Evaluation of Hand-Scaling Skills of Dental Hygienist Students: Identification of Contact Between Hand-Scaler Blade Tip and Tooth Surface

Dental hygienist students require a self-learning simulator to learn the correctness of hand-scaling techniques. An essential technique in hand scaling is the maintenance of contact between the tip of the hand-scaler blade and the tooth. However, imaging-based methods cannot effectively reveal this contact because the gingiva and buccal mucosa conceal the blade. Therefore, this study aimed to propose a method to identify the appropriate contact state of the blade with the tooth by using an inertial measurement unit (IMU) attached to the hand scaler and a force sensor attached to the target tooth. The hand-scaling motion was measured in an experiment in which participants were instructed to use the tip or middle of the blade to contact the tooth. The contact state of the blade, whether it was the tip or the middle, was identified using 18 features, including the average and standard deviation of nine dimensions of force, acceleration, and angular velocity with a support vector machine (SVM). The results showed that the model using all 18 features could classify the contact state with an accuracy of 97.1%. Furthermore, the accuracy was 95.9% with the 12 features from IMU alone, which was not significantly different from the accuracy with 18 features. The accuracy was 88.8% with six features from the force sensor alone. These results indicate that the IMU alone can identify the correct contact state, highlighting the possibility of creating a realistic simulator for training dental hygienists in evaluating the blade-contact state.


I. INTRODUCTION A. BACKGROUND
Hand-scaling is the process of removing tartar and other dirt on tooth surfaces by mechanical manipulation using the tip The associate editor coordinating the review of this manuscript and approving it for publication was Carmen C. Y. Poon . of a hand-scaler blade (hereafter referred to as tip) [1]. It is one of the most important tasks for dental hygienists and is difficult to master because it requires precise tip operation by moving the hand scaler to remove the firmly adhered deposits on the tooth surface. An overview of the hand-scaling motion is presented below.  • The hand scaler is grasped like a writing instrument, as shown in Fig. 1. The ring finger of the hand holding the hand scaler is placed on an adjacent tooth and used as a fulcrum.
• A 1-2-mm tip of the hand-scaler blade is used, as shown in Fig. 2. The tip is kept in contact with the tooth surface, and up and down strokes of 2 to 3 mm are applied along the tooth surface, as shown in Fig. 3.
• Fig. 4 shows the motion of forearm rotation, which is the basic operation of the hand scaler. Forearm pronation and supination correspond to the motion used when lowering the blade to the stroke start position and when the blade is raised to the stroke end position, respectively. Stable and repetitive motions are ideal.
Dental hygienist training institutes spend considerable time on training dental hygienist students to acquire the correct hand-scaling technique. Through repetitive training using an oral cavity model, students practice this technique, which is difficult to master. Although one-on-one guidance is the most effective approach because some aspects of the technique are difficult to convey verbally, teachers may be unable to spare time to attend to all students' practices. Furthermore, oral cavity models have no function to automatically identify the correctness of hand scaling when students are selflearning. Without the ability to judge the correctness of their own techniques, students are unable to practice efficiently. Thus, while learning hand scaling, practice without constant feedback regarding the technique is inherently inefficient.
A simulator that also evaluates the student's skill can offer significant advantages for students to improve their technique. Simulation has been widely used in the education of medical professionals [2]. Simulation practice is  considered particularly important in the acquisition of surgical skills because it allows sufficient practice before students are allowed to treat patients [3], [4]. Consequently, students can learn from their mistakes without the fear of harming the patient [5]. Thus, simulation education can serve as an efficient and ethical method for training students to deliver safe medical care to patients [6], [7]. The scope of simulation training has been expanding and is now also being used to evaluate skills [8] because it allows detailed feedback and objective evaluation [9]. Therefore, we focused on a function implementable with an actual self-learning simulator that provides feedback on the correctness of student's hand-scaling motions and thereby improves their technique. Among the important aspects of the safety and efficiency of hand scaling, the use of the tip is particularly important, as shown in Fig. 2 [1]. The tip may damage gingiva when it moves away from the tooth surface [10], and therefore, the use of the middle of the handscaler blade (hereafter referred to as middle), as shown in Fig. 2, is incorrect and a typical example of a dangerous action performed by students. Teachers should guide students to use the tip instead of the middle when the blade is in contact with the tooth surface. However, beginner students experience difficulty in evaluating this point while focusing on other aspects of the technique. Thus, automatic evaluation and feedback on this aspect could greatly facilitate handscaling practice.

B. RELATED RESEARCH
Previous research has included training systems using virtual reality (VR) [11] and imaging technologies [12]. In research using VR technology, the purpose is to build a simulator for tooth scaling training using VR and evaluate its VOLUME 10, 2022 training effect. Hand scaling can be performed on virtual intraoral teeth displayed on a head-mounted display (HMD). It includes a motion input part that imitates the handle of a hand scaler, allowing the practice of hand scaling motions. The HMD also displays the magnitude and direction of the force exerted by the tip of the hand scaler on the teeth, as well as the appropriate force magnitude and direction. Prior research examined the effects of training using this simulator based on the subjective evaluations of experienced participants and the number of successful cases of artificial tartar removal before and after training. Comparing the group using this simulator with the group that practiced removing the paint on a painted tooth model, which is the conventional practice method, showed that this simulator was more effective.
Research using imaging technology, the camera captures the hand-scaling motion using an AR marker attached to the rear end of the hand-scaler grip and a mannequin that simulates the head with the model of the oral cavity. The system estimates the motion of the tip from the image information of the AR marker attached to the rear end of the handscaler grip. A virtual hand scaler is displayed on top of the camera image of the mannequin. For example, the back teeth are concealed by the mannequin's cheek and not visible on the camera image. During hand scaling of the back teeth, the tip of the virtual hand scaler is to be displayed on the mannequin's cheek. Because the tip may be concealed by the gingiva and buccal mucosa during the operation, as shown in Fig. 5, a camera may not adequately visualize the contact or the absence of contact between the tip and the tooth surface.

C. LIMITATIONS OF EXISTING TECHNOLOGIES
The correct motion pattern is created relatively easily because haptic feedback from the hand scaler handle and visual information are uniquely determined. However, creating realistic and practical feedback for incorrect motion patterns is difficult because their motion patterns are myriad.
We considered the possibility of extending the camerabased system using imaging technology. In this approach, precise information on the hand scaler and oral cavity model would be required to identify the state of contact. Although such information can be obtained by acquiring the shape in advance [13], [14] or by using 3D reconstruction from images [15], [16], both methods have limitations. Prepreparation is not realistic because the shapes of oral cavity models and blades often change from the pre-modeling shapes by wear and deformation caused by practice and aging. On the other hand, with 3D reconstruction methods, it is difficult to achieve the required accuracy. The precision of the 3D reconstruction is less practical (0.5 mm) [12]. Blind spots on cameras have also become a major problem in the implementation phase.

D. PURPOSE
Therefore, we considered the use of an inertial measurement unit (IMU) as an alternative approach for contact identification. IMUs have been widely used in skill assessment in the field of sports [17], [18], [19] as well as the medical field such as surgery [20]. IMUs can allow recognition and measurement of minute motion. We thought the use of an IMU would be promising for hand-scaling evaluation as well.
Thus, this study proposed a method to classify tip or middle contact based on hand-scaler motion and the force on the target tooth. We measured the motion using an IMU attached to the hand scaler and the force on the tooth by a force sensor attached to the target tooth. We focused on motion because we considered that the difference between the correct and incorrect techniques would be reflected in the motion. The force on the tooth surface was measured by a force sensor because hand-scaling force is known to relate to an important factor in the effectiveness of removing dirt [21], [22].
We asked students to perform hand scaling on a specific tooth to verify the identification of the contact state of the tip using the system. From the IMU attached to the hand scaler and the force sensor attached to the target tooth, we extracted the features of motions and force. The contact state was then classified using a support vector machine (SVM) based on these features. To implement functions with immediate feedback on a small, low-spec, single-board computer, we verified whether it is possible to reduce the number of features. Hand scaling involves several technical aspects. In the future, the simulator may be equipped with functions to determine them simultaneously. In the present study, we aimed to maximally reduce the processing involved in data measurement, feature calculation, SVM, and memory in this function. Therefore, to determine whether the sensor configuration could be simplified, we examined whether the accuracy achieved with both sensors could be maintained while using either sensor alone. We further examined the minimum number of features within one sensor that would maintain an accuracy similar to that obtained with the two-sensor configuration.

A. HAND-SCALING SIMULATOR
The participants performed hand scaling using the simulator, as shown in Fig. 6 (a). The buccal mucosa was removed to observe the blade of the hand scaler during the experiment. The simulator was created based on our previous study [23]. The hand-scaling motion was measured using an IMU (Meta-Motion R; MbientLab Inc.), and the force applied to the tooth was measured using a force/torque sensor (Mini40 SI-80-4; ATI Industrial Automation Inc.), as shown in Fig. 7. The IMU (Weight including case: 19.5 g) was attached to the end of the gripping portion of the hand scaler, as shown in Fig. 7 (a). The IMU can measure the motion of the hand scaler along three axes of acceleration and three axes of angular velocity. When the hand scaler was held upright and viewed from the tip, the Z-axis was vertical to the direction of the blade, vertical to the gripping portion, and pointing toward the right direction. Although the blade could not always be within the field of view, a camera (Endoscope Borescope Inspection Camera, KKmoon) was installed to capture images of the blade for reference in the analysis.
The force applied to the target tooth was measured using a force sensor that could evaluate force along three axes of force and three axes of torque, as shown in Fig. 7 (b). We used the data for the three axes of force in this study. A space was created between the target tooth and model frame to prevent contact between the target tooth (A2A-739-#46, Nissin Dental Products Inc.) and model frame. The force was transmitted to the sensor through the shaft. This was used to measure only the force transmitted from the hand scaler. The X-axis of the force sensor was in the buccal direction; the Y-axis was directed posteriorly from the target tooth; and the Z-axis was in the maxillary direction.
The hand scaler used was an original standard doubleheaded 13/14 #7 Gracie curette (Length: 17.0 cm, Weight: 21.0 g, Hu-Friedy Mfg. Co., LLC), which is commonly adopted in training schools. Each participant used a new hand scaler to ensure uniform blade sharpness. The target tooth was also replaced with a new tooth for each measurement condition and each participant because the tooth surface was worn down by hand scaling.

B. PARTICIPANTS
The participants were students in the self-learning practice phase of hand scaling. We aimed to collect data from students' actual practice of hand scaling. Eighteen first-year students (18 females; age, 19.1 ± 0.8 years) from dental hygienist training schools in Japan participated in the experiment, including eight participants from Training School A, five from Training School B, and five from Training School C. One participant's data were excluded because it showed unusual characteristics for the force measurement values. This was caused by the fingernail touching the target tooth during hand scaling, because of which the contact force of the fingernail was also measured by the force sensor. This is an exceptional result, since the fingernails of people who perform hand scaling are usually not as long as hers.
In the analysis, we used only the data measured from the other 17 participants. All participants had attended lectures on hand-scaling and received practical training at training schools before participating in the experiment. The textbook used was the same for each school, while the content of the lessons was unique and followed the guidelines of each school.
This study was approved by the Ethics Review Committee of Nara Institute of Science and Technology. All participants provided written informed consent before participating in the experiment.

C. HAND-SCALING TASK
The participants performed hand scaling on a specific tooth as shown in Fig. 6 (b) called the lower right first molar in the hand-scaling simulator shown in Fig. 6 (a). We asked them to perform the task based on the assumption that the tooth had buccal mucosa. They performed seven consecutive handscaling strokes as a single task. We asked them to perform basic operations such as holding the hand scaler, moving the blade on the tooth surface, and operating the hand scaler, as shown in Figs. 1, 2, 3, and 4 [1].

D. EXPERIMENTAL FACTORS
The experimental factor was the contact position of the blade, which was categorized as tip contact or middle contact, as shown in Fig. 8. Tip contact, wherein the tip is in contact with the tooth surface, indicated the correct technique. In middle contact, the tip is lifted off the tooth surface, which may damage the gingiva or other parts of the tooth. Thus, middle contact reflected an incorrect technique. Participants performed the hand-scaling task ten times for each of the tip and middle contacts.

E. SENSOR MEASUREMENT VALUES
Measurements were acquired from the IMU attached to the hand scaler and force sensor attached to the target tooth.    Measurements of acceleration, angular velocity, and force were acquired for each of the three axes (nine dimensions in total). Table 1 shows the correspondence of each axis in the forearm rotation motion (Fig. 9).

F. ANALYSIS METHOD 1) PREPROCESSING
The force, acceleration, and angular velocity data were synchronized using synchronous signals of the trigger motion, as shown in Fig. 10. Based on the Z-axis force, the motion section between the maximum values of the first and seventh strokes was extracted and used for the analysis, as shown in Fig. 10. This preprocessing was performed because the duration from the synchronous signals to the first stroke varied for each participant.

2) FEATURE EXTRACTION
The average and standard deviation, which represented the magnitude of motion and the scattering of the magnitude, of force, acceleration, and angular velocity in each of the three axes were calculated as features of 18 items, as shown in Table 2. Each feature was calculated for each trial from the sectional data obtained during preprocessing. We calculated 18 features, each for a total of 20 trials per participant (10 trials using the tip and 10 trials using the middle). In comparison with the stable repetitive strokes using the tip, the strokes performed using the middle of the blade showed distinct characteristics. For example, when students used the middle, the blade may slide on the tooth surface, causing a sudden change in the acceleration or angular velocity. The blade may be caught on the tooth surface, thereby increasing the force applied. Therefore, we selected features that reflected the magnitude of motion and scattering of the magnitude.

3) CLASSIFICATION OF THE CONTACT POSITION OF THE BLADE
To identify whether the tip or middle was in the contact with the tooth, we used SVM [24], which can determine an appropriate discriminative boundary even with a small amount of training data. We trained and tested the data separately for each participant to account for differences among participants and verified whether it was possible to classify the scaler blade contact with the tooth based on the extracted features. The SVM was implemented using MATLAB R2021b, MATLAB toolbox statistics, and the Machine Learning Toolbox TM (MathWorks, Inc.). A linear kernel was used, and five-fold cross-validation was performed. Data were standardized as mean = 0 and standard deviation = 1. We attempted cross-validation using all folds for all participants using all six of [0.001, 0.01, 0.1, 1, 10, 1000] as the SVM hyperparameter C, and chose the C that achieved the highest average accuracy. We analyzed the accuracy value that corresponded to the highest average accuracy. Table 3 lists the values of SVM hyperparameter C used in each case.

4) COMPARISON OF CLASSIFICATION ACCURACY
First, we examined whether the model accuracy using features obtained with only one type of sensor was as high as that obtained using both types of sensors together. For this purpose, the accuracy obtained when using 18 features from the IMU and force sensor (hereafter referred to as IMU+F features) as predictors was compared with the accuracy of a model using the 12 features of the IMU alone (hereafter referred to as IMU features) and that of a model using the six features of the force sensor alone (hereafter referred to as F features) by the STEEL test [25].
Furthermore, we examined the minimum number of singlesensor feature combinations that yielded the same level of accuracy as that obtained using IMU+F features to determine whether data processing and memory could be further reduced. The comparison procedure is shown in Fig.11 and below. VOLUME 10, 2022   (i) Models with i (i = 1, · · · , N) features from a singlesensor were developed. N equals 6 and 12 for the force sensor and the IMU, respectively. (ii) A model with the highest accuracy was determined for each model group with i (i = 1, · · · , N) features. The models were referred to as i-highest models for each sensor. (iii) i-highest models and IMU+F model were compared by STEEL test. The minimum number of i in which the null hypothesis is validated with i-highest model was calculated.
A one-tailed test was conducted with the null hypothesis ''average accuracy of 18 predictors (IMU+F features) = average accuracy of N predictors (features)'' and the alternative hypothesis ''average accuracy of 18 predictors (IMU+F features) > average accuracy of N predictors (features)''. A p-value more than 0.05 validated the null hypothesis and rejected the alternative hypothesis, implying that the average accuracy of 18 predictors (IMU+F features) and the average accuracy of N predictors (features) were equivalent.

A. CLASSIFICATION ACCURACY OF THE CONTACT POSITION OF THE BLADE
The average accuracy of each participant according to the SVM is shown in Fig. 12. This figure shows the results for the IMU+F, IMU, and F features as predictors. The average overall participants of the IMU sensor and force sensor, IMU sensor alone, and force sensor alone were 97.1±4.7%, 95.9±6.9%, and 88.8±14.2%, respectively.

B. COMPARISON OF ACCURACY BETWEEN SENSOR COMBINATIONS
The results of the STEEL test comparing the accuracy of 18 predictors (IMU+F features) with the accuracy of 12 predictors (IMU features) and six predictors (F features) are shown in Table 4. No significant difference was observed between the accuracy of the 18 predictors (IMU+F features) and the accuracy obtained with IMU alone (p = 0.52). However, the accuracy obtained with the force sensor alone was significantly different from the accuracy of the 18 predictors (IMU+F features) (p = 0.01).

C. COMBINATION OF MINIMUM SENSOR CONFIGURATION AND MINIMUM NUMBER OF FEATURES
The accuracy obtained with the force sensor alone was significantly different from that of the IMU and force sensor. We show the results of the IMU that was able to examine the minimum combination of features. Fig. 13 shows the average accuracy of each participant using the IMU+F features and one and two IMU features as predictors. Table 5 shows the results of the STEEL test comparing the accuracy of the IMU and force sensor with that of the IMU alone. The accuracy obtained when IMU+F features were used as predictors was significantly different from that obtained when one IMU feature was used as a predictor (p = 0.0004), but not significantly different from that obtained when two IMU features were used as predictors (p = 0.06). Table 6 shows the top three combinations of two IMU features. All these combinations did not significantly differ from the accuracy of IMU + F features. The average acceleration along the X-axis (Aax) was always included in these combinations.

IV. DISCUSSION
Using the linear SVM, the contact between the tip or middle of the hand-scaler blade with the tooth could be classified with an accuracy of 97.1%. Unlike the blind spots associated with the use of a camera, the proposed simulator did not involve this problem and the IMU+F features showed high accuracy. Therefore, we considered this approach to have high potential for clinical application. The accuracy obtained with the force sensor alone was significantly different from that obtained with both the IMU and the force sensor. In contrast, the accuracy obtained using the IMU alone was not significantly different from that obtained with the IMU and the force sensor. Thus, the sensor configuration can be reduced from the combination of an IMU and force sensor to one IMU. Since the force sensor was attached directly to the target tooth of the hand-scaling simulator, space was needed under the teeth to mount the force sensor, and the existing mannequin's structure, which imitates the head, had to be redesigned. These changes in design would not be necessary when a force sensor is not required.
To simplify the data processing process, we also examined the minimum number of IMU features that maintained the same level of accuracy as that obtained with the IMU+F features. The results showed that the accuracy of combinations    of only two IMU features was not significantly different from that of the IMU+F features. This can simplify the data processing process. Using a combination based on the accuracy of the first rank shown in Table 6, the classification can be made by measuring the values for the acceleration X-axis and acceleration Z-axis. The processing required for data measurement, feature value calculation, and SVM can be reduced with two axes than with six axes, and the storage capacity of the data can also be saved. These advantages can make a difference in the implementation stage, e.g., when attempting to install functions on a small low-spec singleboard computer in an oral cavity model.
In addition, the combinations of two IMU features shown in Table 6 always included the average acceleration in the X-axis, suggesting that it is an important data axis that shows differences in the hand-scaling motion and can be used. The X-axis acceleration reflects the magnitude of the vertical motion of the hand-scaler blade along the tooth surface. When students use the middle during hand scaling, the blade often cannot move smoothly and slips on the tooth surface. Consequently, the motion of the blade is very wide, which is different from the stable stroke when using the tip. A similar motion was observed by the participants in the experiment. When such a motion is observed, the system advises reconfirming VOLUME 10, 2022 whether the tip is in contact with the tooth surface, which will help improve the effectiveness of self-learning.
A limitation of this study is that the participants in the experiment performed hand-scaling only at a specific site of a specific tooth. Therefore, the potential of this method for identifying the state of contact between the tip and tooth for a different tooth or part is unknown. To improve the generalizability of the results of this study, future studies should aim to verify whether this approach can be used with different teeth, parts, and hand scalers.

V. CONCLUSION
The purpose of this study was to verify whether the features acquired from the IMU attached to the hand scaler and the force sensor attached to the target tooth can be used to identify whether the tip of a hand-scaler blade is in contact with the tooth. The IMU and force sensor were used to measure motion while avoiding the problem caused by the gingiva or buccal mucosa obstructing the view of the tooth on the camera image. Data were obtained during hand scaling with forearm rotation. We used SVM to classify whether the tip or the middle of the hand-scaler blade was in contact with the tooth during hand scaling by using features from two types of sensor data. The results showed that classification with a high accuracy of 97.1% was possible. This study is the first to identify the contact of the blade of a hand scaler with the tooth by using an IMU and force sensor. Development of a simulator equipped with the contact-classification method proposed in this study will allow classification of and feedback for the contact state between the tip and tooth surface and the application of this approach as an easily usable selflearning simulator. Researcher with Applied Robotics, Microsoft. His research interests include robotics, including learning-from-observation, task and motion planning, feasible motion analysis, 3D-shape modeling and analysis, and physicsbased vision. From 1983 to 1998, he was with the Electrotechnical Laboratory, Ministry of International Trade and Industry, Japan. From 1993 to 1994, he was a Humboldt Research Fellow at the Institute for Real-Time Computer Systems and Robotics, University of Karlsruhe, Germany. He joined the Nara Institute of Science and Technology, Nara, Japan, in 1998, where he is currently a Professor with the Division of Information Science. He is also the Vice President of the Nara Institute of Science and Technology and the Dean of the Graduate School of Science and Technology. His research interests include human-robot interaction, dexterous manipulation, human modeling, and bio-inspired robotics. VOLUME 10, 2022