Enhancing prediction of tooth caries using significant features and multi-model classifier

Background

Maintaining good oral health is crucial for overall well-being and enhancing one’s quality of life. Oral health encompasses the absence of throat cancer, mouth infections, sores, tooth decay, dental loss, gum problems, and related issues leading to impede activities such as chewing, speaking, smiling, biting, and overall psychosocial well-being (Samim et al., 2016). Dental caries, a common oral health problem, is caused by the interaction of oral bacteria and dietary sugars. Bacteria break down sugars into acids, which erode tooth enamel over time (Strużycka, 2014). Repeated exposure to these acids can lead to demineralization, which causes cavities, or localized damage to tooth structure. Dental caries is a chronic disease that affects the hard tissues of the teeth, including enamel, cementum, and dentin. It is caused by the interaction of oral bacteria and dietary sugars. The bacteria produce acids that demineralize the tooth enamel, leading to the formation of cavities (Kudkuli et al., 2020). Over time, the caries can progress inward, destroying the dentin and cementum.

Almost half of the global population is affected by oral diseases in one way or the other, with over 2.2 billion individuals worldwide having dental caries of a permanent nature (James et al., 2018). Dental caries, also known as tooth decay, is a condition caused by the interaction of oral bacteria and dietary sugars. Oral bacteria produce acids like lactic acid by breaking down sugars, which directly impact the tooth’s enamel layer. This acid erosion can gradually create gaps or cavities between teeth. If untreated, these cavities can cause pain, infections, and even lead to tooth loss (Pihlstrom & Tabak, 2005; Selwitz, Ismail & Pitts, 2007; Ogden et al., 2002). Dental imaging techniques can help to prevent the progression of caries lesions to more serious stages (Gomez, 2015; Metzger et al., 2022; Michou et al., 2022). Early-stage caries, also known as incipient caries, is the first stage of tooth decay. It is caused by acids produced by bacteria in the mouth when they break down sugars. These acids can demineralize the enamel, the hard outer layer of the tooth. The outer tooth layer may still be intact at this stage, making the damage difficult to see, even with some imaging techniques. The identification and description of specific areas of a tooth are facilitated by the use of named surfaces and directions that are designated according to their location. Smooth surfaces are the flat areas of a tooth’s crown, occlusal surfaces are the chewing surfaces of posterior teeth, and proximal surfaces are the sides of teeth adjacent to neighboring teeth. Each of these surface types has specific characteristics and considerations in terms of dental care and hygiene. Proximal surfaces, which are the sides of teeth adjacent to neighboring teeth, can be particularly vulnerable to caries. However, early-stage caries is a crucial time for preventive action, as it is the most easily reversible stage of tooth decay. The accuracy of diagnosis relies on the skill of the dentists and the quality of the diagnostic equipment employed (Topping & Pitts, 2009).

While dental caries are preventable and treatable, they often correlate with tooth loss and discomfort, particularly in advanced stages. Effective and timely treatment hinges on the early identification of these caries. Visual examination and probing of visible dental cavities are carried out using dental probes and handheld mirrors. These conventional techniques are valuable in identifying easily accessible yet partially concealed caries (Fejerskov, Nyvad & Kidd, 2015). Kielbassa et al. (2006) examined the existence of cavities within proximal caries lesions using varying levels of magnification. However, for hidden or inaccessible lesions, X-ray radiography plays an irreplaceable role in diagnosis. Periapical, bitewing, and panoramic X-rays are commonly utilized radiographs in clinical dental practice. Periapical and bitewing X-rays detail a specific oral area by focusing on it, while panoramic X-rays capture the entire maxillofacial region (Kaur et al., 2017). Bitewing radiography is commonly used to detect caries lesions and their depths with its high sensitivity and specificity (Schwendicke, Tzschoppe & Paris, 2015). However, it cannot provide a complete assessment of all mouth lesions in a single attempt. On the other hand, Panoramic imaging provides a single, two-dimensional image of the entire upper and lower jaws, as well as the surrounding structures, such as the teeth, gums, and bone. This allows dentists to visualize the entire oral cavity in one view, which can be helpful for diagnosing and planning treatment for a variety of dental conditions (Rushton, Horner & Worthington, 1999). Due to its cost and efficacy, panoramic imaging is a widely used and accepted radiological tool for diagnosis, clinical dental disease screening, and treatment evaluation.

Various methods have been explored for the detection of dental caries in previous studies. These include the trans-illumination method (Datta, Chaki & Modak, 2019), the international caries detection and assessment system (Sinton et al., 1997), calibrated diaphragm computed tomography (Kaur et al., 2017; Schwendicke, Tzschoppe & Paris, 2015), and quantitative light-induced fluorescence (Zantner, Martus & Kielbassa, 2006). Additionally, research has been conducted on panoramic radiology procedures to improve their performance for dental caries diagnosis through image processing techniques (Abreu, Tyndall & Ludlow, 2001). Computer-aided caries detection using X-rays has also been investigated (Neuhaus et al., 2015; Alammari et al., 2013; Tracy et al., 2011). In one study, Alammari et al. (2013) utilized the Logicon Caries Detector (LCD) with modified computer-aided design software, and the effectiveness of density analysis was confirmed as an auxiliary tool to assist dentists in detecting and organizing caries based on user feedback. Advanced dentinal diagnostic tools are utilized to detect caries, enabling dentists to diagnose dentinal cavities at a significantly earlier stage compared to traditional and conventional approaches.

In recent decades, there has been a growing interest among scientists in utilizing machine learning techniques for the detection of dental diseases. Traditionally, the evaluation of radiographs and lesion detection is performed physically and subjectively by operators or experts. However, when dealing with large volumes of image data, this task can become tedious and prone to misinterpretations. Computer-aided diagnostics (CAD) using machine learning gaining the attention of many health professionals for the rapid and precise detection of dental diseases. This study investigated the use of CAD systems for the detection of dental disease. The study made the following contributions.

• The study leverages a more robust feature-based dataset, addressing the limitations of previous research on dental caries detection.

• The feature extraction process is enhanced by using principal component analysis (PCA) and chi-square ( $X{}^2$) to improve data insights.

• Introducing a hybrid model combining PCA features and ensemble classifiers including Extreme Gradient Boosting (XGB), Random Forest (RF), and Extra Trees Classifier (ETC).

• A voting classifier using PCA features outperformed other methods and shows potential in improving early diagnosis and treatment outcomes for effective dental caries detection.

The rest of this article is arranged in the following order. The Related Work provides a review of the relevant literature on the current study. The Method and Materials describes the dataset, methodology, and machine learning models used in this study. The findings of the article are presented in the Results and discussed in the Discussion section. Finally, the Conclusion concludes the study and outlines future research directions.

Related work

The combination of data mining and machine learning represents a powerful tool for addressing various challenges. Analyzing potentially large medical data manually is particularly difficult because of its extensive feature vector. Machine learning has emerged as a crucial technique in numerous application domains, including healthcare (Ahmad, Eckert & Teredesai, 2018). It has demonstrated its significance by offering precise and accurate systems for applications related to medicine, even during sensitive data handling in the medical domain (Kaissis et al., 2020). Similarly, ML models have been successfully utilized in recognizing early-stage hazards associated with conditions such as dental caries.

In a study conducted by Kang, Njimbouom & Kim (2023), a dental caries detection model was proposed. This approach involved integrating the minimum Redundancy Maximum Relevance (mRMR) and GINI index (GINI) algorithms with the Gradient Boosting Decision Tree (GBDT) classifier. By utilizing only a few clinical test features, this method aimed to save time and cost during dental caries screening. The planned method was matched to other recently recommended dental processes. Among the different classifiers, the highest classification performance was demonstrated by GBDT having a reduced feature set. It achieved precision, impressive accuracy, F1-score, and recall values of 99%, 95%, 93%, and 88% respectively. In their research, Thanh et al. (2022) put forward a Deep Learning-based system for detecting dental caries by using photos taken by a smartphone from inside the mouth. They employed four deep learning models, namely RetinaNet, YOLOv3, Faster Region Convolutional Neural Network (R-CNNs), and Single-Shot Multi-Box Detector (SSD), to identify early-stage cavities and caries lesions. This study indicated that YOLOv3 achieved the highest sensitivity value, specifically 87.4%. This suggests that YOLOv3 exhibited strong performance in accurately detecting dental caries lesions in the intraoral photos captured by smartphones.

Lian et al. (2021) offered a system based on deep learning for the detection and grouping of caries. The dataset in the study was self-annotated. The authors applied a convolutional neural network (nnU-Net) to detect caries lesions and DenseNet121 to classify the lesions based on their depths. The performance of DenseNet121 and nnU-Net models was compared with outcomes from six dentists on the test dataset using various evaluation metrics, including Dice coefficient, intersection over union (IoU), recall, accuracy, negative predictive value (NPV), precision, and F1-score. This showed that nnU-Net achieved caries lesion segmentation Dice coefficient and IoU values of 0.663 and 0.785 respectively. The recall and recall rates of nnU-Net were 0.821 and 0.986, respectively. This shows the effectiveness of the nnU-Net model in accurately segmenting caries lesions. Oztekin et al. (2023) utilized panoramic radiograph images for dental caries prediction. They employed deep learning models, specifically DenseNet-121, EfficientNet-B0, and ResNet-50. The study’s results revealed that the deep learning model ResNet-50 achieved an accuracy score of 92% in predicting dental caries.

AL-Ghamdi et al. (2022) proposed a neural search architecture network (NASNet) for tooth caries detection. They also equated the performance of NASNet against AlexNet and CNN prototypes. The results of the study indicated that the proposed NASNet model outperformed the rest of the deep learning models with respect to accuracy. It detected tooth caries with a 96.51% accuracy score, which is by far very impressive. Jader et al. (2018) steered a study using 1500 panoramic X-ray radiographs to develop a tool profile using a convolutional neural network based on mask region. The evaluation of the tool was based on precision, accuracy, recall, F1-score, and specificity as outcome metrics. The results of the investigation showed impressive performance metrics, with recall, F1-score, accuracy, precision, and specificity values of 0.84, 0.88, 0.98, 0.94, and 0.99, respectively. Kielbassa, Mueller & Gernhardt (2009) suggested that combining minimally invasive restoration with extensive remineralization can help prevent tooth decay, reduce the need for future fillings, and save money. This approach is consistent with the principles of minimal intervention dentistry. Similar to dentinal caries that have recently crossed the enamel-dentin border, such lesions necessitate specific treatment decisions. They may require either conventional restorations or a notably less invasive treatment approach, as highlighted in Kielbassa et al. (2020).

Muramatsu et al. (2021) utilized a fourfold cross-validation technique with a dataset having 100 dental radiographs of a panoramic nature to develop an object detection network. The network’s performance was evaluated based on sensitivity and accuracy for tooth detection. The study reported a sensitivity of 96.4% and an accuracy of 93.2% for tooth detection. Raith et al. (2017) employed a convolutional neural network (CNN) architecture along with the PyBrain package to identify teeth. They achieved a performance score of 0.93, indicating a high level of accuracy in accurately identifying teeth using their proposed approach. Kühnisch et al. (2022) utilized a model based on deep learning, specifically a convolutional neural network (CNN), for revealing and classification of caries. They compared the diagnostic performance of their model with expert standards in various scenarios. The outcome of the study suggested that the CNN prototype achieved an accuracy score of 92.5%.

The complete summary of the related work is shown in Table 1.

Materials and Methods

This section explains the methodology and procedures employed in this study in detail. Figure 1 depicts the architecture of the suggested technique. Starting with data retrieval, the approach follows the generation of text by Synthetic Minority Over-sampling Technique (SMOTE) to balance the dataset. Feature extraction is then carried out that involves term frequency (TF), term frequency-inverse document frequency (TF-IDF), and CNN. The data is split for training and testing where the selected machine learning models are utilized for sentiment classification.

Proposed approach

This study aimed to introduce an ML-based detection approach utilizing classifiers for more accurate prediction of tooth caries. Figure 2 depicts the design of the tooth caries detection workflow.

In this approach, the very crucial and prominent step is the feature selection. PCA and $X{}^2$, the two unique techniques used to select the features, are employed for feature fusion. Following this, the dataset is divided into a testing ratio of 30% and a training sets ratio of 70%. For classification purposes, an ensemble model is utilized, which combines the XGB, RF, and ETC classifiers using the soft voting criterion. Recall, accuracy, F1 score, and precision are the performance evaluation metrics of any given model or approach.

Proposed model XGB+RF+ETC

To evaluate tooth caries detection, this study conducted experiments in three different scenarios: (i) using all features, (ii) using $X{}^2$-selected features, and (iii) using PCA-selected features. The dataset was then split into a 70:30 ratio, with 70% allocated to training the models and the rest of the 30% for prototype testing purposes. For the proposed tooth caries detection system, an ensemble approach called XGB+RF+ETC was utilized. Ensemble models are a powerful technique that combines predictions from multiple models to improve accuracy and robustness. Each model within the ensemble has its own strengths and weaknesses, and by combining them, the overall performance is often enhanced. In the case of tooth caries detection, this study suggests employing an ensemble learning model that integrates three popular algorithms: XGB (eXtreme Gradient Boosting), RF (Random Forest), and ETC (Extra Trees Classifier). By leveraging the strengths of these individual models and their ability to handle different aspects of the data, the ensemble approach aims to achieve better tooth caries detection results. The architecture of the proposed voting system is shown in Fig. 3.

The ensemble model has been created by combining the predictions of three different ML algorithms: XGB, RF, and ETC. The general process for constructing an ensemble model involves training multiple models on the same dataset and then merging their predictions. In the case of the XGB+RF+ETC ensemble model, each algorithm is trained individually on the dataset. During training, each model generates predicted probabilities for the different classes of the target variable. These predicted probabilities represent the likelihood of each class for a given observation. To create the final prediction for each observation, the predicted probabilities from each model are combined. One common approach is to calculate a weighted average of the predicted probabilities, where the weights assigned to each model are determined based on their performance on a validation set. Models that demonstrate better predictive performance on the validation set are assigned higher weights in the ensemble combination. By training multiple models and combining their predictions, the ensemble model aims to improve overall performance and reduce the risk of overfitting. This approach leverages the strengths of each individual model and can lead to more accurate and robust predictions for tooth caries detection. The algorithm below details the working of the recommended ensemble model, which can be expressed as:

(1) $$\widehat{p}=argmax\left\{{\displaystyle \sum _i^{n}X}G{B}_i,{\displaystyle \sum _i^{n}R}{F}_i,{\displaystyle \sum _i^{n}E}T{C}_i\right\}.$$

In the case of the ${\sum }_i^{n}XG{B}_i$, ${\sum }_i^{n}R{F}_i$, and ${\sum }_i^{n}ET{C}_i$ models, each of them produces prediction probabilities for each test sample. These probabilities represent the likelihood of each class for a given test case. To make a final prediction, the ensemble model utilizes the soft voting criterion. The soft voting criterion involves aggregating the probabilities generated by all three models. This aggregation process combines the predictions from XGB, RF, and ETC to arrive at a consolidated set of probabilities. Once the probabilities from all three models are combined, a final prediction is made based on these combined probabilities. For a test sample, the class having the highest probability is taken as the predicted class. By employing the soft voting criterion and combining the predictions from multiple models, the ensemble model benefits from the collective insights and strengths of XGB, RF, and ETC. This approach enhances the robustness and accuracy of the final prediction for each test case in the tooth caries detection task.

In the ensemble model, the final class prediction is determined by selecting the class with the highest average probability score among the combined predictions of the classifiers. The individual classifiers in the ensemble generate probabilities for each class in the target variable. These probabilities are combined, typically through averaging, to obtain a single probability for each class across all the classifiers. After combining the probabilities, the class having the highest average probability score is selected as the final forecast.

Results

The results and outcomes of the tooth caries prediction are presented in this section. The machine learning models were implemented using Python 3.8 (Python Software Foundation, Wilmington, DE, USA) and executed within a Jupyter Notebook environment (Project Jupyter, Berkeley, CA, USA). The sci-kit learn and TensorFlow libraries (TensorFlow; Google LLC, Mountain View, CA, USA) were utilized for model development and evaluation. The experiments were conducted on a system running 64-bit Windows 10 (Microsoft Corporation, Redmond, WA, USA), featuring a 7th-generation Core i7 processor (Intel Corporation, Santa Clara, CA, USA) with a clock speed of 2.8 GHz. This information provides context regarding the hardware specifications used for the experiments. To assess the performance of the models, various evaluation metrics were employed. These include accuracy, precision, recall (sensitivity), and F1 score. To evaluate tooth caries detection, this study conducted experiments in three different scenarios: (i) using original features, (ii) using $X{}^2$ features, and (iii) using PCA-selected features.

Comparison of machine learning models for all experiments

To evaluate the efficacy of the suggested system, we performed a comparative evaluation of the performance of various machine-learning models across a range of experiments. The findings revealed a noteworthy improvement in the machine learning models’ performance when utilizing the PCA features in the third experiment. Table 6 offers a comprehensive summary of the achieved outcomes by the ML models in all scenarios, facilitating a comprehensive assessment of their performance. This comparison provides a distinct understanding of the influence and advantages of incorporating PCA features in enhancing the predictive abilities of the models. The comparison of all three types of features results is shown in Fig. 4.

Table 6:

Accuracy comparison of the machine learning models.

Model	With all features	With X2 features	With PCA features
LR	85.77	87.22	90.24
DT	87.24	88.13	92.35
RF	90.65	89.21	94.28
SGD	88.59	89.29	91.54
ETC	90.08	89.24	95.34
XGB	90.51	90.25	94.08
SVC	89.45	85.34	93.86
GNB	82.38	88.37	90.81
VC(XGB+RF+ETC)	92.03	91.52	97.36

DOI: 10.7717/peerj-cs.1631/table-6

Discussion

Prevention and early detection play pivotal roles in dental caries management. Upholding effective oral hygiene practices, such as routine brushing with fluoride toothpaste, flossing, and utilizing antimicrobial mouthwashes, aids in plaque removal and mitigates cavity risk. Moreover, adopting a balanced diet low in sugary substances minimizes teeth’s exposure to acid-producing bacteria. Regular dental check-ups and professional cleanings are imperative for promptly identifying and addressing dental caries. Dentists are skilled at spotting early decay signs, delivering necessary interventions like fillings and offering guidance on oral hygiene practices and dietary adjustments. This research work offers a comprehensive framework for the automated detection of tooth caries. The results demonstrate the effectiveness of machine learning models for tooth caries prediction. The VC(XGB+RF+ETC) ensemble model consistently delivers high accuracy and other evaluation metrics across different feature sets. The performance improvement with PCA-selected features highlights the benefits of feature dimension reduction in enhancing model efficiency without compromising accuracy. These findings have potential implications for dental healthcare, suggesting that predictive models can play a crucial role in identifying potential tooth caries cases early, thereby facilitating timely intervention and treatment. However, further validation and testing on diverse datasets would be essential to ensure the generalizability of these models in real-world scenarios. Future work in this research will focus on combining machine and deep learning models as an ensemble to improve accuracy. Additionally, transfer learning models will be used to augment data and improve training. To further validate the results, we performed k-fold cross-validation and compared the results to state-of-the-art approaches.

Conclusions

This research advances the detection of dental caries by leveraging the potential of ML and CAD. Traditional methods, such as transillumination and quantitative light-induced fluorescence, are valuable but often suffer from subjectivity and manual labor. In response, this study introduces a novel framework that combines a powerful ensemble of machine learning models, including extreme gradient boosting, random forest, and extra-tree classifier, with features extracted through PCA. This work makes three key contributions: first, it presents a robust framework tailored for the accurate and efficient detection of tooth caries. Second, it comprehensively explores various machine learning models, revealing the depth of their capabilities. Third, it emphasizes the importance of feature engineering by comparing PCA and $X{}^2$ extraction methods. The ensemble classifier driven by PCA features achieves outstanding results in terms of accuracy, precision, recall, and F1 score reaching 97.36%, 96.14%, 96.84%, and 96.65%, respectively. This achievement not only highlights the model’s potential but also paves the way for a technology-driven transformation in healthcare. This research has the potential to revolutionize the way dental caries are detected. By reducing subjectivity, improving accuracy, and expediting diagnoses, it can help to improve oral health for people all over the world.