Sporting a virtual future: exploring sports and virtual reality patents using deep learning-based analysis

Data collection and dataset structure

Patent documents related to sports and VR were collected from PATENTSCOPE (https://patentscope.wipo.int). The search keywords included “virtual reality” and “virtual”, combined with sports-related terms such as “sport”, “exercise”, “physical activity”, and “fitness”, along with all sports categories recognized by the International Olympic Committee (http://www.olympics.com). To ensure data reliability, only registered patents that had obtained exclusive intellectual property rights—excluding abandoned, expired, or pending applications—were included in the analysis. As a result, a total of 19,388 patents related to sports and VR technologies were collected, covering a 32-year period from 1994 to 2022.

Comparison and implementation of topic modeling

To identify the most suitable topic modeling technique for patent data, we compared traditional approaches, such as LDA, with a transformer-based model called BERTopic (Grootendorst, 2020). The evaluation was conducted using patent data obtained from the World Intellectual Property Organization (WIPO). Additionally, we examined whether domain-specific, fine-tuned embeddings, such as SportsBERT (Huggingface, 2020), could outperform other topic modeling techniques. This comparative analysis enabled us to determine the most effective approach for extracting meaningful insights from sports patent data. For performance evaluation, we utilized C_V and U_Mass coherence measures. These metrics provided a quantitative assessment of topic coherence and interpretability, reflecting each technique’s ability to capture logically structured and meaningful themes in patent data. By comparing these measures, we aimed to identify the optimal topic modeling technique for patent analysis in the sports domain.

We implemented LDA for topic modeling using the Gensim library (version 4.3.1) to analyze text data. The data were processed using the Natural Language Toolkit (NLTK) (version 3.8.1) by applying tokenization, stop-word removal, and lemmatization using the WordNet lemmatizer. The preprocessed data were then transformed into a bag-of-words representation to create a dictionary and corpus. To determine the optimal number of topics, we iteratively trained LDA models with varying numbers of topics using the LdaModel class from Gensim. The C_V coherence measure was computed using the CoherenceModel class in Gensim to empirically select the most appropriate number of topics. As shown in Fig. 2, the maximum coherence score (C_V = 0.5687) was achieved at four topics. Therefore, we analyzed topic modeling using three topics to determine coherence (Mimno et al., 2011).

Additionally, we evaluated BERTopic, a BERT-based topic modeling technique, for comparison with LDA. Two variations of BERTopic were used:

1. A general-purpose model (BERT-based uncased model).

2. A domain-specific model (SportsBERT).

The text data underwent the same preprocessing steps, such as tokenization and stop-word removal using NLTK. Common generic words, such as “copyright”, “invention”, “figure”, “VR”, and “virtual”, were frequently present but did not contribute significantly to the formation of semantically meaningful topics. Hence, they were added to the stop-word list. In BERTopic, document embeddings were generated using the corresponding BERT models. Topics were extracted through dimensionality reduction via uniform manifold approximation and projection (UMAP) and hierarchical clustering. Apart from the embedding model used to generate document embeddings, all topic modeling techniques were evaluated under the same conditions. The coherence scores of both the general-purpose (BERT-based uncased model) and domain-specific (SportsBERT) models were compared.

The three topic modeling techniques were assessed based on coherence scores. Specifically:

1. C_V quantified the topic similarity.

2. U_Mass measured intra-topic word distances across the three models.

As shown in Table 1, BERTopic with SportsBERT demonstrated higher performance, achieving coherence scores of C_V = 0.6195 and U_Mass = −2.1072, indicating that it effectively captured meaningful and coherent sports-related topics. The topics extracted using SportsBERT provided valuable insights into the latent themes in the selected patent data, enabling a comprehensive analysis. Based on these results, topic modeling with SportsBERT was adopted in this study as a robust, domain-specific approach, enhancing our understanding of the underlying patterns and topics in sports-related text data. An optimal topic model was determined by minimizing intra-topic distance and maximizing inter-topic distance, ensuring well-separated and meaningful topic clusters.

Model robustness and reliability analysis

Given the novel application of a BERT-based model to patent topic analysis, additional experiments were conducted to verify the robustness and reliability of the results. Inspired by previous BERT-based topic modeling research, we performed a series of sensitivity analyses on the SportsBERT topic model to ensure that the findings were not artifacts of parameter choices or random initialization.

1. Topic stability across multiple runs

To assess topic stability, the embedding and clustering process was repeated multiple times using different random seeds. The results showed that core topics remained consistent across runs—the same thematic groupings of patents were observed, and the average topic coherence exhibited only marginal variations. This consistency indicates that SportsBERT’s clustering results are not highly sensitive to random initialization.

2. Performance stability on data subsets

The dataset was split into two subsets based on publication year, and topic modeling was performed independently on each subset. The major topics identified in the full dataset were also present in both subsets, confirming that no single time period disproportionately influenced the results. Additionally, a leave-p-out analysis was conducted by randomly holding out a percentage of patents and rerunning the topic modeling to examine changes in topic structure. The held-out patents were then used as a test set, verifying that they could be correctly assigned to their closest matching topics from the remaining data. This analysis confirmed the generalizability of the identified topics.

3. Parameter sensitivity testing

To further confirm model reproducibility, key parameters were systematically varied, including:

• Number of clusters/topics (k)

• Dimensionality reduction settings

Results indicated that while extremely low or high values of k led to topic merging or over-splitting, a broad range of k values consistently produced the same dominant themes. Similarly, minor modifications in text preprocessing did not substantially alter the identified topics. These findings demonstrate that the SportsBERT-based topic modeling approach is robust—the topics and trends remain stable under various perturbations, confirming their reliability. All analyses were conducted in a computationally reproducible pipeline, and the code and processed dataset have been made available for verification by other researchers. This comprehensive robustness testing, combined with expert validation, ensures that the insights derived from the patent data are both credible and reproducible.

Expert validation of identified topics

To ensure that the topics identified by our models are meaningful and aligned with domain knowledge, we conducted an expert validation study. Five domain experts with strong backgrounds in sports technology and VR were engaged to review and validate the discovered topics. These experts represent both academia and industry: two are university professors specializing in sports technology and computer science with a focus on VR applications, one is a senior patent analyst with over a decade of experience in sports and VR-related patents, one is a VR software developer in the sports training industry, and one is an R&D manager at a sports technology company. All experts hold advanced degrees in relevant fields, including three Ph.D. degrees and two M.Sc. degrees, ensuring their qualifications in interpreting sports and VR innovations.

1. Evaluation criteria

To objectively assess the coherence and relevance of topics produced by LDA and SportsBERT, the experts were provided with a predefined set of evaluation criteria adapted from established patent topic modeling studies within the sports–VR domain. The criteria included the following:

Coherence—Assessing whether patents grouped under a given topic share a clear common theme. For instance, do the title and abstracts of patents in a topic relate to a specific concept such as VR-based athlete training? High coherence implies that the topic’s top keywords and patents are semantically related and form a logical grouping.

Distinctiveness—Determining whether each topic is well differentiated from others. This ensures that topics are not redundant or significantly overlapping. For example, a topic on virtual sports coaching systems should be distinct from one on VR gameplay analytics.

Relevance—Evaluating whether the topic represents a meaningful and significant theme in the sports VR domain, rather than being an artifact of text processing or an insignificant cluster. This criterion checks if a topic is recognizable and useful for domain experts studying sports technology innovations.

2. Expert validation process

Each expert was provided with a summary of each topic and independently assessed it using the above criteria. They assigned qualitative ratings and flagged any unclear or inconsistent topics. Following the independent evaluations, a joint review meeting was conducted, where the experts discussed their findings, resolved discrepancies, and refined topic definitions.

As a result, a finalized set of validated topics was established, ensuring that they were coherent, distinct, and relevant. Additionally, expert feedback was used to assign descriptive labels to each confirmed topic. This expert validation step enhances the objectivity and credibility of the topic modeling results, ensuring that the reported topics provide real-world insights into the technological landscape of sports and VR.

Visualization method for topic area determination in VR sports using a two-step review

Topic modeling provides diverse visual insights, enabling an understanding of similarities and associations between topics. Through visualization tools, we can explore the relationships between topics and assess clustering based on their relevance to sports patent documents. Each topic has distinct characteristics, and patent domains can be identified by analyzing the similarities and associations between topics. Sports encompass a wide range of disciplines, each involving specific equipment, movements, and technological advancements. However, multiple disciplines share similar biomechanical principles and technological applications. Therefore, topic modeling should reveal the application of VR technology across different sports and its unique characteristics. In this study, we employed a two-step review process using various visualization tools to determine representative topic areas. As shown in Fig. 3, the process was structured as follows:

1. Step 1: Topic clustering via distance mapping (blue boxes)

2. A topic distance map was used to identify clusters of topics, forming distinct topic areas within sports patent data.

3. Keywords per topic were then organized into specific topic areas, ensuring that related themes were grouped appropriately.

4. Step 2: Expert review and refinement (green boxes)

5. Domain experts in sports technology examined the original patent documents to verify whether topics were correctly assigned within their respective areas.

6. If an expert review indicated that a topic did not fully align with its assigned area, the topic was reallocated to a closer relevant area in the document map, based on the embedding of the documents forming the topic.

7. The source documents were then analyzed using expert domain knowledge to confirm their placement.

8. If a topic was found to be independent, meaning it did not fit within any established area, it was categorized as a standalone topic.

This systematic visualization and validation process enabled a clear understanding of the convergence between sports and VR technologies, ensuring that topic modeling results were both structured and meaningful.

Regression analysis to explore topical trends in sports patents

Technological development generally follows distinct stages of growth, maturity, and decline over time. Our analysis of available data indicates that VR technologies in sports have undergone various phases of change, growth, and decline since 1994. To examine these trends, we utilized the topic proportion metric, which was calculated by dividing the frequency of a topic in a specific year by its overall frequency derived from topic modeling results. This provided time-series data, enabling a quantitative analysis of topic evolution. We then applied regression models to analyze trends within topic areas, allowing us to:

1. Assess the current state of sports VR technology based on patents classified under each topic.

2. Infer future developments by identifying patterns of growth or decline in topic prevalence.

This approach provides a structured and data-driven perspective on the evolution of VR-related innovations in sports, offering valuable insights for researchers, industry professionals, and policymakers.

Results of topic modeling

To analyze the knowledge structure and trends in VR patents related to sports, we applied SportsBERT-based topic modeling to identify and categorize 23 distinct topics within the dataset. Each topic was grouped into clusters based on its thematic relevance to broader topic areas. Additionally, a two-step visual analysis was performed to ensure coherent topic identification and structuring. The topic modeling results are presented in Table 2, summarizing the 23 identified topics. During topic derivation, the domains of each topic were automatically labeled, reflecting the most significant conceptual linkages between the four most relevant keywords for each topic. To ensure accuracy, the topic identities were further refined through expert evaluation and comprehensive analysis of keywords, source documents, and patent metadata. Furthermore, a time-series regression analysis was conducted to explore temporal trends across the identified topics, providing insights into the evolution of VR applications in sports over time. The following section presents a detailed summary of each extracted topic, highlighting its thematic focus and relevance in the context of sports and VR innovations.

Topic areas based on keyword analysis

The topic areas and distances between topics were primarily derived from the inter-topic distance map, as shown in Fig. 4, which illustrates the spatial relationships among the 23 extracted topics. In this map, each circle represents a topic, and the proximity or overlap of circles indicates thematic similarity. The model considers word distributions, keyword frequencies, and their relative importance to determine inter-topic relationships. A well-structured topic model typically exhibits substantial overlap among circles, reflecting cohesive thematic connections, whereas a poor model results in scattered, isolated topics. In this study, highly related topics were clustered, while distinct topic groups were positioned farther apart in the quadrants.

Topics 0, 1, 12, 17, and 20 overlapped, sharing keywords such as “game”, “camera”, “image”, and “space”. Although fewer topics overlapped, additional common keywords, including “character” and “apparatus”, were also observed. This cluster was tentatively identified as a category related to hardware and software development for VR sports games. However, the topic area did not entirely conform to this interpretation, as some topics exhibited common keywords but belonged to different conceptual domains. Additionally, Topic 0, which encompasses general VR technology, required further evaluation for clearer categorization. By leveraging expert insights in VR sports technologies and applying additional visualization algorithms, we refined the topic clusters to enhance interpretability and accuracy.

Topics 2, 5, 8, 9, 11, and 16 shared keywords such as “system”, “reality”, and “user”. However, this grouping did not fully align with a single theme, highlighting the challenge of assuming that all topics within a cluster are inherently related. Topics 9 and 11 were clearly linked to physical exercise in VR, sharing keywords like “exercise”, “rehabilitation”, “fitness”, and “training”. In contrast, Topic 16, which focused on racing sports, lacked a direct connection to physical exercise, requiring further evaluation. Other topics in this cluster were associated with haptic feedback, sound technologies, and AR, which were identified as distinct themes rather than a cohesive category. Topics 4, 13, 18, and 21 shared the keyword “ball”, with some topics also including “golf” and “surface”. This cluster appropriately grouped sports-related VR patents, particularly those involving swinging movements in sports like golf, baseball, and tennis. The alignment of keywords supported the identification of a well-defined VR sports category. Topics 3, 15, and 22 exhibited partial keyword overlap, including “body” and “reality”. However, this cluster did not reveal a strong thematic identity, as no clear relationships between the topics were identified. Nonetheless, Topics 15 and 22 were closely linked through aerobic exercise applications in VR, such as cycling, running, and walking, while Topic 3 was associated with seating, platforms, and immersive VR integration. Given these distinctions, a clearer classification was required using expertise in VR sports technologies and advanced visualization tools. Topics 6, 7, 10, 14, and 19 shared keywords such as “game” and “player”. These topics were inferred to be related to fundamental technical aspects of VR sports games. While most topics in this cluster aligned well, Topic 19 required further examination using keyword organization, source documents, visualization tools, and expert domain knowledge to ensure accurate classification.

The inter-topic distance map is a clustering approach that visualizes the relationships between topics based on keyword commonality. However, some shared keywords identified by the model may not be directly relevant to a given topic. The map also revealed rare but partially low-relevance topics within clustered inter-topic relations. While topic clusters were generally well-separated across all quadrants, certain clusters appeared ambiguous and required reorganization.

Topic areas based on patent documents

To better understand topic relationships, we used the inter-topic distance map for visualization. However, some topics needed to be reassigned to more appropriate areas. To address this, we employed a more detailed visualization approach, analyzing the distribution of documents within each quadrant to determine their correct topic areas. This approach involved recalculating document embeddings and projecting them into a two-dimensional (2D) space for improved visualization. Additionally, the semantic structure of sentences and phrases within each topic was analyzed to position topics more accurately. Since each document was represented as a node, topics with higher frequencies occupied larger areas in the quadrant. While it was challenging to define precise topic boundaries visually, as seen in the inter-topic distance map, the document-based clustering approach allowed for a more domain-relevant classification of VR sports topics. Figure 5 illustrates the document map, displaying nodes representing all topics. Users could refine the visualization by activating or deactivating topic domains through the right-column controls.

Figure 6 presents the maps for active Topics 2, 3, 9, 11, 15, 16, and 22. While these topics were separated into two clusters in the inter-topic distance map, the document map indicated stronger interrelations, with the topics positioned in close proximity. Notably, Topic 3 (seat_connected_reality_utility) was centrally positioned among the active topics, surrounded by document nodes related to the other topics. This topic comprises patents on physical plates, platforms, and seats designed for VR experiences, ensuring safe and effective integration with head-mounted displays (HMDs). Engaging in cardio or fitness exercises, such as biking or running in a VR environment, requires equipment that supports these movements while interacting with an HMD to reflect user motion. The proximity of Topic 3’s document nodes to those of other topics suggests strong associations between these patented technologies. A similar pattern was observed for Topic 16, where document nodes, though separated by small gaps, clustered closely together. A realistic VR riding experience depends on an appropriate driving environment equipped with specialized hardware, leading to technical overlaps in patent relevance and clustering within the document map. Additionally, Topic 11 (rehabilitation) was positioned adjacent to Topic 9 (exercise training fitness user). While their document nodes did not directly overlap, topic 9 focused on technological advancements in VR-based fitness, which aligns with safety considerations and rehabilitation applications in Topic 11. Expanding the scope, Topic 2 also showed technical relevance to these areas. For users engaged in physical activity within a VR environment, tactile feedback, in addition to visual input, is essential for enhancing realism and immersion. Accordingly, the purpose, necessity, expected impact, and application strategies outlined in the patent documents suggest strong interconnections among VR developments in exercise, rehabilitation, and physical activity.

Figure 7 presents a visual representation of Topics 4, 13, 18, and 21. The document nodes formed a weak cluster, with considerable spacing between topics within a single quadrant. Despite their shared association with sports involving swinging movements, such as golf, baseball, and tennis, the distance between topic nodes suggests substantial differences in the underlying patent technologies. Each topic contained keywords related to movement, equipment, and materials associated with these sports. Topic 4 focused on sensor and measurement technologies for implementing golf simulators, while Topic 18 emphasized physical interactions between users and VR devices. The proximity and partial overlap of nodes between these topics suggest a close technical relationship. Additionally, Topic 14 addressed specialized movement training for specific sports, though it also incorporated varied gaming objectives. Meanwhile, Topic 21 contained patents concerning realistic ball movement simulations in golf VR environments. This topic area, therefore, reflects an interesting interplay between shared elements and technological divergence across VR sports applications. Figure 8 illustrates the document map for Topics 6, 7, 10, and 14. Although Topic 19 was clustered in the inter-topic distance map, it was excluded from this visualization due to its greater distance from the other topics. The remaining topics remained closely clustered, all focusing on online VR sports experiences. The infrastructure of online VR environments includes servers, terminals, and interactive elements such as avatars, online transactions, and multiplayer gaming. As a result, this topic area primarily encompasses patented technologies that apply general VR advancements to sports applications.

Figure 9 presents document maps for Topics 1, 12, 17, and 20. Unlike the inter-topic distance map, where these topics were closely clustered with Topic 0 (operation_scene_control_game), the document map reveals an increased distance between them. The newly organized clusters are centered around visual technologies for VR, incorporating keywords such as “display” and “image”, and hardware technologies, associated with terms like “apparatus” and “camera”. Visual stimulation is one of the most intuitive experiences for users engaging in VR sports. This significance is reflected in the large area occupied by Topic 1 (image_space_display_game). The proximity of Topics 12, 17, and 20 to Topic 1 reinforces the shared thematic identity of this topic area. Figure 10 illustrates document maps for Topics 0, 5, 8, and 19. Topic 0 contains a broad range of patents conceptualizing VR devices and software in general, including applications for sports and VR environments. Since many of its underlying technologies are macroscopic in nature, they can be integrated into patents from other topics. This results in notable clustering in one quadrant, with some individual patent nodes distributed separately. Topic 5 features numerous patent nodes related to sound generation in VR devices and software. While it also includes patents related to display–gesture interaction, the document map indicates that patents focusing on sound signals may represent a distinct technological category. Topic 8 has a high concentration of patent nodes, focusing on the development of AR technologies. Although AR shares technical aspects with VR, it is fundamentally different, as it overlays digital information onto the real world rather than creating a fully virtual environment. This distinction suggests that patent nodes in this topic form an independent category. Topic 19, despite sharing the keyword “VR sports games” with other topics, primarily addresses betting within VR sports environments. Since playing sports in VR and gambling involve entirely different technological principles, the patents in this topic support the independent categorization of this area.

Topic trends

Figure 11 presents the frequency of each topic over time. To assess the number of patent applications per topic, a linear regression analysis was conducted, using the year index as the predictor variable and the proportion of topics in that year as the dependent variable (Jelodar et al., 2019; Paek, Um & Kim, 2021). The study period began in 2010 due to the limited number of patents in earlier years, often zero or negligible. Focusing on this relatively recent period was deemed appropriate for regression analysis to examine trend patterns (Resnik et al., 2015). After estimating the slopes of the linear regression models, the 23 topics were classified into four development categories based on their regression coefficients and statistical significance, as detailed in Table 3:

1. Hot topics: Topics with a steady increase in relative frequency over time (significant positive linear trend, p-value < 0.05).

2. Cold topics: Topics with a steady decline in relative frequency (significant negative linear trend, p-value < 0.05).

3. Warm topics: Topics with an increasing trend, but without statistical significance (positive linearity, non-significant p-value).

4. Cool topics: Topics with a declining trend, but without statistical significance (negative linearity, non-significant p-value).

Key findings:

1. “Sports, exercise, and physical activity” topics were generally hot or warm, except for Topic 15 (bicycle_riding_exercise_wheel), which exhibited a weak declining trend.

2. “Swinging sports” topics (e.g., golf, baseball, tennis) were mostly hot or warm, with the exception of Topic 13 (ball_baseball_tennis_pitching), which showed a slight downward trend.

3. The “online environment” area featured predominantly hot or warm topics, except for Topic 6 (avatar_user_world_environment), which was categorized as cool due to a weak decline.

4. “Visual technology” topics demonstrated hot or warm trends across the board.

5. Among the independent topics, Topics 0, 5, and 8 showed hot or warm trends, whereas Topic 19 exhibited a cool trend, reflecting a weak decline.