Does AI really improve orthodontic treatment? A new RCT says “Maybe”

Developments in AI-assisted orthodontics appear to be gathering pace. These are exciting times for the future of orthodontics; however, there has been little research into the effects of AI on the quality of orthodontic treatment. This new study is a great first step towards high-quality research in this area. 

Before I continue with my discussion of this paper, I would like to point out that I did not post over the last two weeks because I was on a fantastic holiday in the Isle of Skye in Scotland. In addition to giving me a break, the internet access was rather poor where we were staying, in a remote part of the island. 

The authors of this paper noted that digital dentistry is significantly affecting our treatment methods. When these are combined with Artificial Intelligence, there are enormous opportunities to deliver our care. 

A team from China and Malaysia did this study. The journal PLOS ONE published the paper. 

What did they ask? 

They wanted to

“Compare AI-assisted digital orthodontics with conventional fixed appliance therapy using percentage reduction in the PASS score as a primary endpoint.” 

What did they do? 

They conducted a single-centre, parallel-group, randomised controlled trial. The PICO was 

Participants 

140 orthodontic patients with class I malocclusion 

Intervention. 

This was an AI-assisted digital workflow. It comprised a suite of AI-assisted tools embedded within 3D planning software and the Dental Monitoring platform. These tools use machine learning algorithms to automate tooth segmentation, landmark identification, arch form estimation, and preliminary tooth movement simulations. The outputs were then used to inform treatment planning. However, all final diagnostic interpretations and movement prescriptions were determined by the treating orthodontists. 

They manufactured custom brackets chair-side using 3D printing. 

When treating patients, remote monitoring with Dental Monitoring was used. This detected issues such as bracket failure, wire displacement, plaque accumulation, or soft tissue irritation. If threshold-based alerts were triggered, clinicians reviewed the images and arranged earlier clinical visits where necessary. 

Control 

This group of patients was treated with conventional orthodontics. Panoramic radiographs and lateral cephalograms were used for diagnosis and treatment planning, and all patients were treated with 0.022-inch-slot pre-adjusted edgewise appliances. Follow-up appointments were scheduled at four-week intervals. 

A team of board-certified orthodontists, each with at least five years of clinical experience, carried out the treatment. 

Outcome 

The primary outcome was the percentage reduction in the Peer Assessment Rating (PAR) from baseline to treatment completion. 

They conducted a clear sample size calculation, which indicated that 140 participants were needed to take part in the study. 

They used a pre-prepared randomisation sequence with block randomisation and concealed the allocation using sequentially numbered, sealed envelopes. These were prepared by a research assistant not involved in participant recruitment or assessment. The envelopes were opened by a treating clinician only after all baseline measurements had been completed. It was not possible to blind patients or operators to the group allocation, but data were analysed blind. 

Finally, they carried out appropriate univariate and multivariate analyses. 

What did they find? 

140 patients were randomised in a one-to-one ratio to the interventions. All participants received their allocated intervention and completed treatment. 

At the start of treatment, there were no major differences between the two groups. 

When PAR scores were examined, they decreased in both groups, with a lower post-treatment score in the digital and AI group. The mean final PAR score for the AI group was 4.88 points, with a standard deviation of 4.45. For the conventional group, the mean was 7.81, with a standard deviation of 0.7. This resulted in a mean difference of -2.93 (95 CI -3.13 to -2.73). These differences were statistically significant. 

Another way to interpret PAR scores is to consider a 70% reduction a “great change” in score. The proportion of participants achieving this was 82.9% in the AI group and 50% in the conventional group. 

The final step in the statistical analysis was a regression analysis aimed at identifying independent predictors of improvement in PAR scores. The most significant factor was treatment modality, followed by patient age and baseline PAS scores. 

Their overall conclusions were 

“The AI-assisted digital orthodontic workflow produced statistically significant and clinically interpretable short-term improvements in PAR scores compared with fixed appliances. These benefits likely reflect the combined influence of high-precision imaging, customised appliance fabrication, and AI-supported monitoring. They cannot be attributed to AI alone.”

The authors made an important point. The intervention was not limited to AI. In fact, they tested improved imaging, appliance customisation, remote monitoring and workflow integration simultaneously. 

What did I think? 

It is not often that I come across a very well executed and well written publication. This was the case with this study, which was published in a very good open access journal. The study team followed a standardised and clear randomised trial methodology, and the trial was carried out very well. 

Their findings were interesting, showing a difference between the two interventions. As usual, we need to consider whether these are clinically significant. When I looked at the raw PAR scores and the differences in finish between the two groups, the differences were only in the region of 2-3 PAR points. These are not clinically significant, even though they were statistically significant. 

However, it was interesting to see that a much greater proportion of cases treated with AI-assisted mechanics were classified as greatly improved. As a result, their conclusions are fairly robust. 

The study team highlighted some shortcomings in their study. Firstly, they noted that the treatment was assessed at the end of active treatment. I thought this was entirely reasonable and reflected common practice in occlusal index studies. They also noted that it was a single-centre study. While this may reduce generalisability, it still provides us with very useful information. 

When I consider the findings of this study, however, I cannot help but feel that I would have liked to see some information on the overall duration of treatment. As I am sure this would influence the final treatment outcomes. It would also provide us with very useful information on whether AI-assisted treatment was of shorter duration. The study team pointed this out and suggested that larger-scale studies be conducted. This study is an excellent starting point for planning future studies.

I think that we should all look forward to further studies in this interesting area, as it is very likely to influence the future of orthodontic treatment.