Hemodialysis teaching system based on AI intelligent pushing

Abstract

The invention relates to the technical field of hemodialysis teaching systems, and discloses a hemodialysis teaching system based on AI intelligent pushing. The system comprises an operation feature extraction module, an instrument positioning analysis module, an operation step analysis module and a teaching strategy generation module. Operating a feature extraction module to obtain a hemodialysis teaching video stream, and extracting multi-modal features containing an instrument movement track sequence and physiological parameter simulation data; an instrument positioning analysis module calls a pre-training network for positioning and generates a coordinate data set containing instrument types and three-dimensional position coding mapping; the operation step analysis module analyzes the time sequence characteristics by means of an action segmentation model, and generates a result containing step integrity and an error operation mark; and the teaching strategy generation module inputs a result into the intelligent push model, plans a personalized teaching path, outputs a knowledge point sequence containing a student ability model and dynamic teaching elements of operation difficulty parameters, and assists the hemodialysis teaching to be precise and personalized.

Description

Technical Field

[0001] This invention relates to the field of hemodialysis teaching system technology, specifically a hemodialysis teaching system based on AI intelligent push. Background Technology

[0002] Hemodialysis is a vital treatment for end-stage renal disease patients, and the standardization and precision of its operation directly affect patient safety and treatment outcomes. Therefore, the professional development of hemodialysis operators relies on a systematic and efficient teaching system. Currently, hemodialysis teaching primarily employs a combination of traditional offline practical training and fixed video courses, which has several limitations in actual teaching practice. From the perspective of content delivery, existing hemodialysis instructional videos are mostly standardized operating procedures with fixed recordings, which cannot be dynamically adjusted according to the actual learning progress and operational mastery of the trainees. During the learning process, trainees may have misunderstandings about the operation steps of certain specific instruments or the key points of physiological parameter monitoring, but fixed videos cannot address these individual issues in detail, resulting in blind spots in trainees' mastery of key knowledge points. In terms of instrument operation guidance, traditional teaching methods rely heavily on visual observation to assess students' instrument operation, making it difficult to accurately capture subtle deviations in the instrument's movement trajectory. For example, during the connection of hemodialysis tubing, it is difficult to comprehensively and accurately record and analyze details such as whether the instrument's three-dimensional spatial position is accurate and whether its movement trajectory meets the standard operating requirements through manual observation. Consequently, it is impossible to provide students with targeted operational correction suggestions, affecting the efficiency of improving students' operational skills. From the perspective of operational procedure evaluation, existing teaching models lack systematic analytical methods for judging the completeness and correctness of trainees' operational steps. During simulated operations or actual training, trainees may omit steps, reverse the order, or perform incorrect operations. However, teachers can often only conduct an overall evaluation after the operation is completed, and cannot track every step of the operation in real time. It is also difficult to accurately locate the specific type of error and the point at which it occurred. As a result, trainees cannot promptly identify and correct problems in their operations, which is not conducive to forming standardized operating habits. In terms of personalized teaching path planning, traditional hemodialysis teaching adopts a "one-size-fits-all" approach, ignoring the differences in learning foundation, comprehension ability, and learning needs among different students. Some students may grasp the instrument identification stage quickly but have difficulty with physiological parameter analysis; while others may have the opposite. Due to the lack of construction and application of individualized ability models for each student, it is impossible to develop personalized teaching plans for different students, resulting in unreasonable allocation of teaching resources, unmet learning needs of some students, and poor overall teaching effectiveness. Summary of the Invention

[0003] The purpose of this invention is to provide an AI-based intelligent push-based hemodialysis teaching system to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides an AI-based intelligent push-based hemodialysis teaching system, the system comprising: The operation feature extraction module is used to acquire hemodialysis teaching video stream data and extract multimodal operation features from the video stream data. The multimodal operation features include instrument movement trajectory sequences and physiological parameter simulation data. The instrument positioning and analysis module calls a pre-trained instrument recognition network to perform spatial positioning processing on the video stream data, generating an instrument coordinate dataset. The instrument coordinate dataset contains a mapping relationship between instrument type identifiers and three-dimensional position codes. The operation step parsing module, based on the instrument coordinate dataset, calls the action segmentation model to perform temporal feature parsing on the instrument motion trajectory sequence, and generates feature analysis results containing step integrity features and error operation markers; The teaching strategy generation module inputs the feature analysis results into the intelligent push model to plan personalized teaching paths and outputs dynamic teaching elements. The dynamic teaching elements include a sequence of knowledge points and operation difficulty adjustment parameters generated based on the student's ability model. Preferably, the operation feature extraction module includes: The video frame segmentation unit performs keyframe extraction processing on the video stream data to generate multiple operation stage segments and corresponding timestamp indices. The trajectory reconstruction unit extracts the motion optical flow features of each operation stage segment and performs temporal attention weighting processing on the motion optical flow features to generate a weighted trajectory feature vector. The physiological parameter association unit performs similarity matching between the trajectory feature vector and a preset physiological parameter threshold, and labels abnormal operation events. The multimodal fusion unit performs cross-modal alignment processing on the abnormal operation event labels to generate multimodal operation features that include device usage trajectory and physiological change curves.

[0005] Preferably, the instrument positioning analysis module includes: The spatial rasterization unit divides the operation stage segment into a three-dimensional spatial mesh, generating multiple instrument detection areas and corresponding voxel coordinate parameters; The feature encoding unit is used to extract the depth convolution features of each instrument detection area and perform spatial pyramid pooling on the depth convolution features to generate a multi-scale feature tensor. The positioning and calibration unit performs feature matching between the multi-scale feature tensor and a preset instrument template library to filter out candidate instrument regions with acceptable confidence levels. The coordinate optimization unit performs non-maximum suppression processing on the candidate instrument region to generate an instrument coordinate dataset containing the instrument type identifier. Each instrument coordinate contains a normalized position code relative to the teaching equipment reference point.

[0006] Preferably, the operation step parsing module includes: The operation slicing unit performs spatiotemporal slicing processing on the video stream data based on the normalized position encoding in the instrument coordinate dataset to generate multiple operation step segments; The trajectory standardization unit performs velocity normalization and path smoothing on each operation step segment to generate standardized motion analysis input. The step segmentation unit is used to call the bidirectional temporal convolutional subnet in the action segmentation model to divide the action analysis input into stages and generate standard step sequence labels. The error detection unit is used to call the operation rule base in parallel to identify violations of the operation step segments and generate an error operation probability distribution; The feature integration unit performs feature cross-fusion between the standard step sequence labels and the error operation probability distribution to generate feature analysis results that include step completeness scores and error types.

[0007] Preferably, the teaching strategy generation module includes: The competency assessment unit analyzes the current student's historical operation records and generates knowledge weakness weights associated with the feature analysis results. The path planning unit generates a teaching node transition matrix based on the weights of the knowledge weaknesses and prioritizes the standard step sequence labels. The strategy optimization unit uses an adaptive push strategy to select the optimal teaching sequence from the teaching node transition matrix and generates a set of teaching units that includes theoretical explanations, simulated operations, and error reviews. The parameter integration unit logically associates the teaching unit set with the operation difficulty adjustment parameters to generate dynamic teaching elements that conform to the individual learning curve.

[0008] Preferably, the system further includes: The real-time feedback module collects student operation data during the teaching process and generates operation logs that include instrument positioning deviations and abnormal physiological parameter values. The anomaly response module extracts the operational risk features from the operation log, performs pattern matching between the operational risk features and the teaching case library, and generates teaching strategy adjustment instructions. The model update module updates the detection parameters of the instrument recognition network online based on the teaching strategy adjustment instructions; The strategy optimization module injects the updated detection parameters into the intelligent push model and recalculates the knowledge point priority weights in the teaching node transition matrix.

[0009] Preferably, the anomaly response module includes: The event slicing unit divides the operational risk characteristics into time windows to generate multiple risk event segments and corresponding operation video sequences. The root cause analysis unit is used to call the pre-trained risk classification model to trace the source of errors in each risk event segment and generate error classification labels including puncture angle error, abnormal tubing connection, and parameter setting error. The strategy matching unit retrieves teaching intervention templates corresponding to the misclassification labels from the teaching case library and generates a candidate strategy set. The instruction generation unit, based on the matching degree evaluation between the operation video sequence and the candidate strategy set, selects the strategy with the highest confidence to generate the teaching strategy adjustment instruction.

[0010] Preferably, the teaching strategy generation module further includes: The disturbance test unit injects virtual operation deviation parameters before the teaching is pushed out. The virtual operation deviation parameters are used to simulate the positioning offset scenario of the instrument. The stability monitoring unit is used to monitor the adaptation and processing results of the intelligent push model to deviation scenarios and generate teaching stability indicators. The model retraining unit triggers the incremental learning mode of the device recognition network when the teaching stability index falls below a preset threshold. The parameter optimization unit updates the convolutional layer parameters of the instrument recognition network using gradients based on the operational difference data before and after the deviation.

[0011] Preferably, the system further includes: A cross-platform adaptation module is used to build a teaching terminal adapter and to resolve the differences in the operation interfaces of different hardware devices through the teaching terminal adapter. The instruction conversion module converts the dynamic teaching elements into interactive instructions supported by the target device; The context preservation module is used to maintain the logical dependencies and error handling context of the knowledge point sequence during the transformation process; The performance optimization module is used to inject device-matched rendering parameters and generate teaching content data packages that support multi-terminal operation.

[0012] Preferably, the cross-platform adaptation module includes: The rule base construction unit is used to establish a device interface rule base and store operation instruction mapping tables and parameter constraint paths for various terminals. The syntax parsing unit performs abstract syntax tree parsing on the dynamic teaching elements to generate intermediate representation layer data; The instruction replacement unit queries the operation instruction mapping table based on the intermediate presentation layer data to generate a terminal-compatible instruction conversion scheme; The conflict resolution unit performs dependency injection on conflicting parameter constraint paths to generate unambiguous device interaction commands.

[0013] Compared with the prior art, the beneficial effects of the present invention are: The operational feature extraction module acquires hemodialysis teaching video stream data and extracts multimodal operational features from it, including instrument movement trajectory sequences and simulated physiological parameter data. This module overcomes the limitations of traditional teaching methods that rely solely on fixed videos to convey single pieces of information. By combining the dynamic trajectory of instrument movement with simulated physiological parameter data, it provides more comprehensive and richer foundational data for subsequent teaching analysis. The trainee's operational process is no longer an isolated demonstration of actions, but a complete set of information linked to key parameter data. This allows teaching analysis to delve into operational details and data correlations, helping to uncover potential problems and areas for improvement during the operation. The device localization and analysis module utilizes a pre-trained device recognition network to perform spatial localization processing on the video stream data, generating a device coordinate dataset that includes a mapping between device type identifiers and 3D position codes. This module achieves precise localization and recognition of hemodialysis devices in 3D space. Compared to traditional manual observation, which cannot accurately capture device position information, this module can objectively and accurately record the spatial position changes of different types of devices during operation. Whether it's the core components of the hemodialysis machine or auxiliary devices such as tubing and puncture needles, their position information can be precisely encoded, providing a quantitative basis for subsequent judgment of whether device operation is standardized. This shifts the evaluation of device operation from subjective judgment to objective data support, avoiding errors and omissions inherent in manual evaluation. The operation step parsing module, based on the instrument coordinate dataset, uses a motion segmentation model to perform temporal feature analysis on the instrument movement trajectory sequence, generating feature analysis results that include step integrity features and error operation markers. This module can track the entire temporal process of the trainee's operation in real time, associating the instrument movement trajectory with the operation steps. It can not only determine whether the operation steps are complete, avoiding the problem of not being able to monitor missing steps in real time in traditional teaching, but also accurately mark the specific nodes and types of errors. For example, in the preparation operation before hemodialysis treatment, if the trainee skips a disinfection step or makes a sequential error in instrument connection, this module can promptly identify and mark it, allowing the trainee to clearly understand the problems in their operation, facilitating timely adjustments and corrections, and promoting the development of standardized operating habits. The teaching strategy generation module inputs feature analysis results into an intelligent push model for personalized teaching path planning, outputting dynamic teaching elements that include knowledge point sequences generated based on the student's ability model and operational difficulty adjustment parameters. This module breaks away from the traditional "one-size-fits-all" teaching model, fully considering the different learning foundations and operational mastery levels of various students by constructing student ability models. For students with strong instrument recognition abilities but insufficient physiological parameter analysis, the system will push more knowledge point sequences related to physiological parameter interpretation and appropriately reduce the difficulty of instrument recognition-related content. For students who are not proficient in the operation steps, the system will focus on pushing knowledge points related to the operation process and gradually improve their operational proficiency by adjusting the operational difficulty parameters. This personalized teaching path planning can fully meet the learning needs of different students, ensuring the rational allocation of teaching resources and allowing each student to gradually improve their professional skills at a pace suitable for them, avoiding the low learning efficiency caused by the mismatch between teaching content and student abilities in traditional teaching. Attached Figure Description

[0014] Figure 1 This is a timing diagram of the AI-based intelligent push hemodialysis teaching system described in this invention; Figure 2 A flowchart illustrating the operation of the feature extraction module; Figure 3 A flowchart illustrating the work of the instructional strategy generation module. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Please see Figure 1 This invention provides an AI-based intelligent push-based hemodialysis teaching system, the system comprising: This system acquires hemodialysis teaching video stream data and extracts multimodal operational features, including instrument movement trajectory sequences and simulated physiological parameter data. A pre-trained instrument recognition network is then used to perform spatial localization processing on the video stream data to generate an instrument coordinate dataset. This dataset contains a mapping relationship between instrument type identifiers and 3D position codes. Based on the instrument coordinate dataset, a motion segmentation model is used to perform temporal feature analysis on the instrument movement trajectory sequences, generating feature analysis results including step integrity features and error operation markers. Finally, the feature analysis results are input into an intelligent push model for personalized teaching path planning, outputting dynamic teaching elements. These dynamic teaching elements include a sequence of knowledge points generated based on the student's ability model and operation difficulty adjustment parameters. This system achieves fully automated processing from video data acquisition to personalized teaching push, adapting to the learning needs and operational levels of different students.

[0017] Example 1: See Figure 2 The operation feature extraction module receives real-time video stream data from the hemodialysis teaching equipment. The video framing unit uses a hybrid algorithm based on inter-frame difference and optical flow analysis to detect operation stage switching points. When a significant change in the device's motion trajectory or a specific operation gesture is detected, keyframe capture is triggered, generating a sequence of operation stage segments with precise timestamp indexes. Each segment corresponds to a standard step in hemodialysis operation, such as arterial puncture or dialyzer installation. The trajectory reconstruction unit uses the Horn-Schunck optical flow algorithm to calculate the pixel displacement field of adjacent frames within the operation stage segment. Through 3D reconstruction technology, the 2D displacement vector is mapped to the coordinate system of the teaching equipment to form an initial motion trajectory point cloud. Temporal attention weighting processing introduces a weighting mechanism based on motion amplitude, assigning higher weights to high-speed movement areas of the device while suppressing background noise interference. Finally, it outputs a trajectory feature vector characterizing the device's motion law, which includes kinematic parameters such as displacement and acceleration. The physiological parameter association unit incorporates a blood pressure fluctuation model and a hemodynamic simulator. Upon receiving the trajectory feature vector, it initiates real-time parameter mapping, dynamically comparing the device's movement speed with a preset vascular pressure threshold. A sliding time window is used to calculate the dynamic time warping distance between the trajectory features and the physiological parameter curve. When the distance exceeds a safe operating threshold, an abnormal event label is automatically generated, such as "puncture speed exceeded" or "abnormal tubing pressure." The multimodal fusion unit deploys a time axis synchronization engine, aligning the device's movement trajectory with the time series of the physiological parameter curve using a dynamic time warping algorithm. Based on timestamp indexing, it establishes millisecond-level precision cross-modal association, generating a multimodal operational feature matrix that fuses device spatial coordinates, motion vectors, and physiological parameter fluctuations. This matrix serves as the foundational dataset for subsequent analysis.

[0018] When the instrument positioning and analysis module starts, it loads the pre-calibrated spatial coordinate system of the teaching equipment. The spatial rasterization unit uses an octree structure to divide the operating area into three-dimensional voxels. The size of each voxel unit is set to a 5mm cube according to the minimum recognition accuracy of the instrument, and the three-dimensional spatial coordinates of the center point of each voxel are recorded. The feature encoding unit calls the ResNet-50 convolutional neural network to extract the depth features of each voxel region. The spatial pyramid pooling layer uses three grid scales of 1×1, 2×2, and 4×4 to perform multi-level pooling on the convolutional feature map, generating a multi-scale feature tensor that integrates local details and global context. In the feature matching stage, the positioning calibration unit adopts a multi-view geometric constraint strategy, calculates the Euclidean distance between the multi-scale feature tensor and the standard features in the instrument template library. When the matching similarity exceeds the 0.95 confidence threshold, it is judged as a valid recognition, and candidate region proposals containing instrument type and spatial location are generated. The coordinate optimization unit implements a non-maximum suppression algorithm based on overlap analysis, performing 3D intersection-union (IUU) calculations on all candidate regions. When the overlap between two detection boxes exceeds 30%, only the proposal with higher confidence is retained. The final output instrument coordinate dataset contains three key types of information: instrument type coding adopts the International Medical Device Identification System (IMDIS); 3D position coordinates are established in a right-handed coordinate system with the teaching equipment reference point as the origin; and normalized position coding maps the actual coordinates to the [0,1] interval through linear transformation, for example, the coordinates of the puncture needle tip (325,120,80) are converted to (0.72,0.35,0.42). This dataset records the relative positional relationships between instruments through spatial relationship descriptors, providing structured spatial information for subsequent operation step analysis. The entire processing flow is executed end-to-end on the embedded inference device, with the latency from video input to coordinate dataset output controlled within 200 milliseconds, meeting the response requirements of the real-time teaching system.

[0019] Example 2: The operation step parsing module receives the instrument coordinate dataset from the instrument positioning analysis module. This dataset contains instrument type identifiers with timestamps and normalized position codes. The operation slicing unit uses a spatiotemporal cube segmentation algorithm to automatically divide the operation stage boundaries based on the instrument position change rate. When a specific instrument combination or abrupt change in spatial distribution is detected, segment cutting is triggered, generating independent operation step segments such as "tubular pre-flushing stage" or "puncture operation stage." Each segment is accompanied by start and end time markers accurate to milliseconds. The trajectory standardization unit performs a multi-level processing flow on the instrument motion trajectory within each step segment. First, a Savitzky-Golay filter is used to smooth high-frequency jitter in the trajectory. Then, linear interpolation is used to unify trajectories of different durations to a standard time axis. Finally, velocity normalization is performed to map the motion rate to the 0-1 range, outputting a standard trajectory sequence that eliminates individual operation differences. This sequence retains the original motion trend but has a comparable data structure. The step segmentation unit loads a pre-trained action segmentation model. This model employs a bidirectional temporal convolutional architecture, containing three forward convolutional layers and three backward convolutional layers. The bidirectional network simultaneously captures the dependencies between historical and future actions. After inputting a standardized trajectory sequence, it outputs a step boundary probability distribution map. A dynamic threshold segmentation algorithm identifies standard step labels such as "puncture needle positioning," "catheter connection," and "parameter settings." Each label is accompanied by a confidence score and a precise time interval. The error detection unit simultaneously initiates a rule engine scan. The operation rule base stores over 200 hemodialysis operation standard entries, including constraints such as "puncture angle 30-45 degrees," "catheter connection torque threshold," and "physiological parameter safety range." Rule matching employs a real-time inference mechanism. When a violation of spatial constraints or temporal anomalies in the instrument movement trajectory is detected, a violation event is generated, such as identifying "puncture angle deviation exceeding 5 degrees" or "insufficient disinfection operation time." The output includes a probability distribution matrix containing the error type code and the time point of occurrence.

[0020] The feature integration unit deploys a feature fusion pipeline. First, it establishes a spatiotemporal correlation matrix between step labels and error probabilities. Then, it mines the potential correlation between step execution order and error occurrence through a graph convolutional network. Subsequently, a multi-head attention mechanism is used to enhance the features of key error events. Finally, it outputs structured feature analysis results containing three dimensions: a step completeness score calculated as a percentage based on the completeness and correctness of standard step execution; an error type distribution matrix marking the frequency and severity of various errors; and an operation quality heatmap marking high-risk operation intervals on the timeline. This result data package is encapsulated and transmitted in JSON format, containing time-encoded step sequences, a list of error events, and associated instrument coordinate snapshots, providing a parsable operation quality assessment for teaching strategy generation. In the real-time processing flow, when parsing the "dialyzer installation" step segment, the trajectory normalization unit detects that the instrument movement speed exceeds the standard range, the step segmentation model identifies that the step execution time is less than 70% of the standard duration, and the error detection unit triggers the "installation pressure not meeting standard" rule. The feature integration unit associates these three abnormal signals with the same time interval, generating feature analysis results marked with "step completeness score 65 points" and "pressure operation error." The entire processing cycle is controlled within 300 milliseconds, supporting real-time generation of analysis reports during student operations. The teaching system dynamically adjusts its subsequent teaching content delivery strategy based on these reports. The data processing process retains the timestamp alignment of all intermediate results, ensuring precise synchronization between teaching feedback and the original operation video frames.

[0021] Taking a specific hemodialysis teaching scenario as an example, a student is practicing the "arterial puncture" procedure. The operation feature extraction module captures the movement trajectory of the puncture needle and simulated physiological parameters in the video stream in real time. The operation step parsing module starts working. The operation slicing unit automatically identifies that it is currently in the "puncture needle positioning" stage based on the normalized position code (0.72, 0.35, 0.42) and movement pattern of the puncture needle in the instrument coordinate dataset. It then uses the starting point of the instrument's acceleration movement as the segment start point and the moment the velocity returns to zero as the end point, generating an operation step segment with a duration of 3.2 seconds, timestamped in the T1-T2 interval. The trajectory standardization unit immediately processes this segment. First, it uses a Kalman filter algorithm to smooth out hand tremor noise in the puncture needle trajectory, eliminating minor irregular vibrations caused by trainee tension. Then, it uses linear interpolation to standardize the 3.2-second trajectory to a uniform 2.5-second time axis, maintaining the original movement trend while eliminating individual speed differences. Finally, it performs speed normalization, mapping the actual movement speed to the 0-1 range, and outputs standardized trajectory sequence data, which clearly shows the approximation process of the puncture needle from its initial position to the target blood vessel. The step segmentation unit calls a pre-trained bidirectional temporal convolutional model. After inputting the standardized trajectory sequence, the model's forward convolutional layer identifies three sub-steps: "needle insertion preparation," "angle adjustment," and "puncture execution." The backward convolutional layer verifies the rationality of the step order and outputs a step label sequence with confidence scores: "needle insertion preparation" (0.92), "angle adjustment" (0.85), and "puncture execution" (0.78). Each label is marked with precise time boundaries and spatial coordinate ranges. The error detection unit synchronously runs rule matching, activating 32 rules from the "Arterial Puncture" section of the operation rule base, including constraints such as "needle insertion angle 30-45 degrees," "needle holding stability threshold," and "puncture depth limit." The rule engine detects that at T1+1.8 seconds, the puncture needle angle sensor reading is 52 degrees, exceeding the maximum allowable value by 7 degrees, immediately generating an "angle deviation" error event. Simultaneously, at T1+2.1 seconds, it detects that the needle body vibration amplitude exceeds the safety threshold, triggering an "insufficient stability" error flag. The system outputs an error operation probability distribution matrix, showing an error probability of 0.87 in the "angle adjustment" stage and 0.63 in the "puncture execution" stage. The feature integration unit initiates multi-source data fusion, first establishing a mapping between step labels and error events. Graph neural network analysis reveals that the "angle deviation" error mainly occurs in the "angle adjustment" stage and is causally related to insufficient "needle insertion preparation." Subsequently, an attention mechanism is used to strengthen key error features, assigning a weight coefficient of 0.9 to angle deviation and a weight of 0.7 to stability issues.The final structured analysis results included: a step completeness score of 76 (with major deductions in the angle adjustment phase), an error type distribution showing angle-related errors accounting for 68%, and an operation quality heatmap clearly marking T1+1.8 seconds as a high-risk moment on the time axis. The entire analysis process was completed within 320 milliseconds, and the system immediately generated a detailed operation evaluation report: the arterial puncture operation was basically satisfactory in terms of step completeness, but there were obvious angle control problems, especially a serious angle deviation in the middle of needle insertion, and the stability of the entire operation process needs to be improved. This report was transmitted in real time to the teaching strategy generation module to provide data support for subsequent personalized teaching. In the teaching system interface, trainees immediately saw a 3D trajectory replay of their operation, with error moments highlighted in red and accompanied by specific numerical prompts for angle deviation. The system also suggested focusing on strengthening specific training in "needle insertion angle control".

[0022] Example 3: See Figure 3 The teaching strategy generation module receives a feature analysis result data packet from the operation step analysis module. This data includes a step completeness score, an error type distribution matrix, and an operation quality heatmap with time axis annotations. The competency assessment unit activates the student profile engine, retrieves the student's historical operation records from the teaching system database over the past three months, analyzes the error pattern distribution and step execution time characteristics in the records, and generates a knowledge weakness weight vector associated with the current feature analysis results. The weight calculation uses a time decay weighted algorithm, with recently high-frequency error types receiving higher weight values. For example, if a student has three consecutive angle deviations in the "arterial puncture" step, the weight of this error type is increased to 0.87. The path planning unit constructs a teaching node transition matrix based on the weight vector. The matrix elements represent the logical connection strength between knowledge points, and the transition probability is calculated using the following formula: Where: P ij W represents the probability of switching from knowledge point i to knowledge point j. j S represents the weight of the weak points of knowledge point j. ij This represents the logical coherence coefficient between knowledge points i and j in the hemodialysis operation process, where N is the total number of knowledge points. After the matrix is ​​generated, the standard step sequence labels are dynamically sorted. When a sudden increase in the error weight of the "pipeline connection" step is detected, the priority of that knowledge point is promoted to the first position in the sequence.

[0023] The strategy optimization unit deploys an adaptive push strategy based on Q-learning, selecting the path with the highest cumulative reward value from the teaching node transition matrix. The reward function comprehensively considers the importance of knowledge points, historical mastery, and the urgency of current errors, generating a teaching sequence containing three stages: the theoretical explanation unit uses a 3D animation library to demonstrate the operating principles of the equipment; the simulation operation unit generates a virtual operating environment with force feedback; and the error review unit compares the student's actual operation recording with a standard video in a split-screen format. The parameter integration unit converts operation difficulty adjustment parameters into specific teaching parameters through a rule mapping table. When the student's operation fluency is below a threshold, the resistance of the equipment movement in the simulation operation is automatically reduced and operation guidance markers are added. Finally, a dynamic teaching element package that conforms to the individual learning curve is output. This data package contains a configurable teaching content sequence and a corresponding difficulty parameter configuration file.

[0024] The perturbation testing unit injects a set of virtual operation deviation parameters before the teaching push. These parameters include instrument spatial coordinate offset ±20mm, operation speed fluctuation ±30%, and random perturbation of physiological parameter readings ±15%, simulating scenarios of hemodialysis machine malfunction or student operational errors. The stability monitoring unit runs Monte Carlo simulations, monitoring the output changes of the intelligent push model under 100 perturbation scenarios, recording the consistency level of the teaching sequence adjustment scheme, and generating a teaching stability index value in the 0-1 range. When the index value falls below the 0.75 threshold three times consecutively, the model retraining unit activates the incremental learning mode of the instrument recognition network, collecting operation data from the virtual deviation scenarios to construct a fine-tuning dataset. The parameter optimization unit calculates the L2 norm difference of the instrument recognition feature map before and after the deviation, and updates the network convolutional layer parameters using a stochastic gradient descent algorithm with momentum. The learning rate is set to 1 / 10 of the initial value, and after three iterations, the network's recognition accuracy for the new offset pattern is restored to a normal level. The entire optimization process is executed asynchronously in the background, ensuring that the teaching push is not affected by parameter updates.

[0025] Taking the feature analysis results received by the teaching strategy generation module from the operation step parsing module as an example, the results show that the trainee scored 76 points in the completeness of the steps in the most recent arterial puncture practice, with the main error types being angle control deviation (probability of occurrence 0.87) and insufficient operation stability (probability of occurrence 0.63). The competency assessment unit immediately retrieved the trainee's operation record database from the past two weeks and found that similar angle deviations occurred 7 times in historical operations, and the frequency of occurrence showed an upward trend. The system calculated the weight value of the angle control weakness point to be 0.86 and the stability control weight value to be 0.72, forming a knowledge weakness point weight vector [angle control: 0.86, stability: 0.72, puncture depth: 0.35, disinfection standard: 0.21]. The path planning unit constructed a teaching node transition matrix based on the weight vector. Each element in the matrix represents the transition probability between different knowledge points. The transition probability between the angle control knowledge point and the stability knowledge point was set to 0.78, and the transition probability between the angle control knowledge point and the puncture depth knowledge point was set to 0.45. The system reorders the standard procedure sequence, prioritizing angle control training first, followed by stability training, generating a new teaching sequence: angle control theory explanation → stability simulation training → puncture depth consolidation practice → complete operation assessment. The strategy optimization unit employs a state value function-based push strategy, selecting the path with the highest cumulative reward value from the teaching node transition matrix, prioritizing angle control training on this path. The system generates a scheme containing three teaching units: the theory explanation unit uses 3D animation to demonstrate the correct needle insertion angle of 30-45 degrees; the simulation operation unit uses a force feedback device to provide real-time angle deviation prompts; and the error review unit performs a split-screen comparative analysis of the student's 52-degree deviation operation with the standard operation. The parameter integration unit adjusts the operation difficulty based on the student's current performance, tightening the angle tolerance range in the simulation operation from ±5 degrees to ±3 degrees, while adding a real-time angle indicator to the operation interface, ultimately outputting a dynamic teaching element package containing the teaching content sequence and difficulty parameters. The disturbance testing unit injects virtual deviation parameters before teaching implementation, simulating an unexpected scenario where the puncture needle suddenly slips, causing a 15-degree angle deviation, testing the system's response capability under sudden conditions. The stability monitoring unit recorded how the intelligent push model handled such anomalies, finding that the system could promptly pause operations and prompt corrective measures, achieving a satisfactory teaching stability index of 0.82. The model retraining unit remained on standby, ready to initiate the network parameter update process if the index fell below 0.75. The entire strategy generation process was completed within 400 milliseconds, and students received personalized teaching plans immediately after completing the practice. The system interface recommended starting with 15 minutes of angle control-specific training, including 3D angle perception exercises and real-time feedback simulation operations. The teaching unit dynamically adjusted the focus based on the students' weaknesses, providing more detailed visual guidance and tactile feedback in the angle control section, while appropriately reducing the number of repetitions in the well-mastered disinfection procedures section.This dynamic adjustment mechanism based on real-time assessment ensures that teaching resources are focused on the areas most in need of improvement, thereby increasing learning efficiency.

[0026] Example 4: The real-time feedback module continuously collects student operation data during the teaching process. A pressure sensor records the force change curve during puncture, an infrared positioning device captures the deviation between the instrument's spatial coordinates and the standard path, and a physiological parameter simulator outputs simulated data of blood pressure and blood flow rate. This data is packaged at a frequency of 30 frames per second to generate an operation log, which includes fields such as timestamps, instrument offset, and physiological parameter readings. The anomaly response module activates the risk feature extraction engine to identify continuous abnormal patterns from the operation log, such as persistently excessive puncture force or abnormal fluctuations in the blood pressure curve. Subsequently, the feature vectors are matched with 500 historical cases in the teaching case library for similarity. The matching algorithm employs an improved dynamic time warping technique, ultimately outputting a teaching strategy adjustment instruction containing content adjustment suggestions.

[0027] The event slicing unit performs intelligent time window segmentation on detected operational risk features. When a "puncture angle deviation" risk event is detected, an analysis window is formed by extending 5 seconds forward and backward from the event point. The unit automatically extracts the operation video sequence within this time period and marks keyframes, generating an analysis dataset containing 10 risk event segments. The root cause analysis unit calls a risk classification model based on a deep neural network. The model takes into account the multimodal features of the risk event segments (including motion trajectory, pressure curve, and image features), extracts spatial features through three convolutional layers, captures temporal dependencies through a bidirectional LSTM layer, and outputs error attribution results, such as attributing an abnormal blood pressure fluctuation to "insufficient puncture depth" or "excessive tubing torsion." The strategy matching unit retrieves similar cases from the teaching case library. The search criteria include three dimensions: error type, severity, and student level. The unit returns the top 5 most matching teaching intervention templates to form a candidate strategy set. The instruction generation unit uses a weighted scoring mechanism to evaluate candidate strategies, calculating three matching indicators for each strategy: operation scenario similarity (based on image feature comparison), error pattern consistency (based on operation parameter analysis), and historical teaching effectiveness score (based on past application data). The table below shows the strategy evaluation results for a "pipe connection leak" incident. See Table 1.

[0028] Table 1: Evaluation Results of Teaching Strategy Matching Strategy Number Intervention methods Scene similarity Error Consistency Historical ratings Overall Score STG-202 3D animation demonstration 0.87 0.92 4.8 0.89 STG-156 Virtual hands-on training 0.93 0.85 4.5 0.86 STG-309 Error Video Comparison 0.78 0.94 4.9 0.85 STG-041 Step-by-step slow-motion demonstration 0.85 0.79 4.7 0.82 STG-278 Expert video explanation 0.75 0.88 4.6 0.80 Based on the scoring results, the STG-202 strategy is selected to generate teaching strategy adjustment instructions. These instructions include a triple adjustment scheme: immediately interrupting the current operation, pushing a 3D animation demonstration, and reducing the difficulty of subsequent operations. Upon receiving the teaching strategy adjustment instructions, the model update module initiates the online learning process. First, it extracts the instrument coordinate data related to the erroneous operation from the current operation log, constructing a fine-tuning dataset containing 200 samples. Then, it incrementally updates the region detection parameters of the instrument recognition network, using a moving average algorithm to adjust the detection threshold parameters. The strategy optimization module injects the updated network parameters into the intelligent push model, recalculating the priority weights of knowledge points in the teaching node transition matrix. When frequent "tubular connection" errors are detected, the weight of this knowledge point is increased from 0.6 to 0.9, while simultaneously increasing the frequency of related theoretical explanations. The entire anomaly response cycle is completed within 800 milliseconds, enabling real-time dynamic adjustment of the teaching strategy. When processing an arterial puncture operation for a student, the real-time feedback module detects that the puncture needle trajectory continuously deviates from the standard path by more than 15°, and the operation log records an instrument offset of 22mm. The anomaly response module triggers a high-risk event warning. The event slicing unit extracts an 8-second video clip before and after the offset occurs. The root cause analysis model outputs a "needle holding posture error" judgment. The strategy matching unit returns three targeted training plans. The instruction generation unit selects "handheld instrument correction simulation training" as the optimal strategy. The model update module collects the offset data to update the instrument trajectory recognition parameters. The strategy optimization module prioritizes the "puncture instrument operation" knowledge point to the highest level. The system then pushes dynamic teaching elements containing a 3D demonstration and simulation training of the needle holding technique.

[0029] Example 5: The cross-platform adaptation module constructs a teaching terminal adapter during the initialization phase of the teaching system. This adapter, as core middleware, loads the device configuration file library. The configuration file contains the input / output characteristics and display parameters of different hardware platforms. For example, when a student is detected using a touch tablet device, the adapter automatically records that the device's single-point touch accuracy is ±1.2mm and the maximum supported resolution is 2560×1600. The instruction conversion module receives the dynamic teaching element package output by the teaching strategy generation module. This data package contains a sequence of knowledge points and operation difficulty parameters. During the conversion process, it first parses the interaction capability matrix of the target device. For example, it recognizes that the VR headset device supports 6-DOF controller operation, while the desktop only supports keyboard and mouse input. Then, it maps the abstract teaching instructions into device-specific interaction logic.

[0030] When the system needs to present a 3D demonstration of "dialysis machine installation" on a mobile terminal, the instruction conversion module queries the operation instruction mapping table in the device interface rule base. Finding that the mobile terminal does not support the mouse hover preview function of the PC, it automatically replaces the interactive instruction with a "long press to show details" touch control scheme. The context preservation module maintains the teaching logic chain during conversion. For the "review of pipeline connection errors" teaching unit, it preserves spatial location context information on the VR device to ensure that students can point to specific error locations using the controller; while on the desktop, it converts it to mouse click highlighting to maintain the accuracy of error location. The rule base construction unit dynamically updates the device interface rule base. When new AR glasses are connected to the system, it automatically collects their gesture recognition range and display field of view parameters, adding entries including: mapping a two-hand pinch gesture to an instrument selection operation, a 1-second gaze confirmation instruction at the center of the field of view, and limiting the maximum number of faces of a renderable model to 500,000. When processing the dynamic elements of "Venvenipuncture Teaching," the syntax parsing unit decomposes them into a hierarchical structure of an abstract syntax tree: the root node is the operation demonstration type, and child nodes contain elements such as instrument objects (puncture needles), motion parameters (angle 35°), and physiological feedback (blood flow simulation), generating platform-independent intermediate representation layer data. The instruction substitution unit matches the target device capabilities based on the intermediate representation layer data. For example, it converts the "force feedback" node in the syntax tree into: outputting a 5N resistance simulation on devices that support force feedback gloves, and converting it into a visual vibration effect on ordinary devices. When the conflict resolution unit detects a conflict between the AR glasses' field of view limitation and the complex model rendering requirements, it initiates a dependency injection process, automatically reducing the model detail level and adjusting the layout of teaching elements to ensure that key operation prompts are located in the center of the field of view. The performance optimization module injects rendering parameters according to the device type. For smartwatches, it limits the teaching animation frame rate to 15fps and compresses model textures to 256×256 resolution; for high-performance workstations, it enables 4K textures and physically accurate rendering. The final generated teaching content data package includes adaptive resource packages, such as automatically encapsulating lightweight video tutorials to replace real-time rendering when transmitted to mobile terminals, providing VR devices with high-precision interactive models.

[0031] In the deployment case of the hemodialysis teaching system, when students learned the "puncture operation" unit using a tablet, the system detected the device as a 12.9-inch iPad Pro on the iOS platform. The rule base matched the display parameters to a resolution of 2732×2048. The instruction conversion module mapped the mouse drag operation on the PC to a two-finger rotation and zoom gesture. When the teaching progressed to the "pressure monitoring error review" stage, the context preservation module maintained the synchronization between the pressure curve and the timeline of the operation video, ensuring that when sliding the timeline on the touch device, the video frames and data curves remained precisely linked. The performance optimization module optimized the number of faces in the 3D dialyzer model from 200,000 to 80,000 based on the device's GPU capabilities, adjusted the rendering quality parameters to medium, and ultimately controlled the size of the generated teaching package to within 15MB, achieving loading within 2 seconds. The teaching terminal adapter demonstrated its core value in the mixed reality device scenario. When students used HoloLens2 for simulation operations, the adapter parsed its spatial anchor point positioning system and gesture interaction interface, mapping the "instrument pickup" command to an air grab gesture and converting "parameter adjustment" into the voice command "increase blood flow rate". The conflict resolution unit detected a 40-degree field-of-view limitation on the device and automatically repositioned the teaching prompts to the center of the field of view, preventing key instructional content from exceeding the visible range. The entire adaptation process is completed in the background, and students experience no difference when switching learning progress across devices, as the teaching data package is automatically synchronized and adapted to the characteristics of the new device.

[0032] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A hemodialysis teaching system based on AI intelligent push, characterized in that, include: The operation feature extraction module is used to acquire hemodialysis teaching video stream data and extract multimodal operation features from the video stream data. The multimodal operation features include instrument movement trajectory sequences and physiological parameter simulation data. The instrument positioning and analysis module calls a pre-trained instrument recognition network to perform spatial positioning processing on the video stream data, generating an instrument coordinate dataset. The instrument coordinate dataset contains a mapping relationship between instrument type identifiers and three-dimensional position codes. The operation step parsing module, based on the instrument coordinate dataset, calls the action segmentation model to perform temporal feature parsing on the instrument motion trajectory sequence, and generates feature analysis results containing step integrity features and error operation markers; The teaching strategy generation module inputs the feature analysis results into the intelligent push model to plan personalized teaching paths and outputs dynamic teaching elements. The dynamic teaching elements include a sequence of knowledge points and operation difficulty adjustment parameters generated based on the student's ability model.

2. The AI-based intelligent push-based hemodialysis teaching system according to claim 1, characterized in that, The operation feature extraction module includes: The video frame segmentation unit performs keyframe extraction processing on the video stream data to generate multiple operation stage segments and corresponding timestamp indices. The trajectory reconstruction unit extracts the motion optical flow features of each operation stage segment and performs temporal attention weighting processing on the motion optical flow features to generate a weighted trajectory feature vector. The physiological parameter association unit performs similarity matching between the trajectory feature vector and a preset physiological parameter threshold, and labels abnormal operation events. The multimodal fusion unit performs cross-modal alignment processing on the abnormal operation event labels to generate multimodal operation features that include device usage trajectory and physiological change curves.

3. The AI-based intelligent push-based hemodialysis teaching system according to claim 2, characterized in that, The instrument positioning and analysis module includes: The spatial rasterization unit divides the operation stage segment into a three-dimensional spatial mesh, generating multiple instrument detection areas and corresponding voxel coordinate parameters; The feature encoding unit is used to extract the depth convolution features of each instrument detection area and perform spatial pyramid pooling on the depth convolution features to generate a multi-scale feature tensor. The positioning and calibration unit performs feature matching between the multi-scale feature tensor and a preset instrument template library to filter out candidate instrument regions with acceptable confidence levels. The coordinate optimization unit performs non-maximum suppression processing on the candidate instrument region to generate an instrument coordinate dataset containing the instrument type identifier. Each instrument coordinate contains a normalized position code relative to the teaching equipment reference point.

4. The AI-based intelligent push-based hemodialysis teaching system according to claim 3, characterized in that, The operation step parsing module includes: The operation slicing unit performs spatiotemporal slicing processing on the video stream data based on the normalized position encoding in the instrument coordinate dataset to generate multiple operation step segments; The trajectory standardization unit performs velocity normalization and path smoothing on each operation step segment to generate standardized motion analysis input. The step segmentation unit is used to call the bidirectional temporal convolutional subnet in the action segmentation model to divide the action analysis input into stages and generate standard step sequence labels. The error detection unit is used to call the operation rule base in parallel to identify violations of the operation step segments and generate an error operation probability distribution; The feature integration unit performs feature cross-fusion between the standard step sequence labels and the error operation probability distribution to generate feature analysis results that include step completeness scores and error types.

5. The AI-based intelligent push-based hemodialysis teaching system according to claim 4, characterized in that, The teaching strategy generation module includes: The competency assessment unit analyzes the current student's historical operation records and generates knowledge weakness weights associated with the feature analysis results. The path planning unit generates a teaching node transition matrix based on the weights of the knowledge weaknesses and prioritizes the standard step sequence labels. The strategy optimization unit uses an adaptive push strategy to select the optimal teaching sequence from the teaching node transition matrix and generates a set of teaching units that includes theoretical explanations, simulated operations, and error reviews. The parameter integration unit logically associates the teaching unit set with the operation difficulty adjustment parameters to generate dynamic teaching elements that conform to the individual learning curve.

6. The AI-based intelligent push-based hemodialysis teaching system according to claim 5, characterized in that, Also includes: The real-time feedback module collects student operation data during the teaching process and generates operation logs that include instrument positioning deviations and abnormal physiological parameter values. The anomaly response module extracts the operational risk features from the operation log, performs pattern matching between the operational risk features and the teaching case library, and generates teaching strategy adjustment instructions. The model update module updates the detection parameters of the instrument recognition network online based on the teaching strategy adjustment instructions; The strategy optimization module injects the updated detection parameters into the intelligent push model and recalculates the knowledge point priority weights in the teaching node transition matrix.

7. The AI-based intelligent push-based hemodialysis teaching system according to claim 6, characterized in that, The anomaly response module includes: The event slicing unit divides the operational risk characteristics into time windows to generate multiple risk event segments and corresponding operation video sequences. The root cause analysis unit is used to call the pre-trained risk classification model to trace the source of errors in each risk event segment and generate error classification labels including puncture angle error, abnormal tubing connection, and parameter setting error. The strategy matching unit retrieves teaching intervention templates corresponding to the misclassification labels from the teaching case library and generates a candidate strategy set. The instruction generation unit, based on the matching degree evaluation between the operation video sequence and the candidate strategy set, selects the strategy with the highest confidence to generate the teaching strategy adjustment instruction.

8. The AI-based intelligent push-based hemodialysis teaching system according to claim 7, characterized in that, The teaching strategy generation module also includes: The disturbance test unit injects virtual operation deviation parameters before the teaching is pushed out. The virtual operation deviation parameters are used to simulate the positioning offset scenario of the instrument. The stability monitoring unit is used to monitor the adaptation and processing results of the intelligent push model to deviation scenarios and generate teaching stability indicators. The model retraining unit triggers the incremental learning mode of the device recognition network when the teaching stability index falls below a preset threshold. The parameter optimization unit updates the convolutional layer parameters of the instrument recognition network using gradients based on the operational difference data before and after the deviation.

9. The AI-based intelligent push-based hemodialysis teaching system according to claim 8, characterized in that, Also includes: A cross-platform adaptation module is used to build a teaching terminal adapter and to resolve the differences in the operation interfaces of different hardware devices through the teaching terminal adapter. The instruction conversion module converts the dynamic teaching elements into interactive instructions supported by the target device; The context preservation module is used to maintain the logical dependencies and error handling context of the knowledge point sequence during the transformation process; The performance optimization module is used to inject device-matched rendering parameters and generate teaching content data packages that support multi-terminal operation.

10. The AI-based intelligent push-based hemodialysis teaching system according to claim 9, characterized in that, The cross-platform adaptation module includes: The rule base construction unit is used to establish a device interface rule base and store operation instruction mapping tables and parameter constraint paths for various terminals. The syntax parsing unit performs abstract syntax tree parsing on the dynamic teaching elements to generate intermediate representation layer data; The instruction replacement unit queries the operation instruction mapping table based on the intermediate presentation layer data to generate a terminal-compatible instruction conversion scheme; The conflict resolution unit performs dependency injection on conflicting parameter constraint paths to generate unambiguous device interaction commands.

Images