A method and system for assessing the comprehensive abilities of hospital students based on multi-source data

Abstract

This invention discloses a method and system for assessing the comprehensive abilities of hospital students based on multi-source data, relating to the field of artificial intelligence technology. The method includes: collecting a set of clinical operation data from target students and aligning the clinical operation data set according to time series to generate a time-synchronized clinical operation data set; inputting the ability transfer efficiency value and historical assessment scores into a plasticity calibration model, automatically compensating for deviations between simulated training scenarios and real clinical scenarios based on scenario complexity, and outputting a dynamically calibrated ability score; combining the dynamically calibrated ability score with an ability feature vector set to generate a multi-dimensional ability distribution map and improvement schemes for weak links. This invention, by inputting the ability transfer efficiency value and historical assessment scores into the plasticity calibration model, can adjust synaptic connection weights in real time according to scenario complexity during machine learning, achieving adaptive evolution of parameters and dynamic cognitive balance.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for assessing the comprehensive abilities of hospital students based on multi-source data. Background Technology

[0002] With the development of artificial intelligence technology, machine learning and neural networks have been widely applied in the fields of complex pattern recognition and cognitive modeling. Through multi-source data fusion and deep feature extraction, machine learning models can establish mapping relationships between different tasks in a nonlinear feature space, achieving self-learning and adaptive optimization from data representation to decision prediction. In the field of medical education, the use of methods such as deep learning, federated learning, and transfer learning to model clinical operational behaviors provides important support for the digital expression of clinical competence and also promotes the application of intelligent assessment models in educational evaluation systems.

[0003] However, existing machine learning-based medical education assessment techniques struggle to effectively model the ability transfer relationships between different training scenarios. Traditional deep learning models often employ static weight update mechanisms, neglecting the dynamic coupling between learning complexity and cognitive transfer. This leads to discrepancies in the cognitive mapping between simulated training and real clinical scenarios, resulting in insufficient generalization and adaptive capabilities. Therefore, there is an urgent need for a machine learning neural network with neuroplasticity regulation and scenario complexity perception capabilities to achieve cross-scenario ability transfer quantification and cognitive adaptive calibration. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for assessing the comprehensive abilities of hospital students based on multi-source data to solve the problem that dynamic quantification of cognitive transfer efficiency and adaptive adjustment of neuroplasticity cannot be achieved in medical education scenarios based on machine learning.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, this invention provides a method for assessing the comprehensive abilities of hospital students based on multi-source data. The method includes: collecting a set of clinical operation data of the target students and aligning the set of clinical operation data according to a time series to generate a time-series synchronized clinical operation data set; inputting the time-series synchronized clinical operation data set into a constructed machine learning model to extract diagnostic decision-making logic features, instrument operation trajectory features, and emergency response timeliness features, and fusing them to generate an ability feature vector set; uploading the ability feature vector set to a federated learning node and using a distributed aggregation algorithm to fuse the ability feature vector sets of multiple institutions to generate a standardized global ability space; calculating the ability transfer efficiency value of the target students between simulated training scenarios and real clinical scenarios based on the standardized global ability space; inputting the ability transfer efficiency value and historical assessment scores into a plasticity calibration model to automatically compensate for the deviation between simulated training scenarios and real clinical scenarios based on scenario complexity, and outputting a dynamically calibrated ability score; and combining the dynamically calibrated ability score with the ability feature vector set to generate a multi-dimensional ability distribution map and improvement schemes for weak links.

[0008] As a preferred embodiment of the hospital student comprehensive ability assessment method based on multi-source data described in this invention, the clinical operation data set includes electronic medical record operation sequences, simulation training videos, and equipment monitoring logs.

[0009] As a preferred embodiment of the hospital student comprehensive ability assessment method based on multi-source data described in this invention, the specific steps for generating the time-series synchronized clinical operation data set are as follows:

[0010] The simulated training video is input into the convolutional neural network to identify the surgical instrument conversion node, segment it into an operation stage sequence, and calculate the video motion intensity value of each operation stage in the operation stage sequence to generate a video motion intensity value sequence.

[0011] The sampled data from the equipment monitoring logs are resampled according to the time window of the operation stage sequence to generate a set of stage statistics.

[0012] Map each operation event in the electronic medical record operation sequence to its corresponding stage in the operation stage sequence to generate a mapped operation event sequence;

[0013] Using the sequence of operational phases as the time axis, a set of phase statistics, a sequence of mapped operational events, and a sequence of video action intensity values ​​are integrated to generate a time-synchronized clinical operational data set.

[0014] As a preferred embodiment of the hospital student comprehensive ability assessment method based on multi-source data described in this invention, the machine learning model includes a diagnostic decision feature extractor, an instrument operation feature extractor, and an emergency response feature constructor.

[0015] As a preferred embodiment of the hospital student comprehensive ability assessment method based on multi-source data described in this invention, the specific steps for generating the ability feature vector set are as follows:

[0016] The mapped sequence of operational events is input into the diagnostic decision feature extractor for temporal dependency analysis and weighted focusing of operational events, and the output is a diagnostic decision logical feature vector.

[0017] The video motion intensity value sequence is input into the instrument operation feature extractor to perform spatiotemporal feature extraction and motion pattern recognition, and outputs the instrument operation trajectory feature vector.

[0018] Input the statistical values ​​of the equipment monitoring phase into the emergency response feature constructor, and output the emergency response timeliness feature vector by calculating the alarm response delay rate, the handling continuity index and the physiological parameter stability ratio.

[0019] The diagnostic decision-making logic feature vector, the instrument operation trajectory feature vector, and the emergency response timeliness feature vector are input into the gating fusion mechanism, and the capability feature vector set is output.

[0020] As a preferred embodiment of the hospital student comprehensive ability assessment method based on multi-source data described in this invention, the specific steps for generating the standardized global ability space are as follows:

[0021] Discreteness analysis and information distribution density analysis were performed on the capability feature vector sets uploaded by each medical institution to calculate data quality weight values;

[0022] Semantic matching is performed between the historical clinical case type distribution of each medical institution and the historical global case distribution to calculate the clinical scenario similarity weight value;

[0023] By harmonizing the aggregated weight values ​​of data quality weight values ​​and clinical scenario similarity weight values, the capability feature vector sets of each institution are dynamically integrated to form an initial global capability space.

[0024] The initial global capability space is input into the perturbation optimization process, and a standardized global capability space is generated by minimizing the impact of external disturbances.

[0025] As a preferred embodiment of the hospital student comprehensive ability assessment method based on multi-source data described in this invention, the specific steps for calculating the ability transfer efficiency value of the target student between simulated training scenarios and real clinical scenarios are as follows:

[0026] The standardized global capability space is decomposed into a simulation training capability subspace and a real clinical capability subspace.

[0027] The instrument operation trajectory feature vector is projected onto the simulation training capability subspace to generate the simulation training capability vector, and the emergency response timeliness feature vector is mapped onto the real clinical capability subspace to generate the real clinical capability vector.

[0028] In the Riemannian manifold space, the shortest conversion path from the simulated training capability vector to the real clinical capability vector is constructed, and the distance value of the shortest conversion path is calculated by geodesic integration.

[0029] Extract case urgency index and concurrent task volume from real historical clinical scenarios to obtain scenario complexity adjustment factors;

[0030] An exponential decay calculation is performed on the shortest conversion path distance value and the scenario complexity adjustment factor to generate a capability migration efficiency value.

[0031] As a preferred embodiment of the hospital student comprehensive ability assessment method based on multi-source data described in this invention, the specific steps for outputting dynamically calibrated ability scores are as follows:

[0032] Input the capability transfer efficiency value and historical evaluation score into the plasticity calibration model, and generate scene deviation compensation coefficients through preset synaptic plasticity rules;

[0033] Hyperbolic tangent control is applied to the scene complexity adjustment factor to generate the scene complexity gain coefficient;

[0034] The scene deviation compensation coefficient and the scene complexity gain coefficient are applied to the historical evaluation scores to generate a dynamic calibration capability score.

[0035] As a preferred embodiment of the hospital student comprehensive ability assessment method based on multi-source data described in this invention, the specific steps for generating the multi-dimensional ability distribution map and the weak link improvement scheme are as follows:

[0036] A multidimensional joint probability density model of the capability feature vector set and the dynamic calibration capability score is performed by using a kernel density estimation algorithm to generate a multidimensional capability distribution map.

[0037] In the multidimensional capability distribution map, the coordinates of weak areas with distribution density below a preset density threshold are automatically located using the gradient descent method.

[0038] Input the coordinates of weak areas below a preset density threshold into the constructed medical knowledge graph and reinforcement learning mechanism, and output improvement plans for the weak links.

[0039] Secondly, the present invention provides a hospital student comprehensive ability assessment system based on multi-source data, including a data synchronization module, a feature fusion module, a federated aggregation module, a transfer quantification module, a neural calibration module, and a profile improvement module;

[0040] The data synchronization module is used to collect the clinical operation data set of the target students and align the clinical operation data set according to the time sequence to generate a time-synchronized clinical operation data set.

[0041] The feature fusion module is used to input the time-series synchronous clinical operation data set into the constructed machine learning model, extract diagnostic decision logic features, device operation trajectory features and emergency response timeliness features, and fuse them to generate a capability feature vector set;

[0042] The federated aggregation module is used to upload the capability feature vector set to the federated learning node, and use a distributed aggregation algorithm to fuse the capability feature vector sets of multiple institutions to generate a standardized global capability space.

[0043] The transfer quantification module is used to calculate the ability transfer efficiency value of the target student between the simulated training scenario and the real clinical scenario based on the standardized global ability space.

[0044] The neural calibration module is used to input the ability transfer efficiency value and historical assessment score into the plasticity calibration model, automatically compensate for the deviation between the simulated training scenario and the real clinical scenario according to the scenario complexity, and output a dynamic calibration ability score.

[0045] The profile improvement module is used to combine the dynamic calibration capability score with the capability feature vector set to generate a multi-dimensional capability distribution map and a weak link improvement plan.

[0046] The beneficial effects of this invention are as follows: By inputting the ability transfer efficiency value and historical evaluation scores into the plasticity calibration model, the synaptic connection weights can be adjusted in real time according to the scenario complexity during machine learning, achieving adaptive evolution of parameters and dynamic cognitive balance. The plasticity calibration model is constructed based on the principle of biological synaptic plasticity. By introducing a complexity adjustment factor and a time-dependent weight function, the learning process exhibits self-regulating characteristics in a high-dimensional cognitive space. Under multi-scenario input conditions, it can maintain representation consistency and transfer stability, thereby generating a high-precision dynamic calibration ability representation, providing a reliable computational foundation for multi-source cognitive modeling and intelligent assessment in the field of medical education. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of a method for assessing the comprehensive abilities of hospital students based on multi-source data.

[0049] Figure 2 This is a schematic diagram of a hospital student comprehensive ability assessment system based on multi-source data.

[0050] Figure 3 A flowchart for generating a time-synchronized clinical operation data set.

[0051] Figure 4 The flowchart for generating a set of capability feature vectors. Detailed Implementation

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0053] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0054] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0055] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a method for assessing the comprehensive abilities of hospital students based on multi-source data, including the following steps:

[0056] S1. Collect the clinical operation data set of the target students, and align the clinical operation data set according to the time series to generate a time-synchronized clinical operation data set.

[0057] The clinical operation data set includes electronic medical record operation sequences, simulation training videos, and equipment monitoring logs.

[0058] It should be noted that the electronic medical record operation sequence is obtained from the hospital's historical information database and records the target student's operation events on the electronic medical record, such as medical record viewing, diagnosis entry, and medical order issuance. Each operation event is timestamped.

[0059] The simulation training videos are obtained from the simulation training equipment and are recorded as video footer of the target student during simulated surgery or clinical operation training. The videos include time information.

[0060] The equipment monitoring logs are obtained from the sensors of medical equipment (such as monitors and surgical instruments), recording equipment status parameters, alarm events and operation records, and the sampled data is timestamped.

[0061] The simulated training video is input into the convolutional neural network to identify the surgical instrument conversion node, segment it into an operation stage sequence, and calculate the video motion intensity value of each operation stage in the operation stage sequence to generate a video motion intensity value sequence.

[0062] It should be noted that the pre-training process of the convolutional neural network is as follows: Initial training is performed using the historical general image dataset ImageNet. The initial training process adopts the ResNet architecture, optimizing the weight parameters of the convolutional neural network by minimizing image classification error, enabling the convolutional neural network to extract general image features. Then, a medical image dataset containing labeled surgical instrument images is used to fine-tune the pre-trained convolutional neural network. The fine-tuning process optimizes the parameters of the fully connected layers of the convolutional neural network for the surgical instrument recognition task, resulting in the trained convolutional neural network.

[0063] Specifically, the simulated training video is input into a convolutional neural network (e.g., a video classification model based on the ResNet architecture), the video frame content is analyzed frame by frame, the appearance and disappearance times of surgical instruments are identified, and the surgical instrument transition nodes are determined. The surgical instrument transition nodes are defined as the video frames in which the instruments are picked up or put down.

[0064] Based on the surgical instrument switching node, the simulated training video is divided into multiple continuous operation stages. Each operation stage corresponds to a specific instrument operation time period, forming an operation stage sequence. Each operation stage in the operation stage sequence has a start time and an end time.

[0065] Calculate the video motion intensity value for each operation stage in the operation stage sequence, arrange the video motion intensity values ​​of all operation stages in chronological order, and generate a video motion intensity value sequence.

[0066] The video motion intensity value for each operation phase in the operation phase sequence is calculated using the following expression:

[0067] ;

[0068] In the formula, Indicates the intensity value of video motion. This indicates the end video frame number of the current operation phase in the operation phase sequence. This indicates the starting video frame number of the current operation phase in the operation phase sequence. This indicates the total number of video frames included in the current operation phase. This represents the reciprocal of the total number of video frames included in the current operation phase. Indicates the video frame sequence number index. This represents the summation operation, starting from index 1. The video frame begins, Indicates the first Frame and the Average optical flow amplitude between +1 frames.

[0069] The sampled data from the equipment monitoring logs are resampled according to the time window of the operation stage sequence to generate a set of stage statistics.

[0070] Specifically, the sampling data in the equipment monitoring logs is time-series data (such as heart rate, blood pressure, and the number of equipment alarms, with inconsistent sampling frequencies).

[0071] Based on the time window (start time and end time) of each operation stage in the operation stage sequence, the sampled data of the device monitoring log is resampled. For each operation stage time window, all sampled data points within the time window are extracted.

[0072] Statistical values ​​(including average, maximum, minimum and standard deviation) are calculated for all sampled data points within the time window of each operation stage to form stage statistics. The stage statistics corresponding to each operation stage are then integrated to form a set of stage statistics.

[0073] Example: Suppose an operation phase called "vascular anastomosis" occurs between 10:00:00 AM and 10:07:30 AM (lasting 450 seconds). The equipment monitoring log (heart rate section) records 450 heart rate data points (one per second) within those 450 seconds. During processing, these 450 individual heart rate data points are no longer analyzed; instead, the average (e.g., 75 bpm), maximum (e.g., 88 bpm), minimum (e.g., 63 bpm), and standard deviation (e.g., 5.2) of these 450 heart rate data points are calculated. The phase statistics for the "vascular anastomosis" phase include an average heart rate of 75 bpm, a maximum heart rate of 88 bpm, a minimum heart rate of 63 bpm, and a heart rate standard deviation of 5.2.

[0074] Map each operation event in the electronic medical record operation sequence to its corresponding stage in the operation stage sequence to generate a mapped operation event sequence;

[0075] Specifically, each operation event in the electronic medical record operation sequence is timestamped; the timestamp of each operation event is compared with the start and end times of each operation stage in the operation stage sequence to determine the operation stage to which the operation event belongs (if the timestamp of operation event 1 falls within the time window of operation stage 2, then operation event 1 is mapped to operation stage 2).

[0076] Collect all operation events mapped to each operation stage to form a staged operation event sequence. Arrange all operation events in the staged operation event sequence in chronological order to generate a mapped operation event sequence.

[0077] Using the sequence of operational phases as the time axis, a set of phase statistics, a sequence of mapped operational events, and a sequence of video action intensity values ​​are integrated to generate a time-synchronized clinical operational data set.

[0078] Specifically, the operation phase sequence is used as a time axis. For each operation phase, the corresponding phase statistics are extracted from the phase statistics set, the corresponding operation events are extracted from the mapped operation event sequence, and the corresponding video action intensity values ​​are extracted from the video action intensity value sequence. The phase statistics, operation events, and video action intensity values ​​of each operation phase are combined into a data structure. All the data structures of the operation phases are arranged in chronological order to form a time-synchronized clinical operation data set.

[0079] S2. Input the time-synchronized clinical operation data set into the constructed machine learning model, extract the diagnostic decision logic features, device operation trajectory features and emergency response timeliness features, and fuse them to generate a capability feature vector set.

[0080] The machine learning model includes a diagnostic decision feature extractor, an instrument operation feature extractor, and an emergency response feature builder.

[0081] It should be noted that the diagnostic decision feature extractor is built on a recurrent neural network structure and is used to process serialized operation events and capture long-term dependencies.

[0082] The instrument operation feature extractor is built on a one-dimensional convolutional layer structure and is used to process video motion intensity value sequences and extract local spatiotemporal patterns.

[0083] The emergency response feature builder is constructed based on a fusion structure of statistical learning and attention mechanisms, and is used to calculate performance indicators.

[0084] The mapped sequence of operational events is input into the diagnostic decision feature extractor for temporal dependency analysis and weighted focusing of operational events, and the output is a diagnostic decision logical feature vector.

[0085] It should be noted that the pre-training process of the recurrent neural network is as follows: A historical medical operation event sequence dataset is called. This dataset contains operation event sequences extracted from historical electronic medical record operation sequences. Each operation event sequence consists of multiple operation events arranged in chronological order. During training, the recurrent neural network uses the previous operation event as the prediction target for the current time step. The parameters of the recurrent neural network are adjusted to minimize the difference between the predicted operation event and the next actual operation event. After training, the trained recurrent neural network is obtained.

[0086] Specifically, each operation event in the mapped operation event sequence obtains its corresponding operation event embedding vector by querying a predefined operation event embedding table. A recurrent neural network processes each operation event embedding vector sequentially and updates the hidden state, capturing the temporal dependencies between operation events. The hidden states at all time steps are processed by an attention mechanism. This mechanism uses a single-layer feedforward network to perform linear transformations and non-linear activations on each hidden state to obtain an initial score. The initial score is then normalized using a softmax function to generate weight coefficients for each hidden state. These weight coefficients are used to weight and combine the hidden states, and the resulting weighted operation event embedding vector serves as the feature vector for the diagnostic decision logic.

[0087] It should be noted that during the process of recurrent neural network processing operation event embedding vectors, the recurrent neural network maintains a hidden state internally. The hidden state is initialized to a zero vector at the beginning, and the recurrent neural network generates a new hidden state based on the current input operation event embedding vector and the hidden state at the previous time step.

[0088] The video motion intensity value sequence is input into the instrument operation feature extractor to perform spatiotemporal feature extraction and motion pattern recognition, and outputs the instrument operation trajectory feature vector.

[0089] Specifically, the one-dimensional convolutional layer in the instrument operation feature extractor uses multiple convolutional kernels of different widths to slide on the video motion intensity value sequence. Each convolutional kernel extracts local motion patterns at a specific time scale in the video motion intensity value sequence. The feature maps generated by the one-dimensional convolutional layer are passed through pooling layers. The pooling layers retain effective features and reduce feature dimensions. Multiple convolutional layers and pooling layers gradually abstract motion pattern features at different time scales. The global pooling layer converts all motion pattern feature maps into fixed-dimensional instrument operation trajectory feature vectors.

[0090] It should be noted that a specific time scale refers to the range of time segments of a video motion intensity value sequence covered by convolution kernels of different widths. For example, narrow convolution kernels cover short segments to capture subtle motion changes, while wide convolution kernels cover long segments to capture overall motion trends.

[0091] Effective features refer to local patterns with high response intensity in the feature map extracted through convolution operations that can effectively distinguish different instrument operation modes, such as peak features of high-frequency motion segments or specific periodic motion pattern features.

[0092] Input the statistical values ​​of the equipment monitoring phase into the emergency response feature constructor, and output the emergency response timeliness feature vector by calculating the alarm response delay rate, the handling continuity index and the physiological parameter stability ratio.

[0093] The expression for calculating the alarm response delay rate is:

[0094] ;

[0095] In the formula, Indicates alarm response delay rate. Indicates the alarm event index, This indicates the total number of alarm events that occurred during the current operation phase. Indicates the first The absolute time of the first related operation event recorded after the alarm event occurs. Indicates the first The absolute time of the alarm incident This indicates the total duration of the current operation phase.

[0096] The expression for calculating the treatment continuity index is:

[0097] ;

[0098] In the formula, Indicating the consistency index of handling, This indicates the total number of operational events related to alarm handling that occurred during the current operational phase.

[0099] The expression for calculating the physiological parameter stability ratio is:

[0100] ;

[0101] In the formula, Indicates the stable ratio of physiological parameters. This represents the standard deviation of the sampled data for a certain physiological parameter (such as heart rate) during the current operation phase. This represents the average value of the sampled data for the same physiological parameter within the current operational phase.

[0102] The alarm response delay rate, the handling continuity index, and the physiological parameter stability ratio are combined in sequence into a three-dimensional vector, which constitutes the emergency response timeliness feature vector.

[0103] The diagnostic decision-making logic feature vector, the instrument operation trajectory feature vector, and the emergency response timeliness feature vector are input into the gating fusion mechanism, and the capability feature vector set is output.

[0104] Specifically, the diagnostic decision logic feature vector, instrument operation trajectory feature vector, and emergency response timeliness feature vector are projected onto a feature space of the same dimension through a fully connected layer. The projected diagnostic decision logic feature vector and the projected instrument operation trajectory feature vector are concatenated. A gated value vector is generated through a fully connected layer with a Sigmoid activation function. The gated value vector is used to assign weights to the projected diagnostic decision logic feature vector. At the same time, the complement of the gated value vector is used to assign weights to the projected instrument operation trajectory feature vector. The weighted diagnostic decision logic feature vector and the weighted instrument operation trajectory feature vector are added together to form a fused feature vector. The fused feature vector is concatenated with the projected emergency response timeliness feature vector to form a capability feature vector set.

[0105] A superior approach is to construct a machine learning model by inputting a set of time-synchronized clinical operation data. This model achieves collaborative feature learning and semantic-level fusion of multimodal clinical data. Compared with traditional single-channel deep learning methods (such as using only convolutional neural networks to extract static image features from video sequences), the machine learning model introduces a multi-branch neural feature extraction and gating fusion mechanism in its structure. The diagnostic decision feature extractor models the temporal dependence of operation events based on recurrent neural networks, the instrument operation feature extractor captures spatiotemporal trajectory patterns based on convolutional neural networks, and the emergency response feature constructor constructs timeliness weights based on an attention mechanism. This enables dynamic interaction and high-dimensional correlation modeling between different information modalities within a unified representation space.

[0106] S3. Upload the capability feature vector set to the federated learning node, and use a distributed aggregation algorithm to fuse the capability feature vector sets of multiple institutions to generate a standardized global capability space.

[0107] Discreteness analysis and information distribution density analysis were performed on the capability feature vector sets uploaded by each medical institution to calculate the data quality weight value, expressed as follows:

[0108] ;

[0109] In the formula, Indicates the first Data quality weighting values ​​for individual medical institutions Indicates a medical institution index. Indicates the first The coefficient of variation of the capability feature vector set uploaded by each medical institution. Indicates the first Information entropy of the capability feature vector set uploaded by each medical institution.

[0110] Semantic matching is performed between the historical clinical case type distribution of each medical institution and the historical global case distribution to calculate the clinical scenario similarity weight value, expressed as:

[0111] ;

[0112] In the formula, Indicates the first The weighting of clinical scenario similarity among medical institutions. Indicates the first Historical clinical case type distribution vector of each medical institution Represents the historical global case type distribution vector. Indicates the first The magnitude of the distribution vector of historical clinical case types of a medical institution. This represents the magnitude of the historical global case type distribution vector.

[0113] By harmonizing the aggregated weight values ​​of data quality weight values ​​and clinical scenario similarity weight values, the capability feature vector sets of each institution are dynamically integrated to form an initial global capability space.

[0114] Specifically, an aggregation weight value is set for each medical institution, and the capability feature vector set of the corresponding medical institution is weighted. The weighted capability feature vector sets of all medical institutions are summed according to the corresponding dimensions to obtain the aggregated feature vector set. The ratio of the aggregated feature vector set to the sum of the aggregation weight values ​​of all medical institutions is used as the initial global capability space.

[0115] The aggregated weight value, calculated by harmonic calculation of data quality weights and clinical scenario similarity weights, is expressed as follows:

[0116] ;

[0117] In the formula, Indicates the first The aggregate weight value of each medical institution. Represents the harmonic coefficient. Indicates the index for the summation in a loop. This represents the summation of data quality weights for all medical institutions. This represents the complement of the harmonic coefficient. This represents the summation of clinical scenario similarity weights across all medical institutions.

[0118] It should be noted that the harmonic coefficient is derived from prior experience of the federated learning aggregation strategy. The example value is 0.7, which is based on the verification of grid search on historical medical data. The results show that a value of 0.7 can optimally balance the weight distribution relationship between data quality and clinical scenario representativeness.

[0119] The initial global capability space is input into the perturbation optimization process, and a standardized global capability space is generated by minimizing the impact of external disturbances.

[0120] Specifically, the initial global capability space is subjected to random noise conforming to the Laplace distribution through the Laplace mechanism to achieve differential privacy protection, generating a global capability space with added noise. Principal component analysis is then used to reduce the dimensionality of the global capability space with added noise, retaining the main variance components to eliminate the influence of noise and extract effective feature patterns, generating a dimensionality-reduced feature space, which serves as the standardized global capability space.

[0121] It should be noted that the effective feature pattern refers to the feature component that has the largest variance contribution and is retained after dimensionality reduction by principal component analysis, which can best represent the differences in comprehensive abilities among different students.

[0122] S4. Based on the standardized global capability space, calculate the capability transfer efficiency value of the target student between the simulated training scenario and the real clinical scenario.

[0123] The standardized global capability space is decomposed into a simulation training capability subspace and a real clinical capability subspace.

[0124] Specifically, the principal component contribution of all feature dimensions in the standardized global capability space is ranked, and the top principal component dimensions with the highest correlation to instrument operation are selected to form the simulation training capability subspace, while the top principal component dimensions with the highest correlation to emergency response are selected to form the real clinical capability subspace.

[0125] The instrument operation trajectory feature vector is projected onto the simulation training capability subspace to generate the simulation training capability vector, and the emergency response timeliness feature vector is mapped onto the real clinical capability subspace to generate the real clinical capability vector.

[0126] Specifically, the feature vector of the instrument operation trajectory is projected onto the simulation training capability subspace to generate the simulation training capability vector, expressed as:

[0127] ;

[0128] In the formula, Represents the simulated training capability vector. The projection matrix represents the simulation training capability subspace. This represents the feature vector of the instrument's operating trajectory.

[0129] Mapping the emergency response timeliness feature vector to the real clinical capability subspace generates the real clinical capability vector, expressed as:

[0130] ;

[0131] In the formula, Represents the true clinical capability vector. The projection matrix representing the true clinical capability subspace. This represents the feature vector of emergency response timeliness.

[0132] In the Riemannian manifold space, the shortest conversion path from the simulated training capability vector to the real clinical capability vector is constructed, and the distance value of the shortest conversion path is calculated by geodesic integration.

[0133] Specifically, in the Riemannian manifold space, the simulated training capability vector and the real clinical capability vector are respectively regarded as two points on the Riemannian manifold, and the shortest transformation path between the two points is naturally formed by the geodesics inherent in the Riemannian manifold.

[0134] The shortest transformation path distance is calculated using geodesic integration, expressed as:

[0135] ;

[0136] In the formula, This represents the shortest conversion path distance value. This represents the inverse hyperbolic cosine function.

[0137] Extract case urgency index and concurrent task volume from real historical clinical scenarios to obtain scenario complexity adjustment factors;

[0138] Specifically, the case criticality index is extracted from the critical value report database of historical electronic medical records. By statistically classifying the abnormal vital signs records under different disease categories, the average criticality level of each disease is determined. The concurrent task volume is extracted from the operation records of historical equipment monitoring logs. The number of independent operation events occurring in parallel per unit time in each clinical scenario is counted. By linearly combining the case criticality index and the concurrent task volume according to preset weights, the scenario complexity adjustment factor is obtained.

[0139] The expression for calculating the scene complexity adjustment factor is:

[0140] ;

[0141] In the formula, This represents the scene complexity adjustment factor. The weighting coefficients represent the severity index of a case. Indicates the severity index of the case. This represents the mean of the case severity index. The standard deviation of the case severity index The weighting coefficients represent the number of concurrent tasks. Indicates the number of concurrent tasks. This represents the average number of concurrent tasks. The standard deviation of the number of concurrent tasks.

[0142] It should be noted that the preset weights are set based on the feature importance analysis results in historical medical data. The weight coefficient of the case criticality index is set to an empirical value of 0.6 based on the feature importance analysis results in historical medical data. This value is determined by grid search cross-validation, which shows that this weight can best represent the dominant influence of the patient's critical state on the complexity of the scenario. The weight coefficient of the concurrent task volume is set to an empirical value of 0.4 based on the feature importance analysis results in historical medical data. This value is determined by grid search cross-validation, which shows that this weight can reasonably represent the auxiliary influence of workload on the complexity of the scenario.

[0143] An exponential decay calculation is performed on the shortest conversion path distance value and the scenario complexity adjustment factor to generate a capability migration efficiency value.

[0144] ;

[0145] In the formula, This represents the ability transfer efficiency value. This represents the natural exponential function. This represents the attenuation coefficient.

[0146] It should be noted that the attenuation coefficient is optimized based on historical capability migration efficiency data using a grid search method. The example value is 0.3, which is determined by minimizing the prediction error of capability migration efficiency values ​​on historical datasets through cross-validation.

[0147] A superior approach involves introducing machine learning-based nonlinear mapping and manifold geometric inference mechanisms into a standardized global capability space to calculate the capability transfer efficiency of target students between simulated training scenarios and real clinical scenarios. Compared to traditional static scoring or linear regression assessment methods (such as capability comparison analysis based on multiple linear regression models), machine learning models can automatically learn high-dimensional nonlinear relationships between capability features and utilize manifold embeddings and geodesic distance calculations to characterize the transformation paths of capability representations across different scenarios. In this way, machine learning models can not only quantify learners' cognitive transfer efficiency in different scenarios but also dynamically reflect the impact of training difficulty and scenario complexity on the learning state of the machine learning model, thereby achieving adaptive mapping and accurate assessment of cross-domain capabilities and improving the intelligence and scientific rigor of capability transfer modeling.

[0148] S5. Input the ability transfer efficiency value and historical assessment score into the plasticity calibration model, automatically compensate for the deviation between the simulated training scenario and the real clinical scenario according to the scenario complexity, and output the dynamic calibration ability score.

[0149] Input the capability transfer efficiency value and historical evaluation score into the plasticity calibration model, and generate scene deviation compensation coefficients through preset synaptic plasticity rules;

[0150] It should be noted that the plasticity calibration model is a mathematical modeling method that simulates the brain's adaptive learning process. By establishing a dynamic mapping relationship between ability transfer efficiency value, historical assessment score and scenario complexity, it achieves quantitative compensation for the deviation between simulated training scenarios and real clinical scenarios.

[0151] The pre-defined synaptic plasticity rule specifically refers to a rule constructed based on the Hebbian plasticity principle that reflects the positive correlation between the simultaneous activation strength and connection strength of neurons.

[0152] The expression for the scene deviation compensation coefficient is:

[0153] ;

[0154] In the formula, This represents the scene deviation compensation coefficient. This represents the learning rate parameter. This indicates the historical assessment score.

[0155] Hyperbolic tangent control is applied to the scene complexity adjustment factor to generate the scene complexity gain coefficient;

[0156] Specifically, the hyperbolic tangent function is used to perform a nonlinear transformation on the scene complexity adjustment factor, the expression of which is:

[0157] ;

[0158] In the formula, This represents the scene complexity gain coefficient. This represents the hyperbolic tangent function.

[0159] The scene deviation compensation coefficient and the scene complexity gain coefficient are applied to the historical evaluation scores to generate a dynamic calibration capability score.

[0160] Specifically, a linear weighted average is used to generate the dynamic calibration capability score, expressed as follows:

[0161] ;

[0162] In the formula, This indicates the score for dynamic calibration capability.

[0163] S6. Combine the dynamic calibration capability score with the capability feature vector set to generate a multi-dimensional capability distribution map and improvement schemes for weak links.

[0164] A multidimensional joint probability density model of the capability feature vector set and the dynamic calibration capability score is performed by using a kernel density estimation algorithm to generate a multidimensional capability distribution map.

[0165] Specifically, the various dimensions of the capability feature vector set and the dynamic calibration capability score are combined to form a single data point in the multidimensional feature space. In the multidimensional feature space, a Gaussian kernel function is used to smoothly superimpose the distribution contribution of each data point. The probability density distribution of the entire multidimensional feature space is obtained by traversing the grid points. Based on the probability density distribution, contour maps are drawn to form a multidimensional capability distribution map.

[0166] In the multidimensional capability distribution map, the coordinates of weak areas with distribution density below a preset density threshold are automatically located using the gradient descent method.

[0167] It should be noted that the preset density threshold is an empirical value set based on the ability distribution density of historical high-achieving students in the multidimensional ability distribution map. The value range is 0.1-0.3. 0.1 ensures that weak areas below the excellent level are identified and avoids misjudging normal fluctuations as weak points. 0.3 covers the vast majority of weak areas that need improvement, while avoiding missing potential ability defects due to an excessively high density threshold.

[0168] Specifically, a preset density threshold is set in the multidimensional capability distribution map, and gradient descent search is performed starting from a random initial position. The search iterates along the probability density descent direction and stops when a region with a distribution density lower than the preset density threshold is found. The spatial coordinates of all such stopping points are recorded as the coordinates of the weak region.

[0169] Input the coordinates of weak areas below a preset density threshold into the constructed medical knowledge graph and reinforcement learning mechanism, and output improvement plans for the weak links.

[0170] It should be noted that the construction process of the medical knowledge graph is as follows: Entities and relationships are extracted from the historical standardized medical terminology database and historical clinical operation guidelines. The historical standardized medical terminology database contains the ICD disease classification and CPT operation coding system, and the historical clinical operation guidelines contain standardized operation process documents of various medical institutions. Natural language processing is used to automatically identify medical skill entities and their relationships from the historical standardized medical terminology database and historical clinical operation guidelines to form a preliminary graph structure. The preliminary graph is verified and improved through a knowledge fusion process to ensure the accuracy and consistency of knowledge, and then stored as a medical knowledge graph containing medical skill nodes, relationship edges, and their attribute values.

[0171] Specifically, the coordinates of weak areas are mapped to corresponding skill nodes in a medical knowledge graph, which contains prerequisite relationships and training dependencies between medical operation skills. By querying the medical knowledge graph, the specific type of ability deficiency and associated skill nodes corresponding to the coordinates of weak areas are determined, generating a deficiency diagnosis result. A reinforcement learning mechanism performs targeted training on each weak skill node in the deficiency diagnosis result, generating a targeted training strategy. The deficiency diagnosis result and the targeted training strategy are combined to form a personalized improvement plan for weak links.

[0172] This embodiment also provides a hospital student comprehensive ability assessment system based on multi-source data, including: a data synchronization module, a feature fusion module, a federated aggregation module, a transfer quantification module, a neural calibration module, and a profile improvement module; the data synchronization module is used to collect the clinical operation data set of the target students and align the clinical operation data set according to the time series to generate a time-series synchronized clinical operation data set; the feature fusion module is used to input the time-series synchronized clinical operation data set into the constructed machine learning model, extract diagnostic decision logic features, instrument operation trajectory features, and emergency response timeliness features, and fuse them to generate an ability feature vector set; the federated aggregation module is used to aggregate the ability features... The feature vector set is uploaded to the federated learning node, where a distributed aggregation algorithm is used to fuse the capability feature vector sets of multiple institutions to generate a standardized global capability space. The transfer quantization module is used to calculate the capability transfer efficiency value between the target student and the real clinical scenario based on the standardized global capability space. The neural calibration module is used to input the capability transfer efficiency value and historical assessment scores into the plasticity calibration model, automatically compensate for the deviation between the simulated training scenario and the real clinical scenario according to the scenario complexity, and output a dynamic calibration capability score. The profile improvement module is used to combine the dynamic calibration capability score with the capability feature vector set to generate a multi-dimensional capability distribution map and a weakness improvement plan.

[0173] This embodiment also provides a computer device applicable to the hospital student comprehensive ability assessment method based on multi-source data, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the hospital student comprehensive ability assessment method based on multi-source data as proposed in the above embodiment.

[0174] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0175] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the method for assessing the comprehensive abilities of hospital students based on multi-source data as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0176] In summary, this invention, by inputting ability transfer efficiency values ​​and historical evaluation scores into a plasticity calibration model, enables real-time adjustment of synaptic connection weights based on scenario complexity during machine learning, achieving adaptive parameter evolution and cognitive dynamic balance. The plasticity calibration model is constructed based on the principle of biological synaptic plasticity, and by introducing complexity adjustment factors and time-dependent weight functions, it allows the learning process to exhibit self-regulating characteristics in a high-dimensional cognitive space. Under multi-scenario input conditions, it maintains representation consistency and transfer stability, thereby generating a high-precision dynamic calibration ability representation, providing a reliable computational foundation for multi-source cognitive modeling and intelligent assessment in the field of medical education.

[0177] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A hospital student comprehensive ability assessment method based on multi-source data, characterized by: The application relates to a method for dynamically evaluating the ability of a target student in a medical field, and belongs to the technical field of medical education. The method comprises the following steps: a clinical operation data set of the target student is collected, and the clinical operation data set is aligned in a time sequence to generate a time sequence synchronized clinical operation data set; the time sequence synchronized clinical operation data set is input into a constructed machine learning model to extract diagnosis decision logic features, instrument operation trajectory features and emergency response time efficiency features and to fuse the features to generate a capability feature vector set; the capability feature vector set is uploaded to a federal learning node, and a distributed aggregation algorithm is used to fuse the capability feature vector sets of multiple institutions to generate a standardized global capability space, and the specific steps are as follows, the capability feature vector sets uploaded by the medical institutions are analyzed in terms of discrete degree and information distribution density to calculate data quality weight values; the historical clinical case type distribution of each medical institution and the historical global case distribution are subjected to semantic matching to calculate clinical scene similarity weight values; the aggregation weight values of the data quality weight values and the clinical scene similarity weight values are calculated by harmonic calculation, and the capability feature vector sets of the institutions are dynamically fused to form an initial global capability space; the initial global capability space is input into a perturbation optimization process to generate a standardized global capability space by minimizing external interference influence; according to the standardized global capability space, the capability transfer efficiency value of the target student between a simulation training scene and a real clinical scene is calculated, and the specific steps are as follows, the standardized global capability space is decomposed into a simulation training capability subspace and a real clinical capability subspace; the instrument operation trajectory feature vector is projected into the simulation training capability subspace to generate a simulation training capability vector, and the emergency response time efficiency feature vector is mapped into the real clinical capability subspace to generate a real clinical capability vector; a shortest conversion path of the simulation training capability vector to the real clinical capability vector is constructed in a Riemannian manifold space, and the shortest conversion path distance value is calculated by geodesic integral calculation; a case criticality index and a concurrent task amount are extracted from a historical real clinical scene to obtain a scene complexity adjustment factor; exponential decay calculation is performed on the shortest conversion path distance value and the scene complexity adjustment factor to generate the capability transfer efficiency value; the capability transfer efficiency value and a historical evaluation score are input into a plasticity calibration model, the deviation between the simulation training scene and the real clinical scene is automatically compensated according to the scene complexity, and a dynamic calibration capability score is output; 2. The hospital student comprehensive ability examination method based on multi-source data according to claim 1, characterized in that: the dynamic calibration capability score and the capability feature vector set are combined to generate a multi-dimensional capability distribution graph and a weak link improvement scheme.

3. The hospital student comprehensive ability examination method based on multi-source data according to claim 1, characterized in that: The clinical operation data set comprises an electronic medical record operation sequence, a simulation training video and a device monitoring log. The time sequence synchronized clinical operation data set is generated by the following specific steps, the simulation training video is input into a convolutional neural network to identify a surgical instrument conversion node, is segmented into an operation stage sequence, and the video action intensity value of each operation stage in the operation stage sequence is calculated to generate a video action intensity value sequence; the sampling data of the device monitoring log is resampled according to the time window of the operation stage sequence to generate a stage statistical value set; each operation event of the electronic medical record operation sequence is mapped to a belonging stage in the operation stage sequence to generate a mapped operation event sequence; The time sequence synchronization clinical operation data set is generated by integrating the stage statistical value set, the mapped operation event sequence and the video action intensity value sequence along the operation stage sequence as a time axis.

4. The hospital student comprehensive ability examination method based on multi-source data according to claim 1, characterized in that: The machine learning model comprises a diagnosis decision feature extractor, an instrument operation feature extractor and an emergency response feature constructor.

5. The method for hospital student comprehensive ability examination based on multi-source data according to claim 1, characterized in that: The generated ability feature vector set has the following specific steps, The mapped operation event sequence is input into the diagnosis decision feature extractor to perform time sequence dependency analysis and operation event weighted focusing, and a diagnosis decision logic feature vector is output; The video action intensity value sequence is input into the instrument operation feature extractor to perform space-time feature extraction and motion pattern recognition, and an instrument operation trajectory feature vector is output; The device monitoring stage statistical value is input into the emergency response feature constructor to calculate an alarm response delay rate, a treatment continuity index and a physiological parameter stability ratio, and an emergency response timeliness feature vector is output; The diagnosis decision logic feature vector, the instrument operation trajectory feature vector and the emergency response timeliness feature vector are input into a gated fusion mechanism, and an ability feature vector set is output.

6. The hospital student comprehensive ability examination method based on multi-source data according to claim 1, characterized in that: The generated dynamic calibration ability score has the following specific steps, The ability transfer efficiency value and the historical evaluation score are input into a plasticity calibration model to generate a scene bias compensation coefficient through a preset synaptic plasticity rule; The hyperbolic tangent control is performed on the scene complexity adjustment factor to generate a scene complexity gain coefficient; The scene bias compensation coefficient and the scene complexity gain coefficient are applied to the historical evaluation score to generate a dynamic calibration ability score.

7. The hospital student comprehensive ability examination method based on multi-source data according to claim 1, characterized in that: The generated multi-dimensional ability distribution graph and the weak link improvement scheme have the following specific steps, The multi-dimensional joint probability density of the ability feature vector set and the dynamic calibration ability score is modeled through a kernel density estimation algorithm to generate a multi-dimensional ability distribution graph; In the multi-dimensional ability distribution graph, the weak area coordinates with a distribution density lower than a preset density threshold are automatically located through a gradient descent method; The weak area coordinates lower than the preset density threshold are input into the constructed medical knowledge graph and the reinforcement learning mechanism to output a weak link improvement scheme.

8. A hospital student comprehensive ability assessment system based on multi-source data, based on the hospital student comprehensive ability assessment method based on multi-source data according to any one of claims 1-7, characterized in that: The system comprises a data synchronization module, a feature fusion module, a federal aggregation module, a transfer quantization module, a neural calibration module and an image improvement module. The data synchronization module is configured to collect a clinical operation data set of a target student, align the clinical operation data set according to a time sequence, and generate a time sequence synchronization clinical operation data set. The feature fusion module is configured to input the time sequence synchronization clinical operation data set into the constructed machine learning model, extract and fuse diagnosis decision logic features, instrument operation trajectory features and emergency response timeliness features to generate an ability feature vector set. The federal aggregation module is configured to upload the ability feature vector set to a federal learning node, fuse the ability feature vector sets of multiple institutions by using a distributed aggregation algorithm, and generate a standardized global ability space. The data quality weight value is calculated by performing discrete degree analysis and information distribution density analysis on the ability feature vector set uploaded by each medical institution. The clinical scene similarity weight value is calculated by performing semantic matching on the historical clinical case type distribution of each medical institution and the historical global case distribution. An aggregation weight value of the data quality weight value and the clinical scene similarity weight value is calculated, capability characteristic vector sets of each institution are dynamically fused, and an initial global capability space is formed; The initial global capability space is input into a perturbation optimization process, and a standardized global capability space is generated by minimizing external interference effects; The migration quantification module is configured to calculate a capability migration efficiency value of the target student between the simulation training scene and the real clinical scene according to the standardized global capability space, and the specific steps are as follows, The standardized global capability space is decomposed into a simulation training capability subspace and a real clinical capability subspace; An instrument operation trajectory feature vector is projected into the simulation training capability subspace to generate a simulation training capability vector, and an emergency response time efficiency feature vector is mapped into the real clinical capability subspace to generate a real clinical capability vector; A shortest transformation path from the simulation training capability vector to the real clinical capability vector is constructed in a Riemannian manifold space, and a shortest transformation path distance value is calculated by geodesic integral calculation; A case criticality index and a concurrent task quantity are extracted from historical real clinical scenes to obtain a scene complexity adjustment factor; Exponential decay calculation is performed on the shortest transformation path distance value and the scene complexity adjustment factor to generate the capability migration efficiency value; The neural calibration module is configured to input the capability migration efficiency value and the historical evaluation score into a plasticity calibration model, automatically compensate for deviations between the simulation training scene and the real clinical scene according to the scene complexity, and output a dynamic calibration capability score; The image improvement module is configured to combine the dynamic calibration capability score with the capability characteristic vector set to generate a multi-dimensional capability distribution map and a weak link improvement scheme.

Images