Rise of the machines – artificial intelligence in dentistry
Artificial intelligence will fundamentally change the way that healthcare is delivered, and health practitioners need to be aware of these emerging technologies.
If you were like me, your Christmas break was not only consumed with eating, relaxing and spending time with family and friends, but also exploring the wonders of artificial intelligence (AI) with ChatGPT. Launched in November 2022, this AI chatbot has exploded into public consciousness, and hardly a day seems to go by without another news story extolling the wonders and risks that it poses across a range of business sectors.
As an experiment, I used ChatGPT to write half of my monthly column for the Australian Dental Association Victorian Branch magazine in January. It took about one tenth of the time that it would normally take, using just a few brief prompts. I’m sure if I hadn’t confessed to using it in the second half of the column, most readers would have been none the wiser (although I’m not sure whether that says more about my writing ability or that of ChatGPT).
AI will undoubtedly change the healthcare landscape in a range of different ways over the next decade and beyond. This is because fundamentally healthcare is about data. If we think about the typical patient journey - from history and examination through to treatment - clinicians are collecting and recording data points at every step, and this data drives the decision making process.
In order to understand where AI might influence dentistry, it’s instructive to look at perhaps the most fundamental process in patient care - determining a diagnosis. Dental practitioners diagnose using pattern recognition. They look for patterns in the signs and symptoms that a patient presents with through clinical examination, history and tests including radiographs. Diagnosing tooth decay is a process of looking for observable clinical patterns on the tooth surface, and radiographic changes that indicate demineralisation of the tooth enamel and dentine.
How does AI work? Machine learning is essentially pattern recognition, using an algorithm to analyse mountains of data to observe and learn from patterns, and then apply that learning when presented with new information.
A good clinician uses their clinical judgement - their own internal database built from past experience coupled with their understanding of the broader evidence base - to make decisions. Perhaps it is easy to see an obvious problem here in the way that clinicians process the data that they collect as they treat patients. By definition the database of clinical experience that helps to inform decision making is different for each clinician. People attend continuing education courses and read journal articles based on their own preferences - often preferencing areas of clinical practice that they enjoy (building on strengths) rather than focusing on less enjoyable aspects of practice. Confirmation bias shapes the way we interpret evidence, and what we bring back from continuing education courses to implement in practice.
A common trope in dentistry is that if you send a patient to 5 dentists, you end up with 10 different treatment plans - and perhaps not surprisingly there is some research to suggest that is true. On something supposedly as simple as dental tooth decay, dentists will sometimes disagree on the presence, extent or severity of the disease and the intervention required. Not only is there variation between clinicians, but that same variation exists within clinicians – repeat examinations by the same clinicians yield a small but significant variation in diagnostic decisions. As a profession we accept that this variation exists (as it does across all areas of healthcare) and one rationalisation is that there is more than one ideal treatment option for any given patient.
Right now, AI is being trained on relatively small datasets which limits their applicability. Over time though, the expectation is that AI will have much larger datasets to draw from, which means than machine learning will be looking for patterns and refining decision making based on thousand or even millions of pieces of data - much more than any human could ever process.
This highlights one of the potential benefits that AI could bring - overcoming the biases that all clinicians have, and bringing consistency and uniformity to the diagnostic and treatment planning process. This could be in standardising diagnosis from radiographs, or analysing treatment success and failure in large datasets to provide more robust predictions of risk of adverse outcomes.
Of course, this is all dependent on how the AI is trained to look at the data - because biases can be introduced in the creation of algorithms used to interrogate data. Whether this is a result of the humans who choose the data that AI is trained on or the parameters that guide machine learning, the possibility exists that the algorithm will inherit the unconscious biases of the developers. This raises some important practical and perhaps ethical issues in the development, testing and implementation of AI in health care.
Data security is also a significant risk - developers will require access to large datasets of private health information, and in use AI will have access to even more data. It is critical that we ensure that this information is adequately protected from the inevitable cyber risks.
These are all problems that can be overcome, but it is incumbent on those who are researching and developing these new technologies that these risks are front of mind. It is also important for clinicians to take an active interest in understanding both the benefits and risks of these emerging technologies. AI should be seen as an adjunct to clinical practice to help guide decision making and improve patient outcomes.
Inevitably AI will move to direct-to-consumer models, with patients able to take images of their mouth with a smart phone and be provided with a ‘diagnosis’ and recommended treatment options. Without the filter of a clinician, there is a risk that one of the important components of health care – informed consent – will be missing. Will direct-to-consumer AI have sufficient capability to adequately explain diagnosis, prognosis, risks, benefits and treatment options specific to the patient? At some point in time, the answer to these questions is likely to be yes.
One of the key learnings from the pandemic is the accelerated development, uptake and acceptability of various technologies. A decade ago I was involved in researching teledentistry as a way to improve access to dental care for children and residents of aged care facilities. The technology was clunky and as a consequence not particularly acceptable to either dentists or patients, despite the potential benefits. The past few years has seen telehealth become an accepted, and perhaps even preferred, option for many patients and clinicians.
How quickly artificial intelligence will be adopted into dental practice remains to be seen, but if history is any guide, it won’t be long. So prepare yourself, because artificial intelligence is coming (it’s actually already here), and we can’t bury our head in the sand and expect to avoid it.
This week in dental research
A smartphone app that enables remote screening of children’s teeth by dental professionals has passed a feasibility study by researchers at The University of Western Australia. Researchers developed a dental to enable asynchronous dental screening based on dental photographs taken by primary caregivers. Although the sample size was small, the model was found to be feasible and acceptable in detecting dental caries from dental photographs. It has the potential to triage and refer children, particularly in regions that have limited dental care access.
Next week: Sustainability in dental practice.
Last week: If you missed the last few weeks, you can go back and read about Suicidal ideation in Australian dental practitioners and How is public dental care funded in Australia?