators. The protocol also gives an 1846 advantage by fixing the PoW's energy use problem. Fur-1847 thermore, coin burning lowers the number of coins on the 1848 ledger, resulting in a gradual increase in coin value, amount 1849 balancing of currencies on the blockchain, the spending 1850 of unsold coins, and payment of the transaction cost. PoB 1851 protocols can be used by AI systems to urge participants 1852 to keep the value of the underlying judgments. Applica-1853 tions needing a specific degree of precision, a set amount 1854 of clusters or items recognized, for example, can consume 1855 learning models and search trees to keep value over the 1856 ledger.   Handling such massive amounts of data is a significant 1939 challenge. When it comes to dealing with large amounts 1940 of data via blockchain, the task becomes more difficult 1941 because storing data on the blockchain is costly. The 1942 blockchain was initially designed to keep data small in 1943 size, basically the financial transaction information. How-1944 ever, in order to enjoy the benefits of blockchain while over-1945 coming the limitations of data storage capacity, researchers 1946 devised a number of solutions. While many people haven't 1947 considered blockchain's scalability for data storage, oth-1948 ers have focused on storing data in the cloud or in local 1949 databases and linking the address from that storage to 1950 the blockchain. Among the papers we examined for the 1951 review, slightly less than half did not address the major 1952 data storage issue. Authors of [3] [93], [109], [117], [142]. The rest 1958 of the papers proposed using private blockchain or off-chain 1959 storage to handle scalability issues.

1960
The solutions mentioned above to overcome big data issues 1961 are significant, but more research is required to handle a 1962 significant amount of EHRs data.
vate blockchain is at the top of the popularity ranking.
Interoperability refers to a system's capacity to seamlessly 1999 integrate with another system to share critical data. The ease methods that do not result in information blocking [92], [93]. 2006 To begin with, the EHR must have core interoperability. This 2007 enables the entire system to send data to another system while 2008 also receiving data. While the data received will not need to exchange ensures that patient information is provided and 2019 received in a relevant and shareable fashion. Furthermore, 2020 even if the data changes hands, the facts, and meanings will 2021 not be altered.

2022
Third, the EHR must have semantic interoperability, which 2023 allows data to be accurately reorganized and codified so that 2024 any system can receive and interpret the new information. 2025 This means that the language used by one EHR system must 2026 be readable by the next system. This is the highest level 2027 of interoperability possible with significant implications for 2028 patients, clinicians in a health system, and scientists and 2029 researchers who collect data to study patient populations. 2030 Due to the adoption of standardized coding, information is 2031 transferrable and usable at this level. In contrast to stud-2032 ies [69], [115], which lacked the possibility of interoperability 2033 and is not discussed in EMR systems as a result, medical 2034 and health data experts must perform manual inspection and 2035 mapping of predefined ontologies. At the same time, clinical 2036 malpractice is uncontrollable. Furthermore, scalability and 2037 interoperability concerns are at the forefront of current and 2038 future research in this area. The lack of standards for design-2039 ing healthcare applications based on blockchain technology 2040 is revealed by the interoperability challenge.

2041
The second challenge has to do with questions of privacy 2042 and security [115]. The data on the blockchain is spread to 2043 all nodes, resulting in non-compliance with privacy rules and 2044 vulnerabilities. As a result, to protect data privacy and secu-2045 rity, data must be stored off-chain. New privacy technologies, 2046 such as homomorphic and attribute-based encryption, secure 2047 multiparty computation, zero-knowledge proof, obfuscation, 2048 and format-preserving encryption, and may be able to accom-2049 plish data privacy [111]. 2050 Designing using hybrid privacy approaches and leveraging 2051 security-enhancing technology, such as a homomorphic sig-2052 nature, which works better than public-key certificates, could 2053 speed up the different security levels in a system. More sig-2054 nificantly, any malicious attacker can manipulate health data 2055 acquired from hospitals, clinical labs, and patients, rendering 2056 AI learning useless. As a result, utilizing federated learning 2057 mixed with blockchain technology, it is necessary to collect 2058 health data from many sources without any privacy leaks. 2059 Each healthcare organization's central entity is responsible 2060 for any legal difficulties as well as the overall seamless oper-2061 ation of the centralized healthcare systems. A decentralized, 2062 patient-centric system, on the other hand, makes it difficult 2063 to resolve any legal disputes or inconsistencies in the public 2064 blockchain architecture. When personal data is run on con-2065 verging AI and blockchain platforms, for example, copyright 2066 infringement and defamation issues occur.

2067
On the other hand, scalability is the main issue in 2068 blockchain-based healthcare systems [90], [100], [101], 2069 [102], [103], especially when dealing with large amounts of 2070 medical data. Due to the high volume of healthcare data, 2071 it is not feasible to store it on-chain, as this would result in 2072 significant performance degradation.
Due to the consensus method and ledger replication to 2074 all network participants, scalability has always been a con-2075 straint in blockchain networks [137]. In the case of healthcare When blockchain technology is combined with AI in a 2125 variety of real-world healthcare applications, the resulting 2126 systems become more efficient and stable [175]. Machine 2127 learning (ML) and deep learning (DL) are two major branches 2128 of AI that are assisting in the automation of real-world 2129 applications. In the near future, machine learning will be 2130 used in concert with blockchain to manage EHRs. Despite 2131 the difficulties associated with storing, distributing, and 2132 training vital EHR data to design practical applications, 2133 interest among researchers in developing machine learning 2134 and blockchain-based EHR applications has grown tremen-2135 dously [176], [177]. IBM has announced intentions to imple-2136 ment an intelligent blockchain, in which an AI agent performs 2137 various duties such as enforcing laws, improving records, 2138 detecting suspicious activity, and making recommendations 2139 for upgrading smart contracts over a broad network. In the 2140 MATRIX project [178], AI is employed to construct a next 2141 generation blockchain that enables the automated develop-2142 ment of intelligent contracts, enhances protection against 2143 malicious attacks, and enables highly scalable operations. 2144 Various machine learning techniques can be used to detect 2145 fake EHR data, ensuring that only authentic EHRs are main-2146 tained on the blockchain. Deep learning enables the recov-2147 ery and storage of previously damaged scanned medical 2148 records in blockchain for the sake of knowledge enhancement 2149 (e.g., drug analysis and prediction) [179]. Additionally, deep 2150 learning as-a-service (DaaS) is employed on stored EHRs to 2151 accurately forecast future diseases based on current patient 2152 diagnosis reports [180]. Machine learning techniques can 2153 also be employed to protect blockchain networks from large-2154 scale attacks [181]. There are some established projects that 2155 mix AI with blockchain. For example, SingularityNET [182] 2156 focuses on developing AI and blockchain-based networking 2157 for the robot brain, while DeepBrainChain focuses on devel-2158 oping a platform for developing AI algorithms. Additionally, 2159 several machine learning and deep learning-based health-2160 related projects are underway, including the Gamalon project, 2161 TraneAI [183], and Neureal [184].

2163
Due to network congestion and data size, sharing huge vol-2164 umes of EHRs among diverse health care companies is prob-2165 lematic. Recent options for EHR management are limited in 2166 terms of scalability, computing cost, and reaction time. Edge 2167 computing may provide a solution to these difficulties. It can 2168 process a vast amount of data from multiple locations, as edge 2169 computing is comprised of a set of servers/computers [185]. 2170 Researchers in [186] propose using edge computing to extend 2171 cloud services to the network's edge, thereby increasing pro-2172 cessing capacity and device QoS. Edge Processing offers the 2173 advantages of large data storage, extensive networking, and 2174 high computing power, while also enabling secure and reg-2175 ulated scaling for distributed EHR applications. While edge 2176 computing has several drawbacks, including security, vulner-2177 ability to various attacks during message transmission, and 2178 integrity, blockchain-based solutions face several challenges, 2179 including storage, scalability, block size constraints, and 2180 block creation time, all of which can be addressed using edge 2181 computing. Similar approaches for decentralized technology 2182