Surgical Tool Handle Vibration‐Based Drilling State Recognition During Hip Fracture Fixation

Abstract

Objectives

Traditional manual drilling during hip fracture fixation can easily lead to unstable fixation and vascular damage. This study aimed to investigate a safe and easy‐to‐use robot‐assisted method to automatically drill bone and distinguish critical bone drilling states with high accuracy in real‐time for the bone hole‐making process during hip fracture fixation.

Methods

A bone‐drilling robotic system was designed to automatically create holes in the femoral neck. Four fresh pig femurs were drilled at the posterosuperior femoral neck using three modes: “all‐in” (AI), “in‐out‐in” (IOI), and “percutaneous fixation” (PF). A high‐frequency accelerometer captured the generated vibrations of the drill handle, which were then transferred to a personal computer using a data acquisition card. Five bone drilling states are defined, including: “drill idling,” “initial drilling,” “in the cancellous bone,” “out the femoral neck,” and “in the cortical bone.” The harmonic distribution of the vibration signal was extracted by fast Fourier transform (FFT) and used as a critical feature to identify different drilling states. To prove the difference in the harmonic distribution at different drilling states, an independent sample t‐test was used to compare the percentage of the first harmonic amplitude in the first 10 harmonics at each drilling state. A neural network classifier was trained with the frequency spectrum as the input and the drilled state as the output to distinguish the critical bone drilling states with high accuracy in real‐time. The classifier was trained and tested on four specimens to ensure that the surgical robot could accurately identify the five drilling states.

Results

In each specimen, the harmonic distributions of the drilling vibration at different drilling modes were significantly different (p < 0.05). The average recognition accuracies of the drilling state for the four specimens were all higher than 84%. The three defined modes were distinguished with extremely high accuracies. The recognition accuracies of “in the cancellous bone” for specimens 1 to 4 were 83.2%, 84.8%, 92.9%, and 84.7%. The recognition accuracies of “in out the femoral neck” from specimens 1 to 4 are 98.2%, 88.4%, 95.8%, and 88.8%. The recognition accuracies of “in the cortical bone” for specimens 1 to 4 were 94.6%, 80.8%, 95.5%, and 85.8%.

Conclusions

The proposed robot‐assisted method can automatically distinguish five critical bone‐drilling states with high accuracy in real‐time to avoid weak fixation and damage to the lateral epiphyseal artery.

In this manuscript, a safe and easy‐to‐use robot‐assisted method is proposed to automatically drill bone and distinguish critical bone drilling states with high accuracy in real‐time for the bone hole‐making process during the fixation of the femoral neck.

Introduction

Hip fractures have a high incidence worldwide, ¹ and the treatment choice for most hip fractures is operative, ² usually involving in situ fixation with multiple cannulated screws or a sliding hip screw (SHS). ³ , ⁴ , ⁵ For most stable fracture patterns, multiple cannulated screws provide a relatively minimally invasive technique, shorter operative time, and sufficient fixation capacity than SHS. ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ If multiple cannulated screws and SHS are both appropriate for fracture fixation, screws should be used because they are less expensive. Therefore, multiple cannulated screws are crucial for a stable hip fracture.

Three cannulated screws are inserted in a parallel inverted triangle configuration (inferior, posteromedial, and anterosuperior) because they can provide high mechanical stability. ¹¹ , ¹² , ¹³ , ¹⁴ , ¹⁵ Moreover, surgeons usually place the cannulated screws tightly on the posteromedial cortex and try to keep the screws inside the cortical bone, which provides better support and stabilization of the fracture area. ¹⁶ , ¹⁷

For fracture repositioning and fixation stability, surgeons place cannulated screws as close to the posteromedial cortex as possible; however, this operation is often performed using intraoperative fluoroscopy and is based on their own experience. ¹⁸ An experienced surgeon can attach and even attempt to place the cannulated screws into the posteromedial cortex. However, some surgeons lack experience, and the accuracy varies owing to different personal experiences and inconsistencies in operations. If the posteromedial nail is placed completely into the cancellous bone, it will not provide sufficient stability and is predisposed to the collapse of the femoral neck and fixation failure. ¹⁹ Moreover, a deviated nail can damage the surrounding vessels, such as the lateral epiphyseal artery, which can lead to femoral head necrosis.

In recent years, orthopaedic‐assisted surgical robotics has been introduced to improve surgical safety. ²⁰ , ²¹ , ²² , ²³ , ²⁴ Currently, the application of orthopaedic‐assisted surgical robotics is confined to passive system functions, such as navigation and preoperative planning. However, an increasing number of surgeons are complaining about the robot's lack of semi‐active and active functions, such as haptic feedback. ²⁵ , ²⁶ , ²⁷ , ²⁸ Therefore, we designed a safe and easy‐to‐use robot‐assisted method for automatic bone drilling and to distinguish critical bone drilling states with high accuracy in real‐time for the bone hole‐making process during hip fracture fixation.

The contributions of this investigation are as follows: (i) investigate a safe and easy‐to‐use robot‐assisted method for automatic bone drilling; (ii) use the drilling vibration signal to distinguish critical bone drilling states with high accuracy in real‐time for the bone hole‐making process during fixation of the femoral neck; (iii) apply the proposed method to avoid weak fixation and damage to the lateral epiphyseal artery during femoral neck fixation.

Materials and Methods

Drilling Mode Analysis During Fixation of Femoral Neck Fractures

To place the screw, the orthopaedist must first drill holes from the sidewall surface of the femur to the femoral head, as shown in Figure 1A. These holes passed through the entire femoral neck in an inverted triangle configuration. The screw is usually abutted against the cortical bone to improve the stability in the posterosuperior femoral neck. Figure 1B shows three possible drilling modes in the posterosuperior femoral neck: all in the cancellous bone (called “all‐in,” AI), through the cortical bone (called “percutaneous fixation,” PF), and out the femoral neck (called “in‐out‐in,” IOI). The AI mode has a weak fixation construct. Therefore, surgeons should better distinguish between the AI and PF modes to avoid weak fixation. IOI mode may damage the lateral epiphyseal artery, which can cause femoral head necrosis. Therefore, the drilling operation must be stopped immediately when the drill end breaks through and out of the femoral neck to prevent harm to the artery.

FIGURE 1.

Three drilling modes during fixation of femoral neck fractures

Drilling Vibration Signal Acquisition

An experimental setup was designed to acquire the drilling vibration signals, as shown in Figure 2. The bone‐drilling robotic system consists of one rotation joint (TBR100, ZoliX, China) and two translation joints (PSA200, ZoliX, China). The rotation joint (R‐axis) is used to adjust the drilling angle, and two translation joints (X‐axis and Y‐axis) can realize the linear motion of the drill along the adjusted angle. The drill is mounted at the end of the rotation joint by a specially designed installation unit to simulate the surgeon's hand and prevent potential resonance between the drill and robot arm. A high‐frequency accelerometer (A1, PCB, USA) is glued to the surface of the drill handle. The power device (Minimo, Japan) can provide a max speed of 30,000 rounds per minute (RPM) to the drill. Four pig femurs were harvested from 6‐month‐old pigs (weight range, 24–30 kg; two females and two males) and drilled in this study, and each bone was fixed on a vise. The pig model was chosen because of its similarity to human femurs and its ease of acquisition. The bone‐drilling robot was employed to drill each femur three times to contain AI, PF, and IOI drilling modes. The accelerometer captured the generated vibrations of the drill handle at a sampling frequency of 30 kHz. The captured vibrations were then transferred to a personal computer using an NI USB‐4431 data‐acquisition card.

FIGURE 2.

Experimental setup

Vibration Signal Processing

Figure 3 shows the raw drilling vibration signals captured at three different drill modes. Unlike in actual surgery, we can easily use dissections to observe drill states during in vitro drilling. Before the orange vertical line, vibrations were captured during the drilling of the femur. The vibrations after the orange vertical line were captured when drilling the femoral neck and head. The captured vibrations are usually unstable during the initial drilling phase (between the blue and yellow lines). When the drilling vibration reached a stable level, the drill was observed in the cancellous bone of the femoral neck in the three different drilling modes. In the AI mode, the drill is further drilled into the cancellous bone of the femoral neck and head. In the IOI mode, the drill is first drilled out of the femoral neck and then into the cancellous bone of the femoral head. In the PF mode, the drill is further drilled into the cortical bone of the femoral neck and then into the cancellous bone of the femoral head. It can be intuitively seen that during the drilling out of the femoral neck, the absolute vibration amplitude of the drill handle will decrease slightly, and when the drill is in the cortical bone of the femoral neck, the absolute vibration amplitude of the drill handle will have a significant increase.

FIGURE 3.

Raw drilling vibration signals captured at three drill modes

However, the absolute amplitude is not a stable value for distinguishing the drilling state because it is susceptible to interference from the drilling conditions, such as the drilling feed speed. The vibration signal is composed of a series of harmonic components. The frequency of these harmonics is an integer multiple of the drilling rotation frequency. Therefore, in this study, the harmonic distribution was used to identify the drilling characteristics. Fast Fourier Transformation (FFT) is an effective method for analyzing the harmonic distribution of a signal. Figures 4, 5, 6 show the FFT amplitude spectrum of the captured vibrations during three different drilling states: “in the cancellous bone,” “out the femoral neck,” and “in the cortical bone.”

FIGURE 4.

Drilling vibration of 0.1 s and its FFT in the cancellous bone

FIGURE 5.

Drilling vibration of 0.1 s and its FFT out the femoral neck

FIGURE 6.

Drilling vibration of 0.1 s and its FFT in the cortical bone

The sampling frequency F _s of the vibration signal used in this study is 30 kHz, and the duration T _s of each analyzed frame of vibration signal is 0.1 s, so the length of one frame of vibration signal Nis 30000. The vibration signal is represented as S _1 × N  = [S(1) S(2) … S(N)], and the signal after the FFT calculation is recorded as F _1 × N  = [F(1) F(2) … F(N)]. Considering that the rotational frequency of the drill may be attenuated by drilling the bone tissue, the robust i ^th harmonic amplitude A _i can be calculated as follows:

$$r=F_{\mathrm{c}}\times T_{\mathrm{s}}=500\times 0.1=50$$ (1)

$$A_i=\sum _{i=0.5\times i\times r}^{1.5\times i\times r}F\left(i\right)$$ (2)

where F _c is the rotational frequency of the drill (in this study, the rotational speed of the drill is 15000 RPM, i.e., the frequency is 250 Hz), and r is the center retrieval coefficient of the first harmonic.

As shown in Figures 4, 5, 6, when drilling in the cancellous bone of the femoral neck, the pivotal coefficients of the FFT are the first three harmonic amplitudes. When the femoral neck is drilled out, the pivotal coefficients of the FFT are the first‐harmonic amplitudes. Moreover, when the drill is in the cortical bone of the femoral neck, the pivotal coefficients of the FFT are the first 10 harmonic amplitudes. Therefore, the proportion A _re1 of the first harmonic in the first 10 harmonics is used to distinguish the drilling states and can be calculated as follows:

$$A_{\mathrm{re}1}=\frac{A_1}{\sum _{i=1}^{10}{A}_i},$$ (3)

where A _i is the absolute value of i ^th harmonic amplitude.

Drilling Status Recognition Method

A three‐layered back propagation artificial neural network (BP‐ANN) classifier is established to monitor the drilling status in real‐time (0.01 s). The topology of the BP neural network is shown in Figure 7, where X ₁, X ₂,…, X _n are the input values of the BP‐ANN, Y are the predicted drilling states, and w _ij and w _jk are the network weights. w _ij is the connection weight between the i ^th neuron in the input layer and j ^th neuron in the hidden layer, and w _jk is the connection weight between the j ^th neuron in the hidden layer and k ^th neuron in the output layer. Before prediction, the network must first be trained, and the network has the ability of associative memory and prediction through training. Cross‐entropy is used to train the network, and a confusion matrix is used to evaluate the performance of the network. To train the state classifier, the data were further divided into many 0.01 s frames, and each frame of the data was marked with one drilling state.

FIGURE 7.

Three‐layer back propagation artificial neural network

Results

Analysis of Spectrum Difference Between Drilling States

The “lillietest” function of MATLAB is used to check whether the 50 A _re1 data of each drilling state follow a normal distribution. Table 1 shows the normal distribution test results for five drilling states on four specimens, where l is the statistics, cv is the reference, and p > 0.05 is considered normally distributed. A _re1 data for all specimens and different drilling states are normally distributed, indicating that the independent sample t‐test can be applied to analyze the difference in A _re1 for different drilling states.

TABLE 1.

Normal distribution test result for five drilling states on four specimens

Specimen		State 1	State 2	State 3	State 4	State 5
1	l	0.1030	0.1444	0.0868	0.1269	0.0782
	cv	0.1451	0.1451	0.1451	0.1451	0.1451
	L	12.4173	5.2153	4.6852	7.1820	1.8977
	U	12.7979	6.1505	4.9997	7.3847	2.0898
	p	0.1986	0.0105	0.4326	0.0423	0.5000
2	l	0.0694	0.0917	0.1014	0.0734	0.0578
	cv	0.1451	0.1451	0.1451	0.1451	0.1451
	L	17.8812	2.6229	6.5552	15.0212	2.7915
	U	18.4305	2.9973	7.2660	15.4805	2.9692
	p	0.5000	0.3487	0.2184	0.5000	0.5000
3	l	0.1005	0.0914	0.0991	0.0944	0.0803
	cv	0.1451	0.1451	0.1451	0.1451	0.1451
	L	12.9538	13.9656	5.7677	8.2691	2.1016
	U	13.2266	16.6251	6.2319	8.5649	2.3994
	p	0.2294	0.3532	0.2475	0.3093	0.5000
4	l	0.0802	0.0839	0.0886	0.0842	0.1032
	cv	0.1451	0.1451	0.1451	0.1451	0.1451
	L	9.9347	24.4158	3.4174	7.2415	3.7853
	U	10.2981	25.3034	3.5962	7.5091	4.0568
	p	0.5000	0.5000	0.3991	0.4927	0.1965

Table 1 also shows the lower bound L and the upper bound U of the confidence intervals of each drilling state on the four specimens. Figures 8, 9, 10, 11 show the raw data of A _re1, mean ± SD of A _re1, and significant difference analysis results of the five drilling states on each specimen. The drilling states are compared in pairs, and states with no significant differences are highlighted with “NS.”

FIGURE 8.

Raw data of Ar, mean ± SD of Ar, and significant difference analysis results of five drilling states on specimen 1

FIGURE 9.

Raw data of Ar, mean ± SD of Ar, and significant difference analysis results of five drilling states on specimen 2

FIGURE 10.

Raw data of Ar, mean ± SD of Ar, and significant difference analysis results of five drilling states on specimen 3

FIGURE 11.

Raw data of Ar, mean ± SD of Ar, and significant difference analysis results of five drilling states on specimen 4

Every two adjacent states in the same specimen significantly differ in A _re1. Except for state 2, there is a slight standard deviation of A _re1 for the other four states of all four specimens. Moreover, the mean values of A _re1 from large to small are state 1, state 3, state 4, and state 5 for all four specimens. Although A _re1 follows the above‐mentioned qualitative law, it remains difficult to judge the drilling status by setting the threshold value due to the overlapping intervals between the A _re1 values of different specimens. Moreover, state 2 is an unstable state that is prone to misjudgment. Therefore, A _re1 can be an excellent reference to prove the difference in spectral distribution at the five typical drilling states, although without high accuracy and robust index to recognize the drilling state directly.

Automatic Identification of Five Different Drilling States

The above results for A _r in different drilling states indicate that the drilling vibration signals have different spectral distributions in different states. These five states can be effectively identified using spectral information. Therefore, we trained a neural network classifier with the frequency spectrum as the input and drilled state as the output. The model was first trained using a dataset of four specimens. To avoid overfitting problems, data are randomly divided into three parts at ratios of 0.7, 0.15, and 0.15 for training, validation, and testing, respectively. We adjust the connection weights of the network according to errors in the training dataset, and the training progress stops when generalization stops improving on the validation dataset. Finally, the performance of the network will be tested using a test dataset.

Figures 12, 13, 14, 15 show the recall and accuracy of the neural network classifier on the four specimens, where the bottom line is the recall rate, and the rightmost line is the accuracy rate. The training epochs of the neural network classifier for specimens 1–4 were 62, 78, 57, and 72. The recall rates of “in the cancellous bone” for specimens 1 to 4 were 96.3%, 83.7%, 88.0%, and 88.0%. The recall rates of “in out the femoral neck” from specimens 1 to 4 were 97.2%, 76.2%, 93.7%, and 94.4%. The recall rates of “in the cancellous bone” for specimens 1 to 4 were 97.8%, 92.3%, 100%, and 97.7%. The average recognition accuracies of the drilling state for specimens 1 to 4 were 86.8%, 84.7%,95.2%, and 85.8%. Moreover, the critical drilling states can be distinguished with extremely high accuracy. The recognition accuracies of “in the cancellous bone” for specimens 1 to 4 were 83.2%, 84.8%,92.9%, and 84.7%. The recognition accuracies of “in out the femoral neck” from specimens 1 to 4 were 98.2%, 88.4%, 95.8%, and 88.7%. The recognition accuracies of “in the cortical bone” for specimens 1 to 4 were 94.6%, 80.8%, 95.5%, and 88.9%.

FIGURE 12.

Training results on specimen 1. (A) Performance curve; (B) Recall and accuracy of the neural network classifier

FIGURE 13.

Training results on specimen 2. (A) Performance curve; (B) Recall and accuracy of the neural network classifier

FIGURE 14.

Training results on specimen 3. (A) Performance curve; (B) Recall and accuracy of the neural network classifier

FIGURE 15.

Training results on specimen 4. (A) Performance curve; (B) Recall and accuracy of the neural network classifier

Discussion

Because traditional manual drilling during hip fracture fixation can easily lead to unstable fixation and vascular damage. A safe and easy‐to‐use robot‐assisted method to automatically drill bone and distinguish critical bone drilling states with high accuracy in real‐time for the bone hole‐making process during hip fracture fixation is urgently needed. In this study, a safe and easy‐to‐use robot‐assisted method is proposed for automatic bone drilling and to distinguish critical bone drilling states with high accuracy in real‐time for the bone hole‐making process during femoral neck fixation. Four fresh pig femurs were drilled at the posterosuperior femoral neck in three modes, and five bone‐drilling states were defined. We proved the difference in the harmonic distribution at different drilling states using an independent sample t‐test. To distinguish the critical bone drilling states with high accuracy in real‐time, a neural network classifier was trained with the frequency spectrum as the input and the drilled state as the output. The average recognition accuracies of the drilling state for the four specimens were all higher than 84%. Critical drilling states can be distinguished with extremely high accuracy. Our method can potentially prevent weak fixation and damage to the lateral epiphyseal artery. From the working principle, the identification of bone drilling state has a strong correlation with parameters such as bone type and whether it is drilled. Therefore, the proposed robotic bone hole fabrication strategy can actually be applied to improve drilling safety in other orthopaedic surgeries, such as spine and trauma surgeries, and is scalable. In our future work, we will expand our specimen size, such as including cases of osteosclerosis and osteoporosis, to further demonstrate the validity and accuracy of our study.

Challenges of Drilling Task in Traditional Hip Fractures Fixation

In recent years, with the increasing number of patients undergoing hip fracture surgery, internal fixation of hip fractures has become the focus of discussion. Impacted and non‐displaced femoral neck fractures and stable intertrochanteric fractures usually involve in situ fixation with either multiple cannulated screws or SHS. ²⁹ , ³⁰ Compared with SHS, multiple cannulated screw fixation has the advantages of low revision surgery rate, high survival rate, short operation time, less bleeding, low risk of osteonecrosis, and cost saving. ⁵ Therefore, multiple cannulated screw fixation remains an important and safe surgical procedure to stabilize hip fractures. However, multiple cannulated screw fixations are affected by the surgeon's experience. The unstable bare‐handed operation, visual deviation, and degree of fracture reduction make it difficult to ensure the success of a one‐time puncture when the guide pin is punctured. The number of drills and adjusting the guide pin puncture path attempt should be kept to a minimum during the surgery, as they can weaken the cortical and cancellous bone, re‐injure muscles, soft tissues, and bones, increase the degree of surgical trauma, and increase the amount of bleeding in patients. At the same time, prolonging the operation time, extensive use of fluoroscopy, and increasing the exposure time may endanger the patient and operating room staff. Reference indicates that the femoral head's avascular necrosis was observed in 12 cases (23%) and nonunion was observed in five cases (9%) during their follow‐up periods. ³¹

Advantages and Disadvantages of Current Robot‐Assisted Operations in Hip Fracture Fixation

With the development and updating of medical imaging and computer technologies, computer‐assisted orthopaedic surgery has been widely used in joint surgery, spine surgery, and traumatic orthopaedics. ²⁰ The stereotactic technique based on X‐ray or three‐dimensional (3D) CT images can assist surgeons in planning surgeries more precisely, improving surgical efficiency and accuracy, reducing radiation exposure, and reducing patient injury. ³² However, computer‐assisted orthopaedic surgery is not perfect. ²⁴ Current computer‐assisted orthopaedic surgery has limitations in terms of positioning and navigation. Navigation orthopaedist‐assisted robots have significant advantages over traditional freehand operations. However, key intraoperative steps need to be performed by experienced surgeons with their bare hands, which also greatly reduces the safety, reliability, and ease of computer‐assisted orthopaedic surgery. With advances in technology and surgical innovation, an increasing number of surgeons are complaining of the lack of tactile feedback from computer‐assisted and robotic techniques. As a result, considerable research is currently underway to overcome this barrier.

Comparisons With Other Bone Drilling State Identification Methods

The main results of this study can be summarized as follows. In each specimen, the harmonic distributions of the drilling vibration at each of the two adjacent drilling modes were statistically different (p < 0.05). The average recognition accuracies of the drilling state for the four specimens exceed 84%. Moreover, the critical drilling states can be distinguished with extremely high accuracy. The recognition accuracies of “in the cancellous bone” for specimens 1 to 4 were 83.2%, 84.8%, 92.9%, and 84.7%. The recognition accuracies of “in out the femoral neck” from specimens 1 to 4 were 98.2%, 88.4%, 95.8%, and 88.8%. The recognition accuracies of “in the cortical bone” for specimens 1 to 4 were 94.6%, 80.8%, 95.5%, and 85.8%. The proposed robot‐assisted method can potentially prevent weak fixation and damage to the lateral epiphyseal artery.

The bone‐drilling process generates force, sound, and vibration signals that are composed of a series of harmonic signals. Dai et al. sampled the drill vibration signal with an accelerometer, and considering that the drilling sound signal and vibration signal are homologous, they combined these two signals for bone drilling state monitoring and successfully identified four different bone drilling states, as well as the entanglement state when the drill is wrapped by the muscle. ³³ In 2021, they use of vibration signals to simulate human haptic feedback was proposed to perceive and control orthopaedic assisted surgery robots. ³⁴ Hu et al. identified five primary drilling states by processing force signals in the drilling feed direction, including initial state, outer cortical state, cancellous state, transitional state, and inner cortical state. ³⁵ In 2020, Torun et al. proposed a breakthrough detection method based on the closed‐loop control characteristics of the bone drilling process. ³⁶ Our previous study found that there is an approximately linear relationship between harmonic amplitude and bone drilling depth for a specific range of cutting depths, and the bone drilling depth can be estimated using the amplitude of a specific harmonic over a certain range of cutting depths. ³⁷ , ³⁸ , ³⁹ , ⁴⁰ This study extracts and analyzes the vibration signal features during hip fracture nailing based on the FFT transformation and enables the orthopaedist‐assisted surgical robot to accurately identify five states through the deep learning method of the BP‐ANN classifier. Especially during posteromedial cannulated screw placement, it helps surgeons identify “all‐in” (AI), “in‐out‐in” (IOI), and “percutaneous fixation” (PF) in real‐time, avoiding injury to the surrounding blood vessels and nerves, and reducing the burden of the surgeon and the difficulty of surgery.

Strengths and Limitations

The orthopaedist‐assisted surgical robot with learning via a BP‐ANN classifier combined with preoperative navigation planning can sense and control the placement of multiple cannulated screws, which provides a new research direction for the future development of computer‐assisted orthopedic robots and has the potential to further simplify surgical procedures. It can avoid repeated drilling and adjusting the guide pin puncture path and may assist the surgeon in better placement of multiple cannulated screws, especially posteromedial screws, making the operation more fluent. The posterior medial and lateral cortices under the vastus ridge play an important stabilizing role in hip fracture reduction and fixation. The orthopaedist‐assisted surgery robot with deep learning can help surgeons more accurately insert posteromedial screws into the PF position, which can better stabilize the fracture, reduce the amount of bleeding, and make the surgery safer, conducive to fracture healing, and early rehabilitation exercises.

This study has some limitations. First, the study included ex vivo specimens; therefore, the patient's regular movements (e.g., respiratory movements) during the surgery may have an impact on the accuracy of the experimental results. Next, we will conduct experiments on live animals. Second, the femoral neck of healthy pigs was selected for the experiment; therefore, the results of this experiment are only significant for hip fractures in non‐pathological states. The next step is to expand our specimen size. Femoral necks in pathological states such as osteosclerosis and osteoporosis were selected for experiments to further demonstrate the validity and accuracy of our study. A more sophisticated classification method based on the vibration signal of the drill handle is expected to be designed in the future to identify bone types and to distinguish soft tissues, such as muscles, blood vessels, and ligaments. To improve identification accuracy, more in vivo bone drilling experiments must be performed to obtain sufficient data for network training.

Conclusion

This study investigated a safe and easy‐to‐use robot‐assisted method to automatically drill bone and distinguish critical bone drilling states with high accuracy in real‐time for the bone hole‐making process during hip fracture fixation. The recognition accuracies of “in the cancellous bone” for specimens 1 to 4 were 83.2%, 84.8%, 92.9%, and 84.7%. The recognition accuracies of “in out the femoral neck” from specimens 1 to 4 were 98.2%, 88.4%, 95.8%, and 88.8%. The recognition accuracies of “in the cortical bone” for specimens 1 to 4 were 94.6%, 80.8%, 95.5%, and 85.8%. The average recognition accuracies of the drilling state for the four specimens exceed 84%. The proposed robot‐assisted method can potentially prevent weak fixation and damage to the lateral epiphyseal artery. In our future work, we will expand our specimen size, such as including cases of osteosclerosis and osteoporosis, to further demonstrate the validity and accuracy of our study.

Data Availability

Data are available on request from the corresponding author.

Author Contributions

Conceptualization: Yuan Xue, Yu Dai, Ping Lei, and Jianxun Zhang. Data curation: Guangming Xia and He Bai. Formal analysis: Guangming Xia and Wei Liu. Writing–original draft: Guangming Xia and Jianxun Zhang. Writing– review and editing: Guangming Xia, He Bai, Wei Liu, Yuan Xue, Ping Lei, and Yu Dai.

Conflicts of Interest

The authors declare that there are no conflicts of interest regarding the publication of this paper.

Funding Information

National Natural Science Foundation of China (62173190, U1913207).