We aimed to assess patient-, tooth- and treatment-level covariates on the failure of root canal treatments (RT) and to predict failure using machine learning (ML).
Teeth receiving RT at one large university hospital from 2016-2020 with a minimum follow-up of ≥6 months were included. Failure compromised absent radiographic healing and/or the presence of clinical symptoms. Covariates were selected on tooth-, treatment- and patient-level. We used logistic regression (logR) to determine associations in the full dataset, and logR as well as more advanced ML (random forests (RF), gradient boosting machine (GBM) and extremely gradient boosting (XGB)) for predictive modeling (area-under-the-receiver operating characteristic-curve (ROCAUC)) and testing on a separate test dataset.
458 patients (female/male 47.2/52.8%) with 591 permanent teeth were included (overall success rate 79.5%). In logR, tooth-level covariates showed strong associations with failure: Alveolar bone loss 66-100% (ABL, OR 6.48, 95% CI [2.86; 14.89], p<0.001); Periapical index (PAI) score≥4 (OR 4.59, [2.44; 8.79], p<0.001); ABL 33-66% (OR 2.59 [1.49; 4.49], p<0.001); PAI=3 (OR 2.45, [1.43; 4.34], p<0.01); Treatment type "retreatment" (OR 1.77, [1.01; 2.86], p<0.01). On patient level only smoking (OR 2.05, [1.18; 3.53], p<0.05) was significantly associated with risk of failure. For predictive modelling, the predictive power of all models was limited (ROCAUC: logR 0.63, [0.53, 0.73]; GBM 0.59, [0.50, 0.68]; RF 0.59, [0.50, 0.68]; XGB 0.60, [0.50, 0.70]).
Failure of RT was associated mainly with tooth-level covariates. Predicting failure was only limitedly possible, also with more complex ML.
Identifying specific risk factors for failure of RT and predicting the outcome of RT is relevant for treatment planning and informed shared decision-making. The present study found tooth-level factors to be associated with failure. Notably, predicting failure was only limitedly possible.

Copyright © 2021. Published by Elsevier Ltd.

Author