Root canal retreatment: a retrospective investigation using regression and data mining methods for the prediction of technical quality and periapical healing

Abstract Objectives This study aimed to investigate patterns and risk factors related to the feasibility of achieving technical quality and periapical healing in root canal non-surgical retreatment, using regression and data mining methods. Methodology This retrospective observational study included 321 consecutive patients presenting for root canal retreatment. Patients were treated by graduate students, following standard protocols. Data on medical history, diagnosis, treatment, and follow-up visits variables were collected from physical records and periapical radiographs and transferred to an electronic chart database. Basic statistics were tabulated, and univariate and multivariate analytical methods were used to identify risk factors for technical quality and periapical healing. Decision trees were generated to predict technical quality and periapical healing patterns using the J48 algorithm in the Weka software. Results Technical outcome was satisfactory in 65.20%, and we observed periapical healing in 80.50% of the cases. Several factors were related to technical quality, including severity of root curvature and altered root canal morphology (p<0.05). Follow-up periods had a mean of 4.05 years. Periapical lesion area, tooth type, and apical resorption proved to be significantly associated with retreatment failure (p<0.05). Data mining analysis suggested that apical root resorption might prevent satisfactory technical outcomes even in teeth with straight root canals. Also, large periapical lesions and poor root filling quality in primary endodontic treatment might be related to healing failure. Conclusion Frequent patterns and factors affecting technical outcomes of endodontic retreatment included root canal morphological features and its alterations resulting from primary endodontic treatment. Healing outcomes were mainly associated with the extent of apical periodontitis pathological damages in dental and periapical tissues. To determine treatment predictability, we suggest patterns including clinical and radiographic features of apical periodontitis and technical quality of primary endodontic treatment.


Introduction
Root canal treatment has proven to be a predictable procedure with a high success rate. 1 Nevertheless, failures occur in 14-16% of primary endodontic treatments, 2,3 and retreatments account for approximately 30% of the demand for endodontists. 1 The presence of clinical symptoms and/or maintenance/ progression of periapical radiolucency 4 are evaluated for determining treatment success. Moreover, poor technical quality of previous endodontic procedures and loss of coronal sealing 5 are considered when making decisions on endodontic retreatment. In this regard, satisfactory technical outcomes are considered when well-condensed root fillings are achieved along a working length of 0 to 2 mm from the radiographic apex. 6 Previous investigations support the preference for non-surgical retreatment over endodontic surgery and show that late failures are more prone to occur in surgically treated teeth. Simultaneously, a slower healing dynamic could explain the increased success rate over time in non-surgically retreated teeth. 7 The significantly lower healing rate in retreatments, compared to primary endodontic treatments, has been well documented. 1,8 Within this context, persistent microbial infection is one of the major causes of failure. 5 Several studies radiographically assessed the technical quality of root fillings, assuming it may affect the root canal treatment outcome. 9 However, it is necessary to identify the features and patterns that could impact achieving technical quality. This is especially relevant for non-surgical endodontic retreatment, since root canal morphology may be altered, 10 increasing technical complexity.
Interestingly, most studies that evaluate endodontic retreatment success rates include samples with broadly variable characteristics, 4,7 and few studies address whether demographic, technical, anatomical, and pathological features interfere with retreatment predictability. 1,11 Factors such as the presence of periapical radiolucency 8  Individualized strategies for diagnostic purposes and case selection are still required. In this regard, alternative methods for data analysis should be considered to improve technical and healing outcome prediction and thus guide clinical decisions. To date, descriptive statistics and/or logistic regression are the most widely used methods for observational studies in endodontics. 1,8 On the other hand, advances in the field of computer sciences allow improving the ability to record and intelligently analyze large volumes of information using other approaches. 14 Complementary approaches to purely statistical descriptive and predictive studies -including data mining strategies -are still poorly used in the dentistry field, 15 but they may significantly contribute towards knowledge discovery. 14 Knowledge Discovery in Database (KDD) is a widely employed method and has proven to be a great resource for the identification of valid, new, understandable, and potentially useful patterns for several areas of knowledge. [14][15][16][17][18] This study aimed to improve endodontic retreatment case selection by identifying frequent factors and patterns related to technical complexity and periapical healing. Therefore, we used a predictive data mining functionality to complement descriptive and regression analyses. Finally, 321 met the inclusion criteria for assessing the technical outcomes, and 117 had follow-up records and were considered for the evaluation of healing outcomes ( Figure 1).

Methodology
All teeth were treated following the UFRGS specialization program in Endodontics standard protocols. In brief, it is recommended to determine the working length using apex locators. Foraminal patency should be achieved by cervical apical root canal preparation within 1 mm from the root apex. EDTA and 2% sodium hypochlorite are used as chemical auxiliaries of root canal preparation, and intracanal medication with calcium hydroxide is maintained for two weeks prior to root canal filling.

Data collection and preprocessing
A structured electronic chart database (ECD) containing models that comprised all clinical data was developed using PHP programming language, supported by a database model created in a MySQL database management system (DBMS). Information obtained from physical records and radiographs was transferred to the web application. A total of 239 variables related to endodontic diagnosis, retreatment procedures, and follow-up visits were collected.
Unnecessary features (patient identity code, date of appointments) and variables with all missing values were eliminated. Some attributes were integrated, recoded, or calculated to construct new variables. 14,16 Variables included in the study

Medical history and diagnosis
All medical variables were assessed by self-report.
After data preprocessing, six were selected, including age and sex. Cardiovascular disease, hypertension, diabetes, and smoking habit were either present or absent.
Data collected during dental history taking and clinical and radiographic examination comprised 19 variables. Tooth number was recoded for obtaining tooth type and tooth location. Variables related to any clinical sign and/or symptom of periapical disease were integrated and considered either present or absent.
Root resorption, canal deviation, root perforations, separated instrument, and extruded filling material (present/absent) were selected, as well as their location. Some variables were integrated to indicate whether one or more procedural accidents were present or absent. The level of the root filling and the root filling quality were classified as previously described. 8,20 Root canal morphology (RCM) was regarded as altered if the primary endodontic treatment presented a short root filling level (>2 mm from root apex) and/or canal deviation, root perforation, and/or separated instrument. RCM was respected when these features were not observed.
All radiographs were measured by the same

Endodontic retreatment
Data on endodontic retreatment comprised five variables after data preprocessing. The variables related to the occurrence of new procedural accidents (yes/no) and level after root canal retreatment 8 were selected. The number of appointments was recoded (single/multiple appointments). The variables related to the management of procedural accidents were integrated. Satisfactory outcomes were considered when separated instruments were removed or bypassed, canal deviation was bypassed and/or root perforation was sealed, and the original root canal path was accessed, enabling proper canal instrumentation and filling. When the outcomes described were not achieved, the outcome was classified as unsatisfactory.
Radiographs taken after endodontic retreatment were analyzed to determine its technical quality (Kappa =0.78). The classification suggested by the European Society of Endodontology 6 (2006) was adapted.
Endodontic retreatment was technically satisfactory when well-condensed root fillings were achieved along a working length of 0 to 2 mm from the radiographic apex. If procedural accidents were present, the criteria for assessment of their management were applied. When satisfactory management of procedural accidents was not achieved, the technical quality of endodontic retreatment was deemed unsatisfactory.
Retreatment performed on multirooted teeth with at least one root canal that did not meet the criteria for satisfactory technical quality was deemed technically unsatisfactory.

Follow-up visits
Four variables were considered, including coronal sealing (present/absent) and type of dental restoration (definitive/temporary/absent). The follow-up period

Data Mining Predictive Decision Trees
A .csv file containing the dataset was opened in the Waikato Environment of Knowledge Analysis (Weka -Root canal retreatment: a retrospective investigation using regression and data mining methods for the prediction of technical quality and periapical healing J Appl Oral Sci. 2021;29:e20200799 5/14 version 3.7) software, and a new .arff file was later generated to be modeled using the Weka software (www.cs.waikato.ac.nz/ml/weka). J48 classification algorithm was used. 23 For predicting technical quality, a first experiment considered all variables related to medical history and diagnosis. In a second experiment, only variables related to diagnosis were kept, and technical errors of the previous endodontic treatment were grouped into the RCM variable. One experiment was performed for periapical healing prediction, considering all potential risk factors in the data set. For the three experiments, the minimum number of instances per leaf node was set to 7. The accuracy and stability of induced decision trees were provided by Weka and tested using the cross-validation procedures.

Data Distribution
The distribution of prognostic factors for endodontic retreatment related to medical history, diagnosis, endodontic retreatment, and follow-up visits is summarized in Table 1. Molars accounted for the most frequent tooth type (42.67%). Procedural accidents were present in 12.14% of the cases. Root fillings were regarded as unsatisfactory in 73.52% of individuals.
Root resorption was observed in 16 teeth, and all of them were located in the apical third of the root.
New procedural accidents were not observed. After endodontic retreatment, the technical outcome was satisfactory in 65.20%, and 53.84% of preexisting procedural accidents had a satisfactory technical outcome. Follow-up periods had a mean of 4.05 years.
We observed healing in 80.50% of the cases. Coronal sealing was present in 93.69% of individuals.

Regression Method Results
Univariate analysis revealed that tooth type, RC, procedural accidents, canal deviation location, extruded filling material, and RCM were significantly associated with the technical quality of endodontic retreatment (p<0.05) (

Data mining results
Three decision trees were created. The decision tree shown in Figure 2 (tree A) can be read as follows: (1) For canal deviation located in the apical third of the root, the technical quality of endodontic retreatment was unsatisfactory; (2) For absent canal deviation, severe RC, and tooth located in the mandible, the technical quality was unsatisfactory, whereas, for tooth located in the maxilla, the technical quality was satisfactory; (3) For moderate RC and separated instrument, the technical quality was unsatisfactory; (4) A straight RC immediately predicted that technical quality was satisfactory.
After grouping RCM and excluding the variables related to demographic and medical data, the decision tree ( Figure 3 -tree B) can be read as follows: (1) For altered RCM, the technical quality was unsatisfactory; (2) For respected RCM and severe RC, the technical quality was unsatisfactory, whereas, for moderate RC, it was satisfactory; (3) For respected RCM, straight RC, and presence of root resorption, the technical quality was unsatisfactory. When root resorption was absent, the technical quality was satisfactory.
Regarding the periapical healing binary class, the J48 classifier generated one decision tree ( Figure 4 -tree C) that can be read as follows: (1)         (2) For absent canal deviation, severe RC, and tooth located in the mandible, the technical quality was unsatisfactory, whereas, for tooth located in the maxilla, the technical quality was satisfactory; (3) For moderate RC and separated instrument, the technical quality was unsatisfactory; (4) A straight RC immediately predicted that technical quality was satisfactory. The values in parentheses point out the total number of correctly/incorrectly classified instances in each leaf per node (Accuracy of 62.38%) Figure 3-Decision tree B: Prediction of technical quality of non-surgical root canal retreatment, including variables related to demographic data, medical history, and diagnosis -(1) For altered RCM, the technical quality was unsatisfactory; (2) For respected RCM and severe RC, the technical quality was unsatisfactory, whereas, for moderate RC, it was satisfactory; (3) For respected RCM, straight RC, and presence of root resorption, the technical quality was unsatisfactory. When root resorption was absent, the technical quality was satisfactory. The values in parentheses point out the total number of correctly/incorrectly classified instances in each leaf per node (Accuracy of 66.66%) J Appl Oral Sci. 2021;29:e20200799 10/14 data collection on clinical and radiographic records.
To minimize this limitation, we assessed periapical status for a mean follow-up period of 4.05 years and considered clinical information to determine healing or failure. Moreover, besides PAI, we used differences in the periapical lesion area to assess healing outcomes.
Another limitation is that we obtained medical history data using self-reports, which may underreport some general health variables. 30 Thus, the absence of association between systemic conditions and periapical healing must be interpreted with caution.  (2). For definitive dental restoration and satisfactory technical quality of endodontic retreatment, periapical healing was classified as healed. (3) For definitive dental restoration at the follow-up visit, unsatisfactory technical quality of endodontic retreatment, presence of signs/symptoms at the time of diagnosis, and unsatisfactory root fillings in the primary endodontic treatment, periapical healing was classified as failure to heal; but for satisfactory root filling quality of the primary endodontic treatment, periapical healing was classified as healed; (4) For identification of definitive dental restoration at the follow-up visit, unsatisfactory technical quality of endodontic retreatment, absent signs/symptoms, and periapical lesion area greater than 4 mm2 at the time of diagnosis, periapical healing was classified as failure to heal. However, for periapical lesion areas smaller than 4 mm2, periapical healing was classified as healed. The values in parentheses indicate the total number of correctly/ incorrectly classified instances in each leaf per node (Accuracy of 79.66%) Besides, decision tree C (Figure 4) showed that small radiolucencies were likely to heal even when ideal technical outcomes were not achieved, whereas microorganism quantities and virulence were subcritical to sustain periapical inflammation. At odds with previous findings, 10 neither decision trees nor traditional statistics showed an association between endodontic retreatment failure and altered RCM. Unlike the criteria used herein, the investigation included cases with internal root resorption unsealed by former treatment in the altered RCM group and the presence of separate files and inadequate levels of filling in the respected RCM category. Moreover, the referred study 10 observed the impact of preoperative morphological alterations on endodontic retreatment success, but it did not assess the technical quality of endodontic retreatment. Hence, other features related to technical outcomes could have affected healing.
A previous study 39 revealed that poor clinical outcomes might be expected in cases with inappropriate coronal restoration, which was not verified herein.
The small number of samples with absent coronal restoration and the possibility of recent coronal sealing losses might have influenced the healing outcome.
Unfortunately, the period of absent tooth restoration could not be evaluated. An in vitro study found that the amount of saliva penetration through root fillings should be considered clinically significant only in root canals that have been exposed to the oral cavity for at least three months. 40 As previously demonstrated, 11,12  we observed a lower healing rate than premolars for anterior teeth (Table 4). This outcome is not in line with the technical quality of endodontic retreatment, which is probably related to the coexistence of other risk factors not assessed in the current analysis.

Conclusion
We found that the technical quality of endodontic retreatment is associated with several risk factors, including the severity of RC and altered RCM. We should consider procedural accidents to classify case complexity since they are especially relevant in the apical third of the roots. PLA, tooth type, and apical resorption proved to be significantly associated with healing failure. Base on the decision trees, we suggest considering risk factors and patterns combining different variables to define technical complexity and periapical healing predictability. Straight root canals combined with apical root resorption might prevent satisfactory technical outcomes. Large periapical lesions and poor root filling quality in primary endodontic treatment appeared to predispose to treatment failure. Data distribution and missing data did not allow reliable multivariate models to be carried out to estimate the predictors of periapical healing.
Additional future studies should assess variable operators' experience and include larger samples to enable further analysis.

Author disclosure statement
The authors declare no conflicts of interest.