Longevity Recommendation for Root Canal Treatment Using Machine Learning ^†

Abstract

Root canal therapy is a vital dental procedure for salvaging severely decayed or infected teeth, preserving them instead of extracting them, thus averting the risk of reinfection. Nonetheless, the prevalence of root canal treatment (RCT) failure is surprisingly high, potentially leading to painful abscesses and severe infections. This study delves into the multifaceted reasons behind RCT failures and employs support vector machine (SVM) technology to predict treatment longevity. The research dataset comprises 332 manual instances, subjected to rigorous 10-fold cross-validation for testing and accuracy assessment. SVM is employed to categorize failed RCT cases into distinct classes, such as broken instruments, periapical radiolucency, root fractures, vertical root fractures, pulp stones, adequate periodontal support, periapical abscesses, overfilled cavities, and perforated or underfilled cavities. By scrutinizing the interplay between these treatment-failure-causing factors, the system discerns their impact on treatment duration. Comparisons are made with other machine learning models, including logistic regression (LR) and the naïve Bayes classifier (NB), to pinpoint the root causes of RCT failure in terms of accuracy, sensitivity, and specificity. Interestingly, logistic regression emerges as the top-performing model, with an impressive 92.47% accuracy rate. This study investigates the causes of RCT failure and employs SVM to predict treatment longevity, offering crucial insights for addressing this common dental issue. This study’s findings highlight the efficacy of logistic regression for identifying RCT failure causes, providing valuable guidance for improving dental procedures and patient outcomes.

1. Introduction

In dentistry, X-rays of the teeth are used to diagnose any defects or structural changes. However, relying solely on dentists might occasionally slow down the treatment process because detecting abnormalities in X-ray images involves human work, experience, and time [1]. Due to its enhanced accuracy and speed compared to conventional approaches, machine learning is the newest and most cutting-edge technology used for evaluating medical images When it comes to RCT, the overall median survival rate of a tooth that has been treated with RCT is approximately 11.1 years. Twenty-year survival rates for 26% of teeth were seen in numerous situations [2]. According to the findings of past research, the success rate of root canal treatment ranges from 90% to 95% when the highest possible standards are adhered to during the treatment process [3]. Yet, the effectiveness of the treatment that determines longevity might be negatively impacted by a number of different factors. These variables may include treatment via follow-up care [2]. There are many clinical and non-clinical factors that can cause endodontic treatment to fail, including periapical radiolucency, root fractures, damaged teeth, inadequate periodontal support, pulp stones, and periapical abscesses [4,5,6]. It has been observed that the success rate of treating infected teeth decreases when these types of procedural errors occur. Again, poor dental hygiene, advanced age, non-vegetarianism, smoking, excessive alcohol consumption, geographical location, and lack of formal education are all non-clinical variables that can lead to treatment failure [6,7]. Different machine learning approaches have been used in order to identify the factors that are predominantly responsible for the failure of root canal therapy. Things like a broken instrument, an overfilled cavity, periapical abscess, pulp stones, a vertical fracture in the root, a broken tooth, insufficient periodontal support for a perforated root, or an underfilled cavity are examples of potential complications. This method’s accuracy demonstrates that it outperformed the previous technique in terms of identifying RCT failure. The relationships between all significant clinical and non-clinical features of the root canal treatment are another benefit of using this technique.

2. Literature Review

The research by Su-Jin Jeon et al. [8] made use of both panoramic images and CBCT scans. The use of a convolutional neural network (CNN) provides a 95% accuracy rate in indicating the existence of a canal with a C-shape. S.R. Herbst et al. [9] aimed to see if there was a correlation between preoperative risk factors and root filling length (RFL) success rates during orthograde root canal surgeries (RCT). To achieve this aim, gradient boosting machine (GBM), extreme gradient boosting (XGB), logistic regression (logR), random forest (RF), support vector machine (SVM), and decision tree (DT) models were employed to produce predictions about root filling lengths in this investigation. In addition, Lee et al. [10] used deep convolutional neural networks (CNNs), which have an accuracy rate of 88.0% when it comes to detecting and diagnosing dental caries. Dipit Vasdev et al. [11] used a deep neural network (DNN) technique to distinguish between good and unhealthy periapical dental X-ray images with an accuracy rate of 85%. Biomechanical root canal preparation is important for [12,13] RCT effectiveness. Improper preparation might cause the failure of the RTC treatment. Thus, keeping in this mind, Vinod Thakur et al. [14] used ensemble machine learning to estimate the apical extension’s dimensions throughout biomechanical root canal preparation. The present study used ensemble bagged, boosted, and RUSboosted tree classifiers, with ensemble bagged trees having the highest accuracy rate of 94.2%. Furthermore, using the FMEA technique, Mohsen Yazdanian et al. [15] qualitatively and quantitatively assessed the failure modes and their effects. The RCT process contains 19 steps, and the FMEA identified 48 potential failures. File fracture in the root canal (RPN = 324) and apical extrusion of the irrigating solution (RPN = 320) were attributed to the highest failure causes during the RCT procedure. According to the research conducted by Qu et al. [16], the GBM model performs a little better than the RF model when it comes to predicting the outcome of endodontic microsurgery. The investigation in this study was conducted in a carefully monitored laboratory setting, and data from 178 people were used. Additionally, several criteria were taken into account for endodontic microsurgery prognosis prediction, such as lesion size, tooth type, bone defect type, root filling length, root filling density, apical extension of post, age, and sex. M Mustafa et al. [16] examined patients in the Saudi Arabian city of Al-Kharj. In light of this fact [17], the approach that has been presented makes use of machine learning in order to identify the factors that are primarily [18] responsible for the failure of root canal therapy, such as a broken instrument, an overfilled cavity, a perforated root, or an underfilled cavity [4,6,7]. This paper also provides insights into the importance of [19,20,21] these variables for determining the tooth’s survival time after root canal therapy (treatment longevity detection), which is not revealed in many studies.

3. Methodology

The main goal of the system is to find the clinical or non-clinical causes of root canal treatment not working. Consequently, this method uses various machine learning techniques, including SVM, NB, and LR, to detect endodontic problems in RCT, using datasets collected from the studied healthcare institution. This system makes it easier for people [21] who have had root canals to live a better life after treatment. Figure 1 shows the proposed system block diagram for longevity recommendation.

Data acquisition: This study used a dataset containing 332 instances of root canal therapy. This information was used as an input to the system in order to identify the cause and factors involved in root canal treatment failure [22]. The dataset includes various variables or features for each instance, and here are some specific details about these variables and their significance in terms of predicting treatment failure:

Patient Information: This factor could include [23] demographic details, such as age, gender, and possibly medical history. Patient characteristics can influence the success or failure of RCT, as some individuals may be more prone to dental issues.

Tooth Details: information about the specific tooth undergoing RCT, including its location in the mouth (e.g., incisor, molar), type (e.g., premolar, wisdom tooth), and, possibly, its condition prior to treatment (e.g., the level of decay or infection) [24].

Treatment History: This factor could encompass any previous dental treatments or RCT procedures performed on the same tooth. A tooth with a history of multiple RCTs may be more likely to experience treatment failure.

Procedural Details: Information about the RCT procedure itself, such as the technique used, the number of canals treated, and any complications encountered during the procedure. The complexity of the procedure can affect its success.

Complications: Details about any complications or issues arising after RCT, such as broken instruments, periapical radiolucency, root fractures, vertical root fracture stones, overfilled cavities, perforations, or underfilled cavities. These complications are key indicators of treatment failure.

Periodontal Support: Information about the level of periodontal support for the tooth. Adequate periodontal support is vital for the long-term success of RCT.

Outcome/Longevity: The primary outcome variable indicating whether RCT was successful or not, and if not, the duration of its effectiveness before failure. This variable is critical for training predictive models. The significance of these variables lies in their ability to capture various aspects of RCT procedures, patients’ characteristics, and post-treatment complications.

By analyzing these features, this study aimed to identify patterns and correlations that could help to predict the longevity of RCTs and recommend measures to improve treatment outcomes. Researchers likely collected data on these variables from a sample of RCT cases to build a predictive model, and they likely used machine learning techniques, such as support vector machine (SVM) technology, to analyze the dataset and make recommendations for RCT longevity based on these factors.

Pre-processing: The data collected from the healthcare institution contained a lot of noise. Thus, the raw data were preprocessed to remove noise and other unwanted features before the final analysis can begin [25,26].

Feature Extraction: In our quest to maximize the predictive capability of the model for recommending the longevity of root canal treatment, this process entailed the strategic selection or creation of the most pertinent variables (features) from the dataset. This multifaceted procedure began with data collection and preprocessing, where a comprehensive dataset encompassing clinical and non-clinical factors, such as patient demographics, dental history, treatment specifics, and outcomes, was compiled and meticulously cleaned. Subsequently, the pivotal step of feature selection unfolded, with an emphasis on singling out the variables that wield significant influence over the treatment’s durability. Various methodologies came into play, including correlation analysis to gauge the strength of the relationship between each feature and the target variable (longevity of treatment). In tandem, feature importance was assessed through tree-based models, like random forest, illuminating the features that effectively reduce decision tree impurity. Moreover, domain expertise from dental professionals could inform this selection process. For complex datasets, dimensionality reduction techniques, such as principal component analysis (PCA), could be employed to streamline features while retaining salient information. Feature engineering may also come into play, allowing for the creation of new features capturing crucial relationships or interactions. Validation, achieved via techniques like cross-validation, ensured that the chosen features consistently bolster the model’s performance. Ultimately, the SVM model was constructed, integrating the selected features as inputs to predict root canal treatment longevity. This meticulous feature selection and extraction process served to augment the SVM model’s predictive accuracy and yield critical insights into the variables most instrumental in enhancing dental procedures and patient outcomes. Different clinical and non-clinical factors are capable of causing root canal failure. From the given input data, features were extracted. The system employed a total of 23 features [4,5,7,11,12,13] (clinical and non-clinical factors) as listed in

Table 1. Clinical and non-clinical factors.

Clinical Factors	Non-Clinical Factors
Periapical radiolucency	Age
Root fracture	Eating habits
Broken tooth	Uncooperative behavior
Curved root canal	Bad habits
Non-restorable tooth	Demographic area
Broken instruments	Education of patient
Untreated root canals	Chewing habits
Unfilled and missed canals	Longevity
Under filled canals
Insufficient periodontal support
Pulp stones
Vertical fracture in root
Tooth location
Root resorption
Poor oral hygiene

.

Predict the longevity of the treatment: The relationship between all of the essential clinical and non-clinical elements that define the success of the root canal therapy is defined as the longevity of the treatment. Also, the system can predict for how long the treatment will be successful in the future.

Building Machine Learning Models

Following feature extraction and normalization, the data pertaining to root canal therapy were used to train machine learning models like SVM, NB, and LR. The factors of the ideal RCT [21,22,23] or its failures can then be identified by comparing the test data to the training data system. In this scenario, we employed 10-fold cross-validation to ensure that our fitting technique was accurate. The training set had 90% of the total data, and the other 10% was used as test data. Thus, the system was able to identify potential causes of treatment failure. SVM, a versatile machine learning algorithm, offers customizable performance optimization through parameter adjustments.

Key considerations in configuring an SVM model include the choice of kernel function, such as the linear kernel for linearly separable data or the RBF kernel, a versatile option used for capturing intricate relationships within complex datasets, like those shown in [24,25].

4. Result and Discussion

4.1. Data Collection

The data were collected from clinic related to the root canal treatment. Figure 2 represents the overall data distribution and flow of building machine learning models. Overall, 332 samples are used for this study. This study makes use of total 27 attributes, such as clinical and non-clinical parameters, in order to find the cause of root canal treatment failure.

4.2. Statistical Analysis

It was also found that all 27 attributes are more or less responsible for treatment failure. The factors like periapical abscess, underfilled canals, broken instruments, age, a non-restorable tooth, and demographic area might be the main causes, whereas a curved root canal, insufficient periodontal support, and overfilled canals are less responsible, as shown in Figure 3.

4.3. Implementation Details

Support vector machines (SVMs), naïve bias, and simple logistic regression are evaluated using a manual dataset as follows:

Class 0—low class—attributes that are more likely to lead to failure.

Class 1—high class—attributes that are less likely to lead to failure.

4.4. Support Vector Machine (SVM) Classification

The effectiveness of the system was measured with the help of the WEKA (Waikato Environment for Knowledge Analysis) tool. WEKA solves practical data mining challenges using machine learning methods. As the data are linearly separable, a linear kernel is used.

The dataset is then imported into the WEKA tool, where the support vector machine classification method is executed, and it is represented in the classification and confusion matrixes in

Table 2. Classification report based on an SVM.

Class	0	1
0	95	20
1	8	209

and

Table 3. Confusion matrix based on an SVM.

	Class 0	Class 1
Accuracy	91.57%	91.57%
Precision	92.23%	91.27%
Sensitivity	82.61%	96.31%
Specificity	96.31%	82.61%

, and the results are analyzed.

After delivering the dataset as an input to the WEKA tool, it generates a classification 2 report that categorizes the data according to classes shown in the Table 2. The graphic depiction of the data in Table 3 can be found in Figure 4, which demonstrates an accuracy rate of 91.57%.

4.5. Simple Logistic Regression (LR) Classification

The use of simple logistic regression on the given dataset gives the classification of data, as shown in

Table 4. Classification report based on LR.

Class	0	1
0	105	10
1	15	202

.

Thus, the confusion matrix based on the classification report is shown in

Table 5. Confusion matrix based on LR.

	Class 0	Class 1
Accuracy	92.47%	92.47%
Precision	87.50%	95.28%
Sensitivity	91.30%	93.09%
Specificity	93.09%	91.30%

. Figure 5 provides a visual representation of Table 5. It displays an accuracy rate of 92.47%.

4.6. Naïve Bayes Classification

The use of naïve Bayes classifier for the given dataset gives the classification of data, as shown in

Table 6. Classification report based on naïve Bayes.

Class	0	1
0	74	41
1	28	189

.

The Confusion matrix presented in

Table 7. Confusion matrix based on naïve Bayes.

	Class 0	Class 1
Accuracy	79.22%	79.22%
Precision	72.55%	82.17%
Sensitivity	64.35%	87.10%
Specificity	87.10%	64.35%

is result of applying naïve Bayes algorithm. Figure 6 represents a bar chart of performance metrics, with an accuracy rate of 79.22% for root canal detection using naïve Bayes.

Based on these results, it is clear that machine learning models like as logistic regression, naïve Bayes, and the SVM can significantly help to identify the root cause of root canal failure, with accuracy rates of 92.47%, 79.5%, and 91.7%, respectively.

5. Accuracy of Machine Learning Algorithms

So, it has been pointed out that among the three machine learning models, logistic regression outperforms its rivals, as it has the highest accuracy rate of 92.47% compared to the others represented in Figure 7.

6. Conclusions

This study uses three different machine learning models, namely logistic regression, naïve Bayes, and SVM, in order to shed light on the different factors that can cause the root canal treatment to fail. Some of these factors include instrument fracture, overfilling, perforation, root respiration, and underfilling. It has also been shown that, in comparison to other methods, logistic regression has a higher level of accuracy, i.e., 92.47%. Then, we used a support vector machine (SVM) with an accuracy rate of 91.7% to forecast the effectiveness of the treatment over time. Finding the longevity of the root canal treatment enables the doctors to correct the errors in the treatment, if any exist, in real time, which further improves the quality of service experienced by the patient. Despite the fact that logistic regression provides a higher degree of accuracy (92.47%), building the model required more time. The flawed input data also had a negative impact on the model’s performance. Furthermore, the dataset used had very little data, which reduced the prediction’s accuracy. In order to obtain more precise results, the data size must be enhanced in the future. Taking into account the benefits of deep learning, this system will benefit from the use of a deep learning algorithm combined with X-ray image datasets in order to increase the accuracy of root canal failure detection systems.