Introducing ClinicGPT: A custom large language model for institutional dental clinics

1. PROBLEM

It is well‐established that cost is the largest barrier to dental care in America, with many underprivileged populations relying on the subsidized fees of dental schools. In fact, dentistry faces the greatest prevalence of financial barriers to care compared to all other healthcare industries. ¹ In North America, dental schools operate essentially as a substitute for the extremely limited public dental system. ² Although teaching institutions bear a significant patient load (often with month‐long waitlists overrepresented in low socioeconomic background individuals), many suffer from poor efficiency due to various factors including student inexperience, complex administrative processes, and a shortage of qualified faculty. ³ A two‐week observational study performed in the main student clinic at the Schulich School of Medicine and Dentistry (Ontario, Canada) demonstrated that a substantial portion (21% ± 5%) of the average appointment time slot (typically 2–3 h in duration) was spent waiting for instructor assistance, and 15% ± 5% was spent on administrative duties such as paperwork, documentation, and looking up information. ⁴ Only 38% ± 13% of the average appointment duration was spent on actual clinical procedures (Figure 1). ⁴ Dental schools can generally improve productivity through two avenues: (1) by investing in faculty and staff, or (2) by developing tools to increase the efficiency of the existing workforce. Given the high turnover and recruitment challenges faced by dental schools across North America, it is perhaps more practical to pursue the latter option. ³ , ⁵ , ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ Thus, the objective of this study was to use contemporary artificial intelligence to help address the pressing need of optimizing clinic productivity. Ameliorating the operating efficiency of dental schools will translate directly to a better functioning “public” dental system, granting more underserved patients access to life‐changing oral healthcare.

FIGURE 1.

Breakdown of how student time is spent in a sample of 18 appointments in various disciplines over a 2‐week period. Means reported. *Administrative tasks include note taking, filling out paperwork, updating electronic documentation, etc. **Discussions with patient includes only productive discourse, minus small talk or time spent waiting for instructor assistance.

2. SOLUTION

ClinicGPT is a custom large language model (LLM) chatbot developed as a proof‐of‐concept aide for senior dental students in their clinical years. Based on the same foundational technology as ChatGPT, our ClinicGPT model was built on OpenAI's Davinci LLM application programming interface (API). ¹¹ A custom knowledgebase was created using information curated from Schulich Dentistry's staff directory, standard operating protocols, and clinic manuals. These files were indexed with the OpenAI embeddings API and stored in a Pinecone vector database, allowing ClinicGPT to respond to a user's prompt with context specific to the educational institution it was designed for. Additionally, we implemented a dynamic indexing system that allows clinic administration to update this knowledgebase effortlessly and instantly by cc'ing a private email address whenever memorandums are sent to faculty or students. Finally, OpenAI's fine‐tuning API was used to further train the Davinci LLM to produce replies with Schulich‐specific phrasing and mannerisms, such as the addition of a disclaimer to double‐check with supervising clinic instructors whenever ClinicGPT is unsure of an answer. Fine‐tuning efforts are ongoing to enable ClinicGPT to draft appointment records in a similar format to Schulich dental students.

3. RESULTS

After two weeks of preliminary training, ClinicGPT can respond to administrative or protocol‐related questions accurately and with information specific to Schulich dentistry. The chatbot can recall the context of previous questions (Figure 2) and make inferences regarding the topic of the user's discussion. We expect ClinicGPT to be particularly useful in emergent situations, especially when clinic instructors are unavailable (for instance, a needle stick injury after hours, or a patient inquiring about payment plans as a morning appointment runs overtime into lunch break). It is important to mention that ClinicGPT is not designed to replace clinical instructors. Rather, it allows students to self‐triage their questions and prevents burdening the limited number of faculty with non‐clinical questions. This novel proof‐of‐concept clinical LLM serves as a foundation that can be further developed with additional use cases. For instance, our next step for this project includes training ClinicGPT with live pricing data of dental consumables from various suppliers. The power of natural language processing could then be used to find actionable insights, such as the lowest cost option for endodontic files. This functionality would disrupt the traditionally opaque dental supply industry, with pricing information usually hidden behind paywalls (such as the Ontario Dental Association Supply Hub) or distributed through direct correspondence with sales representatives. Transparent pricing is expected to enable dental schools to make data‐driven bulk purchase decisions, as well as increase manufacturer competition and decrease consumable‐related expenses for clinics.

FIGURE 2.

Screenshot of a real‐time demonstration of ClinicGPT's capabilities. Personal information (names, phone numbers, emails) are redacted for privacy.

4. WEBSITE

ClinicGPT can be accessed at kxzdds.com/clinicgpt. The author can be contacted at cv.kxzdds.com or publish.uwo.ca/∼kzhou54.

Zhou KX. Introducing ClinicGPT: A custom large language model for institutional dental clinics. J Dent Educ. 2024;88(Suppl. 3):1979–1981. 10.1002/jdd.13348