Transforming orthopaedics with AI: Insights from a custom ChatGPT on ESSKA osteotomy consensus

Abstract

The integration of artificial intelligence (AI) in orthopaedics is no longer a futuristic concept but a present-day reality. Overlooking this technological revolution may result in missing out on a powerful tool in today's highly competitive environment. AI technologies in orthopaedics are transforming the field with tools like predictive analytics, computer vision and natural language processing (NLP). While Predictive analytics uses machine learning to predict patient outcomes, potential post-operative complications and length of hospital stay [5], computer vision enhances imaging diagnostics, automated measurements and has great potential in preoperative planning and post-operative assessments in osteotomies and arthroplasty [7, 24, 25]. NLP is another branch of AI that enables computers to understand, interpret and generate human language. LLMs like ChatGPT are specific types of machine learning in the field of NLP. Since its emergence in November 2022, ChatGPT has gained popularity and there have been fast-paced applications in the field of orthopaedics and sports medicine [2, 10, 14, 21]. The rapid advancements and enhanced capabilities of ChatGPT and similar LLMs have made it challenging to keep up with the latest developments. ChatGPT has been successfully used to generate orthopaedic clinical letters and discharge summaries with acceptable accuracy and readability; however, consistency remains a challenge [4, 23]. The development and creation of orthopaedic-specific language models could improve the quality facilitating its implementation in our daily practice. Several studies have tested the accuracy of ChatGPT in answering patients' questions related to anterior cruciate ligament (ACL) surgery, hip and knee arthroplasty [13, 19, 27]. ChatGPT provided answers with reasonable accuracy but could not replace the expertise of orthopaedic surgeons. While models like ChatGPT have been revolutionary, they lack specialisation, use general knowledge and potentially unreliable sources from the web. A custom ChatGPT addresses this by being tailored to specific needs or applications through training or fine-tuning the model on specialised datasets. These models are gaining more attention and preference compared to the generic forms of ChatGPTs as they are more accurate and less likely to generate false information [3, 6, 16, 18]. There are many LLMs created by different organisations and research groups, making the total number of currently available LLMs unknown. While ChatGPT remains the most popular, other available models offer distinct capabilities and excel in certain aspects, making each more suitable for a particular task. For instance, the Perplexity AI model is more accurate in source citations, reducing hallucinations and fake references compared to ChatGPT which is crucial in research and scientific writing [1]. On the other hand, NotebookLM, Google's research assistant, is a great tool that offers personalised AI based on uploaded documents, knowledge on a website, or a YouTube video. It can accurately answer questions, produce summaries and generate AI-powered podcasts from these notes with a single click. Scispace is another AI-powered platform that assists academic writing. It can perform literature research, create a literature review providing genuine references and generate slide presentations from any document. The ESSKA osteotomy consensus is a comprehensive consensus by world experts in the field of osteotomy. It combines scientific evidence and expert opinion regarding the indications, planning, surgical strategy, rehabilitation and complications of osteotomies around the knee for management of the painful, degenerative varus knee [8, 20]. This comprehensive consensus is invaluable for surgeons looking for evidence-based answers to questions related to this topic. However, getting a quick answer from the 117-page document with more than 450 references can be time-consuming. We have developed a custom ChatGPT tool designed to provide answers specifically from the consensus (https://chatgpt.com/g/g-iYCRdvUEz-knee-osteotomy-expert). This is a valuable tool for orthopaedic surgeons and other healthcare professionals seeking swift evidence-based answers from a trusted source. We tested the answers for accuracy, relevance, clarity, completeness and adherence to the consensus statement document [17]. Unlike generic ChatGPT, this tool provides answers strictly from the consensus and doesn't search the web or use its knowledge. This is achieved by giving ChatGPT specific instructions when setting up the custom model followed by training to improve the answers' quality. Several studies have tested ChatGPT's effectiveness in answering questions related to ACL and other sports injuries; however, custom ChatGPT models are yet to be explored [9, 11, 12, 15, 22, 26]. Although the custom model provides accurate answers, occasionally, it can generate incorrect responses (hallucinations). These models continue to improve in terms of speed, reducing hallucinations and increasing accuracy. For example, shortly after creating that custom ChatGPT another LLM emerged that we used to create an improved version that provides more accurate answers with better consistency and supports the answers with references from the consensus (https://notebooklm.google.com/notebook/febad095-2e99-4031-9e29-86eb16d0d2b7). The future applications of similar custom models in orthopaedics are endless. Similar models can be used to answer patients' common questions about a certain procedure or produce patients' education leaflets with a few clicks. Chatbots may eventually serve as an initial triaging tool before a patient consults a surgeon, once these models are advanced enough to provide accurate guidance to patients. While the use of AI in healthcare and research is of great value, it has its limitations that need to be addressed. A primary concern in adopting these models is patient privacy and information governance. Many of these models store data on cloud platforms, which can expose sensitive patients' information to data breaches. Furthermore, these models can be biased or even produce incorrect information, raising the concern about accountability. Finally, with the rapid advancements in language model technology, it would not be long before these models can analyse data and write high-quality manuscripts. Such developments hold significant potential for education and research applications. Nevertheless, this may blur the boundaries between human and AI authorship. These concerns introduce regulatory and ethical challenges that may necessitate new laws and regulations to address them. AI use in orthopaedics will reshape research, patient diagnostics, patient education and treatment. While its integration is time-saving, cost-efficient and can improve outcomes, it brings its own challenges that will evolve as we implement its use. The authors declare no conflicts of interest. The ethics statement is not available.

Ähnliche Arbeiten