AI Root Canal Healing: Can a Model Call It?

Your dentist finishes a root canal, seals the tooth, and sends you off with the sentence millions of patients hear every year: “We’ll know in about a year whether it healed.” That wait — the slow radiographic verdict on whether the infection at the root tip quietly resolves or festers — is one of dentistry’s quiet blind spots. The whole pitch for AI root canal healing prediction is that a model could call the outcome the day the file comes out. A new study put three of them to the test, and the real story is hiding in the error bars.

It sounds like a leap forward. But the details matter. In the Journal of Endodontics, Dufey-Portilla and colleagues trained three machine-learning models — logistic regression, a random forest, and a support-vector machine — on 276 single-tooth root canals performed by endodontists, each with its healing outcome judged clinically and radiographically 12 to 18 months later. Their question wasn’t “can AI read a scan?” It was harder: can a model predict, in advance, which teeth will fail to heal?

The study, in one glance

The design is refreshingly honest about its limits. This was a retrospective study: the researchers looked back at completed cases, labeled each as favorable or unfavorable healing, and asked the models to separate the two from baseline features. Crucially, unfavorable healing — the clinically important outcome, the one you’d actually want a warning about — occurred in just 18.8% of cases. That imbalance is the whole plot.

What an AI root canal healing model gets right

There is something genuinely useful buried here. The best-performing model, a random forest, reached a ROC-AUC of 0.724 — modest, but better than a coin flip at ranking risky teeth above safe ones. Tooth type and the size of the baseline lesion carried the most predictive weight, which squares with what endodontists already know: large periapical lesions on complex teeth are the ones to watch. A tool that nudges a clinician to book a tighter follow-up for exactly those cases is plausible — and, much like the AI pulpotomy models we examined earlier and the other work in our endodontics & restorative coverage, it reads as a decision aid, not a decision.

So why did the ‘accurate’ model catch nothing?

Here is the catch that should stop every AI headline in its tracks. A model can score 80.4% on accuracy by doing something useless: predicting “this will heal” for every single tooth. Because failures are rare — only 18.8% of cases — blind optimism is right 81.2% of the time. Accuracy: 80.4%. Failures detected: zero. The authors flag this explicitly, and it may be the single most important sentence in the paper.

It’s a smoke detector calibrated never to go off: technically right 99 nights out of 100, and catastrophically wrong on the night that matters.

Even the tuned models have a ceiling. With oversampling and SMOTE applied to the training data only, the support-vector machine pushed its precision-recall score for failures (PR-AUC) to 0.535 and sensitivity to 54.5% — better, but still missing nearly half of failures at default settings. And it all rests on 276 cases from a single setting, scored retrospectively, with no external validation — a model that looks sharp on its home data has told us little until it is tested elsewhere.

What happens when you force it to catch failures?

You can tune a model to stop missing failures — but statistics sends the bill elsewhere. When the researchers lowered the decision threshold to raise the alarm more readily, failure-detection (sensitivity) jumped to 90.9%. Impressive, until you see the cost: specificity fell to 48.9%, and the positive predictive value dropped to 30.3%. In plain terms, the aggressive model caught almost every failing tooth, but roughly two out of every three teeth it flagged as “at risk” were actually going to heal just fine. Who decides which error is worse — a missed failure, or a false alarm that triggers an unnecessary retreatment? That is not a question the algorithm can answer.

Does a single-center model work on your mouth?

A prediction model is only as representative as the patients it learned from. These 276 cases came from one group of endodontists, with their referral mix, their imaging equipment, and their definition of “healed.” A practice serving a different population might see the model’s already-modest accuracy slip further. Until these tools are validated across clinics and communities, the quiet risk is that AI prognosis works best for the patients who already receive the best care, and worst for everyone else.

Frequently asked questions