From 2019 to 2022, the global community faced challenges posed by the COVID-19 pandemic. In response, governments worldwide and the World Health Organization (WHO) collaborated extensively to reduce the spread of the virus. An increased demand for trustworthy information sources and accurate health guidance arose during this global health crisis. However, the surge in these informational needs overlapped with the rapid spread of misinformation and false news through social media platforms, leading to widespread public confusion.

The WHO used the term “infodemic” to describe the spread of misinformation during the pandemic [1]. They emphasized the potential threat that such misinformation posed to national epidemic prevention policies. Trust in incorrect or misleading information could result in adverse health behaviors and noncompliance with health policies, worsening the pandemic's challenges.

While the distribution of COVID-19 vaccines contributed to the gradual control of the pandemic, the virus persisted, giving rise to postinfection symptoms known as long COVID, confirmed in at least 10% of people who contracted the virus [2]. Additionally, instances of reinfection after initial recovery were observed, with research from the US Department of Veterans Affairs indicating increased risks of mortality, hospitalization, and postsymptomatic conditions for reinfected patients [3].

Despite the diminishing immediate threat of COVID-19, the ongoing risks associated with long COVID and reinfection make it important to retain public attention on COVID-19–related policies and information. The challenges of fake news and misinformation persist as the world transitions into a postpandemic era coexisting with the virus. Specifically, issues related to long COVID and reinfection continue to be crucial points for misinformation. Therefore, the timely and accurate identification and classification of such misinformation is critical.

Throughout the COVID-19 pandemic, some studies used machine learning and deep learning techniques to address the challenge of detecting fake news and misinformation.

Patwa et al [4] collected COVID-19–related texts from publicly available fact-checking websites and social media platforms. Their approach involved term frequency–inverse document frequency (TF-IDF) for feature extraction and applying various machine learning algorithms, such as logistic regression, support vector machines, decision trees, and gradient boosting, for the binary classification of fake news. Das et al [5] employed pretrained language models (PLMs), including robustly optimized bidirectional encoder representations from transformers pretraining approach (RoBERTa) and XLNet, for preprocessing and training on the same dataset. By combining predictions from multiple models through voting, they achieved admirable results in the CONSTRAINT 2021 COVID-19 Fake News Detection competition.

Paka et al [6] argued that relying solely on textual features might be insufficient for accurate fake news classification. To address this, they gathered COVID-19–related tweets from Twitter (now known as X), incorporating additional data such as the number of likes for a tweet, URL links, and each poster’s follower count. Introducing a multifeature classification approach for fake news using a cross-stitch unit combined with a long short-term memory architecture enhanced the classification accuracy.

Furthermore, research teams focusing on the Chinese language used deep learning frameworks like recurrent neural network, convolutional neural network, and transformers to classify COVID-19 fake news in Chinese text [7]. These endeavors highlight the global commitment to addressing the infodemic and giving accurate and reliable information.

Additionally, the emergence of large language models (LLMs) based on transformers, such as ChatGPT [8], has become prominent in recent years. These models, capable of understanding natural language and interacting with users, hold the potential to contribute to the development of more robust tools for combating the spread of false news across various domains.

Beyond COVID-19–specific studies, recent research in other domains has demonstrated the effectiveness of machine learning in sentiment analysis and fake news detection. For example, one study proposed a highly efficient technique for polarity classification of X (formerly known as Twitter) data related to COVID-19 fake news [9]. This work applied 5 machine learning classifiers—support vector machine, logistic regression, k-nearest neighbor, decision trees, and random forest—to predict whether news was fake or real, thereby completing the natural language processing (NLP) cycle from data corpus to classification. In another study, researchers examined consumer behavior toward online shopping using machine learning [10]. They utilized a count vectorizer to tokenize text documents and build vocabularies, subsequently applying classification models, including k-nearest neighbor, random forest, and support vector machine (SVM), to analyze the sentiment scores of the product reviews. A third study focused on opinion mining regarding politics and inflation using a Roman Urdu dataset sourced from Kaggle [11]. The researchers experimented with various text processing techniques and classification algorithms (naive Bayes, Bayes Net, KStar, decision tree, and random forest), demonstrating how attribute selection and preprocessing can improve accuracy. These diverse studies underscore the versatility and potential of machine learning approaches in various contexts, reinforcing and complementing efforts to address the challenges posed by COVID-19 misinformation.

While numerous studies have explored text classification for fake news detection, many have primarily focused on traditional machine learning techniques or earlier deep learning models. However, few studies have compared these approaches with state-of-the-art LLMs that offer advanced natural language understanding capabilities. Moreover, although ensemble methods have been applied in some cases, there remains a lack of investigation into sophisticated ensemble strategies that leverage model confidence rankings and nonlinear fusion techniques to further improve classification performance. This study aims to bridge these gaps by providing a comprehensive comparison between conventional approaches and LLMs and proposing a novel ensemble framework that integrates fuzzy rank-based fusion with the Gompertz function.

Considering advancements in deep learning technologies and NLP, this study investigates the performance of various deep learning models in detecting fake news. The objective is to provide a scientific and efficient method for fake news detection in the postpandemic era. Texts about long COVID and reinfection were collected from open-source databases and through web crawling, followed by a preprocessing phase to clean and refine the data. Next, various machine learning and deep learning models were trained and evaluated based on their performance after preprocessing. Finally, a fuzzy rank-based ensemble approach combined multiple models. The performance of this ensemble method was then compared with the state-of-the-art LLM methods.

The proposed method achieved an F₁-score of 96.03%, which can significantly help classify misinformation in real time. The results also demonstrate the effectiveness of language models in distinguishing misinformation.

The main contributions of this study are threefold. First, an in-depth analysis of public datasets related to long COVID was conducted, revealing distinct distribution patterns between genuine and fake articles, which provides valuable insights into the nature and propagation of misinformation. Second, a systematic comparison of traditional text classification methods and state-of-the-art PLMs was carried out, delineating the strengths and limitations of each approach in the context of COVID-19 misinformation detection. Third, a novel ensemble method that combines a fuzzy rank approach with the Gompertz function was developed to enhance the classification performance of language models, resulting in a more robust and accurate detection framework.

The rest of this paper is organized as follows. The next section outlines the methodology, including data collection, preprocessing, analysis, machine learning and deep learning models used, proposed ensemble technique, and implementation details. This is followed by the presentation of experimental results and a comparison of different approaches, as well as several real-case inference examples. Finally, the discussion section highlights the principal findings, acknowledges the study’s limitations, and concludes with a summary of key insights.

The method comprises 4 key stages: data collection, data preprocessing, data analysis, and modeling. Initially, information was collected from diverse publicly available open sources. Given the inherent inconsistency of internet-based open sources, the gathered data underwent a preprocessing phase to enhance their cleanliness. After the preprocessing, a foundational analysis was undertaken to better understand the dataset’s characteristics. Finally, various deep learning models were trained and compared with other text classification methods. We introduced a fuzzy ranking method with the Gompertz function, which adjusts weights based on the confidence scores of each classifier to generate final predictions for each sample. This fusion of ensemble learning and the Gompertz function offers a fresh perspective on prediction methodologies.

Ensuring the absence of duplicate entries in the merged open-source datasets was crucial. To achieve this, we cross-referenced the data with existing open-source datasets. Any identified duplicate samples were eliminated to reduce redundancy, thereby preventing potential impacts on subsequent model training and analysis performance.

During the preprocessing phase, social media posts and articles, known to include emojis and external links (URLs) frequently, underwent processing to enhance their suitability for distinguishing between genuine and fake news. The tweet-preprocessor package was used to eliminate emojis and URLs from the texts because of their low distribution in classification.

All the labels were encoded as “0” (stands for genuine) and “1” (stands for fake), respectively. The distribution of our dataset is illustrated in Figure S1 in Multimedia Appendix 1. As the public data sources exhibited a bias toward the “fake” category, an imbalanced label distribution was observed in the dataset. Therefore, we employed stratified sampling, allocating 10% of the data for testing and 90% for training. This sampling approach ensures consistent label proportions in both the test and training sets, thereby preventing the potential insufficiency of “genuine” samples that may arise from random sampling.

The preprocessing pipeline followed the aforementioned sequence. First, using pandas, duplicate entries were identified and removed, retaining one instance per unique sample. Second, emojis and URLs were removed with tweet-preprocessor to enhance textual clarity and clean the text data. Third, we carried out label encoding. Labels were encoded as “0” for genuine and “1” for fake news. Fourth, we split the dataset. Given the imbalanced nature of our data, which had a higher number of “fake” samples, we employed stratified sampling (using scikit-learn) to ensure consistent label proportions. Specifically, 90% (2634/2927) of the data were allocated for training, and 10% (293/2927) were allocated for testing.

Ensemble learning combines the strengths of individual models to yield predictions that outperform any contributing model. This study proposed an approach to enhance prediction performance by incorporating the Gompertz function into ensemble learning techniques.

We employed a fuzzy rank-based ensemble technique [31], where the confidence of each classifier in its predictions was given priority for each test case. This differed from traditional ensemble methods like the average or weighted average rules, which assign predefined fixed weights to classifiers. Moreover, the reparameterized Gompertz function was used to compute the fuzzy ranks of each pretrained model for detection. Incorporating state-of-the-art PLMs like RoBERTa, DeBERTa, and XLNet further improved our approach. These PLMs bring advanced language understanding abilities to the ensemble, contributing to its robust performance. Afterward, the predictions of the 3 models were fused. Figure 1 outlines the process of the proposed method.

The data were randomly split using the scikit-learn [32] package. This study employed 5-fold cross-validation to train the SVM and determine the best-performing model for the final evaluation. The PLMs were fine-tuned using Hugging Face checkpoints, with the AdamW optimizer [33] and cross-entropy loss. The learning rate was set to 2e-5, and the training was carried out for 20 epochs. HAN was trained for the same number of epochs with its default settings. The training procedure was repeated 5 times with different random initial weights, and checkpoints of models were selected for final evaluation based on the highest validation F₁-score [34]. HAN was trained on the Tesla T4, while all other deep models were trained on the RTX A5000.

No ethical review was required for this study because it did not involve human participants.

This study presents a novel approach for detecting misinformation on the long-term effects of COVID-19 by combining state-of-the-art PLMs with a fuzzy rank-based ensemble method that incorporates the Gompertz function. The experimental results demonstrate that using language models, particularly XLNet, for fake news detection outperforms traditional TF-IDF features combined with SVM or deep models like HAN. Some insights can be observed in Table 4, which shows a trend where models with more parameters tend to achieve better classification accuracy. Models such as BERT, RoBERTa, DeBERTa, and XLNet, which possess significantly more parameters than HAN, demonstrated superior performance across various evaluation metrics. Although XLNet had fewer parameters than some models in the BERT series, it still outperformed them in the experiment. This result suggests that classification effectiveness depends not only on the number of parameters but also on the model architecture, training methods, and optimization techniques. XLNet’s success can be attributed to its permutation language modeling approach, which enhances bidirectional context understanding while reducing some limitations in BERT. The fuzzy rank-based ensemble method further enhanced performance by dynamically weighting individual model predictions based on their confidence levels for each test case. This adaptive fusion resulted in an accuracy of 93.52% and an F₁-score of 96.03%.

Additionally, using LLMs’ embedding models led to superior performance than traditional TF-IDF features. In this study, the GPT embedding model performed slightly better than Gemini, possibly due to differences in the length of the embedding vectors. GPT defaulted to 1536, while Gemini defaulted to 768. Despite yielding acceptable outcomes, directly using GPT-4 fell short compared to training SVM on vector-transformed training data using embedding models. This result implies that LLMs still benefit from training data in fake news detection tasks. Although the LLMs’ embedding models showed remarkable performance in the experiment, accessing this kind of embedding model via an application programming interface incurs charges. However, the fuzzy method can be combined with open-source PLMs to achieve even better results. Moreover, compared to using a single language model or soft voting method with predefined weights, the fuzzy fusion-based technique allowed us to determine ensemble model weights for each test case, resulting in superior performance.

One limitation of the study is the presence of data imbalance, which suggests a potential bias toward the prevalence of fake information on the internet. Addressing this issue would require gathering a more extensive and up-to-date dataset of genuine information to achieve a better balance and representativeness in the training data.

This study provides a comprehensive investigation into the detection of COVID-19–related misinformation by leveraging advanced deep learning techniques. An in-depth analysis of open-source datasets related to long COVID revealed distinct distribution patterns between genuine and fake articles, offering valuable insights into the nature and propagation of misinformation. A systematic comparison between traditional text classification methods and state-of-the-art PLMs highlighted the strengths and limitations of each approach when applied to misinformation detection. Furthermore, the development of a novel ensemble method that integrates a fuzzy rank approach with the Gompertz function significantly enhanced the accuracy and robustness of the text classification.

By developing an ensemble method that integrates fuzzy ranking with the Gompertz function, this research introduces a novel perspective on enhancing classification stability and accuracy. The proposed approach moves beyond simple majority voting or static weight assignments by incorporating dynamic confidence-based fusion, offering a refined framework for complex classification tasks in NLP.

Beyond its methodological contributions, this study holds practical relevance for addressing real-world challenges. The proposed detection system, which performs well using text-only inputs, demonstrates strong potential as a scalable tool for the real-time monitoring of internet misinformation. By helping to distinguish between credible and misleading information, the system may support public health efforts, reduce confusion among the public, and contribute to more transparent digital communication in health-related discourse.