Nursing competency assessment system based on multi-channel deep learning model

Abstract

The application discloses a kind of nursing competency assessment systems based on multi-channel deep learning model, system includes scene driving and script management module, multi-modal data acquisition module, data convergence and time synchronization module, main control processing and model inference module and result output and storage module.System with first-aid task script drive key emergency event, in the process that nursing staff executes task, synchronously collect nursing operation behavior sequence, physiological response sequence and nursing action quality evaluation sequence, construct unified time axis after being input respectively stacked LSTM or Bi-LSTM extract feature and fuse score;Simultaneously based on event occurrence time point, determine recognition time point, action start time point and physiological recovery time point, form emergency identification-reaction-stress resistance series index link, and quality statistics directly participate in total score calculation, output scene level and comprehensive competency score and playback anchor point.The system realizes the objective, interpretable, traceable evaluation of nursing competency in comprehensive first-aid scene.

Description

Technical Field

[0001] This invention belongs to the field of nursing competency assessment, specifically relating to a nursing competency assessment system based on a multi-channel deep learning model. Background Technology

[0002] Emergency nursing work is characterized by its suddenness, short time window, high workload, long response chain, and high requirements for teamwork. Nurses in emergency situations not only need to correctly perform key procedures (such as chest compressions, ventilation coordination, hemostasis and immobilization, establishing intravenous access, and monitoring connections and parameter observation), but also need to quickly identify, promptly call for help, and rapidly initiate treatment when a patient's vital signs change abruptly, while maintaining operational quality and process consistency under high-pressure conditions. Therefore, scientifically evaluating nurses' competence in emergency scenarios is a crucial foundation for nursing training, job entry requirements, and quality management.

[0003] Current methods for assessing nursing competence largely rely on manual observation, checklist scoring, or OSCE (Organizational Symptom Checklist-Completeness Assessment), which suffer from problems such as strong subjectivity, inconsistent scoring scales, difficulty in quantifying the sequential behavior throughout the process, and difficulty in objectively characterizing "recognition-response-recovery" abilities during critical emergencies. On the other hand, some equipment- or sensor-based assessment schemes often focus only on a single skill point or a single data source (e.g., assessing CPR quality solely based on compression depth / frequency, or reflecting stress levels solely based on heart rate changes), failing to cover the real needs of cross-scenario, multi-event, multi-action, and multi-dimensional quality constraints in clinical emergency care. They also struggle to establish a correspondence between physiological stress and behavioral management, resulting in insufficient interpretability and reproducibility of assessment results, as well as limited support for training and error correction.

[0004] Therefore, there is an urgent need for a nursing competence assessment system that can achieve objective, interpretable, and traceable assessment based on multi-source time-series data in comprehensive nursing and emergency scenarios, in order to solve the problems of incomplete evaluation dimensions, asynchronous data, lack of event-driven link characterization, and difficulty in standardized application in existing technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a nursing competency assessment system based on a multi-channel deep learning model. This invention can realize an objective, interpretable, and traceable nursing competency assessment system based on multi-source time-series data in a comprehensive nursing emergency scenario, so as to solve the problems of incomplete evaluation dimensions, data asynchrony, lack of event-driven link characterization, and difficulty in standardization application in the existing technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a nursing competency assessment system based on a multi-channel deep learning model, comprising: The scenario-driven and script management module is used to load nursing emergency task scripts and drive task execution. The nursing emergency task scripts include at least standard process constraints and a set of key acute events. The multimodal data acquisition module is used to collect nursing operation behavior data, physiological response data, and nursing action quality data during the nursing staff's performance of emergency care tasks; The data aggregation and time synchronization module is communicatively connected to the multimodal data acquisition module and is used to uniformly timestamp and transmit the nursing operation behavior data, physiological response data, and nursing action quality data; The main control processing and model inference module is communicatively connected to the data aggregation and time synchronization module, and is also controlled by the scene driving and script management module. It is used to establish a unified timeline and construct at least three channels of time-series input, which include nursing operation behavior sequence, physiological response sequence, and nursing action quality assessment sequence. The module extracts time-series features from each channel of time-series input and performs feature fusion to output the scene-level competency score and comprehensive competency score of the nursing staff. The results output and storage module is connected to the main control processing and model inference module, and is used to output and store the competency score and the corresponding key time points.

[0007] The main control processing and model reasoning module is configured to: determine the emergency identification time trec in the physiological reaction sequence based on the nursing emergency task script recording the critical sudden event occurrence time tevent, determine the critical treatment action initiation time tact in the nursing operation behavior sequence, and determine the physiological indicator recovery time trecov in the physiological reaction sequence, thereby forming a series calculation link of emergency identification, response capability and stress resistance capability, and use the output of the series link and the nursing action quality statistics together for competency score calculation.

[0008] Furthermore, the scenario-driven and script management module is configured to: trigger the critical acute events during task execution and change the display or alarm status of monitoring parameters, so that nursing staff can perform treatment actions in the corresponding clinical emergency situation, and at the same time record the event type and the teven of each critical acute event.

[0009] Furthermore, the time step features of the nursing operation behavior sequence include at least the action category code ak, the action duration feature dk, the process step identifier sk, and the process deviation feature rk, wherein the process deviation feature is used to characterize missed steps, out-of-order steps, or timeouts.

[0010] Furthermore, the time step features of the physiological response sequence include at least the baseline normalized heart rate (HR). ~k, baseline-normalized skin electrical activity (EDA) ~ k and its rates of change ΔHRk and ΔEDAk, where HRbase and EDAbase are the resting baselines before the task begins.

[0011] Furthermore, the main control processing and model inference module includes a multi-channel deep learning model, which includes at least three sets of temporal encoders corresponding to the nursing operation behavior sequence, physiological response sequence and nursing action quality assessment sequence, respectively. Each set of temporal encoders is a stacked long short-term memory network LSTM or a bidirectional long short-term memory network Bi-LSTM, used to output the temporal feature vectors v1, v2 and v3 of each channel.

[0012] Furthermore, the emergency identification time trec is determined by the emergency identification probability reaching a threshold, and the emergency identification probability satisfies: Where hk(2) is the encoder hidden state of the physiological response sequence, hk(1) is the encoder hidden state of the nursing operation behavior sequence, [ ⋅∥⋅ ] is vector concatenation; trec is the time corresponding to the smallest time step that satisfies pER(k)≥θERp.

[0013] Furthermore, the responsiveness is determined by the time difference between the critical intervention initiation time tac and the emergency identification time trec, where tac is the moment when an action within the preset critical intervention set AAA first appears after trec in the nursing operation sequence; the responsiveness score satisfies:

[0014] Furthermore, the stress resilience is characterized by the stress response intensity index SP, which further yields a stress resilience score SR. The SP is calculated based on the peak stress amplitude A(j) within the event window and the recovery time Trecov(j) after the action is initiated. in This is the moment when physiological indicators return to the threshold range.

[0015] Furthermore, the nursing action quality statistic Qs is obtained by weighting the time mean values ​​of each quality dimension of the nursing action quality assessment sequence, satisfying the following: Furthermore, the competency score includes a stress resistance modulation factor Ps, and satisfies: ; Where S^proc,s represents the output related to process execution, S^emg,s represents the output related to emergency identification and response, SRs represents the scenario-level stress resistance score, and w1+w2+w3=1.

[0016] Compared with the prior art, the present invention has at least the following beneficial effects: This invention uses multi-scenario emergency rescue task scripts to structure and manage standard procedures, key emergencies, and key treatment actions in emergency rescue tasks. This enables the assessment to cover the complete closed loop of "assessment-identification-treatment-monitoring-reassessment" in nursing emergency care, making it more closely aligned with actual clinical use.

[0017] This invention achieves joint modeling of process execution, action quality, and stress response through multi-channel temporal input of nursing operation behavior sequences, physiological response sequences, and nursing action quality assessment sequences. This avoids biased evaluation caused by a single data source and improves the objectivity and consistency of the scoring.

[0018] This invention constructs a key event-driven cascaded index link in channel 2. Starting from the event occurrence time tevent, it sequentially determines the identification time trec, the action initiation time tact, and the recovery time trecov, and calculates the emergency identification score, reaction ability score, and stress resistance score. This enables the system to distinguish common clinical problems such as "slow identification", "slow action", and "slow recovery", making the evaluation results more interpretable.

[0019] This invention allows the action quality statistic Qs to be directly involved in the calculation of the total score. While ensuring the deep learning modeling capability, it strengthens the rigid constraints on clinical standardization, reduces the risk of score drift caused by relying solely on latent features, and improves the reproducibility and reviewability of the score.

[0020] This invention solves the problem of multi-source asynchronous data fusion by unifying the time axis and time alignment mechanism, making the consistency of three-channel feature fusion better and supporting applications such as training error correction, assessment arbitration and quality traceability. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall deployment of the nursing competency assessment system based on a multi-channel deep learning model according to an embodiment of the present invention. It shows the spatial division relationship between the emergency operation area A and the control calculation area B, as well as the arrangement of modules such as emergency operation objects, scene presentation and interaction, multimodal acquisition, convergence / synchronization gateway, main control processing and model inference, and result output and storage / playback in the two areas and the main data / control directions.

[0022] Figure 2This diagram illustrates the module connection relationships and data / control flow of the system described in the embodiments of the present invention. It shows the connection relationship of the multimodal acquisition module uploading data to the main control processing and model inference module via the aggregation / synchronization gateway, as well as the connection relationship of the control commands and event triggering of the main control processing and model inference module to the scene presentation and interaction module and the emergency operation object module. It also shows the result output and history / configuration interaction relationship between the main control processing and model inference module and the result output and storage / playback module.

[0023] Figure 3 This is a schematic diagram of the control flow of the system described in the embodiments of the present invention, showing the sequential relationship of the process from loading task scripts and establishing a unified time base, device self-test and baseline acquisition, scene operation and key event triggering, data acquisition and time alignment, three-channel input construction, three-channel LSTM / Bi-LSTM encoding, event-driven indicator link calculation, feature fusion and scoring network output, and report generation and storage output.

[0024] Figure 4 This is a schematic diagram of the structure and index chain relationship of the multi-channel deep learning evaluation model described in the embodiments of the present invention. It shows that the nursing operation behavior sequence X(1) of channel 1, the physiological response sequence X(2) of channel 2, and the action quality sequence X(3) of channel 3 are respectively processed by the corresponding encoders E1, E2, and E3 (stacked LSTM / Bi-LSTM) to obtain feature vectors v1, v2, and v3. The event-driven index chain (tevent→trec→tact→trecov) is further formed to output ERs, RCs, SRs and quality statistics Qs. Finally, the fusion vector vfus is constructed and input into the scoring network fθ to output the scene score Stotal,s and sub-indicators, and the comprehensive competency score Stotal is obtained by summarizing. Detailed Implementation

[0025] This embodiment provides a nursing competency assessment system based on a multi-channel deep learning model. It is used to objectively and traceably quantify the comprehensive competency of nursing staff under routine emergency tasks within real clinical emergency workflows or standardized emergency training / assessment processes. The "routine emergency tasks" are not single skill points, but rather continuous processes consistent with actual clinical emergency care, typically including rapid patient assessment, abnormal identification and emergency call coordination, implementation of key interventions, continuous monitoring and review, recording and handover, etc. This system uses a multi-scenario emergency script-driven approach, synchronously collecting the timing of nursing staff's operational behaviors, physiological stress, and action quality throughout the task. It then uses a multi-channel deep learning model for joint modeling, outputting a comprehensive competency score and clinically interpretable sub-competency indicators. These sub-competencies include at least process execution ability, action quality ability, emergency identification and response ability, and stress stability ability.

[0026] In clinical application, this system is preferably deployed in emergency resuscitation rooms, ward resuscitation rooms, or nursing emergency simulation training rooms. To make the evaluation process as close to real operation as possible, this implementation divides the system layout into two spatial subdomains: an "emergency operation area" and a "control and computing area." The emergency operation area is used for nursing staff to perform treatment actions, while the control and computing area is used for scenario-driven operations, data aggregation and synchronization, model inference, and result output. Within the emergency operation area, patient simulators or semi-physical training devices are set up according to the common hospital resuscitation bed layout. Monitoring display terminals and commonly used emergency equipment are placed beside the bed, and emergency carts are located in easily accessible positions for nursing staff, ensuring that the nursing staff's positioning, reach range, retrieval path, and teamwork methods during the evaluation process conform to real clinical habits. The control and computing area houses a main control computer or edge server, connected to the emergency operation area via a local area network or dedicated wireless network to ensure low-latency transmission and synchronization of multimodal data.

[0027] The system in this embodiment includes, in terms of function and connectivity, a scene-driven and script management unit, a multimodal data acquisition unit, a data aggregation and time synchronization unit, a multi-channel deep learning evaluation unit, and a result output and storage unit. The scene-driven and script management unit is deployed on the main control computer in the control computing area and is used to load the emergency rescue scene library, execute task scripts, trigger critical sudden events, and record the event time. Multimodal data acquisition units are distributed throughout the emergency operation area. Acquisition components for collecting nursing action data are preferably positioned around or above the resuscitation bed. For example, cameras / positioning base stations can be installed at the foot or side of the bed to cover the nursing staff's operating range, or events can be extracted directly from the operation logs of equipment such as defibrillators, suction devices, and infusion pumps. Acquisition components for collecting the nursing staff's physiological responses are preferably worn on the nursing staff, such as collecting heart rate data on the chest or wrist, and skin conductance data on the fingers or palm, to minimize interference with aseptic and delicate operations. Acquisition components for collecting action quality data are preferably integrated into a simulation device or training equipment, such as chest compression depth / frequency sensors, ventilation volume sensors, tourniquet tension sensors, and connection correctness detection sensors, enabling action quality to be output in real-time in structured numerical form. The data aggregation and time synchronization unit is preferably located in a gateway device at the edge of the emergency operation area. This unit aggregates multiple sensor data streams, uniformly stamps them with the gateway clock timestamp, caches the data, and then uploads it to the main control computer via Ethernet or a high-speed wireless link. The multi-channel deep learning evaluation unit is deployed on the main control computer, constructing a three-channel time-series input based on a unified timeline and completing inference computation. The result output and storage unit is electrically or network-connected to the main control computer, used to display the comprehensive score, sub-indicators, key segment playback anchor points, and archive the entire process data.

[0028] Regarding connectivity, the emergency operation behavior acquisition components and action quality acquisition components are preferably connected to the data aggregation gateway via wired (USB / Ethernet / serial port) or Wi-Fi, while the physiological sensors are preferably connected to the data aggregation gateway via Bluetooth Low Energy. The data aggregation gateway and the main control computer preferably use an Ethernet connection to ensure stable transmission of continuous physiological and quality data. The main control computer maintains a control connection with the bedside monitoring display terminal. The main control computer changes monitoring parameter curves or triggers alarm prompts according to scripts, and records the occurrence time of critical acute events at the trigger moment. The main control computer maintains a communication connection with the result output terminal, outputting a scoring report after the task is completed and supporting playback verification based on key events. Through the above arrangement and connection, this system forms a closed loop in engineering implementation: "scenario-driven—clinical treatment—data acquisition—time synchronization—model inference—scoring feedback," with clear control and data flow paths, facilitating reproduction and review.

[0029] To meet the clinical needs of "competency testing in comprehensive nursing emergency scenarios," this implementation method pre-configures an emergency scenario library in the script management system. These scenarios cover common emergency situations such as cardiac arrest / ventricular fibrillation management, acute respiratory distress or airway obstruction management, traumatic hemorrhage control and shock identification, anaphylactic shock management, and early identification and emergency call coordination for sudden deterioration in the ward. Each scenario is defined at the script level using a framework of "standard process constraints + critical acute events + corresponding critical handling actions + quality evaluation dimensions." This allows the system to evaluate both whether nurses follow clinical pathways and their ability to identify and respond to critical acute events. During system operation, the main control computer initiates the scenario and triggers critical acute events according to the script, such as a sudden appearance of ventricular fibrillation waveforms accompanied by alarm sounds on the monitor, a sudden drop in blood pressure, a rapid decrease in SpO2, or respiratory arrest. The main control computer simultaneously records the occurrence time t of each event. event (j) This serves as the starting point for calculating subsequent capability indicators. In this event context, nursing staff perform actions such as calling for help, chest compressions, ventilation, defibrillation assistance, hemostasis, establishing access, continuous monitoring, and reporting. The system continuously collects three types of time-series data throughout the entire process.

[0030] Because the sampling frequencies and granularities of the three types of data differ, the system establishes a unified time axis and performs alignment control in the main control computer. At the start of the task, the main control computer generates a global start timestamp T0, sets a unified time step Δt, and constructs a unified time series tk = T0 + kΔt (k = 1, ..., K). Behavioral data is mostly discrete events; the system maps action category, action duration, step number, and deviation information to the corresponding time step. Physiological data is a continuous signal; after denoising and normalization, the system resamples it according to the unified time step. Quality data is usually output as time windows or key action nodes; the system uses hold-up or interpolation to map it to the unified time step. If a cross-channel time misalignment exceeding a preset threshold is found in the vicinity of a key event, the system can perform local correction within the event window using the key event timestamp as an anchor point. If necessary, dynamic time warping is performed to ensure semantic consistency of the three channels near the key event, thereby improving the stability of the fused features.

[0031] After time alignment is completed, the system constructs a three-channel input sequence. The nursing operation behavior channel is denoted as... , where x k (1) It should include at least the action category code ak, action duration or idle time feature dk, process node / step number sk, and process deviation feature rk, used to characterize the nurse's process execution, step sequence, and the timing of key actions. The nurse's physiological response channels are denoted as... , where x k (2) =[HR ~ k,EDA ~ [k, ΔHRk, ΔEDAk]. Before the task begins, the system preferentially collects the resting baseline of nursing staff and obtains HRbase and EDAbase, based on which normalized HR is performed. ~ k=(HRk−HRbase) / (HRbase+ε), EDA ~ k = (EDAk − EDAbase) / (EDAbase + ε), the rate of change term ΔHRk = HRk − HR k−1 ΔEDAk=EDAk−EDA k−1 Used to characterize the rising edge of sudden stress. Nursing action quality channel is denoted as... , where x k (3)=[q1k,q2k,...,qMk] represents a quality score across multiple dimensions. For example, in a cardiopulmonary resuscitation (CPR) scenario, this might include the accuracy of compression depth, compression frequency, rebound adequacy, ventilation volume, and correctness of defibrillation coordination. In a trauma hemostasis scenario, it might include the correctness of compression placement, the effectiveness of tourniquet tension, compliance of reassessment frequency, and the completeness of monitoring. Since the quality of the action directly reflects the degree of operational standardization, this implementation method, in addition to inputting it into the deep learning model, also uses it as an important basis for the overall score in subsequent calculations to avoid insufficient interpretability of the score due to relying solely on latent features.

[0032] The multi-channel deep learning evaluation unit performs temporal encoding on the three-channel sequences using either stacked LSTM or bidirectional LSTM. Taking any channel input xk as an example, its gating update can be represented as ik=σ(Wixk+Uihk−1+bi), fk=σ(Wfxk+Ufhk−1+bf), ok=σ(Woxk+Uohk−1+bo), c~k=tanh(Wcxk+Uchk−1+bc), and updates ck=fk⊙ck−1+ik⊙c~k and hk=ok⊙tanh(ck). The three channels output the hidden state sequence {h k (1)},{h k (2)},{h k (3) Then, through pooling or attention convergence, fixed-length feature vectors v1, v2, and v3 are obtained, where v1 focuses on reflecting the temporal dependence of the process and steps, v2 focuses on reflecting the dynamics of physiological stress, and v3 focuses on reflecting the evolution pattern of action quality over time.

[0033] To ensure that Channel 2 clinically reflects the closed-loop logic of "emergency identification, responsiveness, stress resistance, and stress response," this implementation constructs event-level indicators driven by critical acute events and strictly links these indicators through temporal relationships. For the j-th critical acute event, its occurrence time is t. event (j) After an event occurs, the system calculates the emergency identification probability p based primarily on the hidden state of channel 2, combined with the context of channel 1. ER (j) (k)= When the probability first reaches the threshold θ ER When to determine identification Recognition delay Corresponding to the speed at which a sudden change in a patient's condition occurs and the nursing staff completes the identification process in clinical practice, the system calculates an emergency identification score based on this speed. The faster the recognition, the higher the score.

[0034] Following emergency identification, the system evaluates the efficiency of nurses' response in translating identification into effective intervention. The script pre-sets a set of key intervention actions A for each event. (j) This set aligns with clinical guidelines; for example, ventricular fibrillation events include actions such as "calling for assistance, initiating chest compressions, connecting electrode pads and preparing for defibrillation," while hypotensive shock events include actions such as "establishing intravenous access, preparing for fluid resuscitation, continuous monitoring and reporting." The system searches for the moment when the critical intervention action first occurs after the identification time in the behavior sequence of Channel 1. And calculate the reaction delay. Since the starting point of the reaction delay is the recognition time t rec (j) This indicator is naturally linked to the emergency identification indicator, and can distinguish between "slow identification" and "slow action." The system further calculates a response capability score. The faster you react, the higher your score.

[0035] In high-pressure emergency care, the ability of nurses to recover quickly and remain stable after initiating critical procedures is a crucial indicator of their stress resilience. This implementation method anchors the intensity of the stress response and recovery time to the moment t is initiated. act (j) This allows the stress resilience index to continue sequentially along the "recognition-response" chain. The system first calculates the peak stress amplitude within a short window after the event-to-action transition. Then, after the action is initiated, find the recovery moment when the physiological indicators fall back to the threshold range. Thus, the recovery time is obtained. The stress response intensity index is defined as follows: The stress resistance score is defined as follows: At this point, a continuous chain is formed on the timeline. In terms of scoring indicators The interconnectedness of these parameters allows channel 2 to reflect both the magnitude of the pressure and the clinical ability to quickly achieve effective treatment and rapid recovery under pressure.

[0036] When a scenario contains multiple key events, the system averages the event-level metrics to obtain the scenario-level metrics. , This is to avoid the influence of a single, accidental event on the overall evaluation. Simultaneously, the action quality sequence in Channel 3 forms quality statistics and directly participates in the scoring; the system calculates the average value of each quality dimension. And form scene quality score ωm can be set according to clinical criticality. For example, compression depth, frequency, and rebound can be given higher weights in cardiopulmonary resuscitation, while hemostasis effectiveness and monitoring integrity can be given higher weights in hemostasis and shock scenarios. By directly incorporating Qs as an explicit quantity into the comprehensive score, it can be ensured that "action quality" is not diluted by implicit features, thus better meeting the rigid requirements of clinical standardization.

[0037] In the feature fusion and comprehensive scoring stage, the system constructs a fusion vector by combining the three-channel time-series feature vector with the channel 2 concatenation index and the channel 3 quality statistics. The data is then input into a scoring network to obtain process execution-related components and emergency response-related components, such as the output S^ proc,s With S ^emg,s (Normalized to 0-1). To reflect the clinical pattern that "stress resistance stability affects overall performance," this implementation method introduces a stress resistance modulation factor. This means that when caregivers recover slowly and experience significant fluctuations under high pressure, their overall score will be reasonably lowered even if they complete part of the task. The scenario-based comprehensive competence score is thus defined as follows: Where w1, w2, and w3 are the weights for process, quality, and emergency response capabilities, satisfying w1 + w2 + w3 = 1. It can be seen that the calculation chain from event triggering to the total score is continuous: the event occurrence time determines the identification time, the identification time determines the action initiation time, the action initiation time determines the recovery time, the recovery time determines the stress resistance score, and the stress resistance score participates in the total score through a modulation factor, thus forming a closed loop of "event—identification—response—stress resistance—total score".

[0038] In comprehensive nursing emergency assessment, a single task typically includes multiple scenarios. For example, it might begin with early identification and emergency call coordination for sudden deterioration in the ward, followed by cardiac arrest management, and then trauma hemostasis and transport monitoring. The system calculates Total,s for each scenario according to the script, and then aggregates the results by scenario weight πs to obtain the overall comprehensive competency score. The scenario weight can be set according to the scenario complexity or the number of key events to ensure that the overall score is consistent with the competency structure requirements of clinical emergency positions.

[0039] When outputting results, the system not only outputs the total score but also the chain of evidence that can be used for clinical training and assessment arbitration. The system can present the results of each key event in the report. and corresponding The system simultaneously presents the time curves and trough locations of quality dimensions, enabling instructors to clearly determine whether nurses' problems stem from slow recognition, slow initiation of actions, or unstable quality maintenance under pressure. For scenarios requiring teamwork, the system can incorporate emergency calls, collaborative actions, and handover records as part of the Channel 1 behavioral characteristics and Channel 3 quality dimensions, ensuring the report meets the real evaluation needs of clinical "collaborative treatment." The system archives all process data and model outputs for easy quality control review and capability development tracking.

[0040] Through the above implementation methods, this system, driven by multi-scenario clinical emergency scripts, achieves synchronous perception of nursing staff's operational behavior, physiological stress, and action quality, and realizes fusion modeling in a multi-channel deep learning model. At the same time, by clearly defining the interconnected indicator links, it incorporates emergency identification, response ability, and stress resistance into the total score calculation, thereby achieving an objective quantitative assessment of the comprehensive competence of nursing staff in comprehensive nursing emergency scenarios, and meeting the requirements of clinical training and assessment for interpretability, traceability, and consistency.

Claims

1. A nursing competency assessment system based on a multi-channel deep learning model, characterized in that, include: The scenario-driven and script management module is used to load nursing emergency task scripts and drive task execution. The nursing emergency task scripts include at least standard process constraints and a set of key acute events. The multimodal data acquisition module is used to collect nursing operation behavior data, physiological response data, and nursing action quality data during the nursing staff's performance of emergency care tasks; The data aggregation and time synchronization module is communicatively connected to the multimodal data acquisition module and is used to uniformly timestamp and transmit the nursing operation behavior data, physiological response data, and nursing action quality data; The main control processing and model inference module is communicatively connected to the data aggregation and time synchronization module, and is also controlled by the scene driving and script management module. It is used to establish a unified timeline and construct at least three channels of time-series input, which include nursing operation behavior sequence, physiological response sequence, and nursing action quality assessment sequence. The module extracts time-series features from each channel of time-series input and performs feature fusion to output the scene-level competency score and comprehensive competency score of the nursing staff. The result output and storage module is connected to the main control processing and model inference module, and is used to output and store the competency score and the corresponding key time points. The main control processing and model reasoning module is configured to: determine the emergency identification time trec in the physiological reaction sequence based on the nursing emergency task script recording the critical sudden event occurrence time tevent, determine the critical treatment action initiation time tact in the nursing operation behavior sequence, and determine the physiological indicator recovery time trecov in the physiological reaction sequence, thereby forming a series calculation link of emergency identification, response capability and stress resistance capability, and use the output of the series link and the nursing action quality statistics together for competency score calculation.

2. The nursing competency assessment system based on a multi-channel deep learning model according to claim 1, characterized in that, The scenario-driven and script management module is configured to trigger the critical acute events during task execution and change the display of monitoring parameters or alarm status, so that nursing staff can perform treatment actions in the corresponding clinical emergency situation, and at the same time record the event type and the teven of each critical acute event.

3. The nursing competency assessment system based on a multi-channel deep learning model according to claim 1, characterized in that, The time step features of the nursing operation behavior sequence include at least the action category code ak, the action duration feature dk, the process step identifier sk, and the process deviation feature rk, wherein the process deviation feature is used to characterize missed steps, out-of-order steps, or timeouts.

4. The nursing competency assessment system based on a multi-channel deep learning model according to claim 1, characterized in that, The time-step features of the physiological response sequence include at least the baseline normalized heart rate (HR). ~ k, baseline-normalized skin electrical activity (EDA) ~ k and its rates of change ΔHRk and ΔEDAk, where ； HRbase and EDAbase are the resting baselines before the task begins.

5. The nursing competency assessment system based on a multi-channel deep learning model according to claim 1, characterized in that, The main control processing and model inference module includes a multi-channel deep learning model. The multi-channel deep learning model includes at least three sets of temporal encoders corresponding to the nursing operation behavior sequence, physiological response sequence and nursing action quality assessment sequence, respectively. Each set of temporal encoders is a stacked long short-term memory network LSTM or a bidirectional long short-term memory network Bi-LSTM, used to output the temporal feature vectors v1, v2 and v3 of each channel.

6. A nursing competency assessment system based on a multi-channel deep learning model according to claim 1 or 5, characterized in that, The emergency identification time trec is determined by the emergency identification probability reaching a threshold, and the emergency identification probability satisfies: ； Where hk(2) is the encoder hidden state of the physiological response sequence, hk(1) is the encoder hidden state of the nursing operation behavior sequence, [ ⋅∥⋅] is vector concatenation; trec is the time corresponding to the smallest time step that satisfies pER(k)≥θERp.

7. The nursing competency assessment system based on a multi-channel deep learning model according to claim 1, characterized in that, The responsiveness is determined by the time difference between the critical intervention initiation time tac and the emergency identification time trec, where tac is the moment when an action within the preset critical intervention set AAA first appears after trec in the nursing operation sequence; the responsiveness score satisfies: 。 8. The nursing competency assessment system based on a multi-channel deep learning model according to claim 1, characterized in that, The stress resistance is characterized by the stress response intensity index SP, which is further used to obtain the stress resistance score SR. The SP is calculated based on the peak stress amplitude A(j) within the event window and the recovery time Trecov(j) after the action is initiated. in This is the moment when physiological indicators return to the threshold range.

9. The nursing competency assessment system based on a multi-channel deep learning model according to claim 1, characterized in that, The nursing action quality statistic Qs is obtained by weighting the time mean values ​​of each quality dimension of the nursing action quality assessment sequence, and satisfies the following: 。 10. A nursing competency assessment system based on a multi-channel deep learning model according to claim 1, characterized in that, The competency score includes a stress resistance modulation factor Ps and satisfies: ； Where S^proc,s represents the output related to process execution, S^emg,s represents the output related to emergency identification and response, SRs represents the scenario-level stress resistance score, and w1+w2+w3=1.

Images