Method for evaluating fixed state of fixed dental crown restoration, and medical examination device

Abstract

Provided is a method for evaluating the fixed state of a fixed dental crown restoration, the method including: applying elastic wave stimulation to a fixed dental crown restoration; acquiring, as a time-history waveform, a response waveform of the fixed dental crown restoration receiving the elastic wave stimulation; evaluating frequency response characteristics from the acquired time-history waveform; inputting the frequency response characteristics to a trained anomaly detection model to obtain an anomaly score as an inference result; and determining a fixed state of the fixed dental crown restoration from the anomaly score. The anomaly score relates to a state in which the fixed dental crown restoration is fixed to an abutment tooth or an implant body.

Description

Methods for evaluating the fixation status of fixed crown restorations, and examination devices.

[0001] This disclosure relates to a method for evaluating the fixation status of fixed dental crowns and to an examination device.

[0002] Fixed crowns (FPDs) are fixed restorations attached to natural teeth or implants, applied to one or more teeth for the purpose of restoring tooth structure or prosthetically replacing missing teeth due to tooth structure loss or tooth loss. If the treatment planning, design, fabrication, or placement is not done properly, early complications may occur, potentially causing irreversible damage to the tooth or supporting structure. The main cause of complications in FPDs is considered to be reduced retention. Other complications include fracture of the junction, fracture of the porcelain, fracture of the abutment tooth, and gaps that develop between the retainer or pontic and the abutment tooth. Such fractures can lead to the development of secondary caries or the loss of abutment teeth.

[0003] Degradation of the cement used in dental crowns is a frequent clinical problem. Because this degradation often progresses gradually, a clear diagnosis is difficult in the early stages, posing a significant clinical challenge. Particularly in bridges, when some retainers lose cement fixation while others remain secure, it becomes difficult to clinically assess the overall fixation of the restoration. Such conditions can lead to secondary caries, mechanical tooth damage, periodontal disease, and even tooth loss, requiring early diagnosis and appropriate intervention. Signs of decreased FPD retention are evaluated through clinical examinations such as visual inspection and palpation, radiographic examinations, electronic mobility tests, and occlusal force analysis; however, none of these methods are sufficient to detect subtle changes in cement adhesion or early-stage fixation with high accuracy. Furthermore, in implant FPDs, in addition to decreased cement fixation, early detection of screw loosening is also a crucial issue.

[0004] Patent Document 1 discloses a dental medical device capable of sharing information on measurement results of the stability of an implant among different devices. This document discloses a configuration in which a measurement auxiliary body Ia made of a magnetic material is attached to the upper part of an implant, and by exciting the measurement auxiliary body Ia, the implant and the measurement auxiliary body Ia are vibrated integrally, and the resonance frequency of the measurement auxiliary body Ia is detected. Since the implant and the measurement auxiliary body Ia vibrate integrally, the resonance frequency of the implant can be grasped by detecting the resonance frequency of the measurement auxiliary body Ia. Then, based on the resonance frequency, a value (ISQ value) of implant stability, which is an index indicating the stability of the implant, is calculated according to a control program. However, the technology described in Patent Document 1 targets the evaluation of the stability of the implant body embedded in bone, and does not target the evaluation of the fixation state of FPDs including single crowns and bridges and the detection of minute changes in the cement fixation state, nor does it aim to non-invasively and early evaluate the fixation state of FPDs.

[0005] Resonance frequency analysis (RFA) is a method capable of non-invasively obtaining information on the internal state of a structure and is used in various fields including dentistry. In the dental field, a method for evaluating the stability of an implant using RFA is known, and an example thereof is the "Osstell Mentor" device. This device excites a "SmartPeg" attached to the implant body with a screw and calculates a stability index based on the resonance frequency thereof. However, since SmartPeg is designed on the premise of being attached to the implant body, it is difficult to screw-fix it to an FPD such as a crown or a bridge, and it could not be applied to the evaluation of the fixation state of an FPD.

[0006] The inventors have been studying a new RFA device equipped with a 3D acceleration sensor. This device can detect stability and looseness without screwing the sensor into the implant. Furthermore, it has been shown that the cement fixation state of tooth-supported FPDs can be evaluated using RFA and finite element analysis. However, accurately evaluating the resonant frequency characteristics of FPDs related to cement fixation loss based solely on conventional manual analysis and empirical judgment remains difficult, and challenges in terms of objectivity and reproducibility still exist.

[0007] Japanese Patent Publication No. 2020-146072

[0008] The reduced retention of flat panel implants (FPDs) is a clinically frequent issue. However, diagnostic methods for early detection of abnormalities or loosening remain unestablished. While new RFA devices and finite element analysis have been attempted, objectively and accurately determining the resonant frequency characteristics of FPDs due to cement fixation loss remains a challenge. To address this challenge, there has been a need for a non-invasive and early evaluation method and diagnostic device for assessing the fixation status of FPDs attached to abutment teeth or implants.

[0009] This disclosure was made to solve the problems described above, and aims to provide an evaluation method for easily and accurately evaluating the fixation state of fixed dental crowns, and an examination device that enables such evaluation.

[0010] The method for evaluating the fixed state of an FPD according to this disclosure comprises the steps of: applying elastic wave stimulation to the FPD; acquiring the response waveform of the FPD to the elastic wave stimulation as a time history waveform; calculating frequency response characteristics based on the time history waveform; inputting the frequency response characteristics into a trained anomaly detection model to perform inference and obtain an anomaly score; and determining the fixed state of the FPD based on the anomaly score. According to this evaluation method, by applying machine learning techniques to the vibration response acquired by resonant frequency analysis (RFA), it becomes possible to objectively extract features related to the fixed state of the FPD and quantitatively determine the quality of the fixed state based on numerical indicators such as an anomaly score.

[0011] Other features of this disclosure are outlined below.

[0012] The fixation status of fixed crown restorations can be easily and accurately evaluated.

[0013] This is a diagram showing an example of the experimental process. This is a photograph of a mandibular model. This is a diagram showing the direction of the acceleration sensor and vibration stimulus. This is a diagram showing the type of cement fixation and the measurement site. This is a diagram showing the response frequency obtained using the Fast Fourier Transform (FFT) for the waveform of each accelerometer. This is a diagram showing the structure of a random forest classifier. This is a diagram showing the neural network structure for supervised learning. This is a diagram showing the architecture of an autoencoder. This is a diagram showing the relationship between loss value and epoch. This is a graph showing feature importance. This is a graph showing feature importance. This is a graph showing feature importance. This is a graph showing feature importance. This is a diagram showing the analysis results of the autoencoder. This is a photograph of a jawbone model with two implants. This is a photograph of a custom-made surgical guide. This is a diagram showing an acceleration sensor fixed to the abutment. This is the result of feature importance analysis for screw loosening detection in a conical joint. This is the result of feature importance analysis for screw loosening detection in a conical joint. This is a diagram showing the frequency response anomaly score in a conical joint configuration. This is the result of feature importance analysis for screw loosening detection in an internal tri-channel joint. This is the result of feature importance analysis for screw loosening detection in an internal tri-channel connection. This figure shows the frequency response anomaly score in an internal tri-channel connection. This is the result of feature importance analysis for screw loosening detection in an external hexagonal connection. This is the result of feature importance analysis for screw loosening detection in an external hexagonal connection. This figure shows the frequency response anomaly score in an external hexagonal connection. This figure shows an embodiment of a model experiment. This figure shows an example of an acceleration waveform obtained from a healthy implant. This figure shows the results of FET analysis. This figure shows the anomaly scores for healthy data and loosened data. This is a photograph showing the human experiment. This figure shows the acceleration waveform of the bridge. This figure shows the frequency spectrum.

[0014] Embodiments. An example of a method for evaluating the fixation state of a fixed dental crown restoration (FPD) according to the present disclosure is described. Sensor Fixation The sensor is fixed to the FPD. Here, fixation includes mere contact. The type of FPD is not particularly limited and may be a single or connected dental crown restoration. For example, a small piezoelectric 3D acceleration sensor can be used, but any other type of sensor that can measure up to about 0 to 10,000 Hz can be used. In addition to piezoelectric acceleration sensors, magnetostrictive sensors and non-contact laser vibrometers can also be used. For example, the sensor can be fixed to the top or side surface of the FPD using double-sided tape, adhesive, clips, etc.

[0015] Vibration of dental crowns: The FPD is vibrated, for example from the side, to provide vibration stimulation to the FPD. An exciter can be used to provide vibration stimulation. In one example, the exciter is attached to the FPD to provide vibration stimulation to the FPD. In another example, the exciter is brought into contact with the jaw or cheek, indirectly providing vibration stimulation to the FPD through the human body. Examples of exciters include impact devices such as automatic impact devices, exciters, and bone conduction speakers, but other exciters that can provide frequency sweeps or white noise may also be used. In one example, the exciter is capable of vibrating in the frequency range of 0 to 10,000 Hz. Such exciters include the aforementioned automatic impact devices. Considering that vibration is applied to the FPD by the exciter and the response waveform of the FPD to that vibration stimulation is acquired by a sensor, "vibration stimulation" can be expressed as "elastic wave stimulation".

[0016] Acquisition of response waveform: The response of the FPD caused by the vibration stimulus is recorded by a sensor. For example, the response waveform of the FPD to the vibration stimulus is acquired as a time history waveform. For example, the sampling frequency can be set to 20,000 Hz or higher.

[0017] Frequency Response Characteristics Data Acquisition: The frequency response characteristics are evaluated from the acquired time-history waveform. For example, the frequency response characteristics can be evaluated by performing a Fast Fourier Transform on the time-history waveform. In one example, the frequency bandwidth can be 0 to 10,000 Hz, and the frequency resolution can be approximately 100 Hz. In this disclosure, a frequency spectrum representing the relationship between frequency and amplitude is used as an example, but it may also be converted into a spectrogram image representing time, frequency, and amplitude.

[0018] The frequency response characteristics of AI diagnostics obtained through supervised learning are used as training data to train a machine learning model. If labeled data for healthy and abnormal dental restorations is available in advance, multi-class classification using common supervised learning methods such as neural networks and ensemble techniques becomes possible.

[0019] Unsupervised AI Diagnosis: If data on abnormal FPDs is not available in advance, a good product learning model is constructed by providing only data on the frequency response characteristics of healthy products, and anomaly detection is performed. For example, an autoencoder can be used as an unsupervised learning algorithm to train the anomaly detection model. In this disclosure, an autoencoder using a fully connected neural network is used as an example of a good product learning method, but various unsupervised learning methods, including convolutional autoencoders, can be used.

[0020] Model Enhancement through Transfer Learning: When sufficient training data is unavailable for the aforementioned "supervised AI diagnosis" and "unsupervised AI diagnosis," transfer learning is effective. This involves creating a pre-trained model using similar data and then retraining the model with a small amount of training data. For example, normalizing the frequency spectrum along the frequency axis eliminates dependence on gender, age, etc., and enables diagnosis of a fixed state even without data of a healthy, fixed state. Transfer learning can improve the generalization and discrimination performance of the model.

[0021] The frequency response data described above can be fed into the autoencoder of the "AI diagnosis using unsupervised learning" to calculate the anomaly score. The anomaly score e is defined by focusing on the difference between the input data fi and the reconstructed data gi. By smoothing the input fi and the reconstructed data gi using a moving average, the variability of the diagnosis can be reduced. The width of the moving average can be, for example, around 100 to 1000 Hz. Also, since much of the data is noise, if the absolute amount of fi-gi is smaller than the threshold, it can be set to zero to clearly identify an abnormal FPD. The noise cut threshold can be set to a maximum amplitude ratio of around 0 to 0.3. In this way, the frequency response data is input into the anomaly detection model and the anomaly score is obtained as an inference result.

[0022] Normal / Abnormal Determination The fixation status of the FPD is determined from the abnormality score mentioned above. The determination is, for example, normal / abnormal. Normal indicates that the FPD is properly fixed to the abutment tooth or implant, while abnormal indicates that there is an abnormality in the fixation status of the FPD to the abutment tooth or implant. For example, abnormality scores can be calculated from the training data using the "AI diagnosis using unsupervised learning" model mentioned above, and the abnormality scores corresponding to the top 5% can be set as the abnormality detection threshold.

[0023] For example, the diagnostic device may be equipped with an "AI diagnostic device" that acquires sensor output and outputs a large abnormality score if the retention force of the FPD to the abutment tooth or implant is small. For example, the AI ​​diagnostic device may output feature importance.

[0024] The following provides a more detailed explanation of the method for evaluating the fixation state of FPDs and the examination device, but this disclosure is not limited to these specific examples and encompasses various variations and modifications.

[0025] 1. Materials and Methods for Detecting Abnormalities in Bridge Fixation Figure 1 shows an example of the experimental process. The methodology is divided into clear stages: sample preparation, experimental procedure, data collection, and subsequent data analysis using machine learning techniques. Figure 2 is a photograph of the mandibular model. The in vitro mandibular model consists of a single crown fabricated on the left first molar and a three-unit bridge fabricated from the right second molar to the second premolar, both manufactured using high-gold alloy. In this example, to reproduce the periodontal ligament, the prepared teeth were inserted into sockets of the mandibular model (mandibular E50-512, Nisshin Dental Products Co., Ltd., Kyoto, Japan) using a general-purpose impression material (Fusion II Extra Wash Type, GC Corporation, Tokyo, Japan).

[0026] Between a single crown and its abutment tooth, two cementation conditions were established: a crown that is completely cemented and a crown that is not cemented. In addition, the following four cementation conditions were established between a three-unit bridge and its abutment teeth: Both crowns are firmly cemented. Only the premolar crown is cemented. Only the molar crown is cemented. Neither crown is cemented.

[0027] The RFA system used in this study included the following equipment: data acquisition system (PowerLab ML880, ADInstruments Ltd., Oxford, UK), constant current source (Power Unit 4114B1, Dytran Instruments Inc., California, USA), 3D accelerometer (3133A1, Dytran Instruments), and software package (LabChart 7 for Windows, ADInstruments).

[0028] The 3D acceleration sensor was attached to the occlusal surface of the test tooth using, for example, 0.15 mm thick double-sided tape (NICETACK NW-K15, Nichiban Co., Ltd., Tokyo, Japan). For each cement fixation condition, a tooth and implant stability tester (Periotest Classic, Medizintechnik Gulden e. K., Modautal, Germany) was used to apply 4 Hz vibration excitation in the buccal-lingual direction to the buccal surface of the abutment for 4 seconds, generating 16 impact vibrations at 0.25-second intervals (see Figure 3). Figure 3 shows the 3D acceleration sensor 10 attached to the occlusal surface of the test tooth. The 3D acceleration sensor 10 was used for data acquisition.

[0029] Measurements were performed on a single crown that was fully cemented (Full-Mos) or a single crown that was not cemented (Un-Mos). Measurements for a 3-unit bridge were also performed under the following conditions (measured on molars: MO, measured on premolars: PRE): Both crowns were cemented (Full-MO, Full-PRE); only premolars were cemented (Pre-MO, Pre-PRE); only molars were cemented (Mo-MO, Mo-PRE); and neither crown was cemented (Un-MO, Un-PRE).

[0030] The type of cement fixation (X) and the measurement site (Y) were specified using XY notation. For example, as shown in Figure 4A, vibrations applied to single crowns were designated as Full-Mos or Un-Mos. Mos means a single crown, while Full and Un mean with and without cement, respectively. As shown in Figure 4B, vibrations applied to the second molar of a 3-unit bridge were designated as Full-MO, Mo-MO, Pre-MO, or Un-MO, and vibrations applied to the second premolar were designated as Full-PRE, Mo-PRE, Pre-PRE, or Un-PRE. Here, Full means both the molar and premolar are cemented, Pre means only the premolar is cemented, Mo means only the molar is cemented, and Un means neither the molar nor the premolar is cemented. White crowns indicate that the teeth are not cemented. The arrows indicate the measurement sites.

[0031] When an impulse load was applied, the 3D accelerometer recorded an increase in amplitude exceeding a certain threshold. Next, a Fast Fourier Transform was performed on 128 plots for each spatial dimension (x, y, z) to extract the time history of acceleration. The resulting frequency response function had a resolution of 78.1 Hz and covered a frequency band of 0–4920 Hz. As a result, the machine learning model used 64 parameters for each direction (x, y, z) as features, using a total of 192 parameters (64 parameters × 3 directions) of one-dimensional (1D) data. The frequency response function was normalized to the maximum amplitude to reduce the variation in oscillatory excitations (see Figure 5).

[0032] The training data was expanded by shifting the frequency response function by up to ±2 parameters in the frequency axis direction, increasing the amount of training data fivefold. The parameters at both ends affected by the shift were set to zero for all data. In the context of unsupervised learning, even with identical test setups and conditions, the measured frequency response varies around the resonant frequency. In unsupervised learning, the amount of training data was increased fivefold by copying and shifting the frequency response function, shifting it to a frequency resolution of 78.1 Hz × (-2 to +2) in the frequency axis. This method is effective in improving the performance of unsupervised machine learning with large amounts of training data.

[0033] Each cement fixation condition was labeled as follows: Full-MO (1), Un-MO (2), Pre-MO (3), Mo-MO (4), Full-PRE (5), Un-PRE (6), Mo-PRE (7), Pre-PRE (8). To classify these conditions, two supervised learning methods were employed: a random forest classifier and a neural network classifier. The random forest classifier was implemented using the sklearn.ensemble.RandomForestClassifier module from the Scikit-learn library (version 0.24.1). The model parameters were set to 100 for the number of decision trees, and other settings used the default values ​​of Scikit-learn. Figure 6 shows the structure of the random forest classifier. The neural network classifier was also implemented using the MLPClassifier module from Scikit-learn (version 0.24.1). This model has four hidden layers, with the number of nodes in each layer set to 96, 48, 24, and 12 from the input layer to the output layer, respectively. Figure 7 shows the structure of a supervised learning neural network. Other parameters were set to the default values ​​of Scikit-learn (using the Adam optimizer), with a learning rate of 0.001 and 200 epochs.

[0034] The training dataset was split into 70% and the test dataset into 30%. 80 frequency response functions were generated for each experimental case, and these were randomly split, with approximately 56 assigned to the training dataset and 24 to the test dataset. This splitting was performed using the scikit-learn `train_test_split` module. To evaluate the performance of the machine learning models, confusion matrices and accuracy were calculated using neural network and random forest models. The confusion matrix provides an overview of the model's overall performance by counting the number of true positives, false positives, true negatives, and false negatives. Model accuracy was evaluated by comparing the actual labels in the test data with the predicted labels.

[0035] A feature importance graph was generated using a random forest algorithm, and feature importance scores were calculated. These scores indicate which features or characteristics in the data are most relevant to a particular task. Features with high scores are considered to have a greater impact on the outcome, while those with low scores are considered to have a smaller impact. Visualizing feature importance as graphs and rankings allowed for a clearer understanding of the key factors involved in model decision-making. This information is expected to be useful in identifying critical elements in disease diagnosis, optimizing product design, and understanding the most influential variables in data-driven decision-making processes. In this study, the impact of different cement conditions on the y-direction vibration of a flat panel display (FPD) was evaluated. The feature importance graph visually represents the importance of each feature in the dataset across frequency ranges, proving useful in evaluating complex 3D vibration behavior in bridge and single-crown models.

[0036] Furthermore, an unsupervised learning algorithm was used to identify patterns in the data and group them based on similarity. For this purpose, an autoencoder (AE) architecture consisting of 192 input nodes compressed via three hidden encoding layers with 96, 48, and 24 nodes was employed. Figure 8 illustrates such an autoencoder architecture for adhesion condition evaluation. Decoding was performed by performing the same process in reverse. The activation function used in the model was a normalized linear unit (ReLU), and a sigmoid function was used in the output layer. TensorFlow (version 2.3.0) was used to build and train the model. During training, the AE learned how to match the input and output data. The mean squared error was used as the loss function, the learning rate was set to 0.001, and 500 training epochs were performed. Adam was used as the optimizer, with a batch size of 32 (see Figure 9). The frequency response function was used as a criterion for comparing the input and output data of the AE.

[0037] Next, anomaly scores were calculated using the output frequency response function and input function reconstructed by AE. This applied an unsupervised learning method to detect anomalies in the cement fixing conditions of the FPD. It should be noted that the number of training data instances for Full-Mo and Full-Pre connection conditions differs from the number of data used in supervised learning. In this study, a total of 240 frequency response functions obtained from three sets of tapping tests were used as training data. Using a sufficient amount of training data is important to improve the performance of unsupervised learning and to more effectively identify anomalies in cement fixing conditions.

[0038] To more effectively evaluate the cement fixing conditions for noise reduction and anomaly score calculation, several processes were implemented. First, moving averages of the input and output functions were obtained. A 10-parameter moving average was applied to improve data smoothness. Next, since some frequency ranges may contain only noise, 0.2 was subtracted from the input and output frequency response function values ​​to mitigate its influence. After this process, any negative results were set to zero. After applying these noise reduction processes, the anomaly score e was calculated using the following formula.

[0039]

[0040] Here, fi and gi (1 ≤ i ≤ 192) represent the input frequency response function and the reconstructed output frequency response function, respectively.

[0041] The results showed that both the single-crown neural network and the random forest algorithm were trained using acceleration datasets in the x, y, and z directions. The training results (Table 1) show that both methods achieved a positive confusion matrix and demonstrated high accuracy. Specifically, the neural network achieved 100% accuracy, and the random forest achieved 99% accuracy.

[0042]

[0043] Figure 10 is a graph showing the result of calculating feature importance related to cement fixation conditions using the random forest algorithm. This graph evaluates the frequency response in the x, y, and z directions under cement fixation conditions. Specifically, Figure 10A shows the feature importance of the frequency response in the x direction calculated from the acceleration under single-crown cement fixation conditions. Figure 10B shows the feature importance of the frequency response in the y direction calculated from the acceleration under single-crown cement fixation conditions. Figure 10C shows the feature importance of the frequency response in the z direction calculated from the acceleration under single-crown cement fixation conditions. The results show that the dynamic response in the y direction is greater than the responses in the x and z directions. This suggests that in the single-crown model, the dynamic characteristics in the y direction are more pronounced than in the other directions and play an important role in evaluating cement fixation conditions.

[0044] Training of a supervised machine learning algorithm using only the y-acceleration dataset Using a neural network and a random forest algorithm to train the y-direction acceleration dataset, as shown in Table 2, it was confirmed that both methods achieved high accuracy. Specifically, an accuracy of 97% was obtained for the neural network and 96% for the random forest. Also, favorable confusion matrices were obtained for each model, indicating excellent prediction performance.

[0045]

[0046] This result shows that even when using only the acceleration data in the y-direction, the supervised learning algorithm has the ability to accurately classify the cement fixing conditions of the 3-unit bridge. In other words, a supervised machine learning algorithm using only acceleration data in a specific direction can be used for training the anomaly detection model. The feature importance graph was created using the random forest method. Figure 11 shows the frequency range with the highest score when measurements were taken on the molar. Figure 11 is the feature importance graph of the frequency response function calculated using the y-acceleration data of the 3-unit bridge where the measurement site is the second molar. From Figure 11A, it can be seen that frequencies in the range of 0 to 500 Hz indicate the importance of features in the comparison between the Full-MO condition and the Un-MO condition. From Figure 11B, it can be seen that a frequency of approximately 1000 Hz indicates importance in the comparison between the Full-MO condition and the Pre-MO condition. From Figure 11C, it can be seen that a frequency of approximately 3000 Hz indicates importance in the comparison between the Full-MO condition and the Mo-MO condition.

[0047] Figure 12 shows the frequency range with the highest score when measurements were taken at the premolars. Figure 12 is a feature importance graph of the frequency response function calculated using y-acceleration data of a 3-unit bridge where the measurement site is the second premolar. From Figure 12A, it can be seen that the range of 0 to 500 Hz shows feature importance when comparing the Full-PRE condition and the Un-PRE condition. From Figure 12B, it can be seen that the frequency of approximately 2000 Hz shows importance when comparing the Full-PRE condition and the Mo-PRE condition. From Figure 12C, it can be seen that the frequency of approximately 3000 Hz shows importance when comparing the Full-PRE condition and the Pre-PRE condition.

[0048] The results shown in Figures 11 and 12 indicate that the frequency range of 0–500 Hz is important when evaluating the cement fixation conditions for Un-MO and Un-PRE. On the other hand, for Pre-MO and Mo-PRE cement fixation conditions, the range of 1000–2000 Hz was found to be the most important. Furthermore, the resonant frequency for the Full-MO, Full-PRE, Mo-MO, and Pre-PRE conditions was approximately 3000 Hz, confirming high rigidity under these conditions.

[0049] Supervised machine learning algorithms were trained on a 3D acceleration dataset using a neural network and a random forest algorithm (Table 3). Both approaches yielded positive confusion matrices and demonstrated high accuracy. The neural network achieved an accuracy of 98%, and the random forest achieved an accuracy of 99%.

[0050]

[0051] Using the Random Forest algorithm, the analysis was performed, and as shown in Figure 13, the feature importance for eight cement fixation conditions was evaluated as a function of the frequency response in the x, y, and z directions. Figure 13A is a feature importance graph calculated from the acceleration of the FPD cement state, showing the frequency response in the x direction. Figure 13B is a feature importance graph calculated from the acceleration of the FPD cement state, showing the frequency response in the y direction. Figure 13C is a feature importance graph calculated from the acceleration of the FPD cement state, showing the frequency response in the z direction. Figure 13 shows that the acceleration in the x direction is significantly greater than the acceleration in the y and z directions. This confirms that the x direction has particularly important features in the analysis using 3D acceleration data.

[0052] Figure 14 shows the results of an analysis using an autoencoder (AE) trained with the y-direction frequency response obtained from standard data for Full-MO cementation conditions. This figure shows a comparison of the input and output data for Full-MO and Un-MO conditions. Specifically, the input frequency response function for the Full-MO condition was accurately reconstructed based on the standard data for the Full-MO condition (Figure 14A). On the other hand, the model was unable to accurately reconstruct the input frequency response function when using anomalous data for the Un-MO condition (Figure 14B). Anomaly scores were calculated using Full-MO (Table 4A), Full-PRE (Table 4B), and Full-MO / Full-PRE (Table 4C) conditions for training. This analysis used 1D and 3D acceleration datasets for all cement fixing conditions. The results showed that using 3D acceleration data improved the coefficient of variation (CV) between anomaly scores and increased the ability to detect anomalies.

[0053]

[0054] The data in the table showed that using data from the fully cemented condition yielded the lowest anomaly score, suggesting that this condition was the most stable. On the other hand, the non-cemented and partially cemented conditions yielded higher anomaly scores, indicating that these conditions were more susceptible to anomalies.

[0055] Discussion: The results of this study suggest the potential of machine learning algorithms to rapidly and accurately identify the cement fixation status between FPDs (fixed partial dentures) and their abutments. This method can not only detect cement loss from FPDs but also classify the cement fixation status. This is expected to enable clinicians and researchers to make informed treatment decisions and prevent complications. Random forest is an effective algorithm for estimating the importance of features in a dataset. The importance score of each feature is based on the reduction in impurity (or loss) when the data is split at the nodes of the decision tree. The reduction in impurity is averaged across all trees in the forest, and a single score is assigned to each feature. This makes it possible to identify the most important features, analyze them to gain a deeper understanding of the underlying relationships in the data, and improve the performance of the model.

[0056] In this study, the resonant frequency of the molar crowns in a fully cemented bridge on a mandibular model (Figure 11) was approximately 3000 Hz, consistent with previous research results. This value indicates that the crown or bridge is firmly fixed to the natural tooth. Generally, highly rigid objects resist deformation and require high energy for vibration. Therefore, they tend to vibrate at high frequencies when subjected to external forces or vibrations. On the other hand, in bridges that are not cemented, the resonant frequency of the crown falls within the range of 0 to 500 Hz. This is because the absence of cement requires less energy for vibration, resulting in a tendency to vibrate at low frequencies. Furthermore, in partially cemented bridges, the feature importance score for the uncemented crown was highest at approximately 1000 Hz, indicating that the rigidity of partially cemented bridges is lower than that of fully cemented bridges. However, the resonant frequency of the crowns in partially cemented bridges was similar to that of fully cemented bridges (approximately 3000 Hz), and the peak of the feature importance score showed a similar value. A similar trend was observed in the measurement of the resonant frequency of premolars (Figure 12). When the crown was not fixed with cement, the resonant frequency varied in the range of 0 to 500 Hz. On the other hand, when it was fixed with cement, a maximum of 3000 Hz was recorded. Furthermore, the resonant frequency under MO-PRE conditions was approximately 2000 Hz, exceeding that of Pre-MO conditions (approximately 1000 Hz). This difference is thought to be because the resonant frequency of an object depends on its natural vibration modes, shape, and size. Larger objects tend to have greater mass and lower vibration frequencies, while smaller objects vibrate at higher frequencies because they react quickly to external forces. Position and boundary conditions also affect the resonant frequency. For example, different resonant frequencies are produced depending on whether the object is fixed or released.

[0057] Figures 10 and 13 show that in the bridge model, x-direction acceleration is more important than y-direction acceleration, while in the single crown model, the y-direction feature importance score is the highest. In the single crown model, there are no constraints in the x or y directions, and when an impact load is applied in the y direction, the tooth moves freely in the x direction, resulting in a low importance score in the x direction. On the other hand, in the bridge model, the x direction is constrained, so the tooth cannot move freely, resulting in a high importance score in the x direction. In addition, the constraint in the x direction generates complex 3D vibrations, and the z-direction feature importance is higher than in the single crown model. This result demonstrates the usefulness of machine learning in complex 3D vibration analysis.

[0058] Unsupervised learning is an algorithm that does not require labeled data and identifies patterns in data through clustering, dimensionality reduction, and density estimation. In this study, an autoencoder (AE) was used as the unsupervised learning algorithm. An AE is a neural network that compresses input data into a low-dimensionality and learns the data through reconstruction. When comparing data in a fully fixed state and an unfixed state using an AE, the data in the fully fixed state could be reconstructed accurately, while the data in the unfixed state could not. This suggests that the unfixed state may contain anomalous patterns for the AE.

[0059] Anomaly detection is a field of machine learning that identifies anomalous data points that deviate significantly from the expected behavior of the data. The anomaly score is an index that quantifies whether the data is anomalous, with a higher score indicating a higher probability of anomaly. In this study, the performance of the anomaly detection algorithm was improved by using 3D acceleration data. This is because 3D data comprehensively captures deviations from normal behavior. As a result, the fully fixed state was the most stable, and the number of anomalous data points tended to increase in the unfixed and partially fixed states.

[0060] Identifying the fixation state between the FPD and the abutment using machine learning algorithms is useful in FPD management and serves as a powerful tool to support treatment decision-making in clinical settings. In this study, the highest scores were observed at approximately 3000 Hz for fully fixed states, 1000–2000 Hz for partially fixed states, and 0–500 Hz for unfixed states. Further improvements were obtained by including 3D vibration analysis. Fully cemented fixation resulted in low abnormality scores, while unfixed and partially fixed states resulted in high abnormality scores. The stiffness and vibration characteristics of the FPD differed depending on the measurement conditions and acceleration direction. These results suggest the potential to contribute to improving the accuracy of RFA (Resonant Frequency Analysis) in FPD models and to early abnormality detection. While this disclosure evaluated the fixation state of a specific FPD, it is believed that, in principle, the fixation state and abnormalities can be detected for any FPD using the method described herein.

[0061] 2. In vitro (extracorporeal) investigation of abnormal screw loosening detection in different implant-abutment connection modes. Screw loosening is a common mechanical complication in implant-supported prostheses, typically occurring in 4.3% to 10% of patients after functional loading has been applied. Screw loosening can lead to problems such as implant-abutment instability, pain, increased risk of bacterial invasion, peri-implantitis, and potential marginal bone loss.

[0062] Resonant frequency analysis (RFA) is a non-invasive and objective method widely used in implant dentistry to evaluate implant stability. It measures the vibration frequency of the implant and calculates and displays it as an Implant Stability Quotient (ISQ). RFA readings can be influenced by bone quality, implant design and materials, soft tissue condition, patient-specific factors, and even measurement and environmental variations. Advanced data analysis techniques, particularly machine learning algorithms, can help improve the accuracy and reliability of RFA evaluation. Machine learning algorithms can be broadly categorized into supervised approaches, such as neural networks and random forests, which are trained on labeled data to predict outcomes, and unsupervised approaches, which detect patterns from unlabeled data. However, well-known studies have focused on the retention force of the prosthesis (cement loss) rather than screw loosening between the implant and the abutment. Screw loosening is a common mechanical complication in implant-supported restorations, but its early detection remains difficult due to a lack of non-invasive and quantitative evaluation tools. Therefore, the inventors applied a combination of RFA and machine learning to detect screw loosening in different implant-abutment connection (IAC) configurations. The null hypothesis was that even when RFA is combined with machine learning, it is not possible to reliably detect early screw loosening in different IAC configurations.

[0063] Materials and Methods Three distinctly different implant-abutment connection types were tested: conical connection (NobelParallel CC; Nobel Biocare), internal tri-channel connection (Replace Select Tapered; Nobel Biocare), and external hexagonal connection (Branemark System Mk III; Nobel Biocare). All implants were standardized to a length of 10 mm, with the conical and internal tri-channel implants having a diameter of 4.3 mm, while the external hexagonal implant had a diameter of 4.0 mm. Three jawbone models (EX-7 and 1017A; Nissin) were prepared, and two implants were placed in the first molar region of each model. Figure 15 is a photograph of the jawbone model with two implants. A custom-made surgical guide was used to standardize the implant placement and ensure accuracy. Figure 16 is a photograph of the custom-made surgical guide.

[0064] RFA Equipment and Measurements RFA measurements were performed using a 3-axis piezoelectric pickup sensor (Model 3133A1; Dytran Instruments, Inc.), a power unit (Model 4114B1; Dytran Instruments, Inc.), a PowerLab data acquisition system (Model ML880; Dytran Instruments, Inc.), and a software package (LabChart 7; Dytran Instruments, Inc.). The 3D accelerometer was securely attached to the lingual surface of the abutment cap or jig. Figure 17 is a photograph of the 3D accelerometer fixed to the abutment. In Figure 17A, the direction on the mandibular model is defined. Figure 17B is a plan view of the accelerometer. Vibration was applied buccal-lingually at 4 Hz using an implant stability tester (Periotest Classic; Medizintechnik Gulden), with 16 impulse strikes at 0.25-second intervals. The accelerometer recorded responses along three axes (x: mesiodistal, y: buccolingual, z: implant longitudinal axis) at a sampling rate of 20,000 Hz. To reduce variations due to impact force and angle, the frequency response functions (FRFs) were normalized to the maximum amplitude.

[0065] To simulate various degrees of screw loosening, the abutment screws were first tightened to 15 Ncm (0-degree loosening). Then, they were gradually loosened in 5° increments up to 30°, creating six different loosening conditions: 5°, 10°, 15°, 20°, 25°, and 30°. Under the normal condition (intact condition), each specimen was subjected to 20 vibration cycles (320 total impacts), while under the loosened conditions, each specimen was tested with 4 vibration cycles (384 total impacts). Frequency response functions (FRFs) were recorded for all seven conditions, and resonant frequencies were captured along all three axes.

[0066] Supervised Learning Model The supervised learning model was developed using a random forest algorithm implemented via the machine learning library (Scikit-learn, version 0.24.1; Python Software Foundation) (the random forest architecture is shown in Figure 6). 100 decision trees and default parameters were used. The following two classification tasks were performed: Binary classification: 320 normal samples (0°) were labeled 0, and 384 loosened samples (64 samples each from 6 loosening angles from 5° to 30°) were combined to form label 1. Multiclass classification: 64 samples from each of the 7 conditions (5-degree increments from 0° to 30°), for a total of 448 samples, were classified into labels 0 to 6. The dataset was split 80% for training and 20% for testing. Model performance was evaluated using the Accuracy metric and confusion matrices, and key resonant frequency features related to screw loosening detection were identified by feature importance analysis. To verify reliability, classification using random forests and neural networks was repeated over 200 trials.

[0067] As an unsupervised learning model, a convolutional autoencoder (CAE) was constructed using a deep learning framework (TensorFlow, version 2.8.0; Google LLC) and a neural network library (Keras, version 2.8.0; Google LLC) (see Table 5 for the CAE architecture and Figure 8 for a schematic diagram of the CAE).

[0068]

[0069] CAE performed data compression and reconstruction using a 1D CNN. This included a convolutional layer for extracting local features, a compression layer with a stride of 2, a fully connected layer, and an upsampling layer for increasing the resolution. The ReLU (Rectified Linear Unit) activation function was used for all except the sigmoid function in the final layer. The model was trained to match the input and output using the Mean Squared Error as the loss function. The training was carried out for 500 epochs, with a learning rate of 0.001, the Adam optimizer, and a batch size of 16. The frequency response functions (FRFs) of normal state implants were used as the baseline condition. The dataset was composed of 64 specimens for each of the 7 loosening angles. To cope with data variations, the evaluation was repeated over more than 200 trials. The stability of learning was evaluated by plotting the loss function using 80% of the data for learning and 20% for validation. At that time, the number of epochs was set on the x-axis and the loss value was set on the y-axis.

[0070] Enhanced anomaly detection To enhance anomaly detection, a moving average filter with a window size of 5 consecutive frequency bins (equivalent to approximately 781 Hz) was applied to both the input and output FRFs. The influence of noise was reduced by subtracting 0.1 from the values, and negative results were set to zero. The anomaly score e was calculated using the following equation, similar to the previous equation.

[0071]

[0072] Here, fi and gi (1 < i < 192) are the input and reconstructed output frequency response functions, respectively. The threshold for anomaly determination was defined as the top 10% of the anomaly scores obtained from normal specimens, thereby ensuring approximately 90% confidence in normal classification. Specimens exceeding this threshold were classified as abnormal and were assumed to indicate potential screw loosening. A slight variation in the threshold had little impact on the results, indicating that precise tuning was not essential for the performance of the model.

[0073] The results of this study demonstrate that the proposed method is effective in detecting and classifying different degrees of screw loosening in various implant-abutment connection types. Visual interpretation of the frequency response function (FRF) alone has limited diagnostic value because the spectral changes associated with loosening are subtle and difficult to interpret clinically. Therefore, machine learning was applied to extract highly discriminative features from the FRF data for classification.

[0074] Supervised learning models using random forest and neural network algorithms classified screw looseness based on three-axis vibration patterns in a conical connection configuration. In binary classification, random forest achieved 100% accuracy for the first sample and 99.7% accuracy for the second sample (Table 6), while neural networks achieved 100% and 99.9% accuracy (Table 7).

[0075]

[0076]

[0077] In multi-class classification, the accuracy was 83.7% and 90.1% with random forests (Table 8), and 82.8% and 89.1% with neural networks (Table 9). Feature importance analysis showed that resonant frequencies along all spatial axes were consistently important for distinguishing the severity of loosening. Figure 18 shows the results of feature importance analysis using the random forest algorithm for conical-linked screw loosening detection after testing of intact and abnormal groups, with A for the first sample and B for the second sample. Figure 19 shows the results of feature importance analysis using the random forest algorithm for conical-linked screw loosening detection after testing seven screw loosening conditions, with A for the first sample and B for the second sample.

[0078]

[0079]

[0080] Unsupervised learning was performed using a CAE trained on triaxial frequency response and evaluated across seven screw loosening conditions. Figure 20 shows the detection of frequency response anomaly scores for various screw loosening angles in a conical splice configuration, with A for the first sample and B for the second sample. The anomaly scores shown in Figure 20 indicate the degree of deviation from the normal state. A threshold based on the top 10% of scores for normal samples identified the anomaly. Even a loosening of just 5° consistently exceeded this threshold, demonstrating the high sensitivity of the CAE to subtle changes in implant-abutment stability.

[0081] To evaluate screw loosening, supervised learning analyses using random forest and neural network models were performed on two internal tri-channel coupled samples. In binary classification, both models achieved high accuracy, with random forest achieving 100% and 99.9% (Table 10) and neural network achieving 100% and 99.7% (Table 11).

[0082]

[0083]

[0084] The confusion matrix confirmed reliable separation between normal and relaxed states. In multi-class classification, accuracy was achieved with random forests at 89.8% and 86.7% (Table 12), and with neural networks at 91.6% and 86.7% (Table 13).

[0085]

[0086]

[0087] Figure 21 shows the results of a feature importance analysis using a random forest algorithm for detecting internal tri-channel linked screw looseness after testing of intact and abnormal samples, where A represents the first sample and B represents the second sample. Figure 22 shows the results of a feature importance analysis using a random forest algorithm for detecting internal tri-channel linked screw looseness after testing seven screw looseness conditions, where A represents the first sample and B represents the second sample. These feature importance analyses (Figures 21 and 22) continued to emphasize the resonant frequencies across all axes as the most relevant features for detecting and classifying the severity of looseness.

[0088] Unsupervised learning analysis was performed using a CAE trained on the frequency responses of the X, Y, and Z axes. The CAE evaluated all screw loosening conditions, and the results are shown in Figure 23. Figure 23 shows the detection of frequency response anomaly scores for various screw loosening angles in an internal tri-channel linked configuration, with A for the first sample and B for the second sample. A threshold set at the top 10% of anomaly scores from the normal state played a role in identifying anomalies. In both samples, the 5° loosening condition consistently exceeded this threshold, demonstrating that the CAE is sensitive to subtle deviations from the normal state.

[0089] Supervised learning analysis using random forest and neural network models was performed on two external hexagonal connected samples. In binary classification, the random forest achieved accuracy of 99.4% and 98.3% (Table 14), while the neural network achieved accuracy of 99.8% and 99.6% (Table 15).

[0090]

[0091]

[0092] The confusion matrix confirmed reliable separation between normal and relaxed states. In multi-class classification, accuracy was achieved with random forests at 94.5% and 89.9% (Table 16), and with neural networks at 95.3% and 85.4% (Table 17).

[0093]

[0094]

[0095] Figure 24 shows the results of a feature importance analysis using a random forest algorithm for detecting external hexagonal linked screw looseness after testing of intact and abnormal samples, with A representing the first sample and B representing the second sample. Figure 25 shows the results of a feature importance analysis using a random forest algorithm for detecting external hexagonal linked screw looseness after testing seven screw looseness conditions, with A representing the first sample and B representing the second sample. The feature importance analysis (Figures 24 and 25) consistently showed that the most important features for detecting and evaluating the severity of screw looseness were the resonant frequencies along the X, Y, and Z axes.

[0096] An unsupervised learning analysis was performed using a CAE trained on frequency responses along the X, Y, and Z axes. The anomaly scores for all screw loosening conditions are shown in Figure 26. Figure 26 shows the detection of frequency response anomaly scores for various screw loosening angles in an external hexagonal coupling configuration, with A for the first sample and B for the second sample. A threshold set at the top 10% of the score from the normal state identified the anomaly, and the 5° loosening condition consistently exceeded this threshold.

[0097] Table 18 summarizes the CAE performance over 200 trials for each implant condition. Accuracy is defined as the percentage of correctly detected loosened specimens (384 specimens per trial). The CAE achieved nearly 100% accuracy for all implant types and loosening angles, and was clinically acceptable, albeit slightly lower at 92% for external hexagonal coupling. The coefficient of variation was approximately 7%, demonstrating strong consistency and reproducibility. These results support the potential of this unsupervised approach for reliable screw loosening detection across implant designs. Plotting the training and validation loss functions to confirm model convergence and assess the risk of overfitting showed stable convergence with no divergence between curves, indicating appropriate generalization to unseen data.

[0098]

[0099] Discussion: The null hypothesis that a combination of RFA and machine learning could not reliably detect screw loosening across different implant-abutment connections was rejected. This method provided early, accurate, and non-invasive screw loosening detection in conical, internal tri-channel, and external hexagonal designs. The RFA results revealed patterns specific to each connection type, which are thought to be due to the unique mechanical and structural properties of each implant system. Biomechanical studies have shown that these connections exhibit different torque-displacement responses, supporting the observed variations in vibrational behavior.

[0100] Conical joints employ a tapered design that creates a friction fit between the implant and abutment, providing large surface contact and a self-locking mechanism when fully tightened, thereby enhancing stability. Mechanical analyses using analytical and finite element methods have shown that engagement due to contact pressure and friction determines the pull-out force and loosening torque, and that even slight relative movement reduces friction, leading to decreased stiffness and increased vibrational oscillation. Furthermore, if the abutment is not fully seated, vertical displacement of 22.6 to 62.2 μm can occur, which reduces the engagement of the anti-rotation and anti-flexion functions and impairs stability. When loosening occurs, the taper allows small displacements between the implant and abutment, weakening frictional stability, increasing resonance amplitude, and reflecting a decrease in mechanical stiffness at the interface.

[0101] Internal tri-channel and external hexagonal connections rely on mechanical locking mechanisms such as screws or interlocking channels rather than friction, resulting in less surface contact than conical connections. Stability depends on the engagement of these locking elements and weakens when loosening occurs. Previous studies have reported that internal connections, due to their geometry and large contact area, exhibit larger minute displacements than external hexagonal connections when preload is lost. These minute displacements make internal connections more susceptible to deformation under functional load, amplifying vibration responses detectable by Periotest and resonant frequency analysis. In contrast, external hexagonal designs, with their smaller contact area and reliance on screw stabilization, exhibit less displacement and less pronounced vibration changes after loosening. These differences highlight how the geometry of the implant-abutment connection influences both mechanical stability and vibration behavior under screw loosening.

[0102] The high accuracy in detecting and classifying screw looseness may be due to several factors. Each implant-abutment coupling design exhibits unique resonant frequency patterns under different loosening conditions, providing highly discriminative input information for machine learning. Data from a 3-axis accelerometer captures subtle vibrations across all three axes, providing a rich set of multidimensional features. A random forest estimates feature importance by averaging impurity reductions across the entire tree, identifying key features and improving model performance. In this study, this algorithm achieved perfect accuracy in binary classification and 83% to 92% accuracy in multi-class classification. Feature importance consistently focused on resonant frequencies along the X, Y, and Z axes, supporting its usefulness in detecting varying degrees of looseness.

[0103] CAE provided valuable insights by detecting anomalies associated with screw loosening. Established for unsupervised anomaly detection, this method demonstrated its potential to complement supervised methods without labeled data, identifying deviations from normal conditions even with slight loosening (5°). Trained on X, Y, and Z axis frequency responses, CAE consistently detected subtle vibrational changes across all implant-abutment connection configurations. A threshold set at the top 10% of anomaly scores from normal specimens effectively distinguished between normal and abnormal conditions while minimizing false positives.

[0104] This study provided proof-of-concept evidence that screw loosening can be detected by combining RFA and machine learning. Clear vibration changes were observed as the severity of loosening increased, and the machine learning model demonstrated high classification accuracy. Such a method is also applicable to screw loosening detection in implant-abutment connection types not described above, i.e., connection types other than conical, internal tri-channel, and external hexagonal.

[0105] 3. Abnormality detection method for applying vibrational stimulation to fixed dental crowns using bone conduction speakers. Depending on the location of the fixed dental crown, it may be difficult to insert an exciter into the oral cavity. Also, although jaw models were used in the examples so far, there were cases where it was not possible to place an exciter such as an impactor in the appropriate location in the actual oral cavity. Therefore, we investigated a method of vibrating teeth using bone conduction speakers as an alternative to the tapping vibration method.

[0106] Figure 27 of the model experiment shows an embodiment of the model experiment. A 3D acceleration sensor 52 was attached to the implant of the mandibular model 50. The 3D acceleration sensor 52 was attached to the implant in both a state where the implant was not loose (healthy) and a state where it was loosened by about 20° (loose). A smartphone 58 was connected to the 3D acceleration sensor 52 via a signal conditioner 56. White noise in the range of 1,000-10,000 Hz was reproduced by a bone conduction speaker 54 and the tabletop was vibrated. White noise is a wave that uniformly contains the entire frequency range and has the same characteristics as impact. While impact is a wave that contains the entire frequency range in a short time, white noise is a wave that contains the entire frequency range and is applied continuously. White noise differs from impact in that it has superior vibration accuracy and can generate a steady signal, so a continuous response can be observed. The mandibular model 50 was placed near the bone conduction speaker 54. The bone conduction speaker 54 and the mandibular model 50 are not mechanically fixed together.

[0107] The bone conduction speaker 54 was repeatedly used to induce vibration and then stopped vibration on a desk. Vibration measurements were taken 90 times for the healthy state and 10 times for the relaxed state. In the machine learning study described later, the 80 measurements in the healthy state were used as training data, and the 10 measurements in the healthy state and the 10 measurements in the relaxed state were combined to create test data. The signal from the acceleration sensor 52 was captured by the smartphone 58 through the signal conditioner 56 and saved as an audio file in WAV format. The sampling frequency was set to 48,000 Hz.

[0108] Figure 28 shows an example of an acceleration waveform obtained from a healthy implant. The upper and lower panels in the figure represent the horizontal acceleration in the circumferential and perpendicular directions, respectively. The circumferential direction is the mesiodistal direction, and the perpendicular direction is the buccolingual direction. The vibration and stopping of a bone conduction speaker on a table were repeated about 10 times, and it was confirmed that the amplitude increased during vibration and decreased when stopped. This showed that vibrations can be excited in the implant by a bone conduction speaker. In this experiment, the acceleration in the perpendicular direction was greater than that in the circumferential direction, so in subsequent analyses, only the acceleration waveform in the perpendicular direction was frequency analyzed.

[0109] The response during excitation was extracted from the waveform of the healthy state shown in Figure 28 and subjected to FFT analysis. The number of FFT data plots was 512, and the result of averaging 30 frequency spectra is shown in Figure 29A. From Figure 29A, it was found that the variation from trial to trial was small, and the resonant frequency corresponding to the amplitude peak was 1980 Hz. Next, the frequency spectrum obtained from the loosened implant is shown in Figure 29B. In the loosened state, the frequency response in the 2000-3000 Hz range changed, and the resonant frequency decreased to 1290 Hz.

[0110] The results so far suggest the possibility of detecting implant anomalies by focusing on the resonant frequency, especially in cases of significant loosening of around 20°. Furthermore, we also considered anomaly detection of implants using machine learning (autoencoder). The autoencoder used a fully connected neural network, with the input layer providing the frequency spectrum as one-dimensional data (256 parameters). Through three hidden layers, the dimensionality was reduced to 128, 64, and 32 parameters, and then reconstructed into a 256-parameter frequency spectrum by the reverse process. The anomaly score of the test data (10 healthy data + 10 loose data) was evaluated using a trained model given 80 healthy data points. Figure 30 shows the anomaly scores for the healthy and loose data. The anomaly score on the vertical axis of Figure 30 is the input-output reconstruction error divided by the output data. From Figure 30, the anomaly scores for healthy data Nos. 1-10 were low, while the loose data (Nos. 11-20) showed high anomaly scores. Based on the above, we have confirmed the possibility of detecting implant abnormalities using bone conduction speakers and machine learning, as an alternative to conventional tapping test methods.

[0111] The experiment involved a bone conduction speaker being placed in contact with the jaw and cheek to stimulate fixed dental restorations through the human body. Therefore, an experiment was conducted on a subject with a missing front tooth and a three-unit bridge. Figure 31 shows a photograph of the experiment. The bone conduction speaker 54 was placed against the cheek and repeatedly stimulated and stopped, while an acceleration sensor 52 was placed on the three upper front teeth (bridge) by a dentist. Figure 32 shows the acceleration waveforms obtained by the sensor. The upper and lower sections represent acceleration in the circumferential and perpendicular directions, respectively. As illustrated in Figure 32, the acceleration in the perpendicular direction was not obscured by noise, allowing for the measurement of vibrations excited by the bone conduction speaker. On the other hand, the response in the circumferential direction was small, suggesting that a high-sensitivity sensor would be effective for measuring vibrations in that direction. Furthermore, the quality of the obtained vibrations can be improved by interposing a viscous "contact medium" between the sensor and the tooth.

[0112] The excitation response in the direction perpendicular to the circumference of Figure 32 was extracted, and the frequency spectrum shown in Figure 33 was obtained by FFT analysis. From Figure 33, it was found that the frequency characteristics of the subject's bridge can be clearly measured using a bone conduction speaker and an acceleration sensor.

[0113] Through model experiments and human trials, we proposed a prototype of a medical examination device combining a bone conduction speaker, an accelerometer, and a smartphone. This device allows for vibrational stimulation of fixed dental restorations by contacting the bone conduction speaker with the jaw or cheek, eliminating the need to insert an exciter into the mouth. Furthermore, by attaching a small accelerometer to a rod-shaped material such as a dental mirror, it is expected that vibration characteristics can be easily measured by contacting the accelerometer with the molars.

[0114] The above-mentioned examination device and fixation evaluation method using bone conduction speakers is just one example; other examination devices and fixation evaluation methods can be used. For example, a system with bone conduction speakers, an accelerometer, and a smartphone contributes to the construction of an examination device with reduced total cost. A pre-trained model is implemented on the smartphone, and transfer learning data reflecting the patient's characteristics is added to the smartphone for tuning. The computational resources of the smartphone are sufficient for tuning. By using a device that connects a 3-axis accelerometer to the smartphone, the accuracy of the examination can be further improved.

Claims

1. A method for evaluating the fixation state of a fixed dental crown, comprising: applying elastic wave stimulation to the fixed dental crown; acquiring the response waveform of the fixed dental crown to the elastic wave stimulation as a time history waveform; evaluating the frequency response characteristics from the time history waveform; inputting the frequency response characteristics into an anomaly detection model and obtaining an anomaly score as an inference result; and determining the fixation state of the fixed dental crown from the anomaly score.

2. The method for evaluating a fixed state according to claim 1, wherein the elastic wave stimulation is provided by an exciter.

3. The method for evaluating the fixed state according to claim 1, wherein the elastic wave stimulation is provided by directly vibrating using a striking device or by vibrating a bone conduction speaker in contact with the jaw or cheek.

4. The method for evaluating the fixed state according to claim 1, wherein a three-dimensional acceleration sensor, piezoelectric sensor, or magnetostrictive sensor in contact with the fixed crown restoration, or a non-contact laser vibrometer, is used to acquire the response waveform.

5. The fixed state evaluation method according to claim 1, wherein the anomaly detection model is trained using a supervised machine learning algorithm that uses only acceleration data in a specific direction.

6. The fixed state evaluation method according to claim 1, wherein an unsupervised learning algorithm is used to train the anomaly detection model.

7. The fixed-state evaluation method according to claim 1, further comprising calculating feature importance using a random forest algorithm.

8. The fixed state evaluation method according to claim 7, wherein the importance of the features is displayed as a graph or ranking.

9. The method for evaluating the fixation state according to claim 1, wherein the determination of the fixation state determines whether it is normal or abnormal, and normal indicates that the fixed crown restoration is properly fixed to the abutment tooth or implant body, and abnormal indicates that there is an abnormality in the fixation state of the fixed crown restoration to the abutment tooth or implant body.

10. A diagnostic device for fixed dental crowns, comprising: an exciter that applies elastic wave stimulation to a fixed dental crown attached to the fixed dental crown; a sensor that acquires the response waveform of the fixed dental crown; and an AI diagnostic device that acquires the output of the sensor and outputs a large abnormal score if the retention force of the fixed dental crown to the abutment tooth or implant is small.

11. The medical examination device according to claim 10, wherein the AI ​​diagnostic device outputs feature importance.

12. The medical examination apparatus according to claim 10, wherein the exciter is a striking device or a bone conduction speaker, and the sensor is a three-dimensional acceleration sensor, a piezoelectric sensor, a magnetostrictive sensor, or a laser vibrometer.

Images