Exploratory machine learning–based early post-treatment assessment of willingness to reuse rubber dam isolation after microscopic root canal treatment
Yinyin Cai, Ting Xiang, Qiong Sun, Lixia Guan, Chong ZhaoObjective
To explore factors associated with willingness to reuse rubber dam isolation after microscopic root canal treatment and develop an exploratory machine learning–based early post-treatment assessment model.
Methods
This retrospective cross-sectional study included 306 patients who underwent microscopic root canal treatment with rubber dam isolation from May 2025 to November 2025. The outcome was the willingness to reuse rubber dam isolation at the 1-week follow-up. Forty-seven newly enrolled patients were reserved as a same-center temporal validation cohort, and the remaining 259 patients formed the development cohort, which was divided into training and held-out test sets in a 7:3 ratio. Data regarding 26 demographic, clinical, procedural, nursing-related, and post-treatment experience variables were collected. Preprocessing, recursive feature elimination with cross-validation, and hyperparameter tuning were integrated within a fully nested cross-validation pipeline using the training data only. Six machine-learning models were compared. SHapley Additive exPlanations and generalized additive models were used for exploratory model interpretation. A sensitivity analysis excluding satisfaction level was also performed.
Results
Overall, 246 patients were willing to reuse rubber dam isolation and 60 were unwilling. The Light Gradient Boosting Machine model showed the best exploratory performance, with area under the receiver operating characteristic curve values of 0.939 and 0.983 and F1-scores of 0.94 and 0.96 in the held-out test set and same-center temporal validation cohort, respectively. Recursive feature elimination retained 12 predictors. SHapley Additive exPlanations analysis suggested that satisfaction level and postoperative visual analog scale score were the leading model-associated factors. After satisfaction level was excluded, the area under the receiver operating characteristic curve decreased to 0.820 and 0.840 in the two evaluation cohorts. Given the small temporal cohort, especially the 14 patients unwilling to reuse rubber dam isolation, this estimate was considered preliminary and potentially imprecise.
Conclusion
The model identified satisfaction, postoperative pain, discomfort-related factors, and procedural variables associated with willingness to reuse rubber dam isolation. Prospective multicenter validation is required before clinical implementation.