New Model Predicts Healing After Root Canal Microsurgery

Predicting which periapical lesions will heal after endodontic microsurgery remains difficult. A new study in Scientific Reports, published November 9, 2026, tested a model built for that job. Researchers compared its one-year healing predictions with estimates from five endodontic clinicians. The model performed reasonably, and it did not outperform the specialists.

TL;DR

• Researchers in Hefei, China, built a healing prediction model from 420 endodontic microsurgery patients

• Buccal bone plate height and CBCT bone-loss volume drove most of the model's predictions

• The model's AUC fell from 0.868 in internal testing to 0.703 in later patients

• Five endodontic clinicians scored 0.771 against the model's 0.696, a non-significant difference

What Did The New Endodontic Microsurgery Study Investigate?

Endodontic microsurgery, often called an apicoectomy, removes the infected root tip under magnification. The surgeon then seals the canal with a root-end filling. Most teeth reach this stage after a lesion persists despite earlier treatment. The guide to failed root canal symptoms covers that path.

Patients at this point have often sat through one or two rounds of root canal treatment. Their question at consultation is simple. Is this tooth worth one more attempt, or should it give way to extraction and an implant? A reliable healing estimate would help answer that.

Sheng Feng, Lei Zhang, Anan Wang, and colleagues at Hefei Stomatological Hospital and Anhui Medical University ran the study. The development cohort held 312 patients, with 273 healed and 39 failed cases. A later temporal cohort added 108 patients, for 420 in total.

The team screened 26 candidate predictors and compared six machine-learning models. They selected an L2-penalized logistic regression, a regularized form of a familiar statistical tool. Prediction is its only job. Systems that guide instruments during surgery belong to dental robotics, a separate field.

How Accurately Did The Model Predict Healing?

The study reports discrimination as AUC, which measures how well a model ranks patients who heal above those who fail. A score of 0.5 equals chance, and 1.0 means perfect separation in that sample.

AUC and accuracy measure different things. In the temporal cohort, 95 of 108 patients healed. A model predicting healing for everyone would look 88% accurate and catch zero failures. A 0.703 AUC does not mean 70.3% accuracy.

That imbalance shapes how a clinic would use any such tool. Most teeth heal, so the useful job is spotting the minority that will not. A model needs strong sensitivity for failures to earn a place in case review.

The 0.868 figure comes from the set used to pick the final model, so it likely runs optimistic. The temporal calibration slope was 0.54 against a target of 1.0, which points to overconfident risk estimates.

Which Factors Drove The Model's Healing Predictions?

The authors used SHAP, short for SHapley Additive exPlanations, to show how each variable shaped predictions. Two bone measurements dominated, and both line up with recent clinical research.

Buccal Bone Plate Height

Buccal bone plate height is the bone height on the cheek side of the root. A 2023 meta-analysis in Clinical Oral Investigations pooled 232 cases on this variable. Teeth with over 3 mm of buccal bone healed in 91.4% of cases. Teeth at 3 mm or less healed in 71.4% of cases. The difference did not reach statistical significance.

Clinically, the finding makes sense. A low buccal plate often means the lesion has reached, or nearly reached, the marginal bone. That apicomarginal pattern links the periodontal pocket to the surgical site. Epithelium can then grow into the defect before bone fills it. Surgeons read this on presurgical CBCT sagittal slices. It shapes the flap design, the osteotomy, and any regenerative step.

CBCT Bone-Loss Volume

CBCT bone-loss volume measures bone destruction around the lesion in three dimensions, using CBCT scans. Lesion size was the one hard tissue factor that reached statistical significance in the 2023 meta-analysis. Lesions of 5 mm or less succeeded in 78.4% of cases, against 63.3% for larger ones.

Volume captures more than diameter. A lesion that looks modest on a periapical film can extend deep into the bone buccolingually. A BMC Oral Health cohort tracked volume directly. Lesions shrank by an average of 89.4% at one year or later.

What The Bone Findings Can And Cannot Show

SHAP describes model behavior and cannot show cause. The study does not prove that grafting or preserving buccal bone changes healing odds. Buccal height also drops after osteotomy, so preoperative and postoperative readings differ.

Treatment choices can also act as markers of difficulty. In a study of 143 complicated cases, guided tissue regeneration carried a higher extraction risk. Surgeons tend to reserve membranes for the worst defects. In clinical terms, the membrane probably flags case severity. Any prediction model has to handle that confounding.

Did The Prediction Model Perform Better Than Endodontists?

Five endodontists gave probability estimates for the patients in a paired benchmark. The model reached an AUC of 0.696, and the clinicians reached 0.771. The difference of 0.075 was not statistically significant (DeLong P = 0.186).

The estimates lean toward the clinicians, so the data give no support to claims that software beats specialists. With so few failures, the study also lacked power to prove the two perform equally.

Clinicians also see things a model cannot. During surgery, the resected root surface is stained and inspected under high magnification. Microcracks, missed isthmuses, and leaking root fillings show up there and change the prognosis on the spot. A model built from imaging and record data has no view of that.

Surgeon-controlled steps matter as well. In the 2026 BMC Oral Health cohort, root resection angle predicted outcome. That variable is decided at the chair. In the Dental Reviewed survey on AI in dentistry, only 2% of dentists trusted AI over their own judgment. These results give them no reason to change that view yet.

Why Did Performance Decline In Later Patients?

Temporal validation tests a model on patients treated after the development period. Case mix, imaging protocols, and surgical habits can shift between those periods.

Here, AUC dropped from 0.868 to 0.703. The authors do not tie the drop to one cause. The temporal cohort had only 13 failures, and its confidence interval ran from 0.530 to 0.867. The lower end sits close to chance.

Thirteen failures is a small number for any clinician to learn from, and smaller still for a model. A new root-end material or a new CBCT unit could shift results at that scale. Larger multicenter samples, with calibration checks across operators, are the next step.

Is One Year Long Enough To Judge Healing?

The study defines success at one year, a standard endpoint in endodontic surgery research. Bone does not always finish remodeling on that timeline.

The 2024 Clinical Oral Investigations study followed complicated cases for up to nine years. Success rates kept rising from six months to two years, then held steady from two to nine years. Some one-year failures in a model like this may simply be slow healers.

In practice, an asymptomatic tooth with an uncertain one-year result earns another review at 18 to 24 months. Re-intervention can wait unless symptoms return or the lesion grows. A prediction model trained on one-year labels inherits that uncertainty.

What Could This Mean For Endodontic Treatment Planning?

The model is a research tool and is not ready for routine use. Its predictors include perioperative and postoperative data, so it cannot yet guide a consultation before surgery. Once validated, a tool like this could help with several tasks.

• Setting expectations at the one-year review, when healing may still be incomplete

• Flagging cases with low buccal bone height or large bone-loss volume for closer follow-up

• Giving clinicians a structured way to explain uncertainty to patients

• Guiding future research on individualized surgical risk

Practices can prepare now without any new software. Recording buccal plate height and lesion volume on every pre-surgical CBCT builds a local outcome dataset. That record makes later audits and any future model validation much simpler.

A clear dental treatment plan also documents expected healing, the review interval, and fallback options. The Dental Reviewed planning software does not include this study's model.

What Are The Study's Limitations?

Several limits apply before anyone cites these results.

• Data came from a single hospital in China

• The temporal cohort included only 13 non-healed cases

• Performance dropped in later patients

• Confidence intervals for validation AUC were wide

• Only five clinicians took part in the benchmark

• Predictors include perioperative and postoperative variables

• No trial has tested whether using the model improves patient outcomes

The article is an early accepted manuscript, so figures may change in the final version.

Bottom Line

The Hefei study shows that two bone measurements can support an interpretable healing prediction after endodontic microsurgery. The model kept moderate discrimination in later patients, with an AUC of 0.703. It did not significantly outperform five endodontic clinicians, and its risk estimates looked overconfident. Multicenter validation and a preoperative version come next. Until then, careful CBCT reading, intraoperative inspection, and patient follow-up remain the basis of surgical prognosis.