The role of large language models in dental diagnosis, treatment planning, and prognosis

Introduction

Large language models (LLMs) are advanced deep learning algorithms proficient in executing diverse natu­ral language processing (NLP) tasks, utilizing trans­former models to train on vast datasets [1, 2]. This training enables them to perform functions, such as reco­gnizing, translating, predicting, and generating text and other content [1, 2]. Moreover, LLMs specialize in tasks beyond language, including analyzing protein structures and coding software [1, 2]. Like the human brain, these models undergo pretraining and fine-tuning to master various challenges, such as text classification, question answering, document summarization, and text generation [1, 2]. They have extensive applications across diverse fields, e.g., healthcare, finance, and entertainment, functioning in roles ranging from translation services to powering chatbots and artificial intelligence (AI) assistants [1, 2]. These models’ large parameter counts, similar to accumulated memories, enhance their problem-solving capabilities in both linguistic and non-linguistic tasks [1, 2].
LLMs are built on deep neural networks that consist of multiple layers, including input, hidden, and output (Figure 1). The input layer initiates data processing by first applying tokenization, which breaks textual data into smaller units, i.e., words or sub-words, and then embedding, which converts these tokens into numerical representations, enabling deeper linguistic analysis [3]. In deep learning, the hidden layers, essential in transformer-based architectures, comprise components, such as feedforward networks and normalization layers, which help stabilize learning and address challenges, e.g., the vanishing gradient problem [3]. These deep neural network layers are stacked to support the model’s language-handling capabilities. The output layer processes the data to make predictions or decisions, usually using the softmax function that helps to determine the most likely next word or token by converting the model’s results into probabilities [3].
LLMs have three primary categories: generic, instruction-tuned, and dialog-tuned models. Generic models excel in predicting the next word based on learned language patterns, and are useful in information retrieval [4]. Instruction-tuned models, trained to respond based on specific instructions, are adept at tasks, including sentiment analysis or code generation [5]. Dialog-tuned models are designed for conversational AI, capable of maintaining engaging dialogues through contextually appropriate responses [6].
The landscape of AI and NLP has evolved with the advent of models, such as OpenAI’s, ChatGPT, and Google’s Gemini, marking significant milestones in human-machine interaction [7, 8]. These LLMs are built on a rich history of AI research, tracing back to early innovations, i.e., the ELIZA chatbot in 1966 [9]. Today, LLMs are not only transforming textual interactions, but are also pioneering applications in fields as diverse as medi­cine and dentistry [10]. In dentistry, NLP shows promise by using methods, such as manual analysis, semantic similarity, key word extraction, and regular expressions to extract relevant dental data from patients’ electronic health records, e.g., identifying fillings or tooth decay [11]. AI technologies address various aspects of dental education and practice, from personalized learning and diagnosis to treatment planning and research [12, 13]. The availability of vast amounts of unstructured clinical text data, including electronic health records, medical literature, and patient narratives in the medical and dental fields, has made them invaluable resources for LLMs to process and extract insights from [13].
This review aimed to explore the challenges and opportunities that LLMs present to dentistry, focusing on the potential to enhance patient care through diagnosis, treatment planning, and prognosis. By examining the current state of LLMs and their applications, we seek to shed light on the evolving role of AI in shaping the future of dentistry.

LLMs across dental specialties

The application of machine learning in dentistry, for instance, using statistical techniques to enhance computer systems through experience, leverage these advancements to improve predictive analytics and ope­rational efficiency in dental practices [14]. The core methodology involves algorithms capable of interpreting complex datasets without explicit programming, thus adapting to new patterns in dental health records to optimize decision-making processes [14]. This not only leads to higher diagnostic precision, but also enhances the administrative efficiency of dental practi­ces, ultimately improving patient outcomes and workflow [15].
Recent studies have showcased the value of ChatGPT in medical education [16, 17]. Notably, ChatGPT de­monstrated its capacity to achieve a passing score equivalent to that of a third-year medical student [17]. Furthermore, the latest version of ChatGPT, GPT4, has exhibited a notable competency by successfully passing the dental licensing examinations of both the United States and the United Kingdom, alongside achieving commendable performance across a diverse range of dental disciplines [18].
The application of LLMs within the five main dental specialties (pediatric dentistry, endodontics, oral and maxillofacial radiology, oral and maxillofacial surgery, and periodontics) exhibited diverse applications, ranging from enhancing patient education to improving dia­gnostic capabilities (Figure 2).
LLMs are used in pediatric dentistry to enhance caries detection and improve radiological image analysis [19]. They also assist in predicting oral diseases, promoting health, and educating patients [19]. In a study conducted by Rokhshad et al. [20], various chatbots, including ChatGPT and Google Bard, among others, were evaluated on their ability to respond to true/false questions related to pediatric dentistry. The findings revealed that these chatbots demonstrated lower accuracy rates than pediatric and general dentists, though their performance exceeded that of dental students. This suggests that while chatbots are making strides, they currently may not be advisable for clinical applications within pediatric dentistry.
Similarly, ChatGPT demonstrated an 85.44% consistency level in addressing clinical questions of varying difficulty, sourced from the position statements of the European Society of Endodontology [21, 22]. These questions capture expert consensus on diverse endodontic topics. While maintaining a notable consistency, ChatGPT’s average accuracy was reported as 57.33%, showcasing discernible fluctuations linked to the complexity of questions [21]. Even though these results highlight ChatGPT’s potential utility in endodontics, the current level of performance does not allow it to supplant the clinical decision-making of dental professionals. As AI continues to learn and improve, it is anticipated to become more effective in the field.
Furthermore, in oral and maxillofacial radiology, LLMs assist in determining radiation doses, suggesting appropriate imaging protocols, and creating preprocedural checklists [20]. This technology also aids in report generation, reducing radiologists’ workload and providing educational resources [20]. However, their ability to analyze radiographs and respond to radiographical image-based questions is limited [23]. Mago et al. [24] found that ChatGPT-3 achieved 100% accuracy in identifying radiographic landmarks, but its ability to describe oral and maxillofacial pathologies was limited to their primary or most distinctive radiographic features. In addition, in oral and maxillofacial surgery, LLMs provide quality patient information, but are less reliable for technical training purposes. Surgeons are advised to use it cautiously and as a supplement to their expertise [21]. Balel [22] assessed the proficiency of ChatGPT in answering questions about impacted teeth, dental implants, temporomandibular joint diseases, and orthognathic surgery, categorizing them into patient and technical queries. While ChatGPT’s responses to patient-oriented diagnostic questions were rated highly, indicating its potential as a supportive tool for patient education and diagnosis clarification, its effectiveness in answering complex technical diagnostic questions was less pronounced [22]. This discrepancy underscores the importance of using ChatGPT judiciously in diagnostic contexts, viewing it as a complementary aid to professional medical judgment rather than a standalone diagnostic tool [22].
Moreover, LLMs, such as ChatGPT-4, have shown promise as adjunctive tools in periodontal disease dia­gnosis. They facilitate the provision of information by identifying pertinent topics and creating dialogues simulating patient inquiries, thereby guiding individuals seeking knowledge about their condition [25]. Alan et al. [26] evaluated the quality of ChatGPT-4’s responses to periodontal disease questions, and found that although the overall DISCERN scores indicated good quality for most topics, the information related to treatment options was often poor. Specifically, although the reliability and overall quality sections were rated as good, the treatment choices section frequently scored below the acceptable threshold, indicating a lower quality of information in this critical area [26]. Additionally, NLP demonstrated promising results in automated dental charting using voice-based reporting [27].

Role of LLMs in treatment planning and patient education

LLMs could have a potential role in patient education in dentistry, significantly improving the way medical and dental knowledge is communicated. By translating complex medical and dental technical language into everyday language, these models help improving patient comprehension and engagement with their treatment plans. This is crucial, as some patients may feel uncomfortable with specialized terminology, and misunderstandings can lead to ill-informed decisions or unnecessary concerns [28]. Moreover, LLMs facilitate patient communication by offering translations and summaries, which is particularly beneficial in enhancing the understanding of conditions and treatments among patients [29]. They also address common concerns about dental health and postprocedural care, providing tailored responses based on individual symptoms [30]. Additionally, these models organize patient data and deliver essential guidance for postoperative care, such as pain management after root canal treatments and instructions for postextraction care, which is invaluable for patients whose primary language is not English [31].
Despite these advantages, it is essential to recognize the limitations of LLMs. Even though they show promise for various tasks, such as record-keeping and decision support, they must be used cautiously due to their potential to generate inaccurate or misleading information. This underscores the importance of verifying the reliability and accuracy of information provided by LLMs in healthcare contexts [31]. Furthermore, with culturally and linguistically diverse populations facing unequal access to dental services, LLMs might play a critical role in bridging gaps in oral health outcomes [32], thus underlining their significance in global health equity.
Considering a patient’s unique genetic makeup, AI can suggest treatments that have the potential to be effective [33]. Furthermore, AI can generate clinical insights, combining genomic determinants with information from patient symptoms, clinical history, and lifestyles, facilitating personalized diagnosis and prognostication [33]. Integrating AI and clinical data sources could be valuable in a dental setting. LLMs can automate document comprehension and make treatment plan analysis feasible, which reaps the benefit of large-scale pretraining [26]. They can assist dental practitioners more efficiently in establishing treatment plans tailored to patients’ backgrounds by exa­mining adverse drug reaction patterns linked to various dental procedures and drugs [34]. NLR algorithms can assist in identifying comorbidities by analyzing patient records for common risk factors and symptoms, identification of adverse drug reactions, drug safety surveillance, and patient education [34].

Role of LLMs in prognosis

Prognosis plays a crucial role in dentistry across diverse specialties, offering clinicians valuable insights into the anticipated course and outcome of a patient’s oral health condition [35, 36]. However, as far as our current knowledge extends, there is a lack of literature studying the integration of LLMs in the domain of dental prognosis.
LLMs, equipped with the capability to process extensive textual information and recognize complicated patterns, are potent tools for enhancing both the experiential and analytical facets of prognosis in dentistry. Their ability to synthesize information from various sources, such as clinical records, research literature, and case studies, can empower clinicians with a wealth of knowledge.

Challenges and ethical considerations

The integration of AI and LLMs into dentistry presents significant challenges and ethical concerns [37]. A primary issue is that LLM outputs, shaped by the quality and characteristics of their training data, are proba­bilistic and not absolute truths. This raises concerns about potential biases, which could perpetuate stereotypes and inaccuracies in patient care and education. Ensuring diverse and representative datasets is crucial to mitigate these biases and promote equitable outcomes. Additionally, maintaining the confidentiality and privacy of patient data is paramount due to the extensive data requirements for training LLMs.
Another ethical consideration is the risk of dependency on automated systems, which could potentially devalue dental practitioners’ professional judgment and skills. Balancing the enhancement of clinical practice with AI while preserving the indispensable human element of patient care is essential. Furthermore, the transparency and explainability of decisions made by LLMs, which often operate as “black boxes,” pose challenges to informed consent and patient autonomy, as the rationale behind diagnoses or treatment recommendations may not be easily understood.
Addressing these challenges requires a multidisciplinary approach to develop guidelines and frameworks, ensuring the responsible and ethical utilization of LLMs in dentistry. As this technology advances, ongoing dialogue and research are essential to harness its potential while safeguarding against its risks, making sure that LLMs enhance, rather than undermine, the quality and integrity of dental care.

Conclusions

LLMs have the potential to improve patient education, dental diagnosis, treatment planning, and prognosis, but their full capabilities have yet to be realized due to various limitations. While their future integration into clinical practice could enhance efficiency and accuracy, significant challenges and ethical concerns must first be addressed. These models also come with inherent biases and ethical considerations that need careful evaluation. Given the limited research available in this area, further studies are necessary to understand better the role and impact of LLMs in dentistry, particularly in treatment planning and prognosis.

Disclosures

1. The approval of the Bioethics Committee for the research: Not applicable.
2. Assistance with the article: None.
3. Financial support and sponsorship: Julien Issa is a participant of the STER Internationalization of Doctoral Schools Program from NAWA Polish National Agency for Academic Exchange No. PPI/STE/ 2020/1/00014/DEC/02.
4. Conflicts of interest: None.

References