Non-surgical retreatment is one of those areas in endodontics where experience plays a big role — and still, outcomes can surprise you. Two similar-looking cases don’t always behave the same way. One heals uneventfully, the other doesn’t, even when everything seems to have been done right.
A recent study tried to look at this from a slightly different angle — not how to perform retreatment better, but whether we can predict which cases are more likely to heal before we even begin.
The researchers used artificial intelligence models trained on clinical data from multiple centres. Three models were compared, including logistic regression, random forest, and a machine learning model called XGBoost. Among them, XGBoost performed the best, showing strong accuracy in predicting healing outcomes after retreatment.
What’s interesting is not just that AI performed well, but what it picked up on. The strongest predictors were things we already consider clinically — preoperative periapical lesions, molars, presence of post and core, quality of the coronal restoration, and the number of previous treatments.
None of these are new. But what AI does differently is combine all of these variables and weigh them together, rather than looking at them in isolation. That’s where it becomes useful — not as a replacement for clinical judgment, but as a way to structure it.
The overall healing rate in the study was around 82%, which is in line with what we see in practice. But the more relevant takeaway is this: some cases are inherently higher risk, and identifying them early could change how we plan treatment.
For example, a molar with a pre-existing lesion and compromised coronal seal may not carry the same prognosis as a straightforward retreatment case — even if both appear manageable clinically.
The study does have its limitations. It’s retrospective, follow-up is relatively short, and not every variable that influences outcome — like operator skill or microbial profile — was included. So this isn’t something ready for chairside use yet.
But it does point toward where things are heading.
We’re already seeing AI being used for radiographic interpretation and working length determination. What this adds is the possibility of predictive decision-making — having a tool that helps you assess prognosis before starting treatment.
Because in retreatment, the real challenge is often not the procedure itself.
It’s deciding whether the case is likely to succeed in the first place.
Prediction models require calibration, validation and clinical utility
A model may separate higher- and lower-risk cases in its development data yet perform poorly elsewhere. Before clinical use, report patient selection, missing data, predictor timing, outcome definition, discrimination, calibration and external validation. Data leakage—using information unavailable at the decision point—can create unrealistically strong results.
| Model domain | Question | Clinical safeguard |
|---|---|---|
| Discrimination | Can the model rank risk? | Report AUC with uncertainty |
| Calibration | Do predicted probabilities match outcomes? | Calibration plot and intercept/slope |
| External validity | Does it work in another setting? | Independent multicentre testing |
| Utility | Does it improve decisions? | Compare benefit and harm at thresholds |
Integrate prediction with QPD diagnostic evidence, postoperative pain evidence and dental AI governance.
Frequently asked questions
Can AI decide whether to retreat a tooth?
No. It may support risk estimation after validation, but the clinician and patient make the contextual decision.
Is high model accuracy enough?
No. Calibration, external validation, bias and clinical utility are also essential.
What patient factors still matter?
Restorability, anatomy, lesion status, prior quality, symptoms, periodontal prognosis and patient preference.