Intelligent Dental Handpiece: Real-Time Motion Analysis for Skill Development

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper presents the results of a study of an intelligent dental handpiece developed by the authors. The developed tool is intended for educational purposes, but it can also be useful for specialists. The idea of the development is to include a motion sensor in the instrument with subsequent processing based on machine learning methods for observing and classifying hand movements. Three models were considered in the work. The best result was recorded by a classifier based on a random forest. In general, the methods and tools used in the study are presented in sufficient detail. The results obtained expand and complement the existing achievements in this field of research.
In general, the work corresponds to the subject of the journal, but there are the following comments and questions:
1. In some places of work, it is said about the classification of hand movements (for example, line 6-7), fixation of both movements and behavioral data (for example, line 59-60). Movement is the process of moving through time, and not always along a linear trajectory. In classifiers, the input vector is: identifier, state, time, deviation, roll, pitch, yaw, acceleration, and velocity. The paper does not specify which size of the input vector is being analyzed. If this vector is analyzed at a moment, then this is not entirely correct in relation to the concept of "motion". If a time series is analyzed (i.e., changes in these parameters over time), then the size of the vector (the size of the time window of the analyzed parameters) is not clear. How was the beginning and the end of the movement determined?
2. It is not clear from the work which types (types) of movements are analyzed and classified. It would be clearer if the paper presented examples (time series) of each class of movements.
3. It is not clear from the work how the data was marked up into positive and negative classes – which classification criterion (or methodology). If the movement trajectories are known, is machine learning necessary? Is it possible to use threshold processing of parameter values?
4. The paper talks about the real-time mode, but it does not provide an estimate of the delay time for conversion, transfer to the cloud, classification and output of the result. What sampling rate is used?
Overall assessment: the presented research results are relevant and interesting. However, the article would look better if the issues mentioned in the comments were discussed in more detail.

Author Response

Reviewer 1 Comments

Paper Title: AI-Enhanced Dental Handpiece with Real-Time Motion Tracking for Precision Dental Training

Dear Reviewer,

We sincerely thank you for your thorough review and valuable feedback. We have carefully considered each comment and provide detailed responses below, along with the corresponding revisions made to the manuscript.

• In some places of work, it is said about the classification of hand movements (for example, line 6-7), fixation of both movements and behavioral data (for example, line 59-60). Movement is the process of moving through time, and not always along a linear trajectory. In classifiers, the input vector is: identifier, state, time, deviation, roll, pitch, yaw, acceleration, and velocity. The paper does not specify which size of the input vector is being analyzed. If this vector is analyzed at a moment, then this is not entirely correct in relation to the concept of "motion". If a time series is analyzed (i.e., changes in these parameters over time), then the size of the vector (the size of the time window of the analyzed parameters) is not clear. How was the beginning and the end of the movement determined?

Response: Thank you for your valuable comment. The device captures sensor data at a sampling rate of one sample per second. Each motion segment is evaluated as a short time series comprising consecutive sensor readings. To precisely identify the start and end of each movement, a manual trigger mechanism is employed: the practitioner activates a control button on the handpiece to mark the onset of the motion, and the release of the button designates the motion's conclusion. This method establishes clear temporal boundaries for each motion instance, enabling the model to analyze the complete movement sequence rather than isolated static readings. A modification was made to the manuscript and the data were added to the prototype subsection.

• It is not clear from the work which types (types) of movements are analyzed and classified. It would be clearer if the paper presented examples (time series) of each class of movements.

Response: Thank you for your valuable comment. The device movements are simple and basic which displays the roll tilt deviation with smooth motion, moving naturally as the device tilts left or right

• It is not clear from the work how the data was marked up into positive and negative classes – which classification criterion (or methodology). If the movement trajectories are known, is machine learning necessary? Is it possible to use threshold processing of parameter values?

Response: Thank you for your valuable comment. The data were labeled automatically based on the deviation angle from the target. We defined three ranges of motion as follows:

0°–10°: Normal operation (safe range)

10°–15°: Warning zone (requires attention)

>15°: Critical deviation (task should be stopped).

These thresholds were used for labeling, where safe readings were considered positive and critical deviations negative.

 

Although threshold-based classification can separate the data, it cannot capture the overall motion behavior or hand stability trends over time. Therefore, we used machine learning to analyze temporal patterns, detect subtle fluctuations, and provide a more accurate assessment of the practitioner’s skill level.

• The paper talks about the real-time mode, but it does not provide an estimate of the delay time for conversion, transfer to the cloud, classification and output of the result. What sampling rate is used?

Response: Thank you for your valuable comment. The system operates in real time with a total latency below one second, covering sensor reading, processing, and display.
Data are sampled at around 50 Hz to ensure smooth roll motion feedback without noticeable delay. Actions are taken locally through visual feedback on the screen or auditory feedback via the buzzer. The cloud was used only for data collection and storage, not for any real-time processing or decision-making.

Reviewer 2 Report

Comments and Suggestions for Authors

The paper “AI-Enhanced Dental Handpiece with Real-Time Motion Tracking for Precision Dental Training” presents an interesting application of AI and sensor technology to dental training. The presented concept seems to serve as real-time feedback to students, targeting training activities.

The paper is well written and organized. There are a few issues to deal with before publication.

Perhaps the authors can extend the state of the art and discuss on alternative solutions, such as: https://link.springer.com/chapter/10.1007/978-3-031-59257-7_9

The authors state that they have used 61 practitioners to collect 3720 data records. How can you relate the procedure specifics (for each patient) with the recorded parameters values, so you can evaluate the student activity? Or did they used the same model for training your network?

I don’t see the utility of sub-section 4.1.2, since there are no missing values.

It is not readily understandable how the proposed parameters will be used to asses the students’ skills: Time, Deviation, Roll, Pitch, Yaw, Acceleration, and Velocity.

Regarding the fin-tunning of the selected models, the authors have used hyperparameter tuning with randomized search and cross-validation. Do you think and perhaps comment on the fact that others can be used as well? I.e. feature engineering, regularization, data augmentation, etc.

Author Response

Reviewer 2 Comments

Paper Title: AI-Enhanced Dental Handpiece with Real-Time Motion Tracking for Precision Dental Training

Dear Reviewer,

We sincerely thank you for your thorough review and valuable feedback. We have carefully considered each comment and provide detailed responses below, along with the corresponding revisions made to the manuscript

• Perhaps the authors can extend the state of the art and discuss on alternative solutions, such as: https://link.springer.com/chapter/10.1007/978-3-031-59257-7_9

Response: Thank you for your chapter recommendation. However, the scope of the proposed chapter doesn't align with the proposed work.

 

• The authors state that they have used 61 practitioners to collect 3720 data records. How can you relate the procedure specifics (for each patient) with the recorded parameters values, so you can evaluate the student activity? Or did they used the same model for training your network?

     Response: Thank you for your valuable comment. All tasks performed were standardized drilling exercises that only differed by the target angle. Therefore, a single model was used for all cases. Performance evaluation was based on how many times and by how much the practitioner exceeded the allowed deviation range. This approach allowed consistent assessment of skill level across all participants.

• I don’t see the utility of sub-section 4.1.2, since there are no missing values.(yousef)

Response: Thank you for your observation. The purpose was to eliminates the need for imputation, if you think we are pointing out the obvious, then let us know and we can remove the subsection.

 

• It is not readily understandable how the proposed parameters will be used to asses the students’ skills: Time, Deviation, Roll, Pitch, Yaw, Acceleration, and Velocity.

Response: Thank you for the valuable note. The relationship between tilt and skill is based on the practitioner’s ability to keep a steady hand throughout the procedure. Any roll deviation indicates hand instability, which causes the bur to drift from the target angle and may lead to case failure. Therefore, parameters such as deviation, roll, and pitch directly reflect control and precision. Time, velocity, and acceleration further describe the smoothness and consistency of motion, allowing an objective evaluation of skill level.

• Regarding the fin-tunning of the selected models, the authors have used hyperparameter tuning with randomized search and cross-validation. Do you think and perhaps comment on the fact that others can be used as well? I.e. feature engineering, regularization, data augmentation, etc.

Response:

Thank you for the thoughtful comment. In this study, model fine-tuning was primarily conducted through randomized hyperparameter optimization combined with stratified cross-validation to ensure robust generalization. For the Neural Network model, early stopping and dropout regularization were also implemented to prevent overfitting.

Reviewer 3 Report

Comments and Suggestions for Authors

The research focused on hand movement recognition to facilitate real-time feedback to enhance dental training. The topic sounds interesting and is worth investigating. However, there are many key issues to be addressed before recommending acceptance. Please refer to my comments below.
Comment 1. Paper Title: The title is not precise because one of the core elements is hand movement recognition. Some descriptions (e.g., AI-enhanced dental handpiece and motion tracking) do not fully align with the main theme of the research content.
Comment 2. Abstract:
(a) In later sections, the dataset was not newly collected, but it was extracted from a previous work.
(b) What is the research novelty in terms of the machine learning model, given that motion recognition, hand movement recognition, etc., are well-established research topics?
(c) Compare your work with existing methods (other machine learning and deep learning approaches).
Comment 3. More terms should be added to the “Keywords” to better reflect the scope of the paper.
Comment 4. Section 1 Introduction:
(a) Various commercial simulation tools were summarized and compared. If there exist commercial-grade tools, the reasons for using AI-based approach should be strengthened.
(b) A literature review was missing. Discuss the methodologies, results, and limitations of the existing works on AI algorithms for hand movement or motion recognition.
(c) The last three paragraphs should be revised to strengthen the descriptions of the research merits and contributions. In addition, please discuss how your work outperformed other AI algorithms.
Comment 5. Section 2 Prototype Design:
(a) How did your prototype represent the real-world scenarios? Indeed, it is still not a real and commercial product, so the commercial tools are also good enough. Therefore, a strong justification and sufficient explanation are needed.
(b) Regarding all figures (e.g., Figure 4), zoom in on your file to 200% to confirm that no content is blurred.
(c) The captions of Figures 3 and 4 are not precise.
(d) Figure 4 is too general. It is an incomplete system diagram.
(e) What are the constraints and limitations in your prototype?
Comment 6. Section 3 Dataset:
(a) Why was dataset [23] being chosen? Are there other benchmark datasets being used in the research topic? Please update the content in Reference [23] because the data cannot be found.
(b) Some details were missing, including the number of classes, description of each class, class size, etc.
Comment 7. Section 4 Methodology:
(a) As hand movement or motion recognition, deep learning-based approaches (including the feature extraction process) are very common. Please justify the reasons for using traditional machine learning algorithms.
(b) Repeat your experiment with k-fold cross-validation, choosing k=5 because you used an 80/20 ratio already.
(c) Justify the selections of SVM, LR, and RF. Particularly, how did you fine-tune each model? Subsection 4.2.1 and Table 2 insufficiently shared the details.
Comment 8. Section 5 Experiments:
(a) Other than the linear kernel function, many other non-linear kernel functions must be used. Using the linear kernel function is unfair because it usually achieves a less accurate model.
(b) Results of models when fine-tuning different parameters should be presented.
(c) Compare your results with the existing algorithms. Provide in-text citations for the existing studies being compared.
Comment 8. Discuss the research implications of your work. How can it link to insights, training, etc?
Comment 9. No results related to time complexity are provided to justify “Real-time” as in the paper title.

Author Response

Reviewer 3 Comments

Paper Title: AI-Enhanced Dental Handpiece with Real-Time Motion Tracking for Precision Dental Training

Dear Reviewer,

We sincerely thank you for your thorough review and valuable feedback. We have carefully considered each comment and provide detailed responses below, along with the corresponding revisions made to the manuscript

Comment 1.1 

Paper Title: The title is not precise because one of the core elements is hand movement recognition. Some descriptions (e.g., AI-enhanced dental handpiece and motion tracking) do not fully align with the main theme of the research content.

Response: Thank you for your feedback. The title has been modified.

-------------------------------------------
Comment 1.2

Abstract:
(a) In later sections, the dataset was not newly collected, but it was extracted from a previous work.
(b) What is the research novelty in terms of the machine learning model, given that motion recognition, hand movement recognition, etc., are well-established research topics?
(c) Compare your work with existing methods (other machine learning and deep learning approaches).

Response: Thank you for your comment. For the dataset, no it’s newly made specifically for this research, and the link and DOI lead to our Mendeley repo. There are no similar tool or datasets in the same area of research.

 

Comment 1.3

More terms should be added to the “Keywords” to better reflect the scope of the paper.

Response: Thank you for your feedback. More keywords were added.

Comment 1.4

 Section 1 Introduction:
(a) Various commercial simulation tools were summarized and compared. If there exist commercial-grade tools, the reasons for using AI-based approach should be strengthened.
(b) A literature review was missing. Discuss the methodologies, results, and limitations of the existing works on AI algorithms for hand movement or motion recognition.
(c) The last three paragraphs should be revised to strengthen the descriptions of the research merits and contributions. In addition, please discuss how your work outperformed other AI algorithms.

Response: Thank you for your valuable comment. Other existing research is present in the scope of simulators or augmented reality. This makes our research and product unique and incomparable when it comes to methodology.

Comment 1.5

Section 2 Prototype Design:
(a) How did your prototype represent the real-world scenarios? Indeed, it is still not a real and commercial product, so the commercial tools are also good enough. Therefore, a strong justification and sufficient explanation are needed.
(b) Regarding all figures (e.g., Figure 4), zoom in on your file to 200% to confirm that no content is blurred.
(c) The captions of Figures 3 and 4 are not precise.
(d) Figure 4 is too general. It is an incomplete system diagram.
(e) What are the constraints and limitations in your prototype?

Response: Thank you for your valuable comment.

• There are currently no available commercial tools that can perform the same real-time motion analysis as our proposed device. The developed prototype is capable of understanding and interpreting every movement made by the student. It can be easily attached to any dental handpiece, whether used for preparation or implant procedures, allowing the same natural hand posture and movement as in real clinical scenarios. Therefore, the prototype successfully replicates real-world practice conditions and enables students to train effectively without the need to be physically present in the clinic.
• (C) (d) Figure 4 has been remade and Figure 3 caption has been remade.

(e) The main limitation of the current prototype is its size. Although it does not cause any obstruction during operation, reducing the overall dimensions would make it more comfortable to attach to the handpiece. This improvement would enhance usability and make the device more practical for long training sessions.

Comment 1.6

 Section 3 Dataset:
(a) Why was dataset [23] being chosen? Are there other benchmark datasets being used in the research topic? Please update the content in Reference [23] because the data cannot be found.
(b) Some details were missing, including the number of classes, description of each class, class size, etc.

Response: Thank you for your valuable comment. The dataset in [23] is our created dataset for this research. The dataset was tailored to accommodate the research needs in order to create a model to train the dental students. The dataset subsection has been revised to include the missing details. Specifically, the number of classes, corresponding sample sizes, percentages, and the deviation angle ranges defining each class have been added.

---------------------------------------------------------------
Comment 1.7

 Section 4 Methodology:
(a) As hand movement or motion recognition, deep learning-based approaches (including the feature extraction process) are very common. Please justify the reasons for using traditional machine learning algorithms.
(b) Repeat your experiment with k-fold cross-validation, choosing k=5 because you used an 80/20 ratio already.
(c) Justify the selections of SVM, LR, and RF. Particularly, how did you fine-tune each model? Subsection 4.2.1 and Table 2 insufficiently shared the details.

Response:

(a) It is agreed that deep learning techniques are widely used for motion recognition. In the revised manuscript, a Neural Network (MLP) model has been added to represent a deep learning approach. Traditional machine learning algorithms (RF, SVM, LR) were still included for their interpretability, computational efficiency, and suitability for structured IMU data.
(b) All experiments were repeated using stratified 5-fold cross-validation to ensure robustness and consistency. This change is reflected in the methodology and results sections.

Comment 1.8

 Section 5 Experiments:
(a) Other than the linear kernel function, many other non-linear kernel functions must be used. Using the linear kernel function is unfair because it usually achieves a less accurate model.
(b) Results of models when fine-tuning different parameters should be presented.
(c) Compare your results with the existing algorithms. Provide in-text citations for the existing studies being compared.

Response:

We appreciate this valuable observation. In the revised version, three major SVM kernel functions (Linear, Radial Basis Function [RBF], and Polynomial) were implemented and evaluated. This addition ensures a fair comparison between linear and non-linear decision boundaries.

We present comprehensive CV results showing mean scores, standard deviations, and coefficients of variation for all models, demonstrating performance stability across parameter configurations.

Comment 1.9

Discuss the research implications of your work. How can it link to insights, training, etc?
Response: Thank you for your valuable feedback. The research implications of our work are multifaceted. By embedding motion tracking and AI into a dental handpiece, we offer a novel way to objectively assess and improve manual dexterity in dental training. The system provides real-time feedback, helping learners recognize and correct technique deviations, which enhances both skill acquisition and confidence. Beyond training, the insights from motion data, especially features like deviation and take time which can inform personalized learning paths, curriculum design, and even future developments in robotic-assisted dentistry. We believe this approach bridges the gap between simulation and clinical readiness, making dental education more data-driven and adaptive.

Comment 1.10

No results related to time complexity are provided to justify “Real-time” as in the paper title.

Response: Thank you for your valuable comment. According to the datasheets of the components used in the system, the sensor operates at a sampling rate of 50 Hz, and the microcontroller processes and displays the readings almost instantly.
The total delay from data acquisition to feedback (visual or auditory) is less than 200 ms, confirming that the system performs in real time.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have improved the paper.

Nevertheless, there are a few other things to consider.

First, the authors must acknowledge the limitations of the proposed methodology, since the end-results of the training exercise depends on other parameters as well, such as applied pressure. Which means that only after analysing the result of their work (which must be standard for all trainees), the proposed system could provide an evaluation on the dexterity,  speed, etc. of the student.

I think sub-section 4.1.2 can be removed.

There are some issues with the tables references which should be checked.

Author Response

Reviewer (2)

(1) First, the authors must acknowledge the limitations of the proposed methodology,

since the end-results of the training exercise depends on other parameters as well,

such as applied pressure. Which means that only after analysing the result of their

work (which must be standard for all trainees), the proposed system could provide

an evaluation on the dexterity, speed, etc. of the student.

Response: Thank you for the insightful comments. A detailed paragraph in the

Future Work subsection has been added to discuss the limitations of the

methodology.

(2) I think sub-section 4.1.2 can be removed.

Response: This section has been removed

(3) There are some issues with the tables references which should be checked.

Response: Thank you for the observation. Upon review, all table references have

been revised to ensure that each table is clearly introduced.

Reviewer 3 Report

Comments and Suggestions for Authors

Various key comments have not been addressed. It is noted that the authors responded in several comments that changes were made; however, no changes (or highlighted in yellow) were observed.
Comment 1. Abstract:
(a) Clearly state which kinds of motions (and the number of classes) were considered in your experiments.
(b) Present the names of machine learning algorithms being evaluated and compared.
(c) Provide numeric results in a range because you have experimented with several algorithms.
Comment 2. The terms being used in the “Keywords” do not fully reflect the scope of the paper. Many terms are too general.
Comment 3. Section 1 Introduction:
(a) Paragraph 4: Incomplete content “Table ??”.
(b) Paragraph 4: Define the acronyms VRDTS and IDSS because they first appear in the main text.
(c) The authors made no changes to explain the differences between your and the commercial tools. Perhaps the authors could consider adding the content related to your work in Table 1.
(d) The research novelties were not updated in the last three paragraphs.
(e) A literature review was missing. Discuss machine learning and deep learning algorithms for motion analysis and hand movement detection. Regardless of the applications (whether it is in dentistry), the motion or hand movement recognition shares very similar ideas across different applications.
Comment 4. Section 2 Prototype Design:
(a) Correct the typos “sampling 75”, “time series 76”, “each 77 movement”, etc.
(b) What are the advantages and unique characteristics of your design compared to commercial tools?
Comment 5. Since the authors collected a new dataset, full details must be presented. In addition, provide visualization of what signals were collected.
Comment 6. Without showing the details of the methodology (e.g., how to fine-tune, define grid, define step size, and optimally design each algorithm), the results will be considered insufficient for fair evaluation.

Author Response

Reviewer (3)

Comment 1. Abstract:
(a) Clearly state which kinds of motions (and the number of classes) were considered in your experiments.
(b) Present the names of machine learning algorithms being evaluated and compared.
(c) Provide numeric results in a range because you have experimented with several algorithms.

 

Response: Your constructive feedback on the abstract is greatly appreciated, and the abstract has been modified to apply your feedback.

Comment 2. The terms being used in the “Keywords” do not fully reflect the scope of the paper. Many terms are too general.

Response: The keywords have been modified to reflect the specific scope of the paper

Comment 3. Section 1 Introduction:

(a) Paragraph 4: Incomplete content “Table ??”.

Thank you for your observation, it has been revised.

(b) Paragraph 4: Define the acronyms VRDTS and IDSS because they first appear in the main text.

Response: The manuscript has been revised and acronyms have been defined

(c) The authors made no changes to explain the differences between your and the commercial tools. Perhaps the authors could consider adding the content related to your work in Table 1.

Response: A new row to Table 1 has been added to compare our work to others.

(d) The research novelties were not updated in the last three paragraphs.

Response: Thank you for the insightful comment. The research novelties have been added to the Conclusion section, emphasizing the unique integration of motion-tracking, intelligent feedback, and data-driven performance evaluation in the proposed dental handpiece system.

(e) A literature review was missing. Discuss machine learning and deep learning algorithms for motion analysis and hand movement detection. Regardless of the applications (whether it is in dentistry), the motion or hand movement recognition shares very similar ideas across different applications.

Response: Thank you for the valuable suggestion. In response, a literature review paragraph has been added in the Introduction to discuss recent machine learning and deep learning approaches.

Comment 4. Section 2 Prototype Design:
(a) Correct the typos “sampling 75”, “time series 76”, “each 77 movement”, etc.

Response: Typos have been revised

(b) What are the advantages and unique characteristics of your design compared to commercial tools?

Response: The main advantages of the IDH compared to commercial tools are:

• real-time motion classification,
• sensor-based feedback without bulky haptic hardware,
• machine learning model comparison to provide data-driven foundation for selecting the most accurate and efficient model,
• open dataset contribution,
• cloud-based analytics,
• Compact and cost-efficient design

A detailed paragraph has been added to highlight these features (in the Introduction, right after Table 1)

Comment 5. Since the authors collected a new dataset, full details must be presented. In addition, provide visualization of what signals were collected.

Response: Thank you for the valuable comment. The visualization of signals was added.

Comment 6. Without showing the details of the methodology (e.g., how to fine-tune, define grid, define step size, and optimally design each algorithm), the results will be considered insufficient for fair evaluation.

Response: Thank you for the insightful comment. The full details of the hyperparameter tuning methodology, including fine-tuning procedure, parameter grids, step sizes, and optimization strategy for each algorithm, have been added.