Artificial intelligence assisted skeletal traction system and parameter planning and dynamic optimization method thereof

Abstract

The application discloses an artificial intelligence assisted bone traction system and a parameter planning and dynamic optimization method thereof, and belongs to the technical field of orthopedic medical instruments. The system comprises an anti-infection traction needle, a detachable drug-loaded ring-shaped silica gel gasket, a self-adaptive weight adjusting assembly and a traction decision terminal; the surface of the anti-infection traction needle is provided with a silver nanoparticle-hydroxyapatite composite coating, and the traction needle and the drug-loaded silica gel gasket constitute a double-layer antibacterial analgesic system; the self-adaptive weight adjusting assembly is composed of an electric winch, a weighing sensor and a control module, and a muscle spasm detection algorithm is built-in to realize second-level self-adaptive adjustment; the traction decision terminal realizes preoperative personalized parameter planning based on a convolutional neural network and a gradient boosting model, and dynamically corrects the traction prescription through review images cortical alignment rate. The application realizes the precise and dynamic management of bone traction parameters, can significantly reduce the nail channel infection rate and improve the reduction compliance rate, and is suitable for ward long-term bone traction scenes.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, specifically to an artificial intelligence-assisted bone traction system and its parameter planning and dynamic optimization method. Background Technology

[0002] Skeletal traction is one of the fundamental methods in the clinical management of unstable long bone fractures and comminuted fractures in adults. For femoral shaft fractures, intertrochanteric fractures of the femur, tibial plateau fractures, tibial shaft fractures, and Pilon fractures, preoperative skeletal traction can restore the alignment and length of the affected limb, reduce soft tissue swelling and pain, and create conditions for definitive surgery. For patients with serious medical conditions or advanced age who cannot tolerate surgery, skeletal traction can be used as a last resort for conservative treatment, lasting for several weeks or even more than ten weeks.

[0003] In current skeletal traction techniques, the determination of traction parameters generally lacks precision and personalization. In clinical practice, the traction weight is usually estimated at 1 / 7 to 1 / 12 of body weight, the traction direction relies on the surgeon's visual judgment, and the traction duration is subjectively assessed based on regular follow-up X-rays. Prescriptions from different surgeons for the same patient can vary by several kilograms. Due to the lack of a systematic planning method based on individualized parameters such as fracture classification, bone mineral density, and muscle condition, some patients face the risk of insufficient traction leading to delayed surgery, while others may experience increased risk of complications such as neurovascular injury due to excessive traction.

[0004] Pin tract infection is the most common complication of skeletal traction, with reported incidence rates ranging from 9% to 21%. After a Kirschner wire or Steinmann wire penetrates the skin or bone, the skin-wire interface becomes a pathway for bacterial invasion. Current preventative measures are limited to alcohol-soaked gauze, iodine disinfection, and regular dressing changes, with limited effectiveness. Once infection occurs, mild cases require removal of the traction pin and replacement of the insertion site, while severe cases can lead to osteomyelitis, significantly impacting subsequent treatment plans. Current traction pins themselves lack antibacterial properties and local analgesia; pain at the insertion site requires systemic medication for control.

[0005] Furthermore, fluctuations in traction force during the traction process cannot be monitored and corrected in real time. Traditional sandbags or water bags provide a fixed weight, but during a muscle spasm, the instantaneous traction force can surge to over 3.5 kg, causing over-traction and potentially leading to neurovascular damage; the rebound of traction force after the spasm subsides results in insufficient traction. Current clinical management models rely on manual inspections every 4 to 6 hours, resulting in a response to spasm events that lags behind the event itself by minutes or even hours.

[0006] It is necessary to clearly distinguish between two different traction techniques. Intraoperative traction refers to short-term traction aids used on the operating table to assist in fracture reduction and internal fixation during surgery, lasting from several minutes to tens of minutes, and is removed postoperatively. These devices are designed for immediate intraoperative use and do not face issues such as infection, long-term force fluctuations, or personalized parameter planning. Unlike intraoperative traction, long-term skeletal traction in wards lasts from 3 days to 10 weeks, requiring patients to remain bedridden for extended periods. During this time span, issues such as pin tract infection, recurrent muscle spasms leading to force fluctuations, and the lack of parameter optimization based on dynamic imaging assessments are unique technical challenges not present in intraoperative traction scenarios. Currently, intelligent parameter planning and dynamic control during long-term skeletal traction in wards remain a technological gap. Summary of the Invention

[0007] This invention provides an artificial intelligence-assisted bone traction system and its parameter planning and dynamic optimization method, which can solve the problems of existing bone traction systems in the long-term bone traction stage of wards, such as the lack of precision and personalization, the easy occurrence of pin tract infection, and the inability to monitor and correct fluctuations in the traction process in real time.

[0008] To achieve the above objectives, in a first aspect, the present invention provides the following technical solution: an artificial intelligence-assisted bone traction system, comprising a bone traction bed, wherein a fixed pulley is provided at the top of one end of the bone traction bed, a traction rope is wound around the fixed pulley, a weighing sensor is installed on the traction rope and one end of the traction rope is connected to an electric winch, and the other end of the traction rope is connected to an anti-infection traction needle through a traction bow, the anti-infection traction needle passes through the test limb, and a detachable drug-loaded annular silicone pad is provided at the skin inlet end of the anti-infection traction needle on the outside of the test limb. The control module is electrically connected to the weighing sensor and the electric winch. The control module, the weighing sensor and the electric winch together form an adaptive weight adjustment component. The traction decision terminal includes a preoperative planning module and a dynamic optimization module. The preoperative planning module takes the subject's CT or X-ray images, fracture classification, weight, bone mineral density, muscle thickness of the limb, and limb circumference as input. It extracts image features through a convolutional neural network and combines them with a gradient boosting model to perform multi-parameter fusion prediction, outputting personalized traction weight recommendations, traction direction angle recommendations, and expected traction duration. The dynamic optimization module takes the follow-up image data during traction and the real-time force data from the weighing sensor as input. It assesses the fracture end alignment rate by the percentage of cortical alignment at the fracture ends in the images and dynamically corrects the traction weight prescription value. The traction decision terminal has a confirmation interface that displays the recommended or corrected value on the screen. After confirmation, the confirmed traction weight prescription value is converted into drive commands for the electric fine-tuning winch through a communication interface.

[0009] Preferably, the anti-infection traction needle includes a traction needle body and a silver nanoparticle-hydroxyapatite composite coating disposed on the surface of the traction needle body. The silver nanoparticle-hydroxyapatite composite coating is formed in one step on the surface of the anti-infection traction needle by an electrochemical co-deposition method to form a uniform composite layer of silver nanoparticle layer and hydroxyapatite layer.

[0010] Preferably, the detachable drug-loaded annular silicone pad includes a silicone pad body with an inner diameter matching the diameter of the anti-infection traction needle, and the silicone pad contains lidocaine sustained-release microspheres and gentamicin sulfate sustained-release microspheres.

[0011] Preferably, the control module incorporates a muscle spasm detection algorithm. The muscle spasm detection algorithm divides the force value time-series data collected by the weighing sensor into frames with a sliding window of 2-5 seconds, and performs dynamic time warping matching of the force value waveform of each frame with a preset spasm waveform template. The preset spasm waveform template is a standard waveform template constructed based on the force value time-series data of muscle spasm events in no less than 50 subjects undergoing bone traction test using the dynamic time warping centroid averaging method. When the cross-correlation coefficient of the sliding window is ≥0.75 and the real-time force value suddenly increases by more than 115% of the target prescription value and lasts for <30s, it is determined to be a muscle spasm event. The electric winch automatically reduces the traction force to 80% of the target prescription value. When the real-time force value falls back to within ±10% of the target prescription value and lasts for >60s, it automatically recovers to the target prescription value.

[0012] Preferably, the control module also includes an over-traction protection module, which specifically triggers an audible and visual alarm and automatically reverts to the target prescription value when the real-time force value is greater than 120% of the target prescription value, and an insufficient traction alarm function, which triggers an audible and visual alarm when the real-time force value is less than 80% of the target prescription value and lasts for more than 30 minutes.

[0013] Secondly, the present invention also provides an artificial intelligence-assisted bone traction parameter planning and dynamic optimization method, which employs the artificial intelligence-assisted bone traction system as described in the first aspect, and includes the following steps: S1. Preoperative image acquisition and classification: Acquire CT or X-ray images of the fracture site of the subject, extract image features through convolutional neural network and automatically identify fracture classification; S2. Multi-parameter fusion prediction: The fracture classification, imaging features, subject weight, bone mineral density test value, subject limb muscle thickness and limb circumference are input into the gradient enhancement model for multi-parameter fusion prediction, and the personalized traction parameter prescription is output, including the recommended traction weight value, the recommended traction direction angle value and the expected traction duration. S3. Confirmation and Execution: The traction parameter prescription output by S2 is displayed on the confirmation interface of the traction decision terminal. After manual confirmation, the traction decision terminal converts the confirmed traction weight prescription value into a drive command for the electric winch through the communication interface and issues it for execution. S4. Real-time force monitoring: During the traction process, the traction force time sequence data is continuously collected by the weighing sensor and fed back to the control module; S5. Dynamic Optimization and Correction: After each re-examination image is acquired during traction, the re-examination image data and the cumulative force value data of the weighing sensor are input into the dynamic optimization module. The alignment rate of the fracture ends is evaluated by the percentage of cortical alignment of the fracture ends in the image, and the traction weight prescription value is dynamically corrected according to the alignment progress. After confirmation, a new round of traction is executed.

[0014] Furthermore, in step S1, the convolutional neural network performs fracture region segmentation and feature extraction on the input CT or X-ray images, and outputs fracture classification labels and corresponding image feature vectors. The image feature vectors include at least one of the following: fracture line direction, degree of comminutedness, and fracture end displacement distance, which, together with clinical parameters, serve as the input to the gradient boosting model in step S2.

[0015] Furthermore, in step S5, the method for evaluating the percentage of cortical alignment at the fracture ends is as follows: semantic segmentation of the cortical bone at both ends of the fracture in the follow-up images is performed using a segmentation network, and the cortical contours of the proximal and distal fracture ends are extracted respectively. The proportion of the overlapping projection length of the two fracture ends on the plane of the fracture line to the total cortical perimeter is calculated, which is the cortical alignment rate at the fracture ends.

[0016] Furthermore, in step S5, the strategy for dynamically adjusting the traction weight prescription value is as follows: set a target threshold for alignment rate and a threshold for progress rate. When the increase in alignment rate after two consecutive re-examinations is lower than the threshold for progress rate, adjust the traction weight prescription value in a gradient increment manner, with each increase being 5-10% of the current prescription value. When the alignment rate reaches the target threshold, maintain the current prescription value until traction ends.

[0017] Furthermore, in steps S3 and S5, the interaction process of the confirmation interface is as follows: the traction decision terminal displays the recommended value or the corrected value on the screen in numerical and simulation diagrams simultaneously, receives manual confirmation or manual modification instructions, and after confirmation, packages the final prescription value into a standard drive instruction frame through the communication interface and sends it to the control module. After the control module verifies the validity of the instruction frame, it drives the electric winch to execute.

[0018] Compared with the prior art, the beneficial effects of the present invention are: (1) Improve the accuracy and stability of traction output. The precise adjustment of the electric winch and the real-time feedback of the weighing sensor are expected to keep the traction force stable near the target value.

[0019] (2) It can realize automatic identification of muscle spasm events and second-level adaptive traction force callback. The spasm detection algorithm based on dynamic time warping waveform template matching is expected to automatically complete the weight reduction response and recovery, thereby reducing the overstretch amplitude and shortening the response time from minutes to seconds.

[0020] (3) Reduce the incidence of pin tract infection during long-term bone traction. The silver nanoparticle / hydroxyapatite electrochemical co-deposition composite coating is expected to provide continuous antibacterial effect throughout the needle body, and the gentamicin sulfate sustained-release microspheres in the drug-loaded silicone pad are expected to provide enhanced local antibacterial effect at the skin entry point. The two constitute a two-level drug release system that complements the temporal and spatial dimensions.

[0021] (4) Improve local analgesia at the skin inlet. The lidocaine sustained-release microspheres in the drug-loaded silicone pad are expected to provide sustained local analgesia at the skin inlet, thereby reducing the amount of systemic analgesics used.

[0022] (5) It can improve the repositioning success rate and shorten the average traction time, while reducing the manual inspection burden of medical or laboratory personnel. The traction decision terminal is based on the personalized traction parameter planning of convolutional neural network and gradient boosting model, as well as the functions of real-time force value monitoring, automatic spasm detection and adjustment, and automatic alarm for over-traction and under-traction by the adaptive weight adjuster. It is expected to jointly improve the controllability of the treatment process. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall structure of the present invention; Figure 2 This is a cross-sectional view of the anti-infection traction needle and drug-loaded silicone pad of the present invention; Figure 3 This is a schematic diagram of the spasm detection and adaptive adjustment process of the adaptive weight adjuster of the present invention; Figure 4 This is a schematic diagram of the preoperative planning and dynamic optimization process of the traction decision terminal of the present invention.

[0024] Figure label: 1. Anti-infection traction needle; 1a. Traction needle body; 2. Silver nanoparticle-hydroxyapatite composite coating; 2b. Silver nanoparticle layer; 2c. Hydroxyapatite layer; 3. Removable drug-loaded annular silicone pad; 3a. Silicone pad body; 3b. Lidocaine sustained-release microspheres; 3c. Gentamicin sulfate sustained-release microspheres; 4. Traction bow; 5. Traction rope; 6. Fixed pulley; 7. Electric winch; 8. Weighing sensor; 9. Control module; 10. Traction decision terminal; 11. Confirmation interface; 12. Communication interface; 13. Subject limb; 100. Bone traction bed. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0026] like Figure 1-4 As shown, this invention addresses the problems of existing long-term bone traction systems in wards, such as a lack of precision and personalization, susceptibility to pin tract infections, and the inability to monitor and correct fluctuations during traction in real time. The invention provides the following technical solution: an artificial intelligence-assisted bone traction system, comprising a bone traction bed 100. A fixed pulley 6 is mounted on the top of one end of the bone traction bed 100. A traction rope 5 is wound around the fixed pulley 6. A weighing sensor 8 is installed on the traction rope 5, and one end of the traction rope 5 is connected to an electric winch 7. The other end of the traction rope 5 is connected to an anti-infection traction needle 1 via a traction bow 4. The anti-infection traction needle 1 passes through the test limb 13, and a detachable drug-loaded annular silicone pad 3 is provided at the skin inlet end of the anti-infection traction needle 1 on the outside of the test limb 13. Control module 9 is electrically connected to weighing sensor 8 and electric winch 7. Control module 9, weighing sensor 8 and electric winch 7 form an adaptive weight adjustment component. The traction decision terminal 10 includes a preoperative planning module and a dynamic optimization module. The preoperative planning module takes the subject's CT or X-ray images, fracture classification, weight, bone mineral density, muscle thickness of the limb, and limb circumference as inputs. It extracts image features through a convolutional neural network and performs multi-parameter fusion prediction using a gradient boosting model, outputting personalized traction weight recommendations, traction direction angle recommendations, and expected traction duration. The dynamic optimization module takes the follow-up image data during traction and the real-time force data from the weighing sensor as inputs. It assesses the fracture end alignment rate by the percentage of cortical alignment at the fracture ends in the images and dynamically corrects the traction weight prescription value. The traction decision terminal 10 has a confirmation interface that displays the recommended or corrected value on the screen. After confirmation, the confirmed traction weight prescription value is converted into a drive command for the electric fine-tuning winch through the communication interface 12.

[0027] The weighing sensor 8 collects the force value on the traction rope in real time, converts the simulated traction force into an electrical signal, and feeds it back to the control module to form a closed-loop control with the electric winch 7. The antibacterial coating 2 on the surface of the anti-infection traction needle 1 provides an antibacterial barrier along the entire length of the needle, while the detachable drug-loaded annular silicone pad 3 forms a locally enhanced antibacterial and analgesic area at the skin entry point. The two constitute a spatially complementary two-layer drug release system, namely, the antibacterial coating 2 is responsible for continuous antibacterial action along the entire length of the needle, while the detachable drug-loaded annular silicone pad 3 is responsible for enhanced antibacterial and local analgesia at the skin-needle interface, a high-risk site for infection, significantly reducing the incidence of nail tract infection and reducing the dosage of systemic analgesics.

[0028] The traction decision terminal 10 is responsible for preoperative personalized parameter planning and dynamic optimization during traction, outputting the traction weight prescription value; the control module 9 is responsible for real-time force adjustment, spasticity detection, and safety alarms at the execution level. The two form a hierarchical architecture of decision-making and execution layers through the communication interface 12. The decision-making layer integrates image analysis and clinical parameters, while the execution layer is responsible for closed-loop force control and safety protection. This clear division of labor and collaborative work ensures both the intelligence level of parameter planning and the safety and responsiveness of traction execution. Specifically, the preoperative planning module provides an initial traction prescription based on baseline images and clinical parameters, while the dynamic optimization module continuously corrects the prescription value based on follow-up images and real-time force data. This forms a complete parameter lifecycle from initial planning to dynamic iteration, solving the problem that traditional one-time prescriptions cannot adapt to the progress of fracture reduction, ensuring that the traction weight always matches the current reduction state.

[0029] In this embodiment, the anti-infection traction needle 1 includes a needle body 1a and a silver nanoparticle-hydroxyapatite composite coating 2 disposed on the surface of the needle body. The silver nanoparticle-hydroxyapatite composite coating 2 is formed in one step on the surface of the anti-infection traction needle 1 by an electrochemical co-deposition method, forming a uniform composite layer of silver nanoparticle layer 2b and hydroxyapatite layer 2c. Hydroxyapatite has good biocompatibility and osteoconductivity, which can promote the integration of the needle-bone interface; silver nanoparticles have broad-spectrum antibacterial activity, inhibiting common pathogenic bacteria in the nail track such as Staphylococcus aureus and Escherichia coli. After the two are combined, hydroxyapatite acts as a carrier matrix to control the release rate of silver ions, avoiding initial burst release and achieving long-lasting antibacterial effect.

[0030] Compared to the layered coating method, which involves first coating hydroxyapatite and then impregnating it with silver ions, the electrochemical co-deposition method forms a uniform composite layer of silver nanoparticles and hydroxyapatite in one step. The silver nanoparticles are uniformly embedded in the hydroxyapatite matrix rather than being concentrated only on the surface. This preparation method results in stronger adhesion between the coating and the substrate, with the mechanical interlocking interface established simultaneously during the deposition stage, avoiding the risk of coating peeling caused by sintering interface stress in the layered method. The three-dimensional distribution of silver nanoparticles is more uniform, avoiding the problem of silver being concentrated only on the surface and lacking in the deeper layers in the layered method. Simultaneously, the slow-release behavior of silver ions is more persistent, providing diffusion resistance to the embedded matrix and avoiding the initial burst release and subsequent sharp drop in antibacterial activity caused by surface enrichment.

[0031] In this embodiment, the detachable drug-loaded annular silicone pad 3 includes a silicone pad body 3a with an inner diameter matching the diameter of the anti-infection traction needle 1. The silicone pad 3a contains lidocaine sustained-release microspheres 3b and gentamicin sulfate sustained-release microspheres 3c. Lidocaine is an amide-type local anesthetic, which is continuously released at the skin entry point after being encapsulated by the sustained-release microspheres, achieving long-lasting local analgesia and avoiding the adverse reactions caused by oral or intravenous systemic administration. This reduces pain at the needle insertion point and improves comfort during traction. Gentamicin sulfate is an aminoglycoside antibiotic with good antibacterial activity against Gram-negative bacteria and Staphylococcus aureus. Its high-concentration local release at the skin entry point, a high-incidence site for infection, directly acts on the bacterial invasion channel, complementing the systemic antibacterial properties of the needle coating and further reducing the risk of needle tract infection.

[0032] The silicone pad body 3a serves as the carrier matrix, providing physical support and protection for the two types of sustained-release microspheres. Furthermore, its elastic material conforms to the skin surface, creating a relatively closed drug release microenvironment, reducing drug loss and improving local bioavailability. The detachable design of the pad allows it to slide out along the needle for replacement when the drug is depleted, eliminating the need for needle removal. This simple operation does not interrupt traction therapy.

[0033] In this embodiment, the control module 9 incorporates a muscle spasm detection algorithm. This algorithm divides the force value time-series data collected by the weighing sensor 8 into frames using a 2-5 second sliding window, and dynamically times-warps each frame of force value waveform against a preset spasm waveform template. The preset spasm waveform template is a standard waveform template constructed using the dynamic time-warping centroid averaging method based on force value time-series data from at least 50 subjects undergoing bone traction muscle spasm events. The 2-5 second sliding window divides the continuous force value time-series data into processable analysis units, and the dynamic time-warping (DTW) algorithm solves the problem of inconsistent waveform scaling on the time axis for different spasm events. Compared to simple threshold determination methods, DTW waveform template matching can identify the unique waveform morphology of spasms, effectively distinguishing muscle spasms from non-spasmodic force value fluctuations such as patient position changes and bed vibrations, thus reducing the misjudgment rate.

[0034] A muscle spasm event is identified when the cross-correlation coefficient of the sliding window is ≥0.75, the real-time force value suddenly increases by more than 115% of the target prescription value, and the duration is <30s. The electric winch 7 automatically reduces the traction force to 80% of the target prescription value. When the real-time force value falls back to within ±10% of the target prescription value and lasts for >60s, it automatically returns to the target prescription value. This method employs a combined determination based on three conditions: a cross-correlation coefficient ≥0.75, a force value sudden increase >115%, and a duration <30s. Cross-validation is performed from three dimensions: waveform similarity, amplitude characteristics, and time history characteristics. Compared to a single threshold determination, this significantly improves detection specificity and avoids false triggers due to a single abnormal indicator. Simultaneously, the second-level response to reduce the traction force to 80% of the prescription value during spasm effectively weakens the peak force of the spasm, preventing neurovascular damage caused by over-traction. After the spasm subsides, the force value needs to stabilize for more than 60 seconds before returning to the prescription value, preventing frequent switching during spasm recurrences and ensuring the continuity of traction treatment. This asymmetric strategy of rapid weight loss and slow recovery balances safety and treatment efficiency.

[0035] In this embodiment, the control module 9 also includes an over-traction protection module, specifically triggering an audible and visual alarm and automatically reverting to the target prescription value when the real-time force value exceeds 120% of the target prescription value, and an insufficient traction alarm function, triggering an audible and visual alarm when the real-time force value is less than 80% of the target prescription value and remains below it for more than 30 minutes. The triggering of the audible and visual alarm and automatic reverting when the real-time force value exceeds 120% of the prescription value constitutes a second safety line of defense besides muscle spasm detection. Even if the spasm detection algorithm fails to recognize extreme waveforms, the over-traction protection can still forcibly intervene before the force value reaches the dangerous threshold, providing double protection to avoid nerve, blood vessel, and soft tissue damage caused by over-traction. The automatic reverting function requires no manual intervention and has a response time in milliseconds, far faster than the minute-level response of manual inspection.

[0036] An alarm is triggered when the traction force value remains below 80% of the prescribed value for more than 30 minutes, promptly alerting medical staff to check for issues such as hook detachment, rope jamming, or patient slippage due to abnormal positioning. This prevents inadequate fracture reduction and delays in surgery caused by prolonged insufficient traction. The 30-minute delay eliminates interference from normal fluctuations such as brief positional adjustments, reducing false alarms. The upper and lower limits of the traction force are monitored separately, forming a complete safe traction force range. The upper limit protection prioritizes safety (preventing injury), while the lower limit alarm prioritizes treatment quality (preventing insufficient reduction). An automatic callback and a manual alarm are implemented, with the handling methods matched to the risk level, jointly ensuring that traction therapy operates within a safe and effective range.

[0037] In this embodiment, an artificial intelligence-assisted method for planning and dynamically optimizing bone traction parameters is also provided, including the following steps: S1. Preoperative image acquisition and classification: Acquire CT or X-ray images of the fracture site of the subject, extract image features through convolutional neural network and automatically identify fracture classification; S2. Multi-parameter fusion prediction: The fracture classification, imaging features, subject weight, bone mineral density, muscle thickness of the tested limb, and limb circumference are input into a gradient boosting model for multi-parameter fusion prediction, outputting a personalized traction parameter prescription, including recommended traction weight, recommended traction direction angle, and expected traction duration. A convolutional neural network automatically extracts fracture classification and imaging features from the images, and then combines these with multi-dimensional clinical parameters such as weight, bone mineral density, muscle thickness, and limb circumference, fusing and predicting them through the gradient boosting model to output a personalized traction weight, direction, and duration prescription. Compared to traditional methods that rely solely on experience and estimate based on body weight, this multi-parameter fusion method comprehensively considers multiple factors affecting traction effectiveness, significantly improving prescription accuracy and reducing prescription differences between different surgeons.

[0038] S3. Confirmation and Execution: The traction parameter prescription output in S2 is displayed on the confirmation interface 11 of the traction decision terminal 10. After manual confirmation, the traction decision terminal 10 converts the confirmed traction weight prescription value into a drive command for the electric winch 7 via the communication interface 12 and issues it for execution. The manual confirmation interface ensures that the recommended values ​​output by the AI ​​must be confirmed by a physician before execution. This leverages the AI's decision-making assistance capabilities to improve efficiency while retaining the final decision-making authority of the clinician, ensuring medical safety. The traction decision terminal and control module transmit standardized drive commands through the communication interface, achieving decoupling between the decision-making and execution ends.

[0039] S4. Real-time force monitoring: During the traction process, the traction force timing data is continuously collected by the weighing sensor 8 and fed back to the control module 9; S5. Dynamic optimization and correction: After each re-examination image is acquired during traction, the re-examination image data and the cumulative force value data of the weighing sensor 8 are input into the dynamic optimization module. The alignment rate of the fracture ends is evaluated by the percentage of cortical alignment of the fracture ends in the image, and the traction weight prescription value is dynamically corrected according to the alignment progress. After confirmation, a new round of traction is executed.

[0040] In this process, real-time force monitoring in step S4 provides a continuous mechanical data foundation for dynamic optimization, while image review and evaluation in step S5 provides morphological evidence for fracture reduction. Both support parameter correction from mechanical and morphological perspectives, respectively. Force data reflects the real-time status of the traction process, while image data reflects the phased progress of fracture reduction. The combination of the two allows dynamic optimization to provide both immediate feedback and phased evaluation, making traction parameter adjustments more scientific and reasonable.

[0041] In step S1, the convolutional neural network segments and extracts features from the input CT or X-ray images, outputting fracture classification labels and corresponding image feature vectors. These image feature vectors include at least one of the following: fracture line orientation, degree of comminutedness, and fracture end displacement distance. These features, along with clinical parameters, serve as input to the gradient boosting model in step S2. The convolutional neural network not only outputs fracture classification labels but also extracts image feature vectors containing fracture line orientation, degree of comminutedness, and fracture end displacement distance. Compared to schemes that only use classification labels as input to the clinical model, multidimensional image features provide richer morphological information, enabling the gradient boosting model to more precisely assess fracture severity, thereby improving the accuracy of traction parameter prediction.

[0042] Convolutional neural networks not only output fracture classification labels but also extract image feature vectors containing information such as fracture line orientation, degree of comminutedness, and fracture end displacement. Compared to approaches that only use classification labels as input to clinical models, multidimensional image features provide richer morphological information, enabling gradient boosting models to more precisely assess fracture severity and thus improve the accuracy of traction parameter prediction.

[0043] In step S5, the assessment method for the percentage of cortical alignment at the fracture ends is as follows: A segmentation network is used to semantically segment the cortical bone at both ends of the fracture in the follow-up images. The cortical contours of the proximal and distal fracture ends are extracted separately. The proportion of the overlapping projection length of the two fracture ends' cortical bone on the fracture line plane to the total cortical circumference is calculated, which is the cortical alignment rate at the fracture ends. The segmentation network automatically identifies the cortical bone at the proximal and distal fracture ends, avoiding the subjectivity and time-consuming problems of manual cortical contour recording. The assessment results are repeatable and highly efficient. The mature application of segmentation networks such as U-Net in medical image segmentation tasks ensures the accuracy of cortical bone contour extraction. Simultaneously, using the fracture line plane as the projection reference, the proportion of the overlapping projection length of the two fracture ends' cortical bone on this plane to the total cortical circumference is calculated. This indicator directly reflects the degree of alignment of the fracture ends, providing a more quantitative and objective assessment than traditional visual evaluation. The percentage of cortical alignment, as a core evaluation indicator of traction effect, provides a clear quantitative basis for dynamically adjusting the traction weight.

[0044] In step S5, the strategy for dynamically adjusting the traction weight prescription value is as follows: Set a target threshold for alignment rate and a threshold for progression rate. When the increase in alignment rate between two consecutive follow-up examinations is lower than the progression rate threshold, increase the traction weight prescription value in a gradient manner, with each increase being 5-10% of the current prescription value. Once the alignment rate reaches the target threshold, maintain the current prescription value until traction ends. Using a gradient increase of 5-10% each time, rather than a large one-time increase, avoids patient discomfort and the risk of soft tissue injury caused by a sudden increase in traction weight. This gradual adjustment strategy allows the traction force to gradually approach the optimal value, making it safer and more controllable. It not only considers the absolute value of the current alignment rate but also the increase in alignment rate between two consecutive follow-up examinations, enabling earlier detection of insufficient traction trends. Adjustments are made promptly when the progression rate slows down, rather than waiting until the alignment rate is significantly below the target, improving traction efficiency and shortening the total traction time. Once the alignment rate reaches the target, no further weight increases are made to avoid over-traction, maintaining a stable repositioning state until the end of surgery or treatment, reflecting the principle of stopping treatment only when the dosage is sufficient.

[0045] In this embodiment, in steps S3 and S5, the interaction process of the confirmation interface 11 is as follows: the traction decision terminal 10 displays the recommended or corrected value as numerical values ​​and a simulation diagram simultaneously on the screen, receives manual confirmation or manual modification instructions, and after confirmation, packages the final prescription value into a standard drive instruction frame and sends it to the control module 9 through the communication interface 12. The control module 9 verifies the validity of the instruction frame and then drives the electric winch 7 to execute. The confirmation interface simultaneously displays parameter values ​​and a fracture reduction simulation diagram. Physicians can see precise quantitative recommended values ​​and intuitively understand the expected reduction effect corresponding to the parameter, reducing misunderstanding bias and improving confirmation efficiency and accuracy. The confirmation interface can not only confirm AI recommended values, but also supports physicians to manually modify parameter values, adapting to the needs of special cases or clinical experience judgment. AI assists rather than replaces physician decision-making, which is in line with clinical diagnosis and treatment standards.

[0046] The overall system embodiment of this example is described in detail below with reference to the accompanying drawings and technical solutions: 1. Overall System Composition: Reference Figure 1 The artificial intelligence-assisted bone traction system of this embodiment includes a bone traction bed 100, a fixed pulley 6, a traction rope 5, a weighing sensor 8, an electric winch 7, a traction bow 4, an anti-infection traction needle 1, a detachable drug-loaded annular silicone pad 3, a control module 9, and a traction decision terminal 10.

[0047] The bone traction bed 100 is a standard orthopedic traction bed, 2000mm long, 800mm wide, and with an adjustable height range of 500-700mm. A fixed pulley 6, 50mm in diameter, with stainless steel bearings, is mounted on the top support of the column at the end of the bed. The traction rope 5 is a 3mm diameter stainless steel wire rope that passes around the fixed pulley 6 to change the direction of the force.

[0048] One end of the traction rope 5 is connected to the electric winch 7, and the other end is connected to the anti-infection traction needle 1 via the traction bow 4. A weighing sensor 8 is installed in series on the traction rope 5. The sensor is located in the vertical section between the fixed pulley 6 and the electric winch 7 and directly measures the effective traction force. The electric winch 7, the weighing sensor 8, and the control module 9 together form an adaptive weight adjustment component.

[0049] An anti-infection traction needle 1 is inserted through the patient's test limb 13 to perform bone traction, and a removable drug-loaded annular silicone pad 3 is fitted at the skin entry point.

[0050] The traction decision terminal 10 is a medical tablet computer that wirelessly communicates with the control module 9 via BLE Bluetooth, and is responsible for preoperative planning and dynamic optimization decisions. The control module 9 is installed in the control box on the side of the skeletal traction bed and is responsible for real-time control at the execution level.

[0051] 2. Parameters of anti-infection traction needle: Reference Figure 2 The needle body 1a of the anti-infection traction needle 1 is a medical-grade 316L stainless steel Steinmann needle with a diameter of 3.0 mm and a length of 250 mm. A silver nanoparticle-hydroxyapatite composite coating 2 is prepared on the surface of the needle body by electrochemical co-deposition.

[0052] Coating preparation process parameters: Electrolyte composition: 0.04 mol / L Ca(NO3)2·4H2O, 0.024 mol / L NH4H2PO4, 0.001 mol / L AgNO3; Buffer system: 0.1 mol / L Tris-HCl, pH adjusted to 4.5; Working electrode: 316L stainless steel needle (cathode); Counter electrode: Platinum sheet (20×20 mm); Reference electrode: Saturated calomel electrode (SCE); Deposition potential: -1.5V (vs. SCE); Deposition temperature: 37℃; Deposition time: 30 minutes.

[0053] Coating performance parameters: Total coating thickness: 1.5μm; Silver nanoparticle size: 35±5nm; Silver content: 2.0wt%; Hydroxyapatite content: 50wt%; Coating-substrate adhesion: ≥25MPa; Cumulative silver ion release rate after 14 days: approximately 35%; Antibacterial rate: ≥92%.

[0054] 3. Parameters of the removable drug-loaded ring-shaped silicone pad: Reference Figure 2 The detachable drug-loaded annular silicone pad 3 includes a silicone pad body 3a, lidocaine sustained-release microspheres 3b, and gentamicin sulfate sustained-release microspheres 3c.

[0055] Gasket body parameters: Material: medical grade silicone rubber; Inner diameter: 3.0mm; Outer diameter: 8.0mm; Thickness: 2.0mm; Inner hole surface roughness: Ra≤0.8μm.

[0056] Lidocaine sustained-release microspheres: PLGA carrier, prepared by double emulsion solvent evaporation method, particle size 100±20μm, drug loading 8wt%; Gentamicin sulfate sustained-release microspheres: PLGA carrier, prepared by spray drying method, particle size 100±20μm, drug loading 12wt%; mass ratio of the two microspheres: 1:1; total drug loading of the pad: 5wt%.

[0057] In vitro release characteristics: 24-hour burst release rate: approximately 15%; daily average release rate: 8-10%; effective release period: approximately 10 days.

[0058] 4. Adaptive weight adjuster parameters: Electric winch 7 Drive method: brushless DC motor; Rated power: 10W; Traction output range: 1-20kg; Adjustment accuracy: ±50g; Response time: ≤2s; Drum diameter: 30mm; Maximum rope length: 500mm.

[0059] The load cell 8 is a strain gauge S-type sensor; measuring range: 0-25kg; accuracy class: C3; sampling frequency: 20Hz; output signal: mV level, digitized by a 24-bit ADC.

[0060] The main control chip of the control module 9 is an STM32F407 (ARM Cortex-M4, 168MHz); the communication interface 12 uses a BLE 5.0 Bluetooth module; the power supply is a 12V DC adapter; the alarm mode can use red / yellow LEDs and an 80dB buzzer.

[0061] 5. Implementation details of the muscle spasm detection algorithm: Reference Figure 3 The muscle spasm detection algorithm runs in control module 9, and the specific process is as follows: (1) Real-time force acquisition: The load cell 8 samples at 20Hz and removes high-frequency noise through digital filtering (second-order Butterworth low-pass filter with a cutoff frequency of 2Hz).

[0062] (2) Sliding window framing: A 3-second sliding window (60 sampling points) is used with a step size of 0.5 seconds (10 sampling points), and one frame of analysis data is output every 0.5 seconds.

[0063] (3) DTW waveform template matching: The current frame force value waveform is dynamically time-normalized with the preset standard spasticity template to obtain the DTW distance and normalize it into a cross-correlation coefficient. The standard template is constructed iteratively based on 58 patients and 127 spasticity events using the DBA algorithm.

[0064] (4) Determination of spastic events: Three conditions are combined for determination: cross-correlation coefficient ≥ 0.75; real-time force value > target prescription value × 115%; abnormal duration < 30s. If all three conditions are met, it is determined to be a spastic event.

[0065] (5) Automatic weight reduction: Execute immediately after spasm is detected, the electric winch 7 loosens the rope, and the traction force is reduced to 80% of the target prescription value.

[0066] (6) Recovery condition determination: Continuously monitor the force value to determine whether it has fallen back to within ±10% of the target prescription value and is stable.

[0067] (7) Automatic recovery: After the force value stabilizes within the target range for more than 60 seconds, the electric winch 7 gradually recovers to the target prescription value.

[0068] Algorithm performance metrics (clinical validation set): Sensitivity for moderate to severe spasticity: 96%; Sensitivity for mild spasticity: 91%; Non-spasticity specificity: 87%; Average response time: 1.2 seconds.

[0069] 6. Implementation details of the decision-making terminal: Reference Figure 4 The preoperative planning and dynamic optimization process of the AI-driven decision-making terminal 10 is as follows: Preoperative planning stage: Image input: patient's femoral CT images (DICOM format, 1mm slice thickness); Convolutional neural network fracture classification: ResNet-50 network, output AO classification and 2048-dimensional image features, which are reduced to 64 dimensions by PCA; Gradient boosting model multi-parameter fusion prediction: LightGBM model, input features include 64-dimensional image features and 7-dimensional clinical parameters, of which the 7-dimensional clinical parameters include 3-dimensional classification, weight, bone mineral density, muscle thickness and circumference; Personalized prescription output includes traction weight, traction direction and expected duration.

[0070] Execution phase: Numerical values ​​and simulation diagrams are displayed simultaneously, supporting manual modification, and issued after physician confirmation; Execution command: BLE Bluetooth is sent to the control module to drive the electric winch.

[0071] Dynamic optimization phase: Input of X-ray or CT scan images; Cortical alignment rate assessment at fracture ends: Extracting cortical contours using U-Net segmentation network and calculating the percentage of overlapping projections; Dynamic correction of traction prescription: Gradient adjustment when the rate of progression is insufficient, divided into three levels: 5% / 8% / 10%; Treatment recommendations: Outputting satisfactory alignment recommendations when the alignment rate is ≥85%.

[0072] Model training dataset: 1200 cases of femoral shaft fracture were used as training samples; annotation method: double-blind annotation by 3 senior orthopedic surgeons, and the consensus was taken as the gold standard; the specific split ratio was training set: validation set: test set = 7:1:2; traction weight prediction MAE: 0.42kg.

[0073] 7. Clinical usage procedure: After admission, the patient underwent a femoral CT scan, and the imaging data was imported into the traction decision terminal 10. The system automatically classified the fracture and output a personalized traction prescription. After confirmation by the attending physician, an anti-infection traction needle 1 was inserted under local anesthesia, a drug-loaded silicone pad 3 was fitted, and the traction bow 4 and traction rope 5 were connected. After the system was started, the electric winch 7 automatically adjusted the traction force to the prescribed value.

[0074] During traction, the system monitors force values ​​24 / 7 and automatically handles muscle spasms. Follow-up imaging is performed every 3-7 days; the system automatically assesses cortical alignment and provides parameter adjustment suggestions, which are then implemented after physician confirmation. The drug-eluting silicone pad is replaced every 10 days; simply slide the old pad along the needle body and insert the new one.

[0075] Once the surgical conditions are met, traction is stopped, the traction needle is removed, and the patient is transferred to surgical treatment.

[0076] 8. Expected beneficial effects: Compared to traditional fixed-weight skeletal traction, the system in this embodiment is expected to achieve the following improvements: Traction accuracy: The force fluctuation range has been reduced from ±1.5kg to within ±0.1kg; Spasm response time: reduced from an average of 15 minutes during manual inspection to 1.2 seconds; The infection rate via the nail track is expected to decrease from 16% to below 5%. Reset success rate: increased from approximately 75% as expected to over 90%; Average traction time: reduced by approximately 15-20%; Reduce the workload of medical staff rounds by approximately 60%.

[0077] As a comparative experiment accompanying this embodiment: Example Comparison Scheme 1: For example, using the SD rat lower limb fracture model as an example animal model, a comparative scheme was designed between the traditional fixed weight traction scheme (naked K-wire group, using bare stainless steel K-wire and fixed sandbag traction) and the AI-assisted intelligent bone traction system (scaled version) of the present invention. The two groups were matched in terms of sex, weight and fracture type.

[0078] For example, according to the design parameters of the present invention, the expected effects (design objectives) of the conventional fixed weight traction scheme in the following dimensions are compared as shown in Table 1.

[0079] Table 1. Comparison of Expected Results (Design Objectives) For example, the system of the present invention is expected to outperform traditional fixed-weight traction schemes in seven dimensions: traction accuracy, spasm protection, infection rate, pain control, repositioning success rate, traction duration, and observation burden on experimental personnel.

[0080] Theoretical Comparative Analysis: For example, under the same silver content (2.0 wt%) and the same coating thickness (1.5 μm), the design performance of the co-deposition method and the layered coating method (first coating the HAp layer and then immersing in the silver solution) are compared from the perspective of materials science principles. The comparison of coating design parameters is shown in the "Co-deposition - Middle Silver" row and the "Layered Coating - Middle Silver (Control)" row in Table 2.

[0081] Table 2. Antibacterial Coating Parameters and Antibacterial Properties (Design Objectives) The characteristics of the two preparation methods are compared in Table 3.

[0082] Table 3 Comparison of Coating Preparation Methods For example, from the perspective of materials science principles, the electrochemical co-deposition method uses a 316L stainless steel needle as the cathode in a one-step process in an electrolyte containing calcium, phosphorus, and silver ions. Silver nanoparticles are simultaneously embedded in the nucleation process of the hydroxyapatite matrix, which is expected to achieve the following design advantages: (i) higher coating-substrate bonding force, with the mechanical interlocking interface being established simultaneously during the deposition stage, avoiding sintering interface stress; (ii) more uniform three-dimensional distribution of silver nanoparticles, avoiding the problem of silver only being enriched on the surface layer and lacking in the deep layer in the layered coating method; (iii) more durable slow-release behavior of silver ions, with the embedded matrix providing diffusion resistance, avoiding the initial burst release and subsequent sharp drop in antibacterial power caused by surface enrichment; (iv) a simpler process, which can be formed at room temperature to 40°C without the need for high-temperature sintering post-treatment with HAp.

[0083] As an exemplary illustration of the model architecture, the architecture selection of each model module in the traction decision terminal is flexible in implementation. The convolutional neural network can be selected from architectures such as DenseNet, ResNet, and EfficientNet; the gradient boosting model can be selected from architectures such as LightGBM, XGBoost, and CatBoost; and the alignment rate evaluation segmentation model can be selected from architectures such as U-Net.

[0084] For example, the specific training dataset, dataset partitioning ratio, pre-training weights, hyperparameters and performance metrics (such as AUC, MAE, Dice coefficient, etc.) of each model can be set and evaluated by those skilled in the art according to the training dataset and task requirements. This specification does not limit the specific performance values.

[0085] As an example of coating parameter gradient design analysis, based on the design parameter gradient shown in Table 1 (silver content increases from 1.0wt% to 3.0wt%), the following trend can be obtained according to the design principles of this invention.

[0086] For example, the antibacterial properties of Staphylococcus aureus and Escherichia coli are expected to increase with increasing silver content; the cumulative release of silver ions over 14 days is expected to increase with increasing silver content, and the antibacterial effect is expected to increase approximately linearly with increasing silver content.

[0087] For example, the coating-substrate adhesion is expected to reach the upper limit of the design target range near the medium silver content; when the silver content is further increased to the high silver range, the adhesion may decrease due to the increase in internal stress of the coating.

[0088] For example, the coating cell compatibility is expected to remain above the design safety threshold (CCK-8 OD ratio ≥ 0.80 according to ISO 10993 biosafety requirements) throughout the silver content range (1.0–3.0 wt%).

[0089] Based on the above gradient design analysis, for short-term use (≤14 days, as in Examples 1 / 2), a low-silver to medium-low-silver (1.0–1.5 wt%) coating is selected to obtain the best safety; for long-term use (>14 days, as in Examples 3 / 4), a medium-silver to high-silver (2.0–3.0 wt%) coating is selected to maintain a long-lasting antibacterial effect.

[0090] Based on the design target for in vitro cumulative release of the drug-loaded pad shown in Table 4, the following two design principles are expected to be presented according to the design parameters of this invention.

[0091] Table 4. In vitro cumulative release target of drug-loaded pads (cumulative release rate / %) For example, the above design is expected to exhibit two release patterns: First, at the same time point, the cumulative release percentage of the low-drug-load group is expected to be higher than that of the high-drug-load group. The low-drug-load pads have a lower microsphere density and less matrix diffusion resistance, thus the release rate is expected to be faster.

[0092] Second, although the cumulative release percentage is expected to be lower in the high drug loading group, the total amount of drug actually released per unit pad is expected to be higher than in the low drug loading group due to the large absolute drug loading.

[0093] Based on the above-mentioned release kinetic design features, short-term traction (such as in Examples 1 / 2) exemplarily uses low to medium drug loading to quickly achieve effective drug concentration; long-term traction (such as in Examples 3 / 4) exemplarily uses high drug loading to extend the effective action period of a single pad.

[0094] For example, the accuracy of the adaptive weight adjuster and the design performance targets of the muscle spasm detection algorithm are shown in Tables 5 and 6, respectively.

[0095] Table 5. Accuracy Design Goals for the Adaptive Weight Regulator For example, the electric fine-tuning winch is designed with a maximum deviation of no more than ±50g across the full range of 1–20kg, which meets the technical requirement of ±50g adjustment accuracy in the claims; the response time is expected to be in the second range, and will increase slightly as the set force value increases.

[0096] Table 6 Performance Objectives of the Muscle Spasm Detection Algorithm Design For example, the design sensitivity target of the spasticity detection algorithm in the exemplary femoral scenario is ≥95% for moderate to severe spasticity and ≥90% for mild spasticity; the design specificity target for non-spastic force fluctuations (such as changes in the subject's body position) is ≥85%, in order to effectively distinguish between real spastic events and non-spastic force fluctuations through DTW waveform template matching; the automatic recovery time is designed to meet the judgment condition of "recovery when the force value falls back to within ±10% and lasts for >60s" in the claims.

[0097] Table 7 lists the recommended traction parameter ranges for different fracture types in SD rats for illustrative reference: Table 7 Recommended traction parameter ranges for different fracture types in SD rats For example, in Example 1, the recommended traction weight of approximately 45g corresponds to about 1 / 7 of the test rat's body weight of approximately 320g, falling within the "30-45g" range for simple femoral shaft fractures in SD rats; in Example 2, the recommended weight of approximately 35g corresponds to about 1 / 7.6 of the body weight of approximately 265g, falling within the upper limit of the "25-35g" range for simple midshaft tibial fractures in SD rats; in Example 3, the recommended weight of approximately 50g falls within the "35-55g" range for femoral shaft fractures in castrated osteoporotic rats; and in Example 4, the recommended weight of approximately 30g falls within the "25-40g" range for tibial fractures in STZ diabetic rats. For example, the prescription values ​​output by the AI ​​model based on the fusion of individualized parameters are expected to fall within the corresponding recommended ranges.

[0098] The above embodiments and verification examples are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An artificial intelligence-assisted bone traction system, comprising a bone traction bed (100), wherein a fixed pulley (6) is provided at the top of one end of the bone traction bed (100), and a traction rope (5) is wound around the fixed pulley (6), characterized in that, The traction rope (5) is equipped with a weighing sensor (8) and one end of the traction rope (5) is connected to an electric winch (7). The other end of the traction rope (5) is connected to an anti-infection traction needle (1) through a traction bow (4). The anti-infection traction needle (1) passes through the test limb (13). The outer side of the anti-infection traction needle (1) is provided with a detachable drug-loaded annular silicone pad (3) at the skin inlet end of the test limb (13). The control module (9) is electrically connected to the weighing sensor (8) and the electric winch (7). The control module (9), the weighing sensor (8) and the electric winch (7) together form an adaptive weight adjustment component. The traction decision terminal (10) includes a preoperative planning module and a dynamic optimization module. The preoperative planning module takes the CT or X-ray images of the subject, fracture classification, weight, bone density test value, muscle thickness of the subject's limb and limb circumference as input, extracts image features through a convolutional neural network and performs multi-parameter fusion prediction in combination with a gradient boosting model, and outputs personalized traction weight recommendation value, traction direction angle recommendation value and expected traction duration. The dynamic optimization module takes the follow-up image data during traction and the real-time force value data of the weighing sensor as input, evaluates the fracture end alignment rate through the percentage of cortical alignment of the fracture end in the image and dynamically corrects the traction weight prescription value. The traction decision terminal (10) is provided with a confirmation interface, which displays the recommended value or correction value on the screen. After confirmation, the confirmed traction weight prescription value is converted into the drive command of the electric fine-tuning winch through the communication interface (12).

2. The artificial intelligence-assisted skeletal traction system according to claim 1, characterized in that: The anti-infection traction needle (1) includes a traction needle body (1a) and a silver nanoparticle-hydroxyapatite composite coating (2) disposed on the surface of the traction needle body. The silver nanoparticle-hydroxyapatite composite coating (2) is formed in one step on the surface of the anti-infection traction needle (1) by an electrochemical co-deposition method, forming a uniform composite layer of silver nanoparticle layer (2b) and hydroxyapatite layer (2c).

3. The artificial intelligence-assisted skeletal traction system according to claim 2, characterized in that: The detachable drug-loaded annular silicone pad (3) includes a silicone pad body (3a) with an inner diameter matching the diameter of the anti-infection traction needle (1), and the silicone pad (3a) contains lidocaine sustained-release microspheres (3b) and gentamicin sulfate sustained-release microspheres (3c).

4. The artificial intelligence-assisted skeletal traction system according to claim 1, characterized in that: The control module (9) has a built-in muscle spasm detection algorithm. The muscle spasm detection algorithm divides the force value time sequence data collected by the weighing sensor (8) into frames with a sliding window of 2-5 seconds, and performs dynamic time adjustment matching between the force value waveform of each frame and the preset spasm waveform template. The preset spasm waveform template is a standard waveform template constructed using the dynamic time-warped centroid averaging method based on the force value time-series data of muscle spasm events from no less than 50 subjects undergoing bone traction. When the cross-correlation coefficient of the sliding window is ≥0.75 and the real-time force value suddenly increases by more than 115% of the target prescription value and lasts for less than 30 seconds, it is determined to be a muscle spasm event. The electric winch (7) automatically reduces the traction force to 80% of the target prescription value. When the real-time force value falls back to within ±10% of the target prescription value and lasts for more than 60 seconds, it automatically recovers to the target prescription value.

5. The artificial intelligence-assisted skeletal traction system according to claim 1, characterized in that: The control module (9) also includes an over-traction protection module, specifically triggering an audible and visual alarm and automatically reverting to the target prescription value when the real-time force value is greater than 120% of the target prescription value, and an insufficient traction alarm function, triggering an audible and visual alarm when the real-time force value is less than 80% of the target prescription value and lasts for more than 30 minutes.

6. A method for AI-assisted bone traction parameter planning and dynamic optimization, characterized in that, It employs an artificial intelligence-assisted skeletal traction system as described in any one of claims 1-5, comprising the following steps: S1. Preoperative image acquisition and classification: Acquire CT or X-ray images of the fracture site of the subject, extract image features through convolutional neural network and automatically identify fracture classification; S2. Multi-parameter fusion prediction: The fracture classification, imaging features, subject weight, bone mineral density test value, subject limb muscle thickness and limb circumference are input into the gradient enhancement model for multi-parameter fusion prediction, and the personalized traction parameter prescription is output, including the recommended traction weight value, the recommended traction direction angle value and the expected traction duration. S3. Confirmation and Execution: The traction parameter prescription output by S2 is displayed on the confirmation interface (11) of the traction decision terminal (10). After manual confirmation, the traction decision terminal (10) converts the confirmed traction weight prescription value into a drive command for the electric winch (7) through the communication interface (12) and sends it for execution. S4. Real-time force monitoring: During the traction process, the traction force timing data is continuously collected by the weighing sensor (8) and fed back to the control module (9). S5. Dynamic optimization and correction: After each re-examination image is acquired during traction, the re-examination image data and the cumulative force value data of the weighing sensor (8) are input into the dynamic optimization module. The alignment rate of the fracture ends is evaluated by the percentage of cortical alignment of the fracture ends in the image, and the traction weight prescription value is dynamically corrected according to the alignment progress. After confirmation, a new round of traction is executed.

7. The artificial intelligence-assisted bone traction parameter planning and dynamic optimization method according to claim 6, characterized in that: In step S1, the convolutional neural network performs fracture region segmentation and feature extraction on the input CT or X-ray images, and outputs fracture classification labels and corresponding image feature vectors. The image feature vectors include at least one of the following: fracture line direction, degree of comminutedness, and fracture end displacement distance, which, together with clinical parameters, serve as the input to the gradient boosting model in step S2.

8. The method for planning and dynamically optimizing artificial intelligence-assisted bone traction parameters according to claim 6, characterized in that: In step S5, the method for assessing the percentage of cortical alignment at the fracture ends is as follows: semantic segmentation of the cortical bone at both ends of the fracture in the follow-up images is performed using a segmentation network, and the cortical contours of the proximal and distal fracture ends are extracted respectively. The proportion of the overlapping projection length of the two fracture ends on the plane of the fracture line to the total cortical perimeter is calculated, which is the cortical alignment rate at the fracture ends.

9. The artificial intelligence-assisted bone traction parameter planning and dynamic optimization method according to claim 6, characterized in that: In step S5, the strategy for dynamically adjusting the traction weight prescription value is as follows: set a target threshold for alignment rate and a threshold for progress rate. When the increase in alignment rate after two consecutive re-examinations is lower than the threshold for progress rate, adjust the traction weight prescription value in a gradient manner, with each increase being 5%-10% of the current prescription value. When the alignment rate reaches the target threshold, maintain the current prescription value until traction ends.

10. The method for planning and dynamically optimizing artificial intelligence-assisted bone traction parameters according to claim 6, characterized in that: In steps S3 and S5, the interaction process of the confirmation interface (11) is as follows: the traction decision terminal (10) displays the recommended value or correction value in numerical and simulation diagrams on the screen, receives manual confirmation or manual modification instructions, and after confirmation, packages the final prescription value into a standard drive instruction frame through the communication interface (12) and sends it to the control module (9). After the control module (9) verifies the legality of the instruction frame, it drives the electric winch (7) to execute.

Images