Hemodialysis teaching system based on AI intelligent pushing

Abstract

The application relates to the technical field of hemodialysis teaching systems, and discloses a hemodialysis teaching system based on AI intelligent pushing. The system comprises four modules, namely, an operation feature extraction module, an instrument positioning analysis module, an operation step analysis module and a teaching strategy generation module. The operation feature extraction module acquires a hemodialysis teaching video stream and extracts multi-modal features containing instrument motion trajectory sequences and physiological parameter simulation data; the instrument positioning analysis module calls a pre-trained network for positioning and generates a coordinate data set containing instrument type and three-dimensional position encoding mapping; the operation step analysis module analyzes time sequence features by using a motion segmentation model and generates a result containing step integrity and error operation labels; and the teaching strategy generation module inputs the result into an intelligent pushing model, plans a personalized teaching path, outputs dynamic teaching elements containing a knowledge point sequence of a student ability model and operation difficulty parameters, and helps to realize precise and personalized hemodialysis teaching.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hemodialysis teaching system, in particular to a hemodialysis teaching system based on AI intelligent pushing. BACKGROUND

[0002] As an important treatment method for maintaining the life of patients with end-stage renal disease, hemodialysis treatment is directly related to the safety and effectiveness of patient treatment in terms of operation standardization and accuracy. The professional ability training of hemodialysis operators depends on a systematic and efficient teaching system. Current hemodialysis teaching mainly adopts the mode of traditional offline practical training combined with fixed video courses, which has many limitations in actual teaching process.

[0003] From the perspective of teaching content delivery, existing hemodialysis teaching videos are mostly fixed recording standardized operation processes, which cannot be dynamically adjusted according to the actual learning progress and operation mastery of students. In the learning process, students may have understanding deviations on the operation steps of certain specific instruments or the monitoring points of physiological parameters, but fixed videos cannot focus on explaining these individualized problems, resulting in blind spots in the mastery of key knowledge points by students.

[0004] In terms of instrument operation guidance, teachers in traditional teaching rely on naked eye observation to evaluate students' instrument operation, which is difficult to accurately capture the subtle deviations of instrument movement trajectory. For example, in the process of hemodialysis pipeline connection, whether the three-dimensional spatial position of the instrument is accurate and the movement trajectory meets the standard operation requirements, these detailed information is difficult to be recorded and analyzed comprehensively and accurately through artificial observation, and then it is difficult to provide targeted operation correction suggestions for students, affecting the efficiency of students' operation skill improvement.

[0005] From the perspective of operation step evaluation, the existing teaching mode lacks systematic analysis means for the completeness and correctness of students' operation steps. In the process of simulation operation or actual practical training, students may have problems such as omission of operation steps, reversal of order or wrong operation, but teachers can only make overall evaluation after the operation is completed, and cannot track each link in the operation process in real time, and it is also difficult to accurately locate the specific type and occurrence node of the wrong operation, so that students cannot find and correct the problems in the operation in time, which is not conducive to the formation of standard operation habits.

[0006] In terms of individualized teaching path planning, traditional hemodialysis teaching adopts a "one size fits all" teaching mode, ignoring the differences in learning basis, acceptance ability and learning needs of different students. Some students may master the instrument identification part quickly, but have difficulties in physiological parameter analysis; while another part of students may be the opposite. Due to the lack of construction and application of individual ability model of students, it is difficult to develop individualized teaching programs for different students, resulting in unreasonable allocation of teaching resources, and the learning needs of some students cannot be met, and the overall teaching effect is not good. SUMMARY

[0007] The present application aims to provide an AI intelligent push-based hemodialysis teaching system to solve the problems raised in the background art.

[0008] To achieve the above-mentioned purpose, the present application provides an AI intelligent push-based hemodialysis teaching system, which comprises:

[0009] An operation feature extraction module is configured to acquire hemodialysis teaching video stream data and extract multi-modal operation features in the video stream data, wherein the multi-modal operation features include instrument motion trajectory sequences and physiological parameter simulation data.

[0010] An instrument positioning analysis module is configured to call a pre-trained instrument recognition network to perform spatial positioning processing on the video stream data, thereby generating an instrument coordinate data set, wherein the instrument coordinate data set includes a mapping relationship between instrument type identification and three-dimensional position encoding.

[0011] An operation step analysis module is configured to call an action segmentation model to perform time series feature analysis on the instrument motion trajectory sequences based on the instrument coordinate data set, thereby generating a feature analysis result including step integrity features and error operation labels.

[0012] A teaching strategy generation module is configured to input the feature analysis result into an intelligent push model to plan a personalized teaching path, thereby outputting dynamic teaching elements, wherein the dynamic teaching elements include a knowledge point sequence generated based on a student ability model and an operation difficulty adjustment parameter.

[0013] Preferably, the operation feature extraction module comprises:

[0014] A video frame division unit is configured to perform key frame extraction processing on the video stream data, thereby generating a plurality of operation stage segments and corresponding timestamp indexes.

[0015] A trajectory reconstruction unit is configured to extract motion optical flow features of each operation stage segment and perform time series attention weighting processing on the motion optical flow features, thereby generating weighted trajectory feature vectors.

[0016] A physiological parameter association unit is configured to perform similarity matching on the trajectory feature vectors and preset physiological parameter thresholds, thereby labeling abnormal operation event labels.

[0017] A multi-modal fusion unit is configured to perform cross-modal alignment processing on the abnormal operation event labels, thereby generating multi-modal operation features including instrument use trajectories and physiological change curves.

[0018] Preferably, the instrument positioning analysis module comprises:

[0019] a spatial gridding unit configured to perform three-dimensional spatial meshing on the operation stage segment to generate a plurality of instrument detection regions and corresponding voxel coordinate parameters;

[0020] a feature encoding unit configured to extract deep convolution features of each instrument detection region and perform spatial pyramid pooling processing on the deep convolution features to generate a multi-scale feature tensor;

[0021] a positioning calibration unit configured to perform feature matching between the multi-scale feature tensor and a preset instrument template library to filter out candidate instrument regions with confidence scores meeting a standard;

[0022] a coordinate optimization unit configured to perform non-maximum suppression processing on the candidate instrument regions to generate an instrument coordinate dataset containing the instrument type identifier, each instrument coordinate containing a normalized position code relative to a teaching device reference point.

[0023] Preferably, the operation step parsing module comprises:

[0024] an operation slicing unit configured to perform spatio-temporal slicing processing on the video stream data according to the normalized position code in the instrument coordinate dataset to generate a plurality of operation step segments;

[0025] a trajectory standardization unit configured to perform speed normalization and path smoothing processing on each operation step segment to generate a standardized motion analysis input;

[0026] a step segmentation unit configured to call a bidirectional temporal convolution subnetwork in a motion segmentation model to perform stage division on the motion analysis input to generate a standard step sequence label;

[0027] an error detection unit configured to call an operation rule library in parallel to identify illegal actions in the operation step segment to generate an error operation probability distribution;

[0028] a feature integration unit configured to perform feature cross-fusion between the standard step sequence label and the error operation probability distribution to generate a feature analysis result containing a step completeness score and an error type.

[0029] Preferably, the teaching strategy generation module comprises:

[0030] a capability evaluation unit configured to parse historical operation records of a current student to generate a knowledge weak point weight associated with the feature analysis result;

[0031] a path planning unit configured to generate a teaching node transition matrix based on the knowledge weak point weight and to perform priority sorting on the standard step sequence label;

[0032] The strategy optimization unit selects an optimal teaching sequence from the teaching node transition matrix using an adaptive push strategy, and generates a teaching unit set including theoretical explanation, simulation operation, and error review;

[0033] The parameter integration unit logically associates the teaching unit set with the operation difficulty adjustment parameter, and generates dynamic teaching elements that conform to individual learning curves.

[0034] Preferably, the system further comprises:

[0035] The real-time feedback module collects student operation data during teaching execution, and generates an operation log including instrument positioning deviation and physiological parameter abnormal value;

[0036] The abnormal response module extracts operation risk features from the operation log, and performs pattern matching between the operation risk features and a teaching case library to generate a teaching strategy adjustment instruction;

[0037] The model update module updates the detection parameters of the instrument recognition network online based on the teaching strategy adjustment instruction;

[0038] The strategy optimization module injects the updated detection parameters into the intelligent push model, and recalculates the knowledge point priority weight in the teaching node transition matrix.

[0039] Preferably, the abnormal response module comprises:

[0040] The event slicing unit divides the operation risk features into time windows, and generates a plurality of risk event segments and corresponding operation video sequences;

[0041] The root cause analysis unit is configured to call a pre-trained risk classification model to perform error tracing on each risk event segment, and generate error classification labels including puncture angle error, pipeline connection abnormality, and parameter setting error;

[0042] The strategy matching unit retrieves a teaching intervention template corresponding to the error classification label from the teaching case library, and generates a candidate strategy set;

[0043] The instruction generation unit selects a strategy with the highest confidence level based on the matching degree evaluation between the operation video sequence and the candidate strategy set, and generates a teaching strategy adjustment instruction.

[0044] Preferably, the teaching strategy generation module further comprises:

[0045] The perturbation test unit injects a virtual operation deviation parameter before teaching push, and the virtual operation deviation parameter is used to simulate instrument positioning deviation scenarios;

[0046] A stability monitoring unit is configured to monitor the adaptive processing result of the intelligent push model to the deviation scenario and generate a teaching stability index.

[0047] A model retraining unit is configured to trigger an incremental learning mode of the instrument recognition network when the teaching stability index is lower than a preset threshold.

[0048] A parameter optimization unit is configured to perform gradient update on the convolution layer parameters of the instrument recognition network based on the operation difference data before and after the deviation.

[0049] Preferably, the system further comprises:

[0050] A cross-platform adaptation module is configured to construct a teaching terminal adapter and analyze the operation interface difference of different hardware devices through the teaching terminal adapter.

[0051] An instruction conversion module is configured to convert the dynamic teaching element into an interactive instruction supported by a target device.

[0052] A context reservation module is configured to maintain the logical dependency relationship and error processing context of the knowledge point sequence during the conversion process.

[0053] A performance optimization module is configured to inject a device-matched rendering parameter to generate a teaching content data package supporting multi-terminal operation.

[0054] Preferably, the cross-platform adaptation module comprises:

[0055] A rule base construction unit is configured to establish a device interface rule base and store the operation instruction mapping table and parameter constraint path of various terminals.

[0056] A syntax analysis unit is configured to perform abstract syntax tree analysis on the dynamic teaching element to generate intermediate representation layer data.

[0057] An instruction replacement unit is configured to query the operation instruction mapping table based on the intermediate representation layer data to generate a terminal-compatible instruction conversion scheme.

[0058] A conflict resolution unit is configured to perform dependency injection processing on the parameter constraint path with conflicts to generate a device interactive instruction without ambiguity.

[0059] Compared with the prior art, the present application has the following advantages:

[0060] The operation feature extraction module can obtain hemodialysis teaching video stream data and extract multi-modal operation features including instrument motion trajectory sequences and physiological parameter simulation data from the data. This module breaks through the limitation of traditional teaching that only relies on fixed video to deliver single information, combines the dynamic trajectory of the instrument motion with the simulation data of the physiological parameters, and provides more comprehensive and rich basic data for subsequent teaching analysis. The operation process of the student is no longer an isolated action display, but a complete information set associated with key parameter data, which enables the teaching analysis to go deep into the operation details and data association level, helping to find potential problems and improvement space in the operation process.

[0061] The instrument positioning analysis module calls a pre-trained instrument recognition network to perform spatial positioning processing on the video stream data, generating an instrument coordinate data set containing the mapping relationship between instrument type identification and three-dimensional position encoding. This module realizes the accurate positioning and recognition of hemodialysis instruments in three-dimensional space. Compared with the traditional manual observation that cannot accurately capture instrument position information, it can objectively and accurately record the spatial position changes of different types of instruments during the operation process. Whether it is the core component of the hemodialysis machine or the auxiliary instruments such as pipelines and puncture needles, their position information can be accurately encoded, providing a quantitative basis for subsequent judgment of whether the instrument operation is standardized. This enables the evaluation of instrument operation to shift from subjective judgment to objective data support, avoiding errors and omissions in manual evaluation.

[0062] The operation step analysis module, based on the instrument coordinate data set, calls an action segmentation model to analyze the time sequence features of the instrument motion trajectory sequence, generating feature analysis results containing step integrity features and error operation markers. This module can track the entire time sequence process of the student's operation in real time, associate the instrument motion trajectory with the operation steps, not only can judge whether the operation steps are complete, avoiding the problem of step omission in traditional teaching that cannot be monitored in real time, but also can accurately mark the specific nodes and types of error operations. For example, in the preparation operation before hemodialysis treatment, if the student skips a disinfection step or makes a sequence error when connecting the instruments, this module can identify and mark it in time, so that the student can clearly understand the problems in their own operation, facilitate timely adjustment and correction, and promote the formation of standardized operation habits.

[0063] The teaching strategy generation module inputs the feature analysis result into the intelligent pushing model for personalized teaching path planning, and outputs dynamic teaching elements including a knowledge point sequence generated based on the learner ability model and an operation difficulty adjustment parameter. This module breaks the traditional "one-size-fits-all" teaching mode, and fully considers the learning basis and operation mastery of different learners by constructing a learner ability model. For learners who are strong in instrument recognition ability but lack in 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 learners who are not skilled in operation steps, the system will focus on pushing knowledge points related to operation process, and gradually improve the operation proficiency of learners by adjusting the operation difficulty parameter. This personalized teaching path planning can fully meet the learning needs of different learners, reasonably allocate teaching resources, and enable each learner to gradually improve their professional ability at their own learning pace, avoiding the low learning efficiency problem caused by the mismatch between teaching content and learner ability in traditional teaching. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 a timing diagram of the AI intelligent pushing based hemodialysis teaching system described in the present application;

[0065] Figure 2 a flowchart for the operation feature extraction module;

[0066] Figure 3 a flowchart for the teaching strategy generation module. DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0068] Please refer to Figure 1 The present application provides an AI intelligent pushing based hemodialysis teaching system, which comprises:

[0069] The hemodialysis teaching video stream data is acquired and multi-modal operation features are extracted, the multi-modal operation features including instrument motion trajectory sequence and physiological parameter simulation data, then a pre-trained instrument recognition network is called to perform spatial positioning processing on the video stream data to generate an instrument coordinate data set, the data set including a mapping relationship between instrument type identification and three-dimensional position coding, based on the instrument coordinate data set, a motion segmentation model is called to analyze the time sequence features of the instrument motion trajectory sequence, generating a feature analysis result including step integrity features and error operation labels, finally the feature analysis result is input into an intelligent push model for personalized teaching path planning, and dynamic teaching elements are output, the dynamic teaching elements including a knowledge point sequence generated based on a student ability model and an operation difficulty adjustment parameter. The system realizes full-process automatic processing from video data acquisition to personalized teaching push, and can adapt to the learning needs and operation level of different students.

[0070] Embodiment 1: refer to Figure 2 The operation feature extraction module receives real-time video stream data from the hemodialysis teaching device, a video frame unit adopts a hybrid algorithm based on inter-frame difference and optical flow analysis to detect operation stage switching points, and triggers key frame interception when detecting that the instrument motion trajectory has a significant change or a specific operation gesture appears, generating an operation stage segment sequence with accurate timestamp index, each segment corresponding to a standard step of hemodialysis operation such as arterial puncture or dialyzer installation. The trajectory reconstruction unit calculates the pixel displacement field of adjacent frames in the operation stage segment using the Horn-Schunck optical flow algorithm, maps the two-dimensional displacement vector to the teaching device coordinate system through three-dimensional reconstruction technology, and forms an initial motion trajectory point cloud; the time sequence attention weighting processing introduces a weighting mechanism based on motion amplitude, assigns higher weight to the high-speed moving area of the instrument, and at the same time suppresses background noise interference, finally outputs a trajectory feature vector representing the motion law of the instrument, the vector including kinematic parameters such as displacement and acceleration. The physiological parameter correlation unit is built-in with a blood pressure fluctuation model and a hemodynamics simulator, which starts real-time parameter mapping when receiving the trajectory feature vector, dynamically compares the instrument motion speed with the preset blood vessel pressure threshold, calculates the dynamic time warping distance of the trajectory feature and the physiological parameter curve using a sliding time window, and automatically labels abnormal event labels such as "puncture speed out of limit" or "pipe pressure abnormality" when the distance value exceeds the safe operation threshold. The multi-modal fusion unit deploys a time axis synchronization engine, aligns the time sequences of the instrument motion trajectory and the physiological parameter curve through the dynamic time warping algorithm, establishes a cross-modal correlation with millisecond-level precision based on the timestamp index, and generates a multi-modal operation feature matrix that fuses instrument spatial coordinates, motion vectors and physiological parameter fluctuations, which is the basic data set for subsequent analysis.

[0071] The instrument positioning analysis module loads the pre-calibrated teaching equipment spatial coordinate system when starting, and the spatial gridding unit divides the operation area into three-dimensional voxels using an octree structure. 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 coding unit calls the ResNet-50 convolutional neural network to extract the deep features of each voxel region. The spatial pyramid pooling layer uses 1x1, 2x2, and 4x4 grid scales to perform multi-level pooling on the convolution feature map, generating a multi-scale feature tensor that integrates local details and global context. The positioning calibration unit uses a multi-view geometric constraint strategy in the feature matching stage, calculates the Euclidean distance between the multi-scale feature tensor and the standard features in the instrument template library, and determines effective recognition when the matching similarity exceeds the 0.95 confidence threshold, while generating a candidate region proposal containing the instrument type and spatial position. The coordinate optimization unit implements a non-maximum suppression algorithm based on overlap analysis, calculates the three-dimensional intersection-over-union of all candidate regions, and only retains the proposal with higher confidence when the overlap of two detection boxes exceeds 30%. The final output instrument coordinate dataset contains three types of key information: instrument type code using the international medical instrument identification system, three-dimensional position coordinates establishing a right-handed coordinate system with the teaching equipment reference point as the origin, and normalized position encoding mapping the actual coordinates to the [0,1] interval through linear transformation, such as converting the puncture needle tip coordinates (325, 120, 80) to (0.72, 0.35, 0.42). This dataset records the relative position relationship between instruments through spatial relationship descriptors, providing structured spatial information for subsequent operation step analysis. The entire processing flow is implemented on an embedded inference device for end-to-end execution, with a delay control of 200 milliseconds from video input to coordinate dataset output, meeting the response requirements of real-time teaching systems.

[0072] Embodiment 2: The operation step parsing module receives the instrument coordinate dataset from the instrument positioning analysis module, which contains instrument type identification and normalized position encoding with timestamps. The operation slice 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 appears or a spatial distribution mutation is detected, the segment cutting is triggered, generating independent operation step segments such as "pipe preflush stage" or "puncture operation stage". Each segment is attached with start and end time markers accurate to milliseconds. The trajectory standardization unit implements a multi-level processing flow on the instrument motion trajectory within each step segment. First, the Savitzky-Golay filter is used to smooth the trajectory high-frequency jitter. Then, linear interpolation is used to unify trajectories of different lengths 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 that uses a bidirectional time series convolution architecture containing three forward convolution layers and three backward convolution layers. The bidirectional network simultaneously captures historical and future action dependency relationships. After inputting the standardized trajectory sequence, it outputs a step boundary probability distribution map. Through a dynamic threshold segmentation algorithm, it identifies standard step labels such as "puncture needle positioning", "catheter connection", and "parameter setting". Each label is attached with a confidence score and accurate time interval. The error detection unit synchronously starts the rule engine scan. The operation rule library stores more than 200 hemodialysis operation specification items, including "puncture angle 30-45 degrees", "pipe connection torque threshold", "physiological parameter safety range", and other constraint conditions. The rule matching uses a real-time inference mechanism. When the instrument motion trajectory violates spatial constraints or time series abnormalities are detected, a violation event is generated, such as identifying "puncture angle deviation exceeding 5 degrees" or "insufficient disinfection operation time". The output contains error type coding and occurrence time point probability distribution matrix.

[0073] The feature integration unit deploys a feature fusion pipeline, first establishes a spatio-temporal association matrix of step labels and error probabilities, mines the potential association between step execution order and error occurrence through a graph convolution network, then uses a multi-head attention mechanism to enhance the features of key error events, and finally outputs a structured feature analysis result containing three-dimensional data: step completeness score is based on the standard step execution completeness and sequence correctness to calculate the percentage value, error type distribution matrix marks the frequency and severity level of each type of error, and operation quality heat map marks the high-risk operation interval on the time axis. The result data package is packaged and transmitted through JSON format, including time-coded step sequence, error event list and associated instrument coordinate snapshot, providing interpretable operation quality evaluation for teaching strategy generation. In the real-time processing process, when parsing the "dialyzer installation" step segment, the trajectory standardization 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 time, and the error detection unit triggers the "installation pressure substandard" rule item. The feature integration unit associates these three abnormal signals to the same time interval, generates a feature analysis result containing "step integrity score 65 points" and "pressure operation error" label. The whole processing cycle is controlled within 300 milliseconds, supporting real-time generation of analysis report during student operation, and the teaching system dynamically adjusts the subsequent teaching content push strategy according to the report. The data processing process retains the timestamp alignment characteristics of all intermediate results, ensuring that the teaching feedback is accurately synchronized with the original operation video frame.

[0074] Taking a specific hemodialysis teaching scenario as an example, a student is practicing "arterial puncture" operation. The operation feature extraction module captures the motion trajectory of the puncture needle and the simulated physiological parameters in the video stream in real time. The operation step analysis module starts working. The operation slice unit automatically identifies that the current is in the "puncture needle positioning" stage according to the normalized position code (0.72, 0.35, 0.42) of the puncture needle in the instrument coordinate data set and its motion mode, and takes the starting point of the acceleration motion of the instrument as the starting point of the segment and the zero velocity moment as the end point, generating an operation step segment with a time length of 3.2 seconds, and the timestamp is marked as T1-T2 interval. The trajectory standardization unit immediately processes this segment. First, the Kalman filter algorithm is used to smooth the hand jitter noise in the puncture needle trajectory, eliminating the slight irregular vibration caused by the student's nervousness; then the 3.2-second trajectory is standardized to a unified 2.5-second time axis through linear interpolation, maintaining the original motion trend but eliminating individual speed differences; finally, the speed normalization processing is performed, which maps the actual motion speed to the range of 0-1, and outputs the standardized trajectory sequence data, which clearly shows the approaching process of the puncture needle from the initial position to the target blood vessel. The step segmentation unit calls the pre-trained bidirectional time convolution model, and after inputting the standardized trajectory sequence, the model forward convolution layer identifies three sub-steps: "needle preparation", "angle adjustment", and "puncture execution", and the backward convolution layer verifies the rationality of the step order, and outputs the step label sequence with confidence score: "needle preparation" (0.92), "angle adjustment" (0.85), and "puncture execution" (0.78), each label is labeled with accurate time boundary and spatial coordinate range. The error detection unit synchronously runs rule matching, and 32 rules in the "arterial puncture" chapter of the operation rule library are activated, including "needle angle 30-45 degrees", "needle holding stability threshold", "puncture depth limit", etc. The rule engine detects that at T1+1.8 seconds, the puncture needle angle sensor reading is 52 degrees, which exceeds the maximum allowed value of 7 degrees, and immediately generates an "angle deviation" error event; at the same time, it detects that the needle body jitter amplitude exceeds the safety threshold at T1+2.1 seconds, triggering an "insufficient stability" error mark. The system outputs the error operation probability distribution matrix, showing that the error probability in the "angle adjustment" stage is 0.87, and the error probability in the "puncture execution" stage is 0.63. The feature integration unit starts multi-source data fusion. First, it establishes the association mapping between step labels and error events, and through graph neural network analysis, it finds that the "angle deviation" error mainly occurs in the "angle adjustment" stage, and has a causal relationship with the previous "needle preparation" not sufficient; then it uses the attention mechanism to strengthen the key error features, giving the angle deviation a weight coefficient of 0.9 and the stability problem a weight of 0.7.The final structured analysis result contains: step integrity score 76 points (main deduction points in angle adjustment stage), error type distribution shows that angle error accounts for 68%, operation quality heat map clearly marks T1+1.8 seconds as high-risk moment on time axis. The whole analysis process is completed within 320 milliseconds, and the system immediately generates a detailed operation evaluation report: this arterial puncture operation is basically qualified in terms of step integrity, but there are obvious angle control problems, especially serious angle deviation in the middle of needle insertion, and the stability of the whole operation process needs to be improved. This report is transmitted to the teaching strategy generation module in real time to provide data support for subsequent personalized teaching. In the teaching system interface, the student immediately sees the three-dimensional trajectory playback of his operation, the error moment is highlighted in red, and the specific value of the angle deviation is prompted, and the system suggests that special training should be focused on "needle insertion angle control".

[0075] Example 3: see Figure 3 The teaching strategy generation module receives feature analysis result data packets from the operation step analysis module, which contains step integrity score, error type distribution matrix and time axis marked operation quality heat map, the ability assessment unit starts the student portrait engine, and retrieves the student's historical operation records in the past three months from the teaching system database, analyzes the error mode distribution and step execution time characteristics in the records, and generates a knowledge weak point weight vector associated with the current feature analysis result. Weight calculation uses time decay weighting algorithm, recent high-frequency error types get higher weight values, for example, a student has angle deviation in "arterial puncture" step for three times, the error type weight is increased to 0.87. The path planning unit constructs a teaching node transition matrix based on the weight vector, and the matrix elements represent the logical association strength between knowledge points. The transition probability is calculated by the following formula:

[0076]

[0077] Where: P ij represents the probability of transitioning from knowledge point i to knowledge point j, W j is the weak point weight of knowledge point j, S ij represents the logical tightness coefficient of knowledge points i and j in the hemodialysis operation process, and N is the total number of knowledge points. After the matrix is generated, the standard step sequence label is dynamically sorted, and when a sudden increase in "circuit connection" step error weight is detected, the priority of this knowledge point is raised to the first position in the sequence.

[0078] The policy optimization unit deploys a Q-learning-based adaptive pushing strategy, selects the path with the highest cumulative reward value from the transition matrix of the teaching node, and generates a teaching sequence containing three-phase teaching units by comprehensively considering the importance of knowledge points, historical mastery level, and current error urgency. The reward function: the theoretical explanation unit calls the three-dimensional animation library to show the principle of instrument operation; the simulation operation unit generates a virtual operation environment with force feedback; the error review unit compares the actual operation video of the student with the standard video in split screen. The parameter integration unit converts the operation difficulty adjustment parameters into specific teaching parameters through a rule mapping table. When the student's operation fluency is lower than the threshold, the instrument movement resistance in the simulation operation is automatically reduced and the operation guide markers are increased. Finally, a dynamic teaching element package that meets the individual learning curve is output, which includes a configurable teaching content sequence and a corresponding difficulty parameter configuration file.

[0079] The disturbance test unit injects a set of virtual operation deviation parameters before teaching pushing, including instrument spatial coordinate offset ±20mm, operation speed fluctuation ±30%, physiological parameter reading random disturbance ±15%, simulation of hemodialysis machine failure or student operation error scene. The stability monitoring unit runs Monte Carlo simulation to monitor the output changes of the intelligent pushing model under 100 disturbance scenarios, records the consistency level of the teaching sequence adjustment scheme, and generates a teaching stability index value in the 0-1 interval. When the index value is below the 0.75 threshold for three consecutive times, the model retraining unit activates the incremental learning mode of the instrument recognition network, collects operation data in the virtual deviation scene to construct a fine-tuning data set. The parameter optimization unit calculates the L2 norm difference of the instrument recognition feature map before and after the deviation, updates the network convolution layer parameters using the stochastic gradient descent algorithm with momentum, and sets the learning rate to 1 / 10 of the initial value. After three rounds of iteration, the recognition accuracy of the network for new offset patterns is restored to normal level. The entire optimization process is executed asynchronously in the background to ensure that the teaching pushing is not affected by parameter updates.

[0080] For example, the teaching strategy generation module receives the feature analysis results from the operation step parsing module, which shows that the student's step completeness score in the last arterial puncture practice is 76, and the main error types are angle control deviation (probability of occurrence 0.87) and insufficient operation stability (probability of occurrence 0.63). The ability assessment unit immediately retrieves the student's operation record database in the past two weeks and analyzes that similar angle deviation has occurred 7 times in historical operations, and the frequency is increasing. The system calculates the weight value of the angle control weak point as 0.86 and the weight value of the stability control as 0.72, forming the knowledge weak point weight vector [angle control: 0.86, stability: 0.72, puncture depth: 0.35, disinfection specification: 0.21]. The path planning unit constructs a teaching node transition matrix based on the weight vector, and each element in the matrix represents the transition probability between different knowledge points. The transition probability between angle control knowledge points and stability knowledge points is set to 0.78, and the transition probability with puncture depth knowledge points is 0.45. The system reorders the standard step sequence, prioritizes angle control training, and generates a new teaching sequence: angle control theory explanation → stability simulation training → puncture depth consolidation practice → complete operation assessment. The strategy optimization unit uses a push strategy based on the state value function to select the path with the highest cumulative reward value from the teaching node transition matrix, which prioritizes angle control training. The system generates a plan that includes three teaching units: the theory explanation unit calls a three-dimensional animation to demonstrate the correct 30-45 degree needle insertion angle, the simulation operation unit configures a force feedback device to provide real-time angle deviation prompts, and the error review unit compares and analyzes 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, tightens the angle tolerance range from ±5 degrees to ±3 degrees in the simulation operation, and adds a real-time angle indicator to the operation interface. Finally, the dynamic teaching element package is output, which includes the teaching content sequence and difficulty parameters. The disturbance test unit injects virtual deviation parameters before teaching to simulate an unexpected scenario where the puncture needle suddenly slides, causing a 15 degree angle deviation. The system tests its response ability in emergency situations. The stability monitoring unit records the intelligent push model's handling of such abnormalities and finds that the system can pause the operation and prompt corrective measures in time, with a teaching stability index of 0.82, meeting the qualified level. The model retraining unit remains on standby and is ready to start the network parameter update process when the index falls below 0.75. The entire strategy generation process is completed within 400 milliseconds, and the student receives a personalized teaching plan immediately after completing the operation practice. The system interface recommends that the student first perform 15 minutes of angle control training, which includes 3D angle perception exercises and real-time feedback simulation operations. The teaching unit dynamically adjusts the focus based on the student's weak point weight, providing more detailed visual guidance and tactile feedback in the angle control section, and appropriately reducing the number of repeated training in the well-mastered disinfection specification section.This dynamic adjustment mechanism based on real-time assessment ensures that teaching resources are concentrated in the areas most in need of improvement, improving learning efficiency.

[0081] In Example 4, the real-time feedback module continuously collects student operation data during the teaching process, the pressure sensor records the force variation curve during the puncture operation, the infrared positioning device captures the deviation value of the instrument spatial coordinates from the standard path, and the physiological parameter simulator outputs the simulation data of blood pressure and blood flow rate. These data are packaged to generate an operation log at a frequency of 30 frames per second, which includes fields such as timestamp, instrument deviation, physiological parameter readings, etc. The abnormal response module starts the risk feature extraction engine to identify continuous abnormal patterns from the operation log, such as continuous puncture force exceeding the limit or abnormal fluctuations in blood pressure curve, and then matches the feature vector with the 500 historical cases in the teaching case library. The matching algorithm uses an improved dynamic time warping technique, and finally outputs a teaching strategy adjustment instruction containing content adjustment suggestions.

[0082] The event slicing unit implements intelligent time window segmentation on the detected operation risk features. When a "puncture angle deviation" risk event is found, an analysis window is formed by expanding 5 seconds forward and backward from the event occurrence point, the operation video sequence in this time period is automatically intercepted and key frames are marked, and an analysis data set containing 10 risk event segments is generated. The root cause analysis unit calls a risk classification model based on deep neural networks, which inputs the multi-modal features of the risk event segments (including motion trajectory, pressure curve, image features), extracts spatial features through three convolutional layers, captures temporal dependencies through a bidirectional LSTM layer, and outputs error tracing results, such as attributing an abnormal blood pressure fluctuation to "insufficient puncture depth" or "excessive pipe twisting". The strategy matching unit retrieves similar cases from the teaching case library, with retrieval conditions including error type, severity, and student level, and returns the top 5 most matched teaching intervention templates to form a candidate strategy set. The instruction generation unit uses a weighted scoring mechanism to evaluate the candidate strategies, calculates three matching degree indicators for each strategy: operation scene similarity (based on image feature comparison), error pattern consistency (based on operation parameter analysis), and teaching effect historical score (based on past application data). The following table shows the strategy evaluation results for a "pipe connection leakage" event, see Table 1.

[0083] Table 1: Teaching strategy matching degree evaluation results

[0084] Strategy Number Intervention Type Scenario Similarity Error Consistency Historical Score Overall Score STG-202 Three-dimensional 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

[0085] According to the scoring results, the STG-202 strategy is selected to generate teaching strategy adjustment instructions, which include a triple adjustment scheme of immediately interrupting the current operation, pushing a three-dimensional animation demonstration, and reducing the difficulty of subsequent operations. The model updating module receives the teaching strategy adjustment instructions and starts the online learning process. First, it extracts the instrument coordinate data related to the error operation from the current operation log to construct a fine-tuning dataset containing 200 samples. Then, it performs incremental updates on the region detection parameters of the instrument recognition network and adjusts the detection threshold parameters using the sliding average algorithm. The strategy optimization module injects the updated network parameters into the intelligent push model and recalculates the knowledge point priority weights in the teaching node transition matrix. When the "pipe connection" error is frequently identified, the knowledge point weight is increased from 0.6 to 0.9, and the push frequency of related theoretical explanation content is also increased. The entire abnormal response cycle is completed within 800 milliseconds, achieving real-time dynamic adjustment of the teaching strategy. When the system processes the arterial puncture operation of 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 abnormal response module triggers a high-risk event warning. The event slicing unit intercepts the video segments 8 seconds before and after the offset occurs, the root cause analysis model outputs the determination result of "incorrect needle holding posture", and the strategy matching unit returns 3 targeted training schemes. The instruction generation unit selects "hand-held instrument correction simulation training" as the optimal strategy. The model updating module collects the offset data to update the instrument trajectory recognition parameters, and the strategy optimization module increases the "puncture instrument operation" knowledge point priority to the highest level. The system then pushes dynamic teaching elements including a three-dimensional demonstration of needle holding techniques and simulation training.

[0086] In the initialization phase of the teaching system, the cross-platform adaptation module constructs a teaching terminal adapter. This adapter loads the device configuration file library as the core middleware. The configuration file contains the input and output characteristics and display parameters of different hardware platforms. For example, when a student uses a touch tablet device, the adapter automatically records the single-point touch accuracy of the device as ±1.2mm and the maximum supported resolution as 2560x1600. The instruction conversion module receives the dynamic teaching element package output by the teaching strategy generation module. This data package contains knowledge point sequences and operation difficulty parameters. During the conversion process, it first analyzes the interactive capability matrix of the target device. For example, it identifies that a VR headset device supports 6-degree-of-freedom handle operation, while a desktop device only supports keyboard and mouse input. Then it maps the abstract teaching instructions to device-specific interactive logic.

[0087] When the system needs to present a "dialyzer installation" three-dimensional demonstration on a mobile terminal, the instruction conversion module queries the operation instruction mapping table in the device interface rule library and finds that the mobile terminal does not support the mouse hover preview function of the PC terminal. The interactive instruction is automatically replaced with a touch control scheme of "long press to display details". The context retention module maintains the teaching logic chain during the conversion. For the "pipeline connection error review" teaching unit, the spatial position context information is retained on the VR device to ensure that the student can point to a specific error position through the handle; and on the desktop, it is converted into a mouse click highlight marker to maintain the accuracy of error positioning. The rule library construction unit dynamically updates the device interface rule library. When a new AR glasses is connected to the system, it automatically collects its gesture recognition range and display field of view angle parameters, and adds new entries including: double-hand pinch gesture mapping to instrument selection operation, gaze at the center of the field of view for 1 second to confirm the instruction, and the maximum number of renderable model faces is limited to 500,000. The syntax analysis unit processes the "intravenous puncture teaching" dynamic elements and decomposes them into a hierarchical structure of an abstract syntax tree: the root node is the operation demonstration type, the child nodes include instrument objects (puncture needle), action parameters (angle 35°), physiological feedback (blood flow simulation) and other elements, and the platform-independent intermediate representation layer data is generated. The instruction replacement unit matches the target device capability based on the intermediate representation layer data, for example, the "force feedback" node in the syntax tree is converted to: output 5N resistance simulation on devices that support force feedback gloves, and convert to visual vibration effect on ordinary devices. When the field of view angle limit of the AR glasses conflicts with the complex model rendering requirement, the conflict resolution unit starts the dependency injection process, automatically reduces the model detail level and adjusts the teaching element layout to ensure that the key operation prompt is located in the center of the field of view. The performance optimization module injects rendering parameters according to the device type. For smart watch terminals, the teaching animation frame rate is limited to 15fps, and the model texture is compressed to 256x256 resolution; for high-performance workstations, 4K texture and physically accurate rendering are enabled. The finally generated teaching content data package contains adaptive resource packages, such as automatically packaging lightweight video tutorials instead of real-time rendering when transmitted to mobile terminals, and providing high-precision interactive models for VR devices.

[0088] In the deployment case of hemodialysis teaching system, when the student learns the unit of "puncture operation" through the tablet computer, the system detects that the equipment is an iOS platform 12.9-inch iPad Pro, the rule base matches the display parameters of 2732x2048 resolution, and the instruction conversion module maps the mouse drag operation on the PC end to the double-finger rotation zoom gesture. When the teaching advances to the "pressure monitoring error review" link, the context retention module maintains the time axis synchronization relationship between the pressure curve graph and the operation video, ensuring that when the time axis is slid on the touch device, the video frame and the data curve remain accurate linkage. The performance optimization module optimizes the three-dimensional dialyzer model surface number from 200,000 to 80,000 according to the GPU capability of the device, and adjusts the rendering quality parameter to medium, finally generates a teaching package volume controlled within 15MB, and realizes loading completion within 2 seconds. The teaching terminal adapter demonstrates its core value in the mixed reality device scene, when the student uses HoloLens2 to simulate operation, the adapter analyzes its spatial anchor positioning system and gesture interaction interface, maps the "instrument pickup" instruction to the air grab gesture, and converts the "parameter adjustment" to the voice command "increase blood flow rate". The conflict resolution unit detects that the field of view angle of the device is limited to 40 degrees, automatically repositions the teaching prompt information to the center area of the field of view, and avoids the key guidance content from exceeding the visible range. The whole adaptation process is completed in the background, and the student has no perceptual difference when switching the learning progress across devices, and the teaching data package is automatically synchronized and adapted to the new device characteristics.

[0089] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. An AI intelligent push-based hemodialysis teaching system, characterized in that, The method comprises the following steps: An operation feature extraction module is used to obtain hemodialysis teaching video stream data and extract multi-modal operation features in the video stream data, wherein the multi-modal operation features include instrument motion trajectory sequences and physiological parameter simulation data; An instrument positioning analysis module is used to call a pre-trained instrument recognition network to perform spatial positioning processing on the video stream data, thereby generating an instrument coordinate data set, wherein the instrument coordinate data set includes a mapping relationship between instrument type identification and three-dimensional position encoding; An operation step analysis module is used to call an action segmentation model to perform time sequence feature analysis on the instrument motion trajectory sequences based on the instrument coordinate data set, thereby generating a feature analysis result including step integrity features and error operation labels; A teaching strategy generation module is used to input the feature analysis result into an intelligent push model to plan an individualized teaching path, thereby outputting dynamic teaching elements, wherein the dynamic teaching elements include a knowledge point sequence generated based on a student ability model and an operation difficulty adjustment parameter; The instrument positioning analysis module comprises: A spatial gridding unit is used to perform three-dimensional spatial grid division on an operation stage segment, thereby generating a plurality of instrument detection regions and corresponding voxel coordinate parameters; A feature encoding unit is used to extract deep convolution features of each instrument detection region and perform spatial pyramid pooling processing on the deep convolution features, thereby generating a multi-scale feature tensor; A positioning calibration unit is used to perform feature matching on the multi-scale feature tensor and a pre-set instrument template library, thereby screening out candidate instrument regions with confidence up to standard; A coordinate optimization unit is used to perform non-maximum suppression processing on the candidate instrument regions, thereby generating an instrument coordinate data set including the instrument type identification, wherein each instrument coordinate includes a normalized position encoding relative to a teaching device reference point; The teaching strategy generation module comprises: An ability evaluation unit is used to analyze historical operation records of a current student, thereby generating a knowledge weak point weight associated with the feature analysis result; A path planning unit is used to generate a teaching node transfer matrix based on the knowledge weak point weight and perform priority sorting on a standard step sequence label; The path planning unit constructs a teaching node transition matrix based on the weight vector, and the matrix elements represent the logical association strength between knowledge points. The transition probability is calculated by the following formula: wherein, Pij represents the probability of transitioning from knowledge point i to knowledge point j, Wj represents the weak point weight of knowledge point j, N represents the total number of knowledge points, and the matrix is generated. After the dynamic sorting of the standard step sequence label is implemented, when a sudden increase in the error weight of the pipeline connection step is detected, the priority of the knowledge point is promoted to the first position of the sequence. A strategy optimization unit is used to select an optimal teaching sequence from the teaching node transfer matrix using an adaptive push strategy, thereby generating a teaching unit set including theoretical explanation, simulation operation, and error review; A parameter integration unit is used to perform logical association processing on the teaching unit set and the operation difficulty adjustment parameter, thereby generating dynamic teaching elements conforming to individual learning curves. 2.The AI intelligent push-based hemodialysis teaching system according to claim 1, characterized in that, The operation feature extraction module comprises: A video frame division unit is used to perform key frame extraction processing on the video stream data, thereby generating a plurality of operation stage segments and corresponding timestamp indexes; A trajectory reconstruction unit is used to extract motion optical flow features of each operation stage segment and perform time sequence attention weighting processing on the motion optical flow features, thereby generating a weighted trajectory feature vector; A physiological parameter association unit is used to perform similarity matching on the trajectory feature vector and a pre-set physiological parameter threshold, thereby labeling an abnormal operation event label; A multi-modal fusion unit is used to perform cross-modal alignment processing on the abnormal operation event label, thereby generating multi-modal operation features including instrument use trajectories and physiological change curves. 3.The AI intelligent push-based hemodialysis teaching system according to claim 1, characterized in that, The operation step analysis module comprises: An operation slicing unit, configured to perform spatio-temporal slicing on the video stream data according to the normalized position coding in the instrument coordinate data set, to generate a plurality of operation step segments; A trajectory normalization unit, configured to perform speed normalization and path smoothing on each operation step segment, to generate a standardized motion analysis input; A step segmentation unit, configured to call a bidirectional time convolution subnetwork in a motion segmentation model to divide the motion analysis input into stages, to generate a standard step sequence label; An error detection unit, configured to call an operation rule library in parallel to identify illegal actions in the operation step segment, to generate an error operation probability distribution; A feature integration unit, configured to cross and fuse the standard step sequence label and the error operation probability distribution, to generate a feature analysis result containing a step completeness score and an error type. 4.The AI intelligent push-based hemodialysis teaching system according to claim 1, characterized in that, Further comprising: A real-time feedback module, configured to collect operation data of a student during teaching execution, to generate an operation log containing instrument positioning deviation and physiological parameter abnormal value; An abnormal response module, configured to extract operation risk features from the operation log, and perform pattern matching on the operation risk features and a teaching case library, to generate a teaching strategy adjustment instruction; A model update module, configured to update detection parameters of the instrument recognition network online based on the teaching strategy adjustment instruction; A strategy optimization module, configured to inject the updated detection parameters into the intelligent pushing model, to recalculate knowledge point priority weights in the teaching node transition matrix. 5.The hemodialysis teaching system based on AI intelligent push according to claim 4, wherein, The abnormal response module comprises: An event slicing unit, configured to divide the operation risk features into time windows, to generate a plurality of risk event segments and corresponding operation video sequences; A root cause analysis unit, configured to call a pre-trained risk classification model to trace errors in each risk event segment, to generate error classification labels containing puncture angle error, pipeline connection abnormality and parameter setting error; A strategy matching unit, configured to retrieve a teaching intervention template corresponding to the error classification label from the teaching case library, to generate a candidate strategy set; An instruction generation unit, configured to select a strategy with the highest confidence based on matching degree evaluation of the operation video sequence and the candidate strategy set, to generate a teaching strategy adjustment instruction. 6.The hemodialysis teaching system based on AI intelligent push according to claim 5, wherein, The teaching strategy generation module further comprises: A perturbation test unit, configured to inject a virtual operation deviation parameter before teaching pushing, the virtual operation deviation parameter being used to simulate instrument positioning deviation scenarios; A stability monitoring unit, configured to monitor adaptive processing results of the intelligent pushing model on deviation scenarios, to generate a teaching stability index; A model retraining unit, configured to trigger an incremental learning mode of the instrument recognition network when the teaching stability index is lower than a preset threshold; A parameter optimization unit, configured to perform gradient update on convolution layer parameters of the instrument recognition network based on operation difference data before and after deviation. 7.The AI intelligent push-based hemodialysis teaching system according to claim 6, characterized in that, Further comprising: A cross-platform adaptation module, configured to construct a teaching terminal adapter, and to analyze operation interface differences of different hardware devices through the teaching terminal adapter; An instruction conversion module, configured to convert the dynamic teaching elements into interactive instructions supported by a target device. A context retention module is configured to maintain logical dependency relationships and error handling contexts of the knowledge point sequence during the conversion process. A performance optimization module is configured to inject rendering parameters matched with the device to generate a teaching content data package supporting multi-terminal operation. 8.The AI intelligent push-based hemodialysis teaching system according to claim 7, characterized in that, The cross-platform adaptation module comprises: A rule base construction unit is configured to establish a device interface rule base, and store operation instruction mapping tables and parameter constraint paths of various terminals; A syntax analysis unit is configured to perform abstract syntax tree analysis on the dynamic teaching elements to generate intermediate representation layer data; An instruction replacement unit is configured to query the operation instruction mapping tables based on the intermediate representation layer data to generate a terminal compatible instruction conversion scheme; A conflict resolution unit is configured to perform dependency injection processing on parameter constraint paths with conflicts to generate unambiguous device interaction instructions.

Images