Am I Hurt?: Evaluating Psychological Pain Detection in Hindi Text Using Transformer-based Models

The International Association for the Study of Pain ( IASP) described pain as “an unpleasant sensory and emotional experience associated with, or resembling that associated with, actual or potential tissue damage” [1]. It is challenging to manage pain since it is a personal, subjective sensation linked to several conditions and presentations [2]. It is often referred to as an unpleasant emotion or sensation that acts as a body's alarm system, warning of impending damage or harm. Even with the extensive pain ratings and scales accessible in pain clinics, Electronic Health Records (EHRs) are a rich database with a large amount of granular pain information [3]. Acute or chronic pain can vary in severity, duration, and quality. A specific injury or disease can cause acute pain, which is often abrupt and brief [4]. On the other hand, chronic pain lasts for a long time—typically 3 months or longer. It might last longer than it usually would or appear for no apparent reason [5].

The perception and feeling of pain can be significantly influenced by psychological factors. Both sensory and cognitive processes are involved in the complicated phenomena of pain. Psychological variables can affect how people perceive pain and how they interpret and react to it. Pain has evident emotional and behavioral effects that have an impact on how chronic issues develop and how therapy works [6]. It is well recognized that psychological processes can trigger the onset of chronic pain issues [7]. Avoiding regular activities, resulting in deterioration of functionality, worsens pain and heightens fear of (re)injury. Anger, distress, fear, anxiety, remorse, disappointment, and depression are common emotional responses to pain [8]. Some of the psychological factors, including empathy, pain anticipation, affection, emotions, attachment, and learning, can activate the neuropathological processes [9]. The patient's ability to control these emotions has an influence on how they affect pain. Various studies have demonstrated a bidirectional relationship between depression and pain: both chronic pain and depression are positively associated with each other's growth. Hope, pain acceptance, and optimism are all positive psychological qualities that influence how one adjusts to chronic pain [10]. The semantic meaning of words is essential to how pain is perceived because pain perception in the brain is a sophisticated system of modulation and processing. One stimulus can influence several others according to the priming effect, a cognitive phenomenon. It is therefore possible that this effect is essential for controlling and sensing pain [11]. People commonly resort to social media as an outlet to express and vent their emotional pain [12]. Natural Language Processing (NLP) is a computer science branch that examines unstructured free text and uses statistical algorithms and computational techniques to extract quantitative data from it.

NLP and Deep Learning (DL) have gained popularity, and there is a growing interest in investigating these technologies to aid in pain diagnosis utilizing textual data. An emerging field of research, pain detection using NLP, uses text data to identify and rate pain events. While traditional techniques of pain evaluation rely on subjective self-reporting by individuals, which are subjective and prone to biases, NLP presents a unique opportunity to extract pain-related information from numerous textual sources, including social media postings, patient records, online forums, and healthcare surveys. Medical keyword searches, diagnosis categorization, and information extraction from clinical notes about disease type and symptoms have all been done using NLP approaches [13]. NLP algorithms may find phrases associated with pain, such as “headache” or “frontal,” in clinical notes or other text sources. NLP models can recognize and extract phrases and concepts associated with pain from unstructured text data by using methods like named entity recognition and keyword extraction. Automated pain detection utilizing NLP methods can offer insightful information and help psychoanalysts and NLP specialists spot probable psychological issues in patients. Models for emotion detection may be created using machine learning methods like logistic regression and DL. The emotions detected using NLP approaches can offer useful information for identifying depression, aiding in pain assessment, and enabling personalized medical care.

Social media platforms have grown into effective venues for people to express their feelings and share personal experiences with an extensive audience. The social media content often reflects the emotional state of users, including instances of pain and distress [14]. To discover linguistic patterns, the social media posts are assessed based on the frequency of specific words or phrases, the conversational tone, and the overall sentiment. When training machine learning models to predict instances of pain or identify depressive episodes, these patterns might be employed as useful features. This method is especially useful in the subject of psycholinguistics, which focuses on understanding how linguistic and mental variables interact. Psycholinguistics researchers have established that people's personalities may be inferred from their written language and writing styles [15], which can disclose information about their emotional health. This data can be used to spot depressive episodes or pinpoint those who are vulnerable to emotional pain. For instance, a person may be categorized as having a more depressed personality trait if they regularly utilize phrases like grief, sadness, hurt, or tears in their postings [16]. On the other hand, those with a more upbeat and outgoing attitude may use more positive language and expressions. Using labeled data with annotations indicating the presence or absence of pain or depression episodes, machine learning models may be developed. These models are capable of recognizing language patterns and relationships between certain words, phrases, or emotional expressions and the existence of pain or depressed symptoms. The algorithms may then be used to predict or detect the episodes of pain. There are various benefits of using social media data for pain detection. First off, it offers a non-intrusive and economical way to regularly check on people's health, especially in situations where direct observation or in-person assessments might not be practical. Additionally, it enables the analysis of large-scale datasets, which makes it easier to discover population-level trends and patterns. Finally, using machine learning models built on this data can assist healthcare practitioners in making informed decisions and enabling early detection of pain, intervention, and customized treatment planning. Most Indian languages are classified as low resource, implying that there is a scarcity of data for training NLP systems. The lack of data makes creating accurate and effective NLP systems for these languages difficult. Indian languages frequently have fewer resources, such as text corpora, labeled datasets, and linguistic tools, compared to the languages with vast resources, such as English. As a result, developing NLP systems for Indian languages requires innovative methods to overcome data scarcity and address the unique linguistic characteristics of these languages. Hindi is a widely spoken language in India and the other neighboring countries that offers a unique context for pain detection research. Pain detection in local or regional languages like Hindi has received very little attention, even though multiple studies have concentrated on it using English textual data. However, the accessibility of large-scale textual data in Hindi, such as social media posts, patient records, healthcare forums, and online surveys, presents chances to investigate NLP and DL strategies particularly adapted to the Hindi language.

The convergence of computational linguistics and emotional health has opened new avenues for understanding human sentiment, particularly through the lens of poetry. Our research introduces “दर्द-ए-शायरी: The Language of Pain,” a novel dataset of Hindi poetry that serves as a rich repository for detecting expressions of pain in textual form. Harnessing the capabilities of deep neural networks with static embeddings, we have pioneered a method that not only identifies but also interprets the nuances of pain with remarkable accuracy and efficiency. This breakthrough has profound implications across multiple domains, ranging from mental health surveillance to enriching cultural studies. Here are some of the real-time applications that exemplify the utility of our research:

These examples illustrate the potential of your research to extend beyond academic settings and provide practical, real-world benefits in various domains. Thus, the significant contributions to this research can be summarized as follows:

The rest of the article is structured as follows: Section 2 discusses the recent advancements in pain detection, with a focus on the textual data. In Section 3, fundamental concepts related to the static and dynamic word embeddings and commonly used DL models in NLP are explored. In Section 4, we delve into the specifics of our curated Hindi pain detection dataset and our proposed model, HindiPainNet, designed for the purpose of pain detection in Hindi textual data. Section 5 summarizes and analyzes our model's experimental outcomes. Section 6 concludes the article by summarizing findings and proposing potential avenues for future research and improvements in this area.

The notion that information is important in the emergence of health and disease aligns with the primary principles of the Biopsychosocial (BPS) model as mentioned in Figure 1. This paradigm proposes that an individual's well-being or sickness is impacted by psychological and social factors in addition to physical changes.

As a result, the BPS model recognizes the interdependence of biological, psychological, and social factors in understanding and treating pain [17]. Language has lately been demonstrated to be useful in interpreting and quantifying a range of pain experience aspects other than qualia or severity by researchers. The psycholinguistic and affective characteristics of pain-related words were examined in order to set appropriate guidelines for future research on painful emotions [18]. Detecting rumors in Arabic tweets poses challenges due to linguistic nuances. Gumaei et al. [19] proposed the XGBoost-based approach, leveraging diverse features, achieving a top accuracy of 97.18% on a public dataset, surpassing recent methods in rumor detection. Using a large-scale dataset comprising 2.5 million surveys and 1.8 million tweets, a study by Aggarwal et al. [20] investigates the prediction of community-level pain using Twitter posts and reveals significant variations in pain expressions across different communities in the United States, highlighting the potential for Twitter-based interventions in community-focused pain management. Sawhney et al. [21] provide SISMO, a hierarchical attention model that considers the ordinal structure of social media suicide risk assessment. It uses soft probability distribution to accommodate for different risk levels and shows good results on real-world Reddit posts that have been annotated by professionals. Patients with chronic pain were evaluated for their reaction to placebo using quantitative language patterns collected from semi-structured interviews [22], while pain disparities in underprivileged communities were discovered using comprehensive text mining of EHRs [23]. Social media posts have also been studied for a variety of pain-related purposes, including tracking patients over time and identifying new pain phenotypes [24], geographically monitoring and characterizing opioid use [25], exploring how pain is socially [26], and identifying population-level increases in pain conditions and symptoms [27]. A study conducted by Kumar and Albuquerque [28] utilized the XLM-R transformer with zero-shot transfer learning for sentiment analysis in resource-poor Indian languages. A study by Caldo et al. [29] investigates the emotional patterns of effective spine pathology web pages, indicating their potential importance in comprehending chronic pain and affecting health-related behaviors. The findings highlight the necessity of examining the BPS components of pain and present ethical problems for digital health information providers. Mullins et al. [30] examined tweets about pain in Ireland over a 2-week period and found that the most common terms were headache (90%) and migraine (66%). The majority of tweets were from women and identifying the dominant category of advice on back pain management. A longitudinal study by Deng et al. [31] used NLP tools to analyze tweets regarding migraines. User behavior profiles were reported and examined, such as tweeting frequencies, popular words, and sentimental presentations. Many expressive tweets had a negative emotion, particularly those with a high frequency and severe sentiment, including the use of profanity. Guo et al. [32] examined the utilization of social networking platforms (Reddit and Twitter) as a valuable resource for studying migraine, including the availability of relevant discussions and the development of a text classification system. The use of DL neural networks and NLP to text data from medical discharge summaries is investigated in the study proposed by Yang et al. [33], with an emphasis on patient phenotyping. The study emphasizes the importance of data quality, quantity, and token selection in obtaining higher performance, particularly in Chronic Pain classification. To examine headache and migraine discussions in Japan, Germany, and France, researchers analyzed social media data from several platforms, revealing linguistic trends, treatment references, and demographic information [34]. Table 1 demonstrates a thorough examination of the latest advancements in the field.

In our research, we showcase the efficacy of using deep neural networks with static embeddings to identify pain in Hindi text data. To capture the temporal connections between words and discern the nuanced patterns related to pain characteristics in text, we employ the IndicBERT model. IndicBERT, a specialized variant of the BERT model for Indic languages such as Hindi, contributes to pain detection in social media posts by comprehending the contextual nuances and linguistic patterns specific to social media phrases. IndicBERT incorporates word embeddings as an integral part of its model architecture. It recognizes pain indicators and emotional language in social media text by capturing the semantic meaning of words and phrases. The incorporation of these embeddings in pain detection models improves their performance by facilitating a better understanding of word meanings within the context of pain-related text.

Previously, NLP models relied on predetermined rules and characteristics to read text, which is restricted when it comes to understanding natural language. However, due to recent advances such as word embeddings and DL, NLP models can now learn from large amounts of data and automatically detect essential patterns in language. NLP assists in analyzing the emotions portrayed in the text so that we can determine if someone is feeling good or negative about pain. It also helps with finding essential pain-related phrases, such as “pain,” “agony,” and “hurt,” which can help us anticipate and interpret pain from text. Overall, NLP facilitates the study and interpretation of text, providing helpful information for pain prediction and management.

The methods used in the curation of the Dard-e-Shayari dataset and the development of the HindiPainNet model are described in depth in this section. Our approach is tailored to push the frontiers of NLP, particularly focusing on Hindi language and making substantial contributions to the domain of pain detection.

The performance evaluation of the HindiPainNet model encompasses a variety of measures, including accuracy, F1-score, precision, and recall. These measures jointly evaluate the model's performance, with F1-score addressing unbalanced datasets explicitly and reflecting both specificity and sensitivity. To construct a benchmark, a comparison with the most recent model is performed. Table 4 presents a comprehensive overview of the proposed model's performance results, demonstrating its effectiveness.

Since this is pioneer research, it is crucial to acknowledge that subsequent studies have primarily concentrated on English datasets gathered from diverse sources, including Electronic Medical Records (EMRs) and social networking platform data. The focus of these studies is on employing machine learning techniques and NLP to detect and analyze pain. These subsequent studies advance our knowledge of pain patterns, improve pain management tactics, and improve healthcare decision-making by using textual material available in EHRs and social media. While the current research focuses on psychological pain detection using Hindi social media textual data, it provides a valuable foundation for future research in similar domains and languages, offering insights that can be applied and extended to other low-source textual datasets.

The performance of the model yielded impressive results, achieving a peak accuracy of 70.5% and having a recall of 76%, indicating its strong ability to correctly identify positive instances. Additionally, the model exhibited an F1-score of 70.037%, indicating the precision of 68.44%. These outcomes highlight the effectiveness of the HindiPainNet model, equipped with IndicBERT, in capturing the crucial patterns and features required for accurate predictions in the specific task at hand. Figure 3 displays the Confusion Matrix of the formerly discussed model.

The confusion matrix visualizes the performance of a model designed to detect pain from a set of predictions. In this context:

The matrix indicates an accuracy of 70.5%, meaning the model correctly identified pain 70.5% of the time. The misclassification rate is 29.5%, reflecting the proportion of instances where the model's pain detection was incorrect. The matrix is a valuable tool for evaluating the effectiveness of pain detection algorithms, highlighting their diagnostic reliability.

Table 5 provides the comparative analysis of various transformer-based models, including BERT, ELECTRA, DeBERTa, XLNet, and RoBERTa, with the proposed IndicBERT model. The table showcases the results obtained by each model, allowing for a comprehensive assessment of their effectiveness in the given task.

The table presents a comparative analysis of various Transformer-based models in the context of the HindiPainNet model, specifically focusing on their performance in pain detection from Hindi text. The models are evaluated based on standard classification metrics: accuracy, F1-score, precision, and recall.

In summary, IndicBERT outperforms other models in all metrics, suggesting it is the most suited for detecting pain in Hindi text. BERT and DeBERTa offer moderate results, while ELECTRA, XLNet, and RoBERTa seem to struggle with this specific task. The low F1-scores of ELECTRA and XLNet, in particular, highlight the challenges these models face in balancing precision and recall for pain detection in Hindi. Figure 4 displays the comparative analysis of the proposed model with different transformer-based models.

While this research has made significant contributions to NLP in the Hindi language, there are some limitations to our study:

In recent times, significant advances have been witnessed in the field of NLP in various areas and languages. This progress has also been extended to the analysis of the Hindi language, where the availability of comprehensive datasets and the development of efficient models have become vital for enabling sophisticated NLP applications. These datasets and models lay the groundwork for advanced language analysis tasks in Hindi, empowering researchers as well as practitioners to explore and leverage the potential of NLP in this particular language. This study introduces two significant contributions: the establishment of the Dard-e-Shayari dataset and the HindiPainNet model. The Dard-e-Shayari dataset is composed of Hindi social media posts that have been carefully classified into “pain” and “no_pain” categories, providing a useful resource for identifying the occurrences of pain based on psychological states of the social media users. This dataset, which contains 1,000 instances, serves as a benchmark for pain detection from Hindi textual data. The HindiPainNet model, leveraging IndicBERT architecture, effectively captures semantic relationships and contextual information. It performs well in pain detection tasks and exhibits the potential for textual data analysis. These contributions significantly advance NLP in the Hindi language and offer valuable resources for researchers and practitioners in the field.

While this research has made significant contributions to NLP in the Hindi language, there are various avenues for further research and development. Expanding the Dard-e-shayari dataset by gathering additional textual data from diverse sources, such as hospitals, therapists, and so forth, and annotations from different domains, can enhance its representativeness and applicability. Additionally, exploring data collection in other regional languages would further broaden the scope of the research. Furthermore, the HindiPainNet model can be enhanced by making architectural modifications, exploring different DL models, or incorporating advanced techniques such as attention mechanisms or upgrading transformer-based architectures. Integrating multiple modalities, such as audio and visual data, in addition to text, can result in comprehensive pain detection models, allowing for a better understanding of human pain influenced by psychological factors. In terms of pain classification, future research could focus on developing models capable of detecting and distinguishing more complex pain features and dimensions. Lastly, the practical deployment of the HindiPainNet model in real-world applications, such as integrating them into healthcare systems, chatbots, virtual assistants, or social media platforms, will provide insights into the effectiveness, usability, and ethical considerations of using these models in clinical and everyday contexts.