Enhancing emerging technology discovery in nanomedicine by integrating innovative sentences using BERT and NLDA

Numerous studies have emphasized the significant impact of the Internet’s rapid expansion on the scientific and technical literature across all domains throughout the year (Nasar et al., 2018; C. Zhang et al., 2022). A report by the international association of scientific, technical, and medical Publishers revealed that the number of published researches has increased by an annual average of 4%–5%. Approximately 2.2 million fresh scientific articles were disseminated in 2016 (Canagarajah, 2022; Khanna et al., 2022). Consequently, the challenge of extracting essential insights from an extensive repository of literary data while disregarding background information has become an important concern. Boegel et al. (2022) characterized this endeavor as Information Extraction (IE). Moreover, numerous studies have highlighted the increasing application of IE techniques in various fields, such as medicine, genetics, and other biological sciences (Seeger et al., 2022).

Among the various methodologies, Arshamian et al. (2013) adopted a rule-based approach to extract core ideas from research papers. In contrast, Gupta and B. Hu et al. (2014) applied sentence matching, dependency trees, and seed rules to identify focal points, techniques, and domains. Machine learning methods have gained prominence in recent years. Schneider et al. (2018) used preposition disambiguation with Naiver-Bayesian classification for entity extraction. I. C. Kim et al. (2014) proposed a support vector machine-based framework that utilized a linear kernel function during training to extract sentence-level key insights. Several scholars have adopted a structured learning approach with residuals to process a corpus of 400 articles from ACM and ACL using various training algorithms, including Perceptron, AdaGrad, and SVM (Bianchi et al., 2016; Peng et al., 2014).

Experts in this field have adopted diverse methodologies for IE. However, existing studies have primarily focused on abstracts, with a limited exploration of full-text articles. Therefore, a significant contribution of this study is the utilization of an advanced deep-learning approach. Herein, the BERT model was used for comprehensive IE from all articles instead of the conventional practice of abstract-based extraction, marking an innovation (Song et al., 2023).

Topic models represent a category of unsupervised machine learning techniques devised to elucidate the fundamental topics inherent in a collection of documents (Gruen & Hornik, 2011; Jelodar et al., 2019; Zhao et al., 2011). They demonstrate exceptional proficiency in condensing a vast corpus of documents into a concise summary represented as a topic or cluster of related terms.

The inception of topic modeling can be traced back to the early 1990s when Letsche & Berry (1997) introduced a Latent Semantic Indexing (LSI) model, demonstrating how latent semantic analysis can be used in the automated indexing and retrieval of documents from vast databases. Subsequently, in 1999, Thomas Hofmann made a significant contribution to the field of topic modeling by introducing Probabilistic Latent Semantic Indexing (pLSI). Latent Dirichlet Allocation (LDA), proposed by Blei et al. in 2001, has emerged as a seminal contribution. LDA draws inspiration from pLSI (Blei et al., 2003a) and develops a unique generative model based on the Dirichlet distribution (Nigam et al., 2000). This model features more comprehensive assumptions regarding text generation than its counterparts, rendering LDA the most widely used probabilistic topic model.

LDA has undergone numerous adaptations and applications in various facets of natural language processing. Jelodar et al. (2019) thoroughly investigated LDA and its variations in their survey of LDA-based topic models. For example, Lafferty and Blei developed a Correlated Topic Model (CTM) to address the challenges of interrelated topics (Blei & Lafferty, 2007). Recently, topic modeling has been improved by integrating Word2Vec and other word-embedding models with innovative natural language processing techniques (Zhang et al., 2018). Hu et al. (2018) presented the Biterm Topic Model (BTM), designed for short texts like tweets and social media posts, demonstrating exceptional performance in short text modeling. Fu et al. (2018) introduced the Embedding Topic Model (ETM), where words and topics are represented as vectors within the embedding space. Thompson (2020) later suggested using the BERT language model to generate topics by applying k-means clustering to tokens derived from BER-extracted context vectors. Furthermore, Churchill & Singh (2021) proposed a Topic-Noise Model that integrated the modeling of topic and noise distributions, resulting in superior noise filtration. They combined a Topic-Noise Discriminator (TND) and LDA to formulate an integrated Topic-Noise Model, known as NLDA, which demonstrated improved consistency compared to LDA. This is supported by quantitative and qualitative assessments of topic interpretability, facilitated by illustrative topics and human assessments (Churchill & Singh, 2023, 2022b).

The NLDA model was used in this study to process the text data IE. This approach effectively filters out extraneous text information, producing more interpretable topics for human understanding. Subsequently, the supporting documents for each topic were identified based on the document-probability matrix generated by this model.

Academic approaches to identifying emerging technologies involve qualitative methods based on specialized knowledge and quantitative text analysis (Bishop, 2006). Technical approaches include inventive problem-solving theory, scenario management, bibliometric analysis, and data mining (Arora et al., 2013; Breitzman & Thomas, 2015; Holmes & Ferrill, 2005; Keenan, 2003a; Li et al., 2019; Noh et al., 2016a). Various studies have focused on developing metrics for emerging technologies from multiple perspectives. Rotolo et al. (2015) refined five attributes to characterize emerging technologies: radical novelty, rapid growth, coherence, prominent impact, and the presence of uncertainty and ambiguity. Scholars have identified emerging technology and emergent topics by analyzing citations and co-citations in scientific data, highlighting the novelty and growth of emerging technologies (Sidaway, 2020; Vatanasakdakul et al., 2023). Other scholars have utilized publication and author counts to identify burgeoning areas in science (Long et al., 1980; Uddin & Khan, 2016; D. Zhao & Strotmann, 2011). Chakraborty & Pradeep (2017) introduced a model for detecting emerging clusters founded on four identifying characteristics: the public sector ratio, scientific index, originality index, and reference index. Other attributes developed by scholars encompass data validity, data availability, implementation cost, ease of use, and methodological adaptability (Dash et al., 2007; Gokhberg et al., 2013; Guderian, 2019; Lee et al., 2018).

Furthermore, frameworks rooted in time series evolution analysis are employed in identifying emerging technologies, aiming to comprehend the dynamic nature of technological progress. For instance, Xu et al. (2021) developed a framework employing thematic models to predict the emergence of new technologies through the analysis of sequential text data. Shen et al. (2020) proposed a deep learning model, SeriesNet, which predicts future technology trends by learning from time series data at multiple ranges and levels, utilizing sophisticated analytical techniques. Mercer and Keogh (2022) developed another predictive model for monitoring outlier behavior in time series, facilitating the discovery of new technologies in industrial and natural sciences. Additionally, Ebadi et al. (2022) combined deep learning with weak signal analysis in a multi-layer quantitative approach to predicting technological trends, including advances in the field of hypersonic motion.

In summary, methods for identifying emerging technologies exhibit significant diversity. In this context, our research has developed discovery metrics for emerging technologies in nanomedicine. These metrics stem from information extracted from technology topic support documents through topic modeling, ultimately facilitating the identification of emerging technologies.

We obtained data from PubMed Central (PMC) and Web of Science (WOS) databases. PMC is an open-access archive containing full-text biomedical and life science journal articles from the National Library of Medicine (NLM) of the National Institutes of Health (NIH) (Roberts, 2001). Currently, it comprises an extensive collection of thousands of journal articles, totaling over 8 million full-text articles, spanning centuries of research in biomedical and life sciences from the late 1700s to the present (Hsiao & Torvik, 2023; Lin et al., 2023). This content comprised articles published in scholarly journals, peer-reviewed author manuscripts, and preprint versions of articles publicly available before peer review (Michaleff et al., 2011). Therefore, we used this database for our comprehensive literature dataset.

The WOS is a valuable resource for citation data in bibliometric analyses (Falagas et al., 2008). Furthermore, WOS offers bibliometric software to produce comprehensive statistical insights (Harzing & Alakangas, 2016). Therefore, this database is frequently used by researchers as a primary resource for extensive bibliometric analysis.

Drawing on the research methods of past scholars, we established the time span for the data collection process as 2013-2023, as this period witnessed a 7.39-fold increase in global annual publications, from 1,851 to 13,683, offering a rich resource for research (Tao et al., 2023). The data collection process consisted of two distinct phases. In the initial phase, a search was conducted in PMC using the keyword “nanomedicine”. Web crawlers were developed using Python Selenium and the Beautiful Soup library to retrieve and analyze HTML content. This process facilitated the extraction of relevant information from the literature, including titles, full texts, and DOI numbers. Notably, we extracted content from full-text articles after the abstract and before the references. The extracted content formed the basis of our text database. Subsequently, we used a similar approach in WOS using the Science Citation Index Expanded (SCI-Expanded) journal categories. The subsequent dataset comprises titles, authors, keywords, publication years, and DOI numbers. In addition, the dataset incorporates supplementary metrics, including the number of citations, citation density (number of citations per year), affiliations, countries of affiliation, journals of publication, and associated impact factors (IF) based on the Journal Citation Reports. This information is brought together in our indicator database.

The preprocessing work performed on the text database and indicator data obtained in the previous step was as follows: First, the two databases were matched based on the DOI numbers of the articles. After matching, 2,083 full-text articles and their corresponding indicators were obtained. The articles were subsequently subjected to sentence processing using the NLTK package. An automated toolkit for sentence splitting has been used over traditional punctuation sentences to more accurately split sentences based on context and obtain precise results (Schmitt et al., 2019; Yao, 2019). Statements containing only formulas and special symbols were removed, and the remaining sentences (totaling 673,642) were stored in a database and traced back to their source literature. We used the criteria described in Table 1 to distinguish between academically innovative and non-innovative sentences.

A sentence database consisting of 2,276 innovative utterances and 7,724 non-innovative utterances was obtained after labeling. This dataset was used to train the upcoming deep-learning models.

It has been empirically demonstrated that fine-tuned BERT models produce superior results across various natural language processing tasks, such as text categorization and automated question-answering (Devlin et al., 2019). The advantages of using BERT include accelerated development, reduced data prerequisites, and improved outcomes (Bello et al., 2023; Reimers & Gurevych, 2019). Therefore, we selected the fine-tuned BERT model to categorize sentences as innovative or non-innovative in the context of IE from an innovative perspective.

To optimize outcomes while saving valuable training time, it is imperative to make an informed decision regarding the pre-trained model. SciBERT was used as the pre-training model in our study, initially trained on a random sample of 1.14 million papers from Semantic Scholar (Lobanova et al., 2023). This corpus comprised 18% papers from computer science and 82% from various biomedical fields. Using the pre-trained parameters of this model as initial weights not only conserve time and resources but also improves performance by transferring its generalized linguistic representations, obtained from extensive text data, to subsequent classification tasks (Shen et al., 2023; Thierry et al., 2023). Consequently, it achieves state-of-the-art performance across a wide spectrum of scientific domain natural language processing (NLP) tasks. Further evaluation details can be found in the associated paper, with the evaluation code and data available in the associated repository.

Our approach to fine-tuning BERT based on the pre-trained model unfolds as follows: Initially, the model was implemented using the Keras package and integrated with the SciBERT pre-training model. Subsequently, we thoroughly preprocessed the input data, involving the disambiguation of text and its conversion into three essential embedding vector representations: word, position, and type embeddings. These embedding vectors are fundamental components that facilitate the BERT model in comprehending the text.

We fine-tuned several hyperparameters during the training phase to ensure training effectiveness. Table 2 indicates the final selection of the hyperparameter values. The learning rate of the optimizer was set to 2e-5, a standard value for fine-tuning the pre-trained models. In order to accommodate longer text sequences, we have broken the limitations of standard BERT. Specifically, the standard BERT model can accommodate a maximum sequence length of 512 tokens. When encountering text that is longer than 512 tokens but still less than 1,024 tokens, our current approach is to truncate the text and retain only the first 512 tokens. This approach is simple and efficient, but may inadvertently result in the exclusion of potentially critical information in the second half of the text. We have therefore adopted a different strategy for excessively long texts with more than 1,024 tokens. In this case, the text is divided into multiple segments, each containing no more than 512 tokens. Each segment is then processed independently. For text shorter than 512 bytes, we use padding techniques to extend the text length to the necessary 512 bytes. The truncation, chunking, and padding processes we employ ensure that the length of the input text is consistent and that the model fully learns the semantic information. The batch size was set as 32 to expedite the training process. Eight iterations were performed to ensure the complete convergence of the model. A fully connected layer was constructed in the final layer of the BERT model for sentence classification, which selected the output category with the highest probability. We used softmax as the activation function instead of the sigmoid function because of scalability concerns. The fully connected layer accepts the multidimensional vectors of the [CLS] tokens obtained after training as input. Unlike pre-training [CLS] tokens, which contain no information, post-training tokens encapsulate information about the entire current utterance (Jeong et al., 2020). To assess the performance and generalization capability of the model, we implemented a ten-fold cross-validation technique. Figure 2 illustrates that the fundamental BERT model consists of 12 transformer layers and 12 attention heads.

Figure 1.

A framework of data acquisition and analysis.

Figure 2.

A framework of the BERT model.

The model with the highest F1 score was selected for predictive purposes during the prediction phase. Following IE, we summarized all statements predicted as innovative sentences through article classification, resulting in refined literature content.

The Noiseless Latent Dirichlet Allocation (NLDA) represents a substantial advancement in topic modeling, particularly in addressing the challenges posed by brevity and noise prevalent in short text data (Churchill & Singh, 2021). NLDA markedly enhances subject modeling by the integration of the traditional latent Dirichlet allocation (LDA) framework with an innovative noise filtering technology, namely the Topic Noise Discriminator (TND) model. Its principal advantage lies in its capacity to simultaneously identify and separate meaningful subject content and noise from the data, thereby generating a more coherent and differentiated subject distribution. The key to the NLDA lies in its capacity not only to identify the subject content but also to construct a detailed noise distribution. Through the careful design of noise distribution construction, random and context-specific noise are systematically filtered out, significantly enhancing the purity and interpretability of the topics (Churchill & Singh, 2023). In particular, NLDA works with any generative topic model through a pre-trained TND noise distribution to enable two-tier analysis, a technique that refines the topic output while retaining the high-quality topic structure inherent in LDA, thereby generating topics that are both diverse and deeply reflective of the underlying data (Churchill & Singh, 2022b).

Detailed architectural depictions of the LDA and TND models are presented in Figure 3. In our notation, D denotes a dataset comprising M documents or posts represented by D ={d₀,d₁,‥,d_M-1}. Each document d is a compilation of N words, expressed as d ={w₀,w₁,…,w_N-1}. A topic, denoted by t, encompasses a set of λ words, expressed as t ={w₀,w₁,…,w_λ-1}, with these words characterized by their coherence and interpretability. The collection of topics is encapsulated in a topic set, T ={t₀,t₁,…,t_k-1}, whereas a complementary noise set, H ={w₀,w₁,…,w_p-1}, is used to detect the presence of the noisy elements.

Figure 3.

LDA (a) and TND (b) graphical models.

In contrast to the comprehensive generative process for LDA elucidated by previous scholars (Blei et al., 2003a), we present a concise overview of high-level generative processes using our notation.

For each document d in D:

Importantly, TND not only generates topics but also provides a valuable noise distribution that can be readily adapted to work with other topic models (Churchill & Singh, 2023). Using a pre-trained topic model that approximates topics through topic word distributions, we can use the pre-trained noise distribution from TND in the integration process to accomplish probabilistic removal of noisy words similar to the process in TND. The generative process for the TND can be summarized as follows:

For each document d in D:

Within NLDA, we combine the noise distribution produced by TND with LDA to develop an LDA variant that exhibits a reduced incidence of noisy words in its topics.

To implement NLDA, the noise distribution H on the dataset D was trained using TND, whereas the LDA model was simultaneously trained on a distinct dataset D², which was not necessarily congruent D. Subsequently, we fused the noise distribution from the TND with the topic word distribution from the LDA to produce a topic set. Similar to the process of distinguishing between a topic word or a noise word, for each topic t within T, we exclude words w_i based on the beta distribution of the frequencies of both the noise and the LDA topic distributions to make noise distributions more transferable to various parameters of LDA. We incorporated a topic weight parameter σ into the beta distribution calculation to downsample or oversample noise distribution. Equation 1 demonstrates how σ is used to scale the noise distribution based on k, the number of topics in the LDA model.1Betaθti+β1,Hi(σ/k){\mathop{\rm Beta}\nolimits} \left( {\sqrt {\theta _t^i + {\beta _1}} ,\sqrt {{H_i}(\sigma /k)} } \right)

Taken together, the NLDA offers a practical and efficient solution to the long-standing challenge of noise in text-subject modeling. NLDA achieves effective separation of noise and subject matter through a refined model design, which enables researchers to more clearly capture and analyze hidden topics in text data, thereby motivating our choice of this model for subject recognition work.

Following the initial IE step, we thoroughly processed the textual data extracted from the paper during this phase of the experimental workflow. This data served as our primary source for conducting comprehensive topic modeling using the NLDA algorithm. Our data preparation protocol involves two critical steps: word segmentation and comprehensive data cleansing. Word segmentation involves dividing the text into individual words or tokens, facilitating subsequent analysis. Data cleansing procedures were used to remove inconsistencies, irregularities, or extraneous elements from the text to ensure the highest data quality.

The NLDA algorithm was implemented using the “gdtm” package, a sophisticated utility for topic modeling tasks. To ensure the reliability and robustness of our findings, the algorithm was executed 20 times. In each run, we varied the number of topics from 2 to 50, covering a wide spectrum of topic configurations. This comprehensive exploration enabled us to evaluate the model’s performance across various topic scenarios systematically.

After identifying the optimal number of topics, we applied the NLDA model to preprocessed textual data under this configuration. This step produced two important outputs: the topic-word matrix, providing insights into the word distribution within each topic, and the document-topic matrix, unveiling the associations between individual documents and topics. These matrices served as the fundamental component of our topic modeling framework, facilitating the extraction of meaningful and interpretable insights from the textual data extracted from the paper.

The NLDA model can reveal significant latent topics by removing noisy words. This prompts the question of how emerging topics can be extracted from these latent ones. Previous studies have revealed that emerging topics generally exhibit three key characteristics: novelty, growth, and impact (S. Xu et al., 2021). This study uses originality and innovation to represent novelty, employs growth rate to reflect growth, and assesses topic strength and significance to evaluate impact. Given that the NLDA model can provide a document-topic probability matrix, we first evaluated η_j, which represents the average weight of each topic j. A document is classified as a supporting document for a given topic when the weight ηj(d)\eta _j^{(d)} of the topic j in the document d exceeds η_j.

During this phase of our study, we systematically annotated a dataset of 10,000 sentences, distinguishing between innovative and non-innovative sentences. This dataset served as the training corpus for our model. Throughout the training, we consistently optimized various hyperparameters to improve the metrics. Table 2 lists the finalized BERT hyperparameters.

Subsequently, a 10 cross-validation process was conducted, during which various metrics, including average accuracy and return rate, were calculated. In addition, we also implemented a variety of traditional machine learning text classification methods and calculated relevant indicators. The quantitative results of these evaluations are shown in Table 3.

These findings unequivocally validate the efficiency of our model in the critical task of identifying innovative sentences, compared with other traditional machine learning classification models, each index of our model is about 20% higher. We use our optimized model to predict innovative sentences within a comprehensive collection of research papers, extracting 101,619 innovative sentences. Following this milestone, we organized these innovative sentences by publication and skillfully integrated publication metadata. We analyzed the distribution of innovative sentences, focusing on two key aspects: the average number of innovative sentences per article and the percentage of innovative sentences per article. Figure 4 illustrates the analytical insights.

Figure 4.

Innovation sentence distribution trends.

Our careful analysis revealed that the number of innovative sentences appearing in the articles and the percentage of innovative sentences were relatively stable. Generally, there were 70–90 innovative sentences in the articles. The time series analysis revealed that the number of innovative sentences peaked at 91 in 2015 and dropped to 68 in 2017. The proportion of innovative sentences to the total number of innovative sentences fluctuated from 30% to 45%, reaching a peak of 44.00% in 2016 and declining to 34.40% in 2017.

Furthermore, we conducted an in-depth cross-country analysis by compiling statistics on the number and proportion of innovative sentences per article. To ensure the robustness of our findings, we excluded countries with insufficient article representation to minimize the impact of potential outliers. Table 4 summarizes the top 20 countries ranked by the number of innovative sentences per article. The United States has emerged as the leader in this regard, with 91 innovative sentences per article. Meanwhile, Mexico has the distinction of having the most innovative sentences per article at 47%.

The “gdtm” package was used in the present study to build an NLDA model. A series of experiments was performed to determine the optimal hyperparameters for the model. Table 5 presents the results.

To assess the significance and relevance of each topic, we initially identified the first ten topic words associated with each topic. Subsequently, words ranked from positions 11 to 40 were generated. In a concerted effort to ensure the accuracy and validity of our findings, we engaged specialists in the field of nanobiomedicine for their expert assessment. Through this collaborative process, we determined that the optimal number of topics in our study was 20. Tables 6a and b indicate the detailed outcomes, including the topic words and their corresponding weights.

These topics are important for encapsulating the key topics that underpin the domain of nanobiomedicine. We distilled the main topics embodied in each topic through expert discussions and consensus.

The meaning of each topic was summarized based on expert discussions, with topic 1 focusing on research related to nanostructures and polymer chemistry, with particular emphasis on nanostructures, polymer chemistry, and delivery control. The research in Topic 2 focuses on bacteria, graphene, and nanocomposites and their applications in drug delivery and biomaterials. Topic 3 highlights the quality and product information of nanomaterials, including the quality assessment, surface properties, and methods for various nanomaterial applications. Topic 4 examined the use of nanoparticles in transfection, cell migration, and signaling pathways, with particular emphasis on cancer research and therapy. Topic 5 discusses ultrafast centrifugation, colloidal dispersion, the preparation of nanomaterials, and experimental methods related to microbiology.

Topic 6 concentrates on antitumor therapy, drug tolerance, and tumor stem cells, emphasizing different therapeutic approaches to tumors. Topic 7 focuses on nanoparticles, hormone release, and nanogels, emphasizing drug delivery and characterization of nanopreparations. Topic 8 covers photothermal therapy, photosensitizers, and the use of nanomaterials in cancer therapy, with a special emphasis on research related to biocompatibility and properties. Topic 9 focuses on antidiabetic nanodrug carriers, insulin delivery, and applications of bioactive molecules, with an emphasis on biomedical research. Topic 10 deals with animal studies, biodistribution, immunohistostaining, and laboratory manipulation, emphasizing protocols in biomedical experiments. Topic 11 focuses on adsorption, surface characterization, and the complex structures of nanoparticles, with particular applications in chemical and physical research.

Topic 12 is concerned with inflammation, cellular responses, and biomaterial interactions and has specific applications in disease research and therapy. Topic 13 covers nanotechnology, vaccine research, and infectious diseases with specific applications in vaccine development and infections. Topic 14 examines antibodies, liposomes, receptors, and cell permeation mechanisms, with particular applications in drug delivery and molecular recognition. Topic 15 discusses transport, deposition, and nanoparticles for materials science applications, focusing on material preparation and physical characterization. Topic 16 describes nanocarriers, hormone release, antitumor drugs for drug delivery, and nanomaterial applications. Topic 17 delves into luminescence, excitation, and spectroscopy, particularly highlighting optical techniques in nanomaterial research and fluorescence applications. Topic 18 deals with endocytosis, cell signaling pathways, and nanoparticles in biology, particularly their role in intracellular processes. Topic 19 discusses methods for dispersing, precipitating, dissolving, and purifying nanoparticles and their applications in chemical synthesis and biomaterials. Topic 20 deals with the permeation properties of nanoparticles in blood circulation, tissue retention, and oncology treatments, emphasizing therapeutics and therapeutic strategies. Collectively, these topics constitute a comprehensive summary of nanomaterial research, covering all aspects of research and application areas.

In addition, Table 7 demonstrates the results of LDA topic modeling using only three metadata elements of papers in the field of nanomedicine: the article title, abstract, and keywords. The dataset comprises 5000 articles, and the number of predefined topics is set to 12 based on perplexity analysis (Baimakhanbetov, 2023; Wang et al., 2019). The topic-keyword distribution in Table 7 appears disorganized, and the specialized vocabulary is scattered, making it difficult to summarize the appropriate semantic information of the topics.

The NLDA model was used in this study as the primary tool for obtaining a document probability matrix. This matrix served as the basis for determining supporting documents for each thematic area under consideration. Subsequently, the NLDA model facilitated the quantification of five distinctive sub-indicators within the context of each thematic area: novelty, innovation, growth, impact, and topic strength. These sub-indicators were subjected to a standardization procedure, resulting in normalized values, which were then used to calculate a comprehensive indicator representing the emergence of each topic.

This composite indicator, termed the “degree of emergence,” was systematically determined using the CRITIC method, encapsulating the multifaceted facets of topic relevance, innovation, and significance. The findings of this analytical process, wherein the degree of emergence of each topic was determined, are presented in Table 7. The tabular representation provides a comprehensive overview of the sub-indicator values and the corresponding total indicators, thereby offering holistic insight into the emergent attributes of each thematic element.

This high-level academic analysis focuses on the emergence and impact of various topics in emerging technology research. Twenty different topics were evaluated based on five sub-metrics (novelty, innovative, growth, impact, and topic strength), which are important for understanding the development and significance of each topic.

In Table 7, we observe distinct trends in various research topics within the field of nanomedicine, as indicated by their performance in different assessment categories. Notably, Topic 8, “ Photothermal Therapy and Nanomaterials in Cancer,” and Topic 12, “ Inflammation and Biomaterial Interactions in Disease Research,” demonstrated the highest levels of “Novelty.” This signifies that they have received significant research attention in recent years due to their innovative contributions.

Furthermore, Topic 5, “ Ultrafast Centrifugation and Nanomaterial Preparation,” received the highest innovation score, highlighting its significant contribution to the field. On the growth trajectory, Topic 4, “ Nanoparticles in Cancer Research and Therapy,” and Topic 8, “A Photothermal Therapy and Nanomaterials in Cancer,” exhibited the highest growth rates. This reflects their rapid expansion and development within the biomedical research landscape.

Our impact metrics revealed intriguing insights. Topic 3, “ Nanomaterial Quality and Surface Properties,” and Topic 12, “ Inflammation and Biomaterial Interactions in Disease Research,” emerged as the highest-impact topics. In addition, Topic 12 obtained the highest topic intensity value, indicating its significant influence on the scientific community. In contrast, Topic 10, “ Animal Studies and Biomedical Protocols,” and Topic 18, “ Endocytosis and Cell Signaling with Nanoparticles,” demonstrated the lowest impact and topic strength, indicating relatively lower research attention in these areas. Disruptive innovations may be necessary for these topics to gain prominence in the field.

The comprehensive index provides an overview of the research landscape. Topic 12, “ Inflammation and Biomaterial Interactions in Disease Research,” and Topic 8, “ Photothermal Therapy and Nanomaterials in Cancer,” claimed the top two positions. This indicates that these research areas are at the forefront of emerging topics. Notably, this surge may be attributed to the profound impact of the COVID-19 pandemic on preventive medicine and infectious disease research, which has resulted in significant advances in vaccine development and infection studies. Concurrently, nanomedicine has made significant advances in the field of cancer treatment in recent years, which is why Topic 8 stands out.

In this study, we identified emerging research topics by selecting those with an emerging degree above the average threshold. Table 9 provides a summary of our final results on emerging theme identification.

To validate our results, we reviewed the bibliometric and review studies of previous scholars on nanomedicine and summarized the emerging technologies identified in the field, as shown in Table 10 (Bragazzi, 2019; Chang et al., 2015; Dundar et al., 2020b; Ledet & Mandal, 2012; Sandhiya et al., 2009; Yeo, 2013). By comparing Table 10 with the emerging technologies we identified, we observed some overlap, demonstrating the credibility of our identification method and results.

In summary, our analysis highlights the dynamic landscape of nanomedicine research, wherein certain specific topics have gained attention owing to their novelty, innovation, growth, and impact. These findings underscore the evolving priorities and emerging areas of significance in the field that have been influenced by recent global events and scientific breakthroughs.

In terms of technology, our approach uses advanced natural language processing models to extract knowledge from a diverse range of textual data more effectively. This approach presents significant advantages in terms of ease and accuracy compared with the traditional machine leaning methods used by scholars for IE. Furthermore, we mitigate the noise word interference encountered in conventional LDA models by using NLDA topic modeling for improved topic recognition, thereby enhancing the interpretability of topic identification methods (Churchill & Singh, 2022a). This shift toward a data-driven approach has improved the efficiency of uncovering critical insights within the vast scientific literature.

Moreover, we adopted a diverse set of evaluation metrics. We have introduced innovativeness and topic strength indicators based on the IE and topic identification outcomes of this study. These multidimensional indicators provide a comprehensive quantitative perspective on the emergence of technology topics. The CRITIC methodology integrates these indicators to provide a holistic measure, thereby furnishing a more nuanced assessment of the emergence level of each technology topic (Diakoulaki et al., 1995). These multidimensional indicators expand the toolkit available for researchers investigating emerging technologies.

These theoretical contributions serve as templates for other researchers, assisting them in systematically discovering and analyzing emerging technologies. In addition to nanomedicine, these contributions have the potential to shape the way emerging trends are studied in various disciplines, thereby contributing to a deeper theoretical comprehension of the evolution of knowledge and technology.

Furthermore, this study offers several practical contributions, particularly in applying these findings (L. Yang et al., 2023). Identifying emerging technologies in nanomedicine provides practical insights for decision-makers, stakeholders, and policymakers. It is important to use technical methods to identify technology topics with a high degree of emergence in practical domains. For example, our study identified “ Inflammation and Biomaterial Interactions in Disease Research “ and “ Photothermal Therapy and Nanomaterials in Cancer “ as topics of notable prominence, implying their importance for further research and investment. Policymakers can use this information to make informed funding decisions, prioritizing areas with a higher potential for scientific progress and societal impact (J. S. Lee et al., 2022). Industry stakeholders can identify untapped market opportunities and focus their research and development efforts on areas with high growth potential.

Moreover, our cross-country analysis offers a global perspective on the distribution of innovative research in nanomedicine. In addition, the analysis of time reveals trends in innovative research in the field over the past decade. This practice promotes international collaboration and facilitates a more efficient allocation of resources.

Although our study provides valuable insights, it is crucial to acknowledge its limitations and identify potential avenues for future research. One notable limitation is our reliance on the existing scientific literature, which may introduce publication biases and language constraints. Furthermore, despite the meticulous manual annotation of the dataset, it remains susceptible to subjectivity and can be time-consuming. Future studies should consider exploring larger datasets, incorporating additional data sources, and automating the annotation process to address and mitigate these limitations effectively.

It is important to note that this research is confined to the domain of nanomedicine, and the identified topics are specific to this field. This methodology can be adapted and extended to investigate emerging technologies in other scientific domains. Furthermore, assessment criteria for emerging topics can be tailored to specific contexts and objectives, facilitating a more precise analysis.

In summary, this study presents a comprehensive framework for identifying emerging technologies, providing theoretical and practical contributions with far-reaching implications. These limitations serve as a roadmap for future research, extending the potential application of this methodology to various scientific disciplines and decision-making contexts.