Big data analytics and sentiment analysis for health misinformation detection: a systematic literature review of influencing factors, implementation challenges and a data-driven adoption framework

1 Introduction

Misinformation is a dominating challenge in the ever-changing digital revolutionary world. In the current age, excessive amount of generated content on digital media has caused concerns related to the accuracy and genuineness of the information being disseminated (Arias et al., 2022; Di Sotto and Viviani, 2022). The creators of misinformation related to health apply different techniques for the promotion of their content to arouse people's sentiments. It has also encouraged the use of sentiment analysis (SA) to capture emotional polarity, attitudes, and opinions expressed in health-related social-media content. Such sentiment-derived information can provide an auxiliary signal for misinformation detection by helping characterize communication patterns and potentially manipulative charged content; however, sentiment analysis alone cannot determine whether a health claim is factually accurate, credible, or misleading (Alonso et al., 2021).

Conceptually, sentiment analysis and misinformation detection represent related but distinct analytical tasks. Sentiment analysis primarily identifies and classifies affective polarity, emotions, attitudes, and opinions expressed in textual and social-media content, whereas misinformation detection seeks to determine whether a claim and information item is inaccurate, misleading, deceptive, and unsupported by available evidence. Consequently, a positive, negative, and neutral sentiment does not establish the truthfulness of a health claim. For example, both factual and misleading health information may be expressed with positive and negative sentiment. Within the context of this review, sentiment analysis is therefore treated as an auxiliary feature that may complement semantic, linguistic, behavioral, temporal, network, source-credibility, and evidence-based approaches to misinformation detection. The actual determination of misinformation requires broader contextual and verification mechanisms rather than sentiment polarity alone.

Digital misinformation on social networks misguides people and negatively affects their sentiments about health related matters (Morita et al., 2023). Authenticity of online health information is difficult due to determine due to the unavailability of quality data (Villanueva-Miranda et al., 2025). Lack of expert system for analyzing huge volumes of social media data, health officials might not detect health misinformation effectively and efficiently (Arora et al., 2026; Cavus et al., 2025; George and Baskar, 2024; Hussain et al., 2023).

The vast amount of online information has brought a paradigm shift in big data analytics (BDA) innovations as it assists to take evidence-based decisions in healthcare sector (Cano-Marin et al., 2023). BDA and natural language processing transformers support the extraction and analysis of sentiment, semantic, linguistic, and contextual signals that can contribute to health misinformation detection (Sharma et al., 2025). Deep learning algorithms and NLP feature extraction allow for nuanced sentiment and emotional tone analysis, enhancing detection frameworks designed to identify deceptive or fabricated user-generated content (Wani et al., 2024). Machine learning techniques are used to develop sentiment models that extract affective and opinion-related features which may contribute to broader misinformation detection systems (Kabir, 2022). Through big data analytics and deep learning algorithms, sentiment-related features can be incorporated into broader frameworks for analyzing and detecting health misinformation (Zamir et al., 2024). BDA can support health professionals in analyzing large-scale health information environments, while sentiment analysis can provide complementary signals within broader misinformation detection and verification processes (Cynthia et al., 2026; Feng et al., 2025; Liu et al., 2019).

Advanced computational techniques can improve sentiment analysis and other contextual representations that contribute to the identification and analysis of health misinformation (Taha, 2025). The integration of big data analytics (BDA) with deep neural network (DNN) proves fruitful in detecting health misinformation from vast amounts of online datasets (Babar et al., 2024). Natural language processing innovations support in reliable sentiment analysis for public health safety through digital informational content analysis (Marques et al., 2022). BDA innovations assist in filtering online datasets for detecting health misinformation to facilitate different stakeholders in order to sustain a healthy and smart city (Azzaoui et al., 2021). Artificial intelligence, neural networks, and fact-checking resources can jointly support misinformation detection, while sentiment analysis may provide additional information about the emotional and attitudinal characteristics of potentially misleading content (Shahzad et al., 2022). Lexicon-based scoring algorithms can characterize sentiment-related signals within health information, which may subsequently be combined with other features for misinformation analysis (Balshetwar et al., 2023; Chatta et al., 2026).

Certain challenges are associated with sentiment analysis on social media data. To accurately identify sentiment of slangs, short texts, emojis, contexts, and linguistic differences from various social media platforms is a prominent challenge (Kansara and Adhvaryu, 2024). The continuous spread of harmful data by spammers on social media causes challenges for healthcare stakeholders to detect online health misinformation (Kansara and Adhvaryu, 2022). Campaigns by different fake content generators on social media to receive public attraction are a critical barrier to the credibility of health related information (Alarabid et al., 2026). Unclear policies about fake online health data lead to serious problems for the incorporation of big data analytics sentiment-driven misinformation detection systems in healthcare information environments (Hussain et al., 2023). Algorithm biases create problems for an efficient and sustainable adoption of big data analytics and sentiment analysis to detect health misinformation (Ahmed and Ahmed, 2023; Malek et al., 2026; Wreyford et al., 2026).

1.1 Statement of the problem

In the current age of rapidly changing technological innovations, health misinformation posted through social media forums, digital communication platforms, and online news portals is a critical challenge. Sentiment analysis (SA) has shown considerable potential for identifying users' emotions, opinions, and attitudes toward health-related information. However, sentiment characteristics alone cannot establish the factuality of health claims. Although BDA and SA can contribute to online health misinformation detection, their effective use requires the integration of sentiment-derived signals with contextual semantics, linguistic characteristics, source credibility, behavioral patterns, temporal information, and evidence-based verification. Therefore, an important research challenge is to understand how BDA and sentiment-related analytical signals can be integrated into broader misinformation detection ecosystems while avoiding the assumption that emotional polarity itself represents evidence of misinformation. The increasing volume, velocity, variety, and veracity of health-related data created on social networking sites have caused challenges to detect health misinformation for making informed decisions related to healthcare. In this context, big data analytics (BDA) is an effective approach to process and analyze heterogeneous datasets in real time, and sentiment analysis (SA) has shown positive prospects in identifying users' emotions, opinions, and attitudes toward health-related information. Although BDA and SA have a potential to detect online health misinformation notwithstanding the existing body of knowledge remains dispersed across diverse disciplines, application domains, and methodologies.

Previous investigations have positively considered the effectiveness of big data analytics, and sentiment analysis approaches in identifying, evaluating, and predicting health misinformation posted through digital platforms. In past, various researchers proposed data science-based approaches and analytical frameworks for detecting health misinformation on social media (Di Sotto and Viviani, 2022; Morita et al., 2023; Zakir Hussain et al., 2023). Similarly, studies have indicated the usefulness of sentiment analysis to examine public perspectives, sentimental responses, and behavioral intentions associated with health misinformation, fake news, and public health crises (Kabir, 2022; Marques et al., 2022; Villanueva-Miranda et al., 2025). Recent investigations have further demonstrated the incorporation of deep learning models, machine learning, transformer architectures, cloud-based systems, and large language models (LLM) to enhance the scalability of misinformation detection systems (Alarabid et al., 2026; Babar et al., 2024; Cavus et al., 2025; Feng et al., 2025; Malek et al., 2026). Although several reviews, surveys, and scientometric studies have examined fake news detection, sentiment analysis, and big data applications independently (Cano-Marin et al., 2023; Kansara and Adhvaryu, 2024; Shahzad et al., 2022), however a comprehensive systematic literature review (SLR) based study focusing on the factors influencing big data analytics and sentiment analysis-based approaches for health misinformation detection has not been conducted yet. It is a pertinent gap in existing body of knowledge that needs to be bridged.

The current study has conducted an in-depth SLR of peer-reviewed journal articles and conference proceedings (n = 30) to consolidate the existing evidence on the factors influencing the adoption of big data analytics and sentiment analysis for health misinformation detection. The study has addressed the need to identify the factors influencing big data analytics and sentiment analysis-based approaches for health misinformation detection, and revealed the challenges associated with the implementation of big data analytics and sentiment-driven misinformation detection systems in healthcare information environments. Based on the study's findings, a data-driven framework has been developed for the sustainable adoption of these technologies for detecting online health misinformation. The study has added a substantial amount of literature in existing body of knowledge by identifying the factors influencing the adoption of BDA and SA for health misinformation detection. It has provided practical implications for policy makers and relevant stakeholders to efficiently adopt and sustain BDA and SA for public health care.

1.2 Objectives of the study

The objectives of the study are as follows:

• To identify the factors influencing big data analytics and sentiment analysis-derived signals within health misinformation detection approaches.
• To reveal the challenges associated with the implementation of big data and sentiment-driven misinformation detection systems in healthcare information environments.

1.3 Research questions of the study

The study's research questions are as follows:

• Which factors influence big data analytics and sentiment analysis-derived signals within health misinformation detection approaches?
• Which challenges are associated with the implementation of big data and sentiment-driven misinformation detection systems in healthcare information environments?
• What is an effective framework to adopt big data analytics and sentiment analysis for health misinformation detection on social media?

2 Materials and methods

2.1 SLR methodology

To address the study's research questions, systematic literature review (SLR) methodology was employed. An SLR is applied using reproducible, transparent, and structured techniques through pre-developed standards. These standards are regarded as protocols that offer insights for all phases of the systematic review based studies to maintain rigor, accuracy, efficiency, and authenticity. These protocols increase reliability, credibility, and validity of research and provide an explicit roadmap for future researchers to conduct similar studies using other dimensions (Ali et al., 2026; Shahzad et al., 2024, 2026; Shahzad and Tariq, 2025).

2.2 Usage of PRISMA

The current study was executed through the Preferred Reporting Items for Systematic Review and Meta-Analyses (PRISMA 2020) guidelines. The review applied the PRISMA framework to ensure methodological transparency, reproducibility, and uniformity. A four-phase flow diagram was displayed to manifest the entire study's selection procedure. The SLR protocol was registered in the Open Science Framework (OSF) and is accessible through: osf.io/wjrm6. Furthermore, the PRISMA 2020 checklist was followed that is shown through Table 1. Each reporting is aligned systematically with the corresponding sections of the review. The required core studies (n = 30) were identified through the criteria of inclusion and exclusion for the final selection to conduct the systematic review on the focused topic.

Table 1. PRISMA 2020 checklist.

2.3 Prospects of SLR method in different fields

The SLR method is applied in several areas of knowledge including agricultural sciences, arts and humanities, behavioral and social sciences, business, economics and administrative sciences, chemical and materials engineering, commerce, computing and information technology, education, electrical, energy and environmental engineering, geo-sciences, health sciences, information and media studies, Islamic studies, law, life sciences, languages, industrial systems engineering, and pharmacy. It offers insightful implications to assist the bodies of knowledge and management. It provides evidence-based insights for informed policy making. The findings drawn from the SLR based studies are considered authentic, valid, effective, and reliable to refine existing systems and services (Iqbal et al., 2023; Khan et al., 2022; Tariq et al., 2026).

Investigators conducting investigations in several disciplines incorporate the methodology of SLR to collect and analyze studies available in worldwide published literature on specific topics. The SLR methodology assists to identify gaps in existing body of knowledge and practice on the basis of data-driven literature. Existing systems, policies, and services are reframed to empower and uplift different institutions (Ahmmad et al., 2025; Butt and Warraich, 2022; Jafree et al., 2021; Mahmood, 2016). In the SLR based studies, there are four stages that are called planning, selection, data extraction, and execution. These stages are elaborated as below systematically.

2.4 Planning

The research team executed planning to conduct an investigation on big data analytics and sentiment analysis for health misinformation detection: a systematic literature review of influencing factors, implementation challenges and a data-driven adoption framework. The following focused objectives of the study were developed:

• To identify the factors influencing big data analytics and sentiment analysis-based approaches for health misinformation detection.
• To reveal the challenges associated with the implementation of big data and sentiment-driven misinformation detection systems in healthcare information environments.

These objectives provided a roadmap to conduct the systematic review on the topic efficiently. Table 2 shows the inclusion and exclusion criteria that were considered for retrieving required studies (journal articles and conference proceedings).

Table 2. Inclusion and exclusion criteria.

2.5 Selection

The research team adopted various strategies to conduct a comprehensive search for the required documents. A comprehensive and structured search strategy was developed from the principal conceptual domains of the review: (i) big data analytics and sentiment analysis, and (ii) health misinformation detection, together with terms representing influencing factors, challenges, determinants, barriers, and drivers. Synonyms and closely related expressions were combined using the Boolean operators “AND” and “OR”. Database- and platform-specific syntax, field codes, controlled vocabulary, and search functions were adapted where necessary to accommodate the technical requirements of individual information sources. Eleven information sources were searched to maximize coverage across computing, health sciences, information science, and multidisciplinary research. These information sources included Google Scholar, Scopus, Web of Science, PubMed, MEDLINE, Ovid MEDLINE, ScienceDirect, Wiley InterScience, PLoS, Global Health, and MDPI. Google Scholar was treated as a scholarly search engine, whereas the remaining sources were searched through their respective bibliographic, publisher, and subject-specific platforms. All searches were conducted during June 2026.

The complete conceptual search strings used for each information source are presented in Table 3. Although the conceptual structure of the search was kept consistent across sources, field codes, controlled vocabulary, and platform-specific syntax were adapted where required. English-language studies published between 1 January 2015 and 31 March 2026 were considered according to the predefined eligibility criteria. The search process retrieved a total of 3,280 records. The use of multiple information sources was intended to improve the breadth of retrieval and reduce dependence on the indexing coverage of any single database or platform. Records identified through the different sources were subsequently subjected to duplication, title and abstract screening, and full-text eligibility assessment before the final set of included studies was established. To identify duplicate records, all retrieved references were imported into Mendeley Reference Manager, where duplicate entries were identified and checked before screening. The identified duplicate records were removed prior to title and abstract screening. This process resulted in the exclusion of 87 duplicate records from the initial 3,280 retrieved records.

Table 3. Complete search strings, information sources, search period, applied limits, and records retrieved.

The applied strategies are displayed through Table 4.

Table 4. Search strategies for retrieving studies.

Table 3 show the search queries and filters that were employed in different digital to retrieve required studies.

Figure 1 shows the complete search process explicitly from the initial retrieval of the studies to the final selection of the most relevant core articles and conference proceedings. It shows that initially 3,280 documents were accessed. From 3,280 documents, 87 duplicate titles were excluded. Four hundred sixty-three studies were discarded due to ineligibility, 329 studies were skipped during the process of screening, and 558 documents were removed during full-text evaluation of the studies. Sixteen non-English language studies were also not included and 1,797 were excluded for not meeting the study's objectives. Finally, the 30 most relevant peer-reviewed research articles and conference proceedings were selected to conduct the systematic literature review (SLR) on the topic.

Although the search retrieved 3,280 records, only 30 studies satisfied all predefined eligibility criteria. The difference between the initial retrieval and final sample should not be interpreted as evidence that the review imposed an arbitrary sample-size restriction. Instead, the substantial reduction resulted from sequential duplication, title and abstract screening, full-text assessment, application of the language and publication-type criteria, and assessment of substantive relevance to the review objectives. The final sample therefore represents the body of evidence that met the pre-specified methodological and topical criteria rather than a predetermined number of studies. Nevertheless, the restrictive nature of the eligibility criteria may have excluded potentially relevant evidence. In particular, limiting the review to English-language peer-reviewed publications and conference proceedings may introduce language and publication bias, while differences in information-source coverage and indexing practices may contribute to selection bias. These limitations were considered when interpreting the findings and developing the data-driven adoption framework.

2.6 Reviewers, screening, study selection and quality assessment

The review process was conducted by the five authors of the study. Following the database searches conducted during June 2026, all retrieved records were compiled and organized for screening. The screening process was conducted in two sequential stages: (i) title and abstract screening, and (ii) full-text assessment against the predefined inclusion and exclusion criteria presented in Table 2. To minimize selection bias, the screening and eligibility decisions were independently undertaken by three reviewers from the author team. Any disagreements concerning the eligibility of a record were discussed among the reviewers and resolved through consensus. Where consensus could not be reached, the remaining two authors were consulted and a final decision was reached through discussion and consensus.

Following study selection, data were extracted using the structured extraction framework presented in Table 5. The extracted information included bibliographic characteristics, geographical context, big data analytics and sentiment-analysis approaches, influencing factors, implementation challenges, and other information relevant to the review objectives. Data extraction was independently undertaken by four reviewers, with discrepancies checked, discussed, and resolved through consensus. The methodological quality of the 30 included studies was assessed using the 10-item CASP-informed critical appraisal framework presented in Table 6. Quality appraisal was conducted by three reviewers. Any differences in appraisal judgments were discussed and resolved through consensus. The quality assessment was used to contextualize the synthesized findings. Studies with methodological limitations were not automatically excluded when they met the predefined eligibility criteria; instead, their methodological strengths and limitations were considered when interpreting the strength and consistency of the evidence.

Table 5. Datasets extracted through 30 studies.

Table 6. CASP checklist for quality appraisal.

Yes = 2, partially = 1, no = 0, barely = 0.5, satisfactorily = 1.5.

Source: Shahzad and Tariq (2025).

2.7 Justification of the final sample and potential selection bias

Although the search retrieved 3,280 records, however the final sample consisted of only 30 studies that satisfied all predefined eligibility criteria. The relatively small final sample was not determined by an arbitrary sample-size target. It resulted from the systematic application of the pre-specified inclusion and exclusion criteria to the complete set of retrieved records. The broad initial retrieval indicates the use of multiple information sources and deliberately inclusive search terms, whereas the substantially smaller final sample shows the specific focus of the review on big data analytics and sentiment analysis in relation to health misinformation detection, including relevant influencing factors and implementation challenges. The studies were not retained merely because they were conceptually relevant to the topic. Records addressing general health misinformation, general social-media analytics, unrelated applications of big data or sentiment analysis, and topics outside the predefined review objectives were excluded when they failed to satisfy the eligibility criteria.

The selection process may be affected by potential publication and selection bias. The review was restricted to English-language, peer-reviewed journal articles and conference proceedings and therefore may have excluded relevant evidence published in other languages or in gray literature, reports, dissertations, books, and other non-indexed sources. In addition, differences in database coverage and indexing practices may have influenced which studies were retrievable. These limitations were considered while interpreting the synthesized evidence. Consequently, the findings should be considered as representing the eligible indexed literature retrieved through the specified search strategy rather than the entire universe of research on big data analytics, sentiment analysis, and health misinformation detection.

2.8 Data extraction and quality appraisal

A structured scorekeeping method was applied to evaluate the methodological quality, relevance, and reporting rigor of the included studies. A predefined 10-item CASP-informed quality assessment framework was developed and applied consistently to all 30 included studies. The assessment criteria examined the clarity of study objectives, appropriateness of the research methodology, sampling procedures, rigor of data collection, consideration of confounding and bias, appropriateness of data analysis, clarity and credibility of results, study clarity, and consistency among the different sections of the study.

Each study was assessed against all predefined criteria. Responses were recorded using a five-category scale: “Yes”, “Satisfactorily”, “Partially”, “Barely”, and “No”. To facilitate systematic comparison and synthesis, the qualitative judgments were converted into numerical scores. A response of “Yes” received 2 points, “Satisfactorily” received 1.5 points, “Partially” received 1 point, “Barely” received 0.5 points, and “No” received 0 points. Thus, each study could obtain a maximum score of 20 points across the 10 assessment criteria. The overall quality percentage was calculated by dividing the total score obtained by the maximum possible score and multiplying the result by 100.

The individual quality-assessment results for all 30 included studies are presented in Table 7. The scoring procedure enabled the methodological strengths and limitations of the included studies to be compared systematically and provided a transparent basis for interpreting the evidence synthesized in the review. As the included literature comprised different types of studies, including empirical, methodological, review, and technical studies, so, the appraisal was applied as a CASP-informed framework rather than being described as a verbatim application of a single design-specific CASP checklist.

Table 7. Quality assessment of the 30 included studies.

Scoring key: QA1–QA10: yes = 2; satisfactorily = 1.5; partially = 1; barely = 0.5; no = 0. Maximum = 20 points.

The appraisal criteria were interpreted in relation to the methodological characteristics and purpose of each study rather than applying identical assumptions mechanically across all study designs. For empirical and observational studies, criteria concerning research objectives, sampling or participant selection, data collection, potential bias or confounding, data analysis, and credibility of findings were assessed according to the procedures reported by the authors. For technical, computational, and algorithmic studies, the corresponding criteria were interpreted in relation to the suitability of data sources, analytical techniques, model development, validation procedures, performance assessment, and reporting transparency. For review, scoping, and scientometric studies, the criteria were interpreted in relation to the clarity of review objectives, transparency of evidence identification and selection, appropriateness of the review or analytical methodology, synthesis procedures, and reporting rigor. For conference-based studies, the same domains were applied according to the methodological information available in the published proceedings paper. This design-sensitive approach retained the same 10 appraisal domains across the evidence base to support comparability while allowing their interpretation to reflect the methodological characteristics of individual studies.

All included studies were independently evaluated by at least two reviewers from the author team to enhance rigor and consistency in the quality-appraisal process. Any discrepancies in scoring were recorded and subsequently resolved through discussion and consensus. Where disagreement persisted, a third reviewer was consulted to make the final decision and minimize the possibility of individual reviewer bias. Inter-rater agreement was assessed using Cohen's Kappa coefficient. A value of κ > 0.75 was interpreted as indicating a high level of agreement among the reviewers in applying the predefined appraisal criteria. This procedure was intended to enhance the transparency, reproducibility, and methodological rigor of the quality-assessment process.

For interpretation purposes, predefined thresholds were established for the methodological quality scores. Studies scoring 75%−100% were classified as having high methodological quality. Studies scoring 50%−74.9% were classified as having moderate methodological quality, and studies scoring below 50% were classified as having low methodological quality. These thresholds were defined operationally for the present review to facilitate consistent interpretation of the appraisal results. The quality categories were used to contextualize and interpret the synthesized evidence rather than as automatic exclusion criteria. Thus, studies with methodological limitations were retained when they satisfied the predefined eligibility criteria, while their methodological quality was considered when interpreting the consistency and strength of the reported evidence.

The quality assessment was used primarily to contextualize the synthesized evidence. Studies that satisfied the predefined eligibility criteria were not automatically excluded due to methodological limitations. The identified strengths and limitations were considered while interpreting the consistency and robustness of the evidence and while developing the data-driven adoption framework. Table 7 displays quality mapping scores of the 30 included studies.

For mapping the selected research articles and conference proceedings systematically (n = 30), a structured form (Table 5) was formulated. The form included the following components:

• Authors with publishing years.
• Geographical territories.
• Journals.
• Factors influencing big data analytics and sentiment analysis-based approaches for health misinformation detection.
• Challenges associated with the implementation of big data and sentiment-driven misinformation detection systems in healthcare information environments.

2.9 Execution and thematic synthesis

The extracted evidence was synthesized using a structured thematic synthesis approach aligned with the two research objectives. The synthesis was conducted in four sequential stages. Firstly, the findings extracted from the 30 included studies were reviewed to identify recurring concepts relating to big data analytics (BDA), sentiment analysis (SA), health misinformation detection, influencing factors, and implementation challenges. Secondly, relevant findings were coded according to their substantive meaning rather than according to the terminology used by individual studies. Thirdly, conceptually related codes were compared, grouped, and organized into descriptive themes representing recurring technological, data-related, organizational, contextual, and implementation-related factors and challenges. Fourthly, the descriptive themes were compared across studies and developed into higher-level analytical categories that directly addressed the research objectives and informed the proposed data-driven adoption framework.

During synthesis, similarities and differences across studies were considered related to research design, geographical context, data sources, analytical approaches, application domains, and methodological quality. Where different studies used different terminology for conceptually similar factors, related concepts were consolidated into common analytical themes while preserving their substantive meaning. Conversely, conceptually distinct factors were retained as separate themes even when they occurred within the same studies. The methodological quality scores were also considered while interpreting the consistency and robustness of the supporting evidence. Because the included studies differed substantially in research design, measures, datasets, contexts, and outcomes, statistical pooling was not considered appropriate. The findings were therefore synthesized narratively and thematically rather than through meta-analysis.

2.10 Evidence traceability and support for major factors

To strengthen evidence traceability, each major factor and implementation challenge identified through the thematic synthesis was linked to the individual included studies from which supporting evidence was extracted. The Results section presents the major themes with citations to the contributing studies. The number and methodological characteristics of supporting studies were considered during interpretation; however, study frequency alone was not treated as an indicator of evidential strength. In the last phase, the extracted datasets were analyzed thematically for addressing the study's research objectives. The following components were considered for interpretation:

• Geographical distribution of studies.
• Research methods applied in the selected studies.
• Factors influencing the adoption of big data analytics (BDA) and sentiment analysis (SA) for health misinformation detection.
• Challenges to incorporate BDA and SA.
• Discussion.
• Framework to adopt BDA and SA for health misinformation detection on social media.
• Implications.
• Limitations and Future Research Directions.
• Conclusion.

3 Results

Table 5 displays an overview of the selected studies. It shows information about authors with publishing years geographical territories, and journals. It also manifests information about factors influencing big data analytics and sentiment analysis-based approaches for health misinformation detection and challenges associated with the implementation of big data and sentiment-driven misinformation detection systems in healthcare information environments.

3.1 Geographical distribution of studies

Findings of the study showed that most of the included studies were produced from India (n = 7). Both USA and Canada contributed three studies each related to factors influencing big data analytics and sentiment analysis based health misinformation detection. The results displayed that the included studies (n = 30) were published from 17 different regions of the world. Figure 2 displays geographical distribution of the selected studies.

3.2 Research methods applied in the selected studies

The findings showed that the experimental method was the most applied methodology in the included studies (n = 10). Literature review methodology was employed in five studies whereas content analysis method was applied in four studies. In the included studies (n = 30), 10 different research methodologies were used. Figure 3 shows the research methods applied in the selected studies.

3.3 Factors influencing big data analytics and sentiment analysis-based health misinformation detection

The findings indicate that advanced deep learning architectures, feature engineering, sentiment-derived signals, real-time analytics, and human-centered verification are important components influencing data-driven health misinformation detection. Importantly, sentiment analysis was identified as an auxiliary analytical signal that can capture emotional, attitudinal, and opinion-related characteristics of health-related content. It does not independently establish whether information is factual or misleading. Reliable misinformation detection therefore requires the integration of sentiment signals with contextual semantics, source and evidence assessment, behavioral or temporal patterns, and other verification mechanisms. These factors are interpreted as follows alternately:

3.3.1 Advanced deep learning architectures

The results showed that advanced deep learning architecture is a key factor to identify health misinformation through deep neural representation learning techniques (Alarabid et al., 2026; Arora et al., 2026; Cavus et al., 2025; Villanueva-Miranda et al., 2025). Large language models contributed contextual reasoning through domain specific fine tuning to reduce reliance on manual datasets (Alarabid et al., 2026; Malek et al., 2026). Machine learning frameworks enhanced semantic abstraction that proved fruitful in strengthening the identification of misinformation in healthcare domain (Babar et al., 2024; Zamir et al., 2024). Semantic representations played a vital role in solving conventional lexical complications (Alonso et al., 2021; Arora et al., 2026; Cavus et al., 2025). Attention mechanisms proved impactful in enhancing contextual feature extraction through priority on clinical relevant textual signals and reduced risks of fake content generated by users (Arora et al., 2026). Multi classification strategies attained rigor by controlling heterogeneous misinformation patterns in healthcare industry (Ahmed and Ahmed, 2023; Kabir, 2022; Zamir et al., 2024). Advanced deep learning architectures significantly detects health misinformation across diverse health environments by improving adaptive learning, contextual reasoning, and semantic generalization (Alarabid et al., 2026; Arora et al., 2026; Villanueva-Miranda et al., 2025).

3.3.2 Feature engineering

The review demonstrated that feature engineering is a valuable factor that positively assists to identify health misinformation. Authentic data, integration, and engineering of underlying datasets proved valuable in detecting health misinformation (Di Sotto and Viviani, 2022; Sharma et al., 2025). High quality data improved classification performance across transformer based frameworks and machine learning models (Babar et al., 2023; Cavus et al., 2025; Di Sotto and Viviani, 2022; Zamir et al., 2024). Feature engineering enhanced analytical coverage through heterogeneous users' interactions at social media and digital health environments (Arias et al., 2022; Cynthia et al., 2026; Morita et al., 2023). Different feature-engineering approaches, including sentiment-score extraction, optimization, and automatic feature selection, can reduce redundant information and provide complementary affective signals for broader misinformation classification (Babar et al., 2024; Balshetwar et al., 2023). The integration of structured and unstructured datasets enabled predictive analytics for detecting health misinformation (George and Baskar, 2024; Sharma et al., 2025). Entity recognition, domain specific health lexicons and semantic normalization accurately extracted health relevant misinformation (Arias et al., 2022; Shahzad et al., 2022). Feature engineering serves as a foundational enabler to efficiently operationalize health misinformation identification (Di Sotto and Viviani, 2022; George and Baskar, 2024; Sharma et al., 2025).

3.3.3 Sentiment-derived intelligence as an auxiliary signal

The review identified sentiment-derived information as an important auxiliary signal within broader health misinformation analytics. Sentiment analysis covers emotional polarity, attitudes, opinions, and affective characteristics expressed in health-related digital content. These signals can complement semantic and contextual representations when analyzing potentially misleading information; however, sentiment polarity alone does not establish the factual accuracy of a health claim. Sentiment analysis integrated with contextual semantic modeling can contribute to misinformation classification by providing information about emotional and attitudinal patterns that may be relevant to the communication and dissemination of health-related content (Alonso et al., 2021; Shahzad et al., 2022). Emotion-aware features may assist in characterizing potentially manipulative or highly affective communication, but such characteristics require interpretation alongside other evidence rather than being treated as direct indicators of misinformation (Alonso et al., 2021).

Sentiment-derived information can also enhance topic-sensitive analysis by helping characterize public reactions to health-related information and identifying patterns of emotional response across digital platforms (Arias et al., 2022). Similarly, health-aware sentiment classification can help distinguish different forms of emotional expression, including potentially harmful or misleading narratives and legitimate expressions of psychological stress; nevertheless, such distinctions require contextual interpretation rather than sentiment polarity alone (Arora et al., 2026). Linguistic and semantic representations remain necessary because misinformation may be communicated through neutral, positive, or negative sentiment. Consequently, contextual language modeling, domain-specific knowledge, user behavior, source characteristics, and other evidence can provide information that sentiment analysis cannot independently establish (Liu et al., 2019; Shahzad et al., 2022). Multilingual and culturally diverse expressions further demonstrate the need to interpret sentiment within its linguistic and social context (Chatta et al., 2026; Zamir et al., 2024). On the whole, the evidence suggests that sentiment analysis is analytically valuable when incorporated with contextual semantics, user behavior, domain-specific health knowledge, and verification mechanisms. It should therefore be considered as a complementary analytical signal within a multidimensional misinformation detection system rather than as a standalone mechanism for determining whether health information is true or false.

3.3.4 Real-time analytics

The findings showed that real time analytics are an important factor that determines operational effectiveness in health misinformation domain (Kansara and Adhvaryu, 2024; Taha, 2025). Cloud based analytics efficiently processed large volumes of health datasets to identify patterns of health misinformation on social media streams (Cavus et al., 2025; Kansara and Adhvaryu, 2024; Taha, 2025). Real time streaming analytics supported immediate identification of health misinformation to effectively respond to public health events (Babar et al., 2024; Azzaoui et al., 2021). Real time integration of steam processing increased analytical capability while maintaining monitoring of fake health information across various digital platforms (Kansara and Adhvaryu, 2022). Cloud based storage and visualization assisted in training infrastructures for large scale sentiment analysis (Cavus et al., 2025; Morita et al., 2023). Internet of Things (IoT) powered ecosystems expanded the scope of social sentiment patterns to ensure public health safety (Azzaoui et al., 2021). The results show that scalable infrastructure is a strategic enabler that facilitates in real time health misinformation detection (Kansara and Adhvaryu, 2024; Morita et al., 2023; Taha, 2025).

3.3.5 Human-centered verification

The study showed that human centered verification enhances the reliability of misinformation detection systems in healthcare domain (Feng et al., 2025; Malek et al., 2026). Domain specific health information enhanced contextual reasoning through structured medical evidence (Feng et al., 2025). Human based verification reduced unsupported inference during automatic claims verification (Malek et al., 2026). Human based verifications enhanced trustworthiness and reliability in sensitive healthcare contexts (Feng et al., 2025; Malek et al., 2026). AI driven support dashboards through the involvement of humans supported in actionable public health intelligence to assist health authorities and policymakers to take evidence based decisions (Hussain et al., 2023). The accuracy of big data and sentiment analytics depends on human supervised decision frameworks apart from algorithmic performance in the current AI dominant age (Feng et al., 2025; Hussain et al., 2023; Malek et al., 2026).

Table 8 illustrates the major and their aligned sub-factors influencing the adoption of big data analytics (BDA) and sentiment analysis (SA) for health misinformation detection.

Table 8. Factors influencing the adoption of BDA and SA for health misinformation detection.

3.4 Challenges associated with the implementation of big data and sentiment-driven health misinformation detection systems

The results displayed that annotation limitations, scalability constraints, domain adaptation challenges, ethical concerns, and sustainability challenges negatively impacted the adoption of big data and sentiment driven health misinformation detection systems. These challenges are elaborated as follows:

3.4.1 Annotation limitations

Annotation limitations are a key challenge to detect health misinformation on social media platforms effectively. Data heterogeneity is a key adoption challenge as health misinformation is scattered across various platforms with diverse contextual, structural, linguistic, and behavioral patterns (Ahmed and Ahmed, 2023; Di Sotto and Viviani, 2022). The shortage of high-quality annotated health-related sentiment and misinformation datasets creates barriers for developing and evaluating integrated detection systems (Arias et al., 2022; Feng et al., 2025). Informal content posted on digital media forums related to health created processing complications in non-English settings (Chatta et al., 2026; Hussain et al., 2023). Sampling and algorithmic biases can reduce the representativeness of sentiment estimates and broader misinformation-detection findings derived from diverse social-media datasets (Cano-Marin et al., 2023; Cynthia et al., 2026). Unstructured datasets, and mixed sentiments increased the possibilities of misclassification (Balshetwar et al., 2023; Kabir, 2022). Manual annotation proved difficult to maintain in a rapidly evolving misinformation in healthcare environments (Morita et al., 2023). Annotation inconsistencies caused problems for effectively detecting health misinformation (Ahmed and Ahmed, 2023; Chatta et al., 2026; Di Sotto and Viviani, 2022).

3.4.2 Scalability constraints

The findings of the study showed that scalability constraints proved a major barrier for large scale adoption of advanced health misinformation detection frameworks (Alarabid et al., 2026; Cavus et al., 2025). Large language models required adequate memory capacity, and processing power for sustainable functioning (Babar et al., 2024; Villanueva-Miranda et al., 2025). Real time analytics produced latency limitations that created complications for quick detection and analytical accuracy (Babar et al., 2024; Kansara and Adhvaryu, 2022). The adoption of cloud based systems increased infrastructure costs and dependency on external resources (Alarabid et al., 2026; Cavus et al., 2025). High dimensional big data environments caused difficulties related to storage optimization and effective model training (Kansara and Adhvaryu, 2024). During health emergencies, resource processing requirements caused complications due to the increase of misinformation volumes in healthcare domain (Hussain et al., 2023; Taha, 2025). Scalability limitations occurred due to the collective effects of streaming data velocity, infrastructure finances, need for real time processing, and model complexity (Alarabid et al., 2026; Cavus et al., 2025; Taha, 2025).

3.4.3 Domain adaptation challenges

The findings showed that domain adaptation challenges cause barriers for effectively detecting health misinformation. Contextual interpretation is a key challenge as misinformation includes emotionally implicit language that may not be effectively managed (Kansara and Adhvaryu, 2024; Zamir et al., 2024). Domain adaptation is required not only because sentiment expressions vary across populations and health contexts, but also because the same sentiment can accompany both factual and misleading information. Therefore, contextual and domain-specific representations are necessary to interpret sentiment signals appropriately within misinformation detection systems (Alonso et al., 2021). Diversity in sentiment expressions across health settings reduced the accuracy ratio of models trained on a single misinformation context (Alonso et al., 2021; Arias et al., 2022). Language related processing complexities negatively affected multilingual detection performance (Chatta et al., 2026; Liu et al., 2019). To differentiate misinformation from authentic mental health expressions caused contextual complication due to the overlapping of linguistic and emotional patterns (Arora et al., 2026). The fast nature of the content posted on social media platforms caused barriers to efficiently identify health misinformation patterns (George and Baskar, 2024; Shahzad et al., 2022). To successfully detect health misinformation on the social media, it was essential to effectively adopt context models capable of managing diverse linguistic patterns and domain specific semantic ambiguity (Arora et al., 2026; George and Baskar, 2024; Kansara and Adhvaryu, 2024).

3.4.4 Ethical concerns

Ethical and privacy concerns are key adoption barriers during the analysis of sensitive patient-generated and mental health-related datasets (Arora et al., 2026; Wreyford et al., 2026). Unreliable automated reasoning and unsupported health recommendations caused challenges for detecting misinformation (Alarabid et al., 2026; Malek et al., 2026). The unavailability of standards reduced transparency and created trust issues in automatic misinformation classification related to health contexts (Alarabid et al., 2026; Villanueva-Miranda et al., 2025). Cyber-security issues were linked to the adoption of big data analytics and sentiment analysis for health misinformation detection (Azzaoui et al., 2021; Taha, 2025). Biases in social media datasets created concerns related to ethical AI adoption within public health decision-making (Cano-Marin et al., 2023; Villanueva-Miranda et al., 2025). Inadequate external knowledge sources caused trust issues in AI based health misinformation detection systems (Alarabid et al., 2026; Malek et al., 2026; Villanueva-Miranda et al., 2025).

3.4.5 Sustainability challenges

Sustainability challenges are linked to the adoption of big data analytics and sentiment analysis for health misinformation detection. Operational integration is a key challenge in AI based misinformation monitoring system into existing healthcare systems (Hussain et al., 2023). The lack of coordination among key stakeholders including healthcare decision makers, policy makers, and technical systems create sustainability issues (Hussain et al., 2023). The unavailability of a comprehensive infodemic surveillance frameworks caused barriers for the continuous misinformation monitoring (Hussain et al., 2023). Architectural complications happened due to integration issues of multiple analytics, storage, classification modules, and visualization (Babar et al., 2024; Morita et al., 2023; Sharma et al., 2025). Cross-platform inconsistencies complicated the adoption of unified context-aware detection frameworks (Liu et al., 2019; Shahzad et al., 2022). The findings demonstrated that long-term sustainability depends on institutional governance, interoperable infrastructure, and integrated public health surveillance ecosystems (Hussain et al., 2023; Morita et al., 2023; Sharma et al., 2025). Figure 4 shows the challenges for the adoption of big data analytics (BDA) and sentiment analysis (SA) for detecting misinformation in health domain.

4 Discussion

The study aimed to identify the factors influencing big data analytics (BDA) and sentiment analysis (SA)-based approaches for health misinformation detection and the challenges associated with implementing big data and sentiment-supported misinformation detection systems in healthcare information environments. The systematic literature review (SLR) synthesized evidence from 30 peer-reviewed studies and conference proceedings. Rather than treating the identified factors as isolated technical elements, the findings indicate that effective health misinformation detection depends on their interaction within a broader data, analytical, human, and governance ecosystem.

A major finding of the review is the transition from conventional feature-based machine learning toward context-aware deep learning, transformer-based architectures, and large language models. Earlier studies emphasized linguistic, statistical, and behavioral features for misinformation classification (Alonso et al., 2021; Liu et al., 2019), whereas more recent investigations increasingly employ contextual representations generated through BERT, transformer models, and large language models (Alarabid et al., 2026; Arora et al., 2026; Cavus et al., 2025; Malek et al., 2026). This progression suggests that the increasing contextual and linguistic complexity of health misinformation has reduced the sufficiency of conventional lexical approaches. However, the evidence does not imply that more sophisticated models necessarily provide reliable detection in all settings. Their performance remains dependent on representative datasets, contextual adaptation, external evidence, and appropriate validation.

Another important finding concerns the role of sentiment analysis. The reviewed literature indicates a growing convergence between sentiment analysis and broader misinformation analytics. Earlier sentiment-oriented studies tended to treat emotional polarity as a relatively independent analytical variable (Alonso et al., 2021; Marques et al., 2022), whereas more recent approaches integrate sentiment-derived information with semantic representations, contextual embedding, and behavioral indicators (Ahmed and Ahmed, 2023; Cavus et al., 2025; Zamir et al., 2024). This finding requires a critical qualification which indicates that the presence of positive, negative, or neutral sentiment cannot independently establish whether health information is factual or misleading. Sentiment is therefore better interpreted as a complementary signal that may help characterize how information is communicated, received, and disseminated. Factuality assessment requires additional evidence, including contextual semantics, source credibility, external knowledge, temporal propagation, user behavior, and verification mechanisms. This distinction is important because treating sentiment classification as equivalent to misinformation detection could lead to conceptual and methodological overstatement.

The review also indicates that data quality may be as important as, or in some circumstances more consequential than, algorithmic complexity. Although several recent studies report strong predictive performance from deep learning and LLM-based architectures (Alarabid et al., 2026; Cavus et al., 2025; Malek et al., 2026), these approaches depend on representative, balanced, accurately annotated, and domain-specific datasets (Di Sotto and Viviani, 2022; Feng et al., 2025; Sharma et al., 2025). Evidence from Morita et al. (2023) and Hussain et al. (2023) similarly highlights the importance of expert-validated information and appropriate data infrastructures. Thus, the findings challenge a purely technology-centered interpretation of misinformation detection which shows that increasing model sophistication cannot fully compensate for weaknesses in data quality, annotation, contextual representation, or domain adaptation. A further contribution of the review is the identification of health misinformation detection as an increasingly dynamic rather than static classification problem. Earlier approaches generally emphasized classification of already disseminated content (Ahmed and Ahmed, 2023; Liu et al., 2019), whereas more recent research emphasizes streaming analytics, continuous surveillance, dashboard-based monitoring, anomaly detection, and predictive assessment (George and Baskar, 2024; Hussain et al., 2023; Morita et al., 2023). This development is particularly relevant to health emergencies, during which misinformation narratives can change rapidly. Nevertheless, continuous or real-time detection introduces additional requirements for computational capacity, infrastructure, model updating, and monitoring. Real-time capability should therefore be regarded as an operational requirement that must be balanced against scalability and resource constraints rather than as an inherent indicator of system superiority.

The implementation challenges identified in this review demonstrate that multilingual communication, sarcasm, implicit emotions, concept drift, platform-specific language, and contextual variation can reduce the transferability of models across datasets and information environments (Ahmed and Ahmed, 2023; Balshetwar et al., 2023; Chatta et al., 2026; Kansara and Adhvaryu, 2024; Zamir et al., 2024). At the same time, transformer architectures and LLMs may require substantial computational resources, storage capacity, cloud infrastructure, and financial investment (Alarabid et al., 2026; Cavus et al., 2025; Taha, 2025; Villanueva-Miranda et al., 2025). Consequently, there is a potential trade-off between analytical sophistication and practical scalability. This issue is particularly relevant to healthcare environments with limited technological and financial resources. It suggests that future research should examine computationally efficient and context-adaptive approaches rather than focusing exclusively on predictive performance.

The findings of this review are broadly consistent with earlier reviews that have emphasized the importance of data science, machine learning, natural language processing, and contextual analysis in health misinformation detection (Di Sotto and Viviani, 2022; Liu et al., 2019). However, the present review extends this literature in several respects. Firstly, whereas previous reviews have largely concentrated on the technological approaches used to identify misinformation, the current synthesis brings together analytical factors and implementation challenges within the same evidence base. This provides a broader perspective on not only what techniques are being used but also the conditions and barriers affecting their implementation. Secondly, the review provides a more explicit interpretation of the relationship between sentiment analysis and factuality assessment. Existing literature has demonstrated the usefulness of sentiment and emotion-related features in analyzing health communication (Alonso et al., 2021; Shahzad et al., 2022), however, the present synthesis emphasizes that sentiment should not be treated as a direct proxy for truthfulness. This distinction assists in clarifying the conceptual boundary between sentiment classification and misinformation detection and addresses an important methodological issue identified across the reviewed studies. Thirdly, compared with reviews that primarily examine misinformation detection as a classification task, the current review emphasizes the movement toward integrated, adaptive, human-supported, and governed detection ecosystems. The combination of heterogeneous data acquisition, intelligent data management, AI-driven analysis, human validation, and adaptive governance is therefore presented as a conceptual synthesis of the evidence rather than simply another classification architecture. This broader perspective is particularly relevant because the challenges identified in the literature extend beyond model accuracy to include annotation, scalability, domain adaptation, privacy, bias, cyber-security, interoperability, and sustainability.

The comparison with existing reviews should be interpreted cautiously. The relatively small number of included studies, methodological heterogeneity, English-language restriction, and absence of quantitative synthesis limit the extent to which the present review can establish that one analytical approach is superior to another. The contribution of the study is therefore primarily integrative and conceptual, rather than a statistical demonstration of comparative effectiveness. On the whole, the evidence suggests that future health misinformation detection research should move beyond isolated algorithmic classification toward systems that combine contextual intelligence, heterogeneous data, sentiment-derived signals, evidence assessment, human supervision, and adaptive governance. The proposed framework in Figure 5 reveals this synthesis by organizing these requirements into an evidence-informed conceptual structure. Its empirical effectiveness, however, remains to be established through future implementation and validation studies.

5 Framework for big data analytics and sentiment analysis in health misinformation detection

Based on the synthesis of the reviewed studies, an evidence-informed conceptual framework for big data analytics (BDA) and sentiment analysis (SA) in health misinformation detection is proposed. The framework was developed by synthesizing the recurring factors influencing BDA- and SA-based misinformation detection and the challenges associated with their implementation. Rather than being derived from a single existing theoretical model, the five components were conceptually developed by grouping the reviewed evidence according to the functional requirements of a health misinformation detection ecosystem. These components comprise: (1) multi-source data acquisition layer, (2) intelligent data management system, (3) AI-driven analytical engine, (4) human validation layer, and (5) adaptive governance.

The derivation of the five components follows a sequential and interconnected logic. Evidence concerning heterogeneous information sources, diverse digital platforms, and the need for broader contextual coverage informed the multi-source data acquisition layer. Findings related to data quality, annotation limitations, feature engineering, multilingual processing, semantic representation, and domain adaptation informed the intelligent data management system. Evidence concerning advanced deep learning, machine learning, large language models, sentiment-derived signals, contextual reasoning, and real-time analytics informed the AI-driven analytical engine. Findings emphasizing human supervision, explainability, expert verification, evidence assessment, and trust informed the human validation layer. Finally, evidence concerning ethical and privacy concerns, bias, cyber-security, scalability, interoperability, sustainability, continuous monitoring, and institutional coordination informed the adaptive governance component. Thus, the proposed framework represents an evidence-informed synthesis of the major technological, analytical, human, and governance requirements identified in the reviewed literature.

The framework is proposed as a conceptual structure informed by the reviewed evidence, rather than as an empirically validated model. The reviewed studies provide support for the relevance of individual factors and mechanisms, but the present review does not empirically test the five components as an integrated framework. Therefore, the framework is intended to organize the evidence and provide a coherent conceptual basis for future implementation and empirical investigation. Its effectiveness, predictive capability, and relationships among individual components require validation through future empirical research.

5.1 Multi-source data acquisition layer

The multi-source data acquisition layer represents the foundational input component of the framework. Its inclusion was derived from evidence concerning data heterogeneity, diverse information environments, cross-platform variation, and the need for contextual coverage identified across the reviewed studies. Health misinformation is disseminated through multiple digital environments, including social media platforms, online news portals, healthcare repositories, public health databases, IoT-enabled health systems, and patient-generated digital content. The reviewed literature consequently indicates the importance of considering diverse sources when constructing datasets for health misinformation analysis (Arias et al., 2022; Azzaoui et al., 2021; Feng et al., 2025; Morita et al., 2023).

This component also responds to the annotation limitations and data heterogeneity challenges identified in the review. Health-related content varies considerably in structure, language, context, and user behavior across platforms. The scarcity of high-quality annotated datasets, multilingual content, informal health communication, sampling limitations, and annotation inconsistencies can constrain the representativeness and reliability of subsequent analysis (Ahmed and Ahmed, 2023; Chatta et al., 2026; Di Sotto and Viviani, 2022; Feng et al., 2025). Accordingly, the framework positions multi-source acquisition as the initial stage through which heterogeneous health information can be systematically collected for subsequent processing and analysis. The role of this layer is therefore not to establish misinformation independently, but to provide sufficiently diverse and contextually relevant information for downstream analytical processes. Its inclusion directly reflects the reviewed evidence that fragmented and heterogeneous health information environments constitute an important consideration in developing BDA and SA-based misinformation detection systems.

5.2 Intelligent data management system

The intelligent data management system was derived from recurring evidence concerning data quality, annotation limitations, feature engineering, multilingual processing, entity recognition, semantic representation, and domain adaptation. The reviewed studies demonstrate that the quality, structure, and contextual representation of data are important considerations for downstream misinformation analysis (Chatta et al., 2026; Di Sotto and Viviani, 2022; Sharma et al., 2025; Zamir et al., 2024). This component incorporates data cleaning, normalization, multilingual processing, feature engineering, entity recognition, semantic enrichment, and contextual preprocessing. These functions respond directly to the challenges identified in the review, including heterogeneous and unstructured datasets, annotation inconsistencies, mixed sentiments, multilingual content, and domain-specific semantic ambiguity (Ahmed and Ahmed, 2023; Balshetwar et al., 2023; Chatta et al., 2026; Di Sotto and Viviani, 2022).

Feature engineering is particularly relevant because the reviewed studies identified sentiment scores, entity-related information, health-specific lexicons, semantic representations, and other extracted characteristics as useful analytical features (Arias et al., 2022; Babar et al., 2024; Shahzad et al., 2022). At the same time, the literature indicates that sentiment-related characteristics cannot be interpreted independently of context. The same emotional expression may occur in both factual and misleading health information. Consequently, this component is designed conceptually to preserve and integrate contextual, linguistic, semantic, and domain-specific information rather than reducing health content to isolated sentiment categories. The intelligent data management system therefore provides the conceptual bridge between heterogeneous data acquisition and subsequent AI-based analysis. Its inclusion displays the reviewed evidence that reliable analytical outcomes depend on the sophistication of algorithms and on the quality, representation, and contextual preparation of the underlying data.

5.3 AI-driven analytical engine

The AI-driven analytical engine was derived from the strongest cluster of evidence concerning advanced deep learning architectures, machine learning, transformer models, large language models, semantic representations, sentiment-derived signals, contextual reasoning, and real-time analytics. The reviewed studies report the use of increasingly advanced computational approaches for analyzing and classifying health misinformation (Alarabid et al., 2026; Arora et al., 2026; Babar et al., 2024; Cavus et al., 2025; Villanueva-Miranda et al., 2025). The analytical engine conceptually integrates machine learning, deep learning, large language models, transformer architectures, sentiment analysis, knowledge graphs, topic modeling, graph-based analytics, and contextual reasoning. The purpose of bringing these approaches is to reveal the multidimensional nature of health misinformation identified in the reviewed literature, where semantic, linguistic, emotional, behavioral, temporal, source-related, and contextual characteristics may all provide relevant analytical information. Within this component, sentiment analysis is explicitly positioned as an auxiliary analytical signal rather than a standalone misinformation detection mechanism. Sentiment-derived information can capture emotional polarity, attitudes, opinions, and affective characteristics of health-related content and may complement semantic and contextual representations (Alonso et al., 2021; Shahzad et al., 2022). However, sentiment polarity alone cannot establish whether a health claim is factual or misleading because misinformation may be expressed using positive, negative, or neutral sentiment. Factuality assessment consequently requires additional contextual reasoning, source assessment, evidence retrieval, domain-specific knowledge, and/or human verification (Liu et al., 2019; Malek et al., 2026).

Real-time analytics are also incorporated into this component because the reviewed studies identify the velocity and continuous emergence of online health information as important considerations for misinformation monitoring (Babar et al., 2024; Kansara and Adhvaryu, 2024; Taha, 2025). However, the review also identified scalability constraints, computational requirements, processing latency, storage demands, and infrastructure costs (Alarabid et al., 2026; Cavus et al., 2025). Therefore, the framework does not assume that real-time or advanced AI analytics can be implemented without adequate technical and organizational resources. The AI-driven analytical engine consequently represents the computational core of the proposed framework, while its outputs remain subject to contextual interpretation and validation through the subsequent human validation layer.

5.4 Human validation layer

The human validation layer was derived from evidence concerning human-centered verification, explainability, expert oversight, evidence assessment, trust, and limitations of automated reasoning. The reviewed studies indicate that automated systems may generate unsupported inferences and unreliable outputs, particularly in sensitive healthcare contexts where incorrect information classification can have serious implications (Feng et al., 2025; Malek et al., 2026). This component incorporates explainable AI, confidence estimation, evidence retrieval, expert verification, retrieval-augmented evidence generation, and human-in-the-loop decision support. These elements were derived from findings emphasizing the importance of human involvement and domain-specific health knowledge in interpreting automated analytical outputs (Feng et al., 2025; Hussain et al., 2023; Malek et al., 2026).

The human validation layer is therefore conceptually positioned as a mechanism through which analytical results can be reviewed, interpreted, and contextualized before they inform sensitive health-related decisions. Human involvement is particularly important where automated systems encounter ambiguous language, conflicting evidence, culturally diverse expressions, or insufficient external knowledge. This layer consequently addresses the ethical and practical limitations of relying exclusively on automated classification. Rather than assuming that human validation automatically guarantees accuracy, the framework recognizes it as an evidence-informed safeguard that may contribute to more transparent and contextually informed decision-making. The actual contribution of this layer to detection accuracy, trust, or decision quality requires empirical assessment in future research.

5.5 Adaptive governance

The adaptive governance component was derived from evidence concerning ethical and privacy concerns, algorithmic bias, cyber-security, scalability, interoperability, sustainability, continuous monitoring, and institutional coordination. These challenges extend beyond the technical performance of individual analytical models and concern the responsible and sustainable operation of health misinformation detection systems. The framework incorporates ethical governance, privacy protection, bias mitigation, regulatory compliance, performance monitoring, model updating, stakeholder coordination, interoperability, and continuous learning. The inclusion of these elements reflects findings that health misinformation detection operates within rapidly changing technological, social, institutional, and regulatory environments (George and Baskar, 2024; Hussain et al., 2023; Morita et al., 2023; Taha, 2025; Villanueva-Miranda et al., 2025; Wreyford et al., 2026).

Adaptive governance responds directly to the sustainability challenges identified in the review. Operational integration among analytical, storage, classification, visualization, and surveillance systems can be complex, while limited coordination among healthcare professionals, policymakers, technical experts, and other stakeholders can constrain long-term implementation (Hussain et al., 2023; Morita et al., 2023; Sharma et al., 2025). Cross-platform inconsistencies and continuously changing misinformation patterns further indicate the need for systems capable of continued monitoring and adaptation. Governance is conceptualized as a cross-cutting component rather than a final stage of the detection process. Ethical, privacy, security, bias, and compliance considerations apply across data acquisition, data management, analytical processing, evidence assessment, human validation, and decision-making.

5.6 Integrated conceptual logical of the framework

The five components collectively provide an integrated conceptual architecture derived from the evidence synthesized in this review. The multi-source data acquisition layer addresses the diversity and heterogeneity of health information. The intelligent data management system addresses data quality, annotation, feature engineering, multilingual processing, and contextual representation. The AI-driven analytical engine addresses computational detection through advanced analytical and sentiment-derived techniques. The human validation layer addresses explainability, evidence assessment, expert oversight, and contextual verification. Adaptive governance addresses ethical, organizational, infrastructural, sustainability, and continuous adaptation requirements.

The components should not be interpreted as isolated linear stages. They are conceptually interdependent. Data quality and contextual representation influence analytical processing. Analytical outputs may require human interpretation. Human validation can generate feedback for system improvement; and governance considerations apply across the entire ecosystem. This integrated structure reflects the multidimensional nature of health misinformation identified across the reviewed studies. The proposed framework is evidence-informed and conceptual, and not an empirically validated effectiveness model. The systematic review provides the evidential basis for organizing the identified factors and challenges into the five components. However, it does not test whether their integration produces superior misinformation detection outcomes. Consequently, no causal and effectiveness claims are made regarding the framework as a whole. Future empirical research should evaluate the framework through prototype implementation, expert assessment, comparative system testing, and field-based studies across different healthcare and digital-information environments. Such validation could examine detection accuracy, contextual reliability, explainability, timeliness, stakeholder trust, scalability, and sustainability.

Figure 5 illustrates the evidence-informed conceptual framework for the adoption and implementation of big data analytics (BDA) and sentiment analysis (SA) in health misinformation detection.

6 Implications

The study has theoretical, managerial, methodological, and social implications. Theoretically, it contributes an integrated perspective by bringing together technological, data-related, human, and governance considerations in health misinformation detection. Rather than treating sentiment analysis as an independent mechanism for determining misinformation, the study positions sentiment-derived information as a complementary signal within a broader evidence-based analytical process. For management and healthcare decision-makers, the findings emphasize the importance of investing not only in advanced analytical technologies but also in high-quality datasets, scalable infrastructure, human expertise, evidence verification, and responsible AI governance. Methodologically, the study provides an evidence-informed conceptual framework that can guide future empirical investigations of BDA and SA-supported health misinformation detection. Socially, improved integration of contextual analysis, evidence assessment, and human validation may contribute to more responsible identification and communication of potentially misleading health information. However, these implications should be considered conceptual as the proposed framework has not yet been empirically validated.

7 Limitations and future research directions

Following are limitations and future research directions of the study:

• The study was conducted through a systematic literature review (SLR) methodology. Future studies can be conducted through the employment of empirical methods i.e., qualitative, quantitative, and mixed methods designs etc.
• Thirty peer-reviewed articles and conference proceedings were selected to conduct the SLR on the topic. Books, book chapters, dissertations and other sources of knowledge were not consulted. Future investigators can consider the inclusion of these sources for increasing sample size of the studies.
• The required studies were selected through digital databases including Web of Science, Scopus, and PubMed. Other sources of knowledge i.e., educational websites, search engines, and institutional repositories were not explored. Future studies can be proceeded using these platforms for offering a broader outlook to readers.
• In future, a similar study can be conducting by applying bibliometric, content analysis, scoping review, and meta-analysis methods for further contributing methodologically.
• The proposed framework through Figure 5 can be empirically tested by adopting different research studies including research articles and dissertations.
• The reviewed literature frequently combines sentiment analysis with broader machine learning, natural language processing, and misinformation detection approaches, which may create conceptual overlap between sentiment classification and misinformation classification. Although the current SLR distinguishes sentiment-derived signals from factuality assessment, the heterogeneity of the included studies may limit direct comparison of the specific contribution of sentiment analysis to misinformation detection. Future empirical studies should separately evaluate sentiment features and other evidence-based features to determine their incremental contribution to misinformation detection performance.
• The review was restricted to English-language publications, which may have resulted in the exclusion of relevant studies published in other languages. This limitation may be particularly important for health misinformation research because misinformation patterns, sentiment expressions, and detection practices can vary across linguistic and cultural contexts. Future reviews should consider including non-English literature to provide a more comprehensive and internationally representative evidence base.
• The possibility of publication and database bias cannot be completely excluded. The review primarily relied on studies indexed in Web of Science, Scopus, and PubMed, and therefore relevant evidence published in other databases, repositories, and less visible publication outlets may not have been captured. Furthermore, studies reporting positive or significant findings may be more likely to be published than studies reporting limited or non-significant findings. Future investigations should broaden database coverage and, where appropriate, incorporate additional sources of scholarly and gray literature to reduce potential retrieval and publication bias.
• The included studies demonstrated considerable methodological and contextual heterogeneity in terms of research designs, datasets, healthcare contexts, digital platforms, analytical techniques, and applications of machine learning, natural language processing, big data analytics, and sentiment analysis. Such heterogeneity limits direct comparison of findings across studies. In particular, sentiment analysis was frequently integrated with broader analytical approaches. Future empirical studies should separately examine sentiment-derived features alongside semantic, behavioral, temporal, source-related, and evidence-based features to determine their specific and incremental contributions.
• The relatively small evidence base of 30 included studies represents another limitation of the present review. Although the included studies provided sufficient evidence to identify recurring factors, challenges, and conceptual relationships, the relatively limited number of studies may constrain the generalizability and comprehensiveness of the findings. Moreover, the heterogeneity of the available evidence did not facilitate a meaningful quantitative synthesis or meta-analysis. Consequently, the present findings were synthesized narratively and conceptually. Future research should consider quantitative synthesis when a sufficiently large and methodologically comparable body of empirical evidence becomes available.

8 Conclusion

This systematic literature review indicates that effective health misinformation detection requires more than increasingly sophisticated analytical algorithms. The evidence suggests that reliable detection depends on the interaction of high-quality and contextually representative data, advanced analytical techniques, sentiment-derived signals, real-time capabilities, human verification, and responsible governance. In particular, sentiment analysis should be understood as a complementary analytical signal rather than a standalone mechanism for determining whether health information is factual or misleading. The review also demonstrates that implementation remains constrained by annotation limitations, scalability, domain adaptation, ethical concerns, and sustainability challenges. These findings indicate that technological sophistication must be accompanied by appropriate data governance, contextual adaptation, human oversight, and organizational capacity. On this basis, an evidence-informed conceptual framework (Figure 5) was developed comprising multi-source data acquisition, intelligent data management, an AI-driven analytical engine, human validation, and adaptive governance. The principal contribution of the study is therefore an integrated conceptual perspective that connects the technological capabilities and implementation requirements of BDA- and SA-supported health misinformation detection. However, because the framework has not been empirically tested and the evidence base is relatively heterogeneous and limited, its effectiveness cannot yet be established. Future empirical research should validate the framework across diverse healthcare and digital-information environments and examine its accuracy, explainability, scalability, adaptability, and practical utility.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Elio Masciari, University of Naples Federico II, Italy

Reviewed by: Nasseh Tabrizi, East Carolina University, United States

Priya Dongare Jadhav, Symbiosis International (Deemed University), India