Nursing risk real-time assessment and intervention method based on multi-modal data fusion

Abstract

The application discloses a kind of based on multi-modal data fusion's nursing risk real-time evaluation and intervention method, belong to wisdom nursing and medical data processing technical field.The method includes: S1.multimodal data real-time acquisition, the physiological, behavior, environment and text modal data of patient are synchronously collected, and desensitization encryption processing is carried out;S2.multimodal data pre-processing, cleaning, standardization, time alignment and feature extraction are carried out;S3.multimodal data deep fusion, in the server or edge node with power, adopt "intermediate layer feature fusion+dynamic loss optimization" strategy to generate fusion feature vector;S4.nursing risk real-time evaluation, based on LSTM model combined attention mechanism calculates comprehensive risk value, and through the false alarm analysis mechanism of combining clinical evaluation scale reduces false alarm rate;S5.personalized intervention scheme generation, from intervention knowledge base matches optimization intervention measure;S6.intervention execution and feedback iteration, form closed loop management and continuously optimize model.The application realizes the accurate identification of nursing risk, real-time early warning and personalized intervention, effectively reduce the incidence of adverse events, applicable to hospital, old-age and home nursing scene.

Description

Technical Field

[0001] This invention belongs to the field of smart nursing and medical data processing technology, specifically involving a method for real-time assessment and intervention of nursing risks based on multimodal data fusion. It is applicable to various scenarios such as hospital clinical nursing, elderly care, and home care, and can realize accurate identification, real-time early warning, and personalized intervention of patient nursing risks. Background Technology

[0002] Nursing risks refer to all unsafe events that may occur during the nursing process, including falls, pressure sores, deep vein thrombosis, aspiration, bed falls, and adverse drug reactions. These events not only worsen the patient's condition and prolong hospital stays but may also lead to medical disputes, seriously affecting the quality of nursing care and patient safety. Statistics show that approximately 60% of adverse nursing events are related to inadequate risk assessment, while standardized assessment can reduce safety events by 40%. Therefore, achieving accurate, real-time assessment and scientific intervention of nursing risks is a core requirement in the field of smart nursing.

[0003] Currently, existing nursing risk assessment and intervention techniques have the following shortcomings, making it difficult to meet actual clinical needs: 1. Limited Data Sources and Severely Biased Assessment: Existing technologies often rely on single-modal data (such as physiological signals or manual assessment scales) for risk assessment, neglecting the correlation between multiple dimensions of data, including patient physiology, behavior, environment, and medical history. For example, some technologies assess risk solely based on physiological indicators such as heart rate and blood pressure, without considering behavioral and environmental factors such as abnormal patient gait or slippery ward floors, leading to significant biases in assessment results and a high likelihood of missed or incorrect assessments. Other technologies rely solely on manually completed assessment scales, which is not only time-consuming and cumbersome but also susceptible to differences in nurses' experience, exhibiting strong subjectivity and poor real-time performance.

[0004] 2. Low level of data fusion and failure to solve the problems of modal heterogeneity and time alignment: Existing multimodal nursing risk assessment technologies mostly use simple feature splicing methods for data fusion, which do not fully consider the heterogeneity of different modal data (physiological signals, behavioral data, environmental data, and text data) (e.g., physiological signals are continuous numerical, nursing records are textual, and environmental data are discrete). Furthermore, they do not solve the problems of inconsistent temporal resolution and asynchronous acquisition of multimodal data, resulting in poor robustness of fused features, failure to accurately reflect the real-time status of patients, and difficulty in improving assessment accuracy.

[0005] 3. Significant assessment lag and lack of dynamic updating capability: Existing technologies mostly conduct periodic assessments (such as every 4-6 hours), which cannot capture the dynamic changes in the patient's condition in real time. When the patient's condition changes suddenly or risk factors change, it is difficult to update the assessment results in a timely manner, which may delay the best intervention time. Some machine learning-based assessment methods have fixed parameters after the model is trained, which cannot be dynamically iterated according to new clinical data and individual patient differences, resulting in insufficient generalization ability.

[0006] 4. Homogeneous intervention measures, lacking personalization and closed-loop management: Existing intervention programs are mostly general templates, failing to be personalized based on individual patient characteristics (such as age, medical history, physical condition, and nursing needs) and risk type, resulting in poor intervention effects; at the same time, most technologies only achieve risk warning, without establishing a closed-loop management mechanism of "assessment-warning-intervention-feedback-iteration", making it impossible to track intervention effects or optimize assessment models and intervention programs based on feedback results.

[0007] 5. High false alarm rate, which can easily lead to alarm fatigue among medical staff: Existing technologies lack an effective false alarm analysis mechanism and do not adequately handle noise interference in multimodal data, resulting in frequent false alarms of warning signals. This not only increases the workload of medical staff, but also reduces their sensitivity to warning signals, affecting the timely handling of high-risk events.

[0008] Therefore, in view of the shortcomings of existing technologies, there is an urgent need for a nursing risk assessment and intervention method that can integrate multimodal heterogeneous data, achieve real-time accurate assessment, provide personalized intervention and form a closed-loop management system. This method would solve the problems of one-sided, lagging and low accuracy assessment, homogeneous intervention and lack of closed-loop management in existing technologies, thereby improving the quality of nursing care and patient safety. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a real-time assessment and intervention method for nursing risks based on multimodal data fusion. Through deep fusion of multimodal data, construction of dynamic assessment models, generation of personalized intervention plans, and establishment of closed-loop management mechanisms, this method achieves accurate identification, real-time early warning, and scientific intervention of nursing risks. It solves the problems of one-sided, delayed, and high false alarm rates in existing technologies, as well as the homogeneity of interventions, thereby improving nursing efficiency and quality and ensuring patient safety. Technical solution

[0010] To achieve the above objectives, the present invention adopts the following technical solution: A method for real-time assessment and intervention of nursing risks based on multimodal data fusion includes the following steps: S1. Real-time acquisition of multimodal data: Construct a multi-source data acquisition system to simultaneously collect patients' physiological modality data, behavioral modality data, environmental modality data, and text modality data. All data carries a timestamp to ensure the temporal correlation of the data. At the same time, sensitive data (such as video behavioral data and medical record text data) is anonymized in real time, and an encrypted transmission protocol is used to achieve secure transmission between the data acquisition end and the processing end, protecting patient data privacy and security and meeting medical data compliance requirements.

[0011] S2. Multimodal data preprocessing: Cleaning, standardizing, time-aligning and feature extraction are performed on the collected various modal data to solve the problems of heterogeneity, time asynchrony and noise interference of multimodal data, and to obtain a standardized single-modal feature set; S3. Deep Fusion of Multimodal Data: This process is executed on servers or edge nodes with sufficient computing power. A fusion strategy of "intermediate layer feature fusion + dynamic loss optimization" is adopted to perform hierarchical fusion of preprocessed single-modal features, generating a fusion feature vector that can comprehensively reflect the patient's nursing risk status, thereby improving the robustness and representativeness of the features. The servers or edge nodes with sufficient computing power must meet the computing power requirements for synchronous operation and real-time fusion of the Transformer model and the LSTM model to ensure the real-time performance of the assessment.

[0012] S4. Real-time assessment of nursing risks: Based on fused feature vectors, a dynamic risk assessment model is constructed. Combined with a preset risk assessment indicator system, the comprehensive risk value and individual risk values ​​of patients are calculated in real time to determine the risk level. The false alarm rate is reduced through a false alarm analysis mechanism. S5. Personalized Intervention Plan Generation: Based on risk level, risk type and individual patient characteristics, personalized intervention plans are matched and optimized from the intervention knowledge base, specifying intervention measures, timing, responsible person and frequency. S6. Intervention Implementation and Feedback Iteration: Push the intervention plan to relevant nursing staff, track the intervention implementation process, collect multimodal data after the intervention, evaluate the intervention effect, and iteratively optimize the risk assessment model and intervention plan based on the feedback results, forming a closed-loop management of "collection-fusion-assessment-early warning-intervention-feedback-iteration".

[0013] 1. Physiological modal data: Collected through wearable devices, monitors, etc., including continuous numerical data such as heart rate, blood pressure, blood oxygen saturation, respiratory rate, body temperature, blood glucose, and heart rate variability. The collection frequency is 1-5 minutes / time to ensure real-time performance. 2. Behavioral modal data: Collected through devices such as millimeter-wave radar, AI cameras, and smart mattresses, including behavioral data such as the patient's activity trajectory, turning frequency, gait characteristics, sleep status, eating status, and changes in body position. Behavioral features are extracted using action recognition algorithms, with a collection frequency of 10-30 seconds / time. 3. Environmental modal data: Collected through sensors in the ward, including environmental parameters such as room temperature, humidity, light intensity, floor dryness / wetness, noise intensity, and bed protection status, with a collection frequency of 1-10 minutes / time; 4. Text Modal Data: Collected through electronic medical record systems and nursing record systems, including patients' medical history, surgical history, allergy history, medication records, nursing records, Braden pressure ulcer assessment scale, Morse fall assessment scale, and other text data. Natural Language Processing (NLP) technology is used to extract structured features.

[0014] S21. Data cleaning: Remove outliers (such as abrupt changes in physiological signals, invalid data caused by equipment failure) and missing values; use interpolation algorithms to fill in missing data; use Gaussian filtering to remove noise in physiological signals; and use anomaly detection algorithms to remove abnormal records in behavioral and environmental data. S22. Data Standardization: Normalize physiological modal data and environmental modal data of different dimensions and map the data to the [0,1] interval; perform word segmentation, part-of-speech tagging and stop word removal on text modal data, and transform text features into numerical features through word embedding technology; standardize action data of behavioral modal data and unify the threshold and feature dimensions of action recognition. S23. Time Alignment: Based on the timestamps of each data point, linear interpolation is used to align multimodal data from different acquisition frequencies to the same time scale (e.g., 1 minute / data point) to solve the problem of asynchronous time for multimodal data and ensure the temporal consistency of the data. S24. Feature Extraction: Different feature extraction methods are used for different modal data. For physiological modal data, time-domain features (mean, variance, peak value) and frequency-domain features (spectral energy, frequency peak value) are extracted; for behavioral modal data, action features (action type, action amplitude, action frequency) are extracted; for environmental modal data, the variation features of environmental parameters (change rate, fluctuation range) are extracted; and for text modal data, keyword features and semantic features are extracted, finally obtaining a standardized single-modal feature set.

[0015] S31. Single-modal feature enhancement: Feature enhancement processing is performed on each single-modal feature set. An autoencoder model is used to extract high-level features of each modality data, and specific features (unique features that correspond only to a single modality) and corresponding features (related features common to multiple modalities) are separated to improve the discriminative power of the features. S32. Intermediate layer feature fusion: A cross-attention mechanism is adopted to extract intermediate layer features of each single modality high-level features. Fine-grained alignment is achieved through token-level cosine similarity calculation. The intermediate layer features of different modalities are dynamically concatenated and input into the Transformer model for feature fusion to enhance the correlation of multimodal features. S33. Dynamic Loss Optimization: A triple loss function is constructed by combining the output layer loss, the intermediate layer fine-grained loss, and the fusion feature loss. The matching label value is dynamically adjusted to mitigate the impact of noisy data on the fusion effect and improve the robustness of the fusion features. S34. Feature fusion screening: Principal component analysis (PCA) algorithm is used to reduce the dimensionality of the fused feature vector, remove redundant features, and retain the core features that can best reflect nursing risks to obtain the final fused feature vector.

[0016] S41. Construct a risk assessment indicator system: covering four major categories of core indicators, including physiological risk indicators (such as abnormal heart rate, blood pressure fluctuations, insufficient blood oxygen, etc.), behavioral risk indicators (such as excessively low turning frequency, abnormal gait, frequent nighttime activity, etc.), environmental risk indicators (such as slippery floors, excessively high / low room temperature, excessive noise, etc.), and text-related risk indicators (such as history of falls, history of pressure ulcers, history of special medications, etc.). Each indicator is assigned a corresponding weight (determined based on the experience of clinical nursing experts and the analytic hierarchy process). S42. Construct a dynamic risk assessment model: Executed on a server or edge node with sufficient computing power, using an improved LSTM (Long Short-Term Memory) model, inputting a fused feature vector, and combining it with a risk assessment index system to calculate the score of each individual risk index and the overall risk value in real time; the model introduces an attention mechanism to focus on features that have a significant impact on nursing risks (such as the respiratory rate of postoperative patients and the gait characteristics of elderly patients) to improve the accuracy of the assessment; unlike the general "LSTM + attention mechanism" algorithm, this model closely integrates with the four-dimensional data system of the nursing scenario, deeply integrating clinical nursing experience with the model algorithm, rather than simply stacking algorithms.

[0017] S43. Risk Level Classification: Based on the comprehensive risk value, nursing risks are divided into 4 levels: no risk (0-20 points), low risk (21-40 points), medium risk (41-60 points), and high risk (61-100 points). Each level corresponds to a specific warning threshold. S44. False Alarm Analysis and Correction: A false alarm analysis model is constructed, inputting patient information, risk type, and confidence vectors for each indicator to determine whether the warning signal is a false alarm. Among them, the fall risk false alarm analysis combines the Morse fall assessment confidence vector, and the pressure ulcer risk false alarm analysis combines the Braden pressure ulcer assessment confidence vector to filter false alarm signals, reduce the false alarm rate, and avoid alarm fatigue among medical staff. This false alarm analysis mechanism is a dedicated design for nursing scenarios, combining clinical standard assessment scales, which is different from the scenario-independent false alarm handling method in general multimodal fusion algorithms. This is one of the core innovations of this invention that distinguishes it from existing technologies.

[0018] S51. Construct an intervention knowledge base: Integrate clinical nursing guidelines, expert experience, and past intervention cases, and classify and store intervention measures according to risk type (falls, pressure ulcers, deep vein thrombosis, etc.), risk level, and individual patient characteristics (age, physical condition, medical history) to form a standardized intervention template library; S52. Personalized matching and optimization: Based on the risk level and risk type obtained in step S4, combined with the patient's individual characteristics (e.g., for elderly patients at high risk of falls, priority is given to matching fall prevention railings, regular assistance, and other measures; for patients at high risk of postoperative pressure ulcers, measures such as air mattresses, regular turning, and nutritional support are matched), a basic intervention plan is matched from the intervention knowledge base; S53. Plan Refinement: Clarify the specific content of the intervention measures, the timing of implementation (e.g., high-risk patients are checked every 30 minutes, and medium-risk patients are checked every hour), the person responsible for implementation (responsible nurse, nursing assistant, etc.), the frequency of implementation and the expected results, and generate personalized intervention plans; for complex high-risk cases, automatically push to the head nurse and clinical experts, and optimize the plan in combination with manual intervention opinions.

[0019] S61. Intervention Push and Execution: Personalized intervention plans are pushed to the terminal devices (mobile phones, nurse station computers) of the corresponding nursing staff through the nursing management system, with clear execution reminders; after the nursing staff executes the intervention, the execution status (execution time, execution effect, patient feedback) is recorded in real time. S62. Intervention effect evaluation: Collect multimodal data within 1-2 hours after intervention, repeat steps S2-S4, calculate the risk value after intervention, compare it with the risk value before intervention, and evaluate the intervention effect (e.g., a decrease in risk value ≥30% is considered effective intervention, and a decrease <10% is considered ineffective intervention). S63. Model and Program Iteration: For several pre-effective cases, add the current intervention cases and related data to the intervention knowledge base to optimize the intervention program template; for several pre-ineffective cases, analyze the reasons for ineffectiveness (such as assessment bias, inappropriate intervention measures), adjust the parameters of the risk assessment model (such as feature weights, early warning thresholds) and the intervention program, and re-execute the intervention. S64. Closed-loop maintenance: Regularly (e.g., monthly) train and update the risk assessment model, incorporating new clinical data and intervention cases to continuously improve the model's assessment accuracy; at the same time, optimize the risk assessment indicator system and intervention knowledge base according to updates to clinical nursing guidelines to ensure the adaptability of the method.

[0020] Compared with the prior art, the present invention has the following significant novelty, inventiveness and practicality: 1. Novelty: Breaking through the limitations of existing single-modal data evaluation technologies, it constructs a four-dimensional multimodal data acquisition system of "physiology-behavior-environment-text" to cover the entire patient care scenario; it adopts a deep fusion strategy of "intermediate layer feature fusion + dynamic loss optimization" to solve the technical challenges of multimodal data heterogeneity, time asynchrony and noise interference, which is different from the existing simple feature splicing and fusion methods; it establishes a false alarm analysis mechanism to effectively reduce the false alarm rate of early warning, avoid alarm fatigue, and make up for the lack of false alarm handling in existing technologies.

[0021] 2. Creativity: This invention proposes a technical solution combining a dynamic risk assessment model with closed-loop management. The model incorporates an attention mechanism and autoencoder feature enhancement technology, and is executed on servers or edge nodes with sufficient computing power to ensure real-time performance. It can capture dynamic changes in the patient's state in real time, enabling dynamic updates of risk assessment and overcoming the shortcomings of existing technologies such as assessment lag and fixed model parameters. Unlike existing general AI algorithms such as "multimodal fusion + Transformer" and "LSTM + attention mechanism," the core innovation of this invention lies in: firstly, constructing a four-dimensional specific data system of "physiology-behavior-environment-text" for nursing scenarios, rather than a general multimodal data system. The system employs several key technologies: first, it integrates basic data; second, it designs a dedicated false alarm analysis mechanism that combines clinical assessment scales (Morse Fall Assessment Scale, Braden Pressure Ulcer Assessment Scale) to address the pain point of high false alarm rates in nursing scenarios; third, it achieves a closed-loop management system of "assessment-early warning-intervention-feedback-iteration," deeply integrating intervention effect feedback with model and program optimization to form a virtuous cycle of continuous iteration. This differs from existing technologies that only achieve assessment and early warning without a closed loop, and also from general algorithms that simply process data without considering nursing scenarios. Personalized intervention programs combine individual patient characteristics and risk types, solving the problem of homogenization in existing intervention programs and improving intervention effectiveness.

[0022] 3. Practicality: The method of this invention can be directly applied to various scenarios such as hospital clinical nursing, elderly care homes, and home care, without the need for large-scale modification of existing nursing equipment. It is compatible with existing wearable devices, monitors, sensors, and electronic medical record systems, resulting in low deployment costs and easy promotion. Simultaneously, through data anonymization and encrypted transmission, patient data privacy is protected, meeting medical data compliance requirements. Execution on servers or edge nodes with sufficient computing power ensures synchronous operation and real-time fusion of the Transformer and LSTM models, enabling real-time assessment and early warning of nursing risks. Early warning response time is reduced to the second level, allowing for early identification of high-risk events and providing medical staff with intervention time, effectively reducing the incidence of adverse events such as falls and pressure ulcers. Personalized intervention plans are clear and operable, reducing the workload of nursing staff and improving nursing efficiency and quality. A closed-loop management mechanism ensures the method can continuously adapt to changes in clinical scenarios, maintaining high accuracy in assessment and intervention over the long term. Combined with a false alarm analysis mechanism specific to nursing scenarios, it further enhances clinical application value, demonstrating extremely high clinical application value. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0024] Figure 1 This is a flowchart illustrating the overall process of the real-time nursing risk assessment and intervention method based on multimodal data fusion according to the present invention. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to specific embodiments, so that those skilled in the art can understand it.

[0026] Example 1: Application of a real-time nursing risk assessment and intervention method based on multimodal data fusion in elderly hospitalized patients. This embodiment targets elderly hospitalized patients (aged ≥ 65 years). These patients have a higher risk of falls and pressure ulcers, and their physiological state fluctuates greatly and their behavioral abilities are weaker. The method of this invention is used to achieve real-time assessment and intervention of nursing risks. The specific steps are as follows: S1. Real-time acquisition of multimodal data: 1. Physiological modality data: Patients wear smart bracelets and bedside monitors to collect heart rate (collection frequency 1 minute / time), blood pressure (collection frequency 5 minutes / time), blood oxygen saturation (collection frequency 1 minute / time), and body temperature (collection frequency 3 minutes / time) in real time, and upload them synchronously to the nursing management system; 2. Behavioral modal data: Millimeter-wave radar and AI cameras are installed in the ward to collect patients' activity trajectories, turning frequency (collection frequency 30 seconds / time), gait characteristics (such as walking speed, stride), and number of times they get up at night. Behavioral features are extracted through action recognition algorithms. A smart mattress is installed under the patient's mattress to collect data on changes in the patient's body position. 3. Environmental modal data: Temperature and humidity sensors, floor wetness and dryness sensors, and light sensors are installed in the ward to collect data on room temperature (5 minutes / time), humidity (5 minutes / time), floor wetness and dryness (1 minute / time), and light intensity (10 minutes / time) in real time. 4. Text modal data: Extract patient medical history (such as hypertension, diabetes history), surgical history, past fall history, pressure ulcer history, and medication records (such as antihypertensive drugs and sedatives) from the electronic medical record system, and extract the data filled in the Braden Pressure Ulcer Assessment Scale and the Morse Fall Assessment Scale from the nursing record system.

[0027] S2. Multimodal data preprocessing: S21. Data cleaning: Remove sudden heart rate values ​​caused by smart bracelet malfunctions (such as heart rate >150 beats / min or <40 beats / min), fill in missing blood pressure data using linear interpolation, remove noise from body temperature signals using Gaussian filtering, and remove invalid behavioral data caused by AI camera obstruction. S22. Data Standardization: Normalize physiological data such as heart rate, blood pressure, and blood oxygen saturation to the [0,1] interval; perform word segmentation and stop word removal on text data such as nursing records and medical history, and convert text features into numerical features through Word2Vec word embedding technology; standardize behavioral data such as turning frequency and gait features, and unify the action recognition threshold; S23. Time alignment: Based on the timestamps of each data point, linear interpolation is used to align all data to 1 minute / data point to ensure the temporal consistency of physiological data, behavioral data, and environmental data. S24. Feature Extraction: Physiological data is extracted to include time-domain features such as mean, variance, and peak value; behavioral data is extracted to include action features such as turning frequency and walking speed; environmental data is extracted to include features such as room temperature fluctuation range and ground wetness / dryness changes; and text data is extracted to include keyword features such as past fall history and special medications, resulting in a standardized single-modal feature set.

[0028] S3. Deep fusion of multimodal data: S31. Single-modal feature enhancement: Use an autoencoder model to extract high-level features of each modality of data, and separate specific features (such as the unique features of body temperature data) from corresponding features (such as the correlation features between heart rate and blood oxygen saturation). S32. Intermediate layer feature fusion: The cross-attention mechanism is adopted to extract the intermediate layer features of each single modality high-level features. Fine-grained alignment is achieved through token-level cosine similarity calculation. The intermediate layer features of physiological, behavioral, environmental and text modalities are input into the Transformer model for dynamic fusion. S33. Dynamic Loss Optimization: Construct a triple loss function (output layer loss + intermediate layer fine-grained loss + fusion feature loss) to dynamically adjust the matching label value and mitigate the impact of noisy data; S34. Feature fusion selection: The PCA algorithm is used to reduce the dimensionality of the fusion feature vector and remove redundant features to obtain the core fusion feature vector.

[0029] S4. Real-time assessment of nursing risks: S41. Risk assessment indicator system: covering physiological risk indicators (abnormal heart rate, blood pressure fluctuation, insufficient blood oxygen), behavioral risk indicators (too infrequent turning over, abnormal gait, frequent nighttime awakening), environmental risk indicators (slippery ground, excessively high / low room temperature), and text-related risk indicators (history of falls, history of pressure sores, use of sedatives). The weight of each indicator is determined through clinical expert experience and analytic hierarchy process. S42. Dynamic Risk Assessment: The improved LSTM model is input with fused feature vectors and an attention mechanism is introduced to focus on features that have a significant impact on the risk of elderly patients, such as gait features and turning frequency. The comprehensive risk value and individual risk values ​​are calculated in real time. S43. Risk level classification: An elderly patient's comprehensive risk score before intervention was 72 points, which was determined to be high risk (fall risk). S44. False Alarm Analysis: Input patient information, fall risk type, and Morse fall assessment confidence vector (including fall history confidence, gait confidence, environmental information confidence, etc.) to determine if the warning signal is a valid warning and there are no false alarms.

[0030] S5. Generation of Personalized Intervention Plans: S51. Match basic intervention protocols for elderly patients at high risk of falls from the intervention knowledge base; S52. Based on the patient's individual characteristics (78 years old, history of falls, unsteady gait, hypertension), optimize the intervention plan as follows: ① Install fall guardrails at the bedside, ensuring they are securely fixed; ② The responsible nurse should check on the patient every 30 minutes, assisting them with getting up and turning over; ③ Lay non-slip mats on the ward floor, keep it dry, and turn on the night light at night; ④ Adjust medication times to avoid dizziness caused by antihypertensive drugs; ⑤ Instruct the patient to wear non-slip shoes and avoid getting out of bed alone. S53. Clearly define the person responsible for the intervention (responsible nurse Li), the frequency of implementation (check every 30 minutes, and inspect fall prevention facilities twice a day), and the expected results (the risk value will be reduced to below 40 points after the intervention).

[0031] S6. Intervention Implementation and Feedback Iteration: S61. The nursing management system pushes the intervention plan to nurse Li's mobile terminal and sets a reminder every 30 minutes; after Li implements the intervention, the implementation status is recorded in real time. S62. Two hours after the intervention, multimodal data of the patient were collected, and steps S2-S4 were repeated. The comprehensive risk score was calculated to be 38 points, the risk level was reduced to low risk, and the intervention was effective. S63. Add this intervention case (intervention measures and effects on elderly patients at high risk of falls) to the intervention knowledge base to optimize subsequent intervention plans for similar patients; S64. Incorporate new elderly patient care data monthly, update risk assessment model parameters, optimize risk assessment indicator weights, and ensure the adaptability of the method.

[0032] After one month of application and verification, the incidence of fall-related adverse events in the ward where the elderly patient was located decreased by 75%, the incidence of pressure ulcers decreased by 60%, the work efficiency of nursing staff increased by 40%, and the intervention effectiveness rate reached 92%, which fully demonstrates the practicality and effectiveness of the method of the present invention.

[0033] Example 2: Application of a real-time nursing risk assessment and intervention method based on multimodal data fusion in postoperative patients This embodiment targets patients after abdominal surgery. These patients face nursing risks such as pressure sores, deep vein thrombosis, and aspiration, and their physiological state is unstable. The method of this invention enables real-time assessment and intervention of nursing risks. The specific steps are the same as in Embodiment 1, with the following key adjustments: 1. Multimodal data acquisition: Increase the acquisition of postoperative drainage volume, pain score (text modal data), and limb range of motion (behavioral modal data); increase the acquisition of ward air cleanliness data in environmental modal data. 2. Risk assessment indicator system: Focus on adding indicators for deep vein thrombosis risk (limb range of motion, coagulation-related physiological indicators), aspiration risk (feeding status, changes in body position), and postoperative infection risk (body temperature fluctuations, drainage fluid status); 3. Personalized intervention plan: For postoperative patients, additional intervention measures such as regular turning (every 2 hours), limb massage, nebulized inhalation (to prevent aspiration), and drainage tube care are added, and the frequency of intervention is adjusted according to the patient's surgical type and recovery status.

[0034] Application verification shows that after adopting the method of the present invention, the incidence of pressure ulcers in postoperative patients decreased by 80%, the incidence of deep vein thrombosis decreased by 70%, aspiration events occurred zero, the average recovery period of patients was shortened by 2 days, and nursing satisfaction increased to over 95%.

Claims

1. A method for real-time assessment and intervention of nursing risks based on multimodal data fusion, characterized in that: Includes the following steps: S1. Real-time acquisition of multimodal data: Construct a multi-source data acquisition system to simultaneously collect patients' physiological modal data, behavioral modal data, environmental modal data, and text modal data, with all data carrying timestamps; S2. Multimodal data preprocessing: The collected multimodal data are cleaned, standardized, time-aligned, and feature extracted to obtain a standardized single-modal feature set; S3. Deep fusion of multimodal data: The fusion strategy of "intermediate layer feature fusion + dynamic loss optimization" is adopted to perform layered fusion of preprocessed single-modal features to generate fused feature vectors; S4. Real-time assessment of nursing risks: Based on fused feature vectors, a dynamic risk assessment model is constructed. Combined with a preset risk assessment indicator system, the comprehensive risk value and individual risk values ​​of patients are calculated in real time to determine the risk level. The false alarm rate is reduced through a false alarm analysis mechanism. S5. Personalized Intervention Plan Generation: Based on risk level, risk type and individual patient characteristics, personalized intervention plans are matched and optimized from the intervention knowledge base, specifying intervention measures, timing, responsible person and frequency. S6. Intervention Implementation and Feedback Iteration: The intervention plan is pushed to relevant nursing staff, the intervention implementation process is tracked, multimodal data after the intervention is collected (data anonymization and encrypted transmission are performed simultaneously during the collection process), the intervention effect is evaluated, and the risk assessment model and intervention plan are iteratively optimized based on the feedback results to form a closed-loop management; the risk assessment model and the multimodal data fusion process are both executed on servers or edge nodes with certain computing power, and the servers or edge nodes with certain computing power need to meet the computing power requirements for the synchronous operation and real-time fusion of the Transformer model and the LSTM model.

2. The method according to claim 1, characterized in that: In step S1, the multimodal data includes: Physiological modal data: including at least one of heart rate, blood pressure, blood oxygen saturation, respiratory rate, body temperature, blood glucose, and heart rate variability; Behavioral modality data: including at least one of the patient's activity trajectory, turning frequency, gait characteristics, sleep status, eating status, and changes in body position; Environmental modal data includes at least one of the following: room temperature, humidity, light intensity, ground wetness / dryness, noise intensity, and bed protection facility status. Text modal data includes at least one of the following: patient medical history, surgical history, allergy history, medication records, nursing records, Braden pressure ulcer assessment scale, and Morse fall assessment scale.

3. The method according to claim 1, characterized in that: In step S2, the data preprocessing includes: S21. Data cleaning: Remove outliers and missing values, fill in missing data using interpolation algorithms, and remove noise using Gaussian filtering; S22. Data standardization: Normalize data of different dimensions, convert text data into numerical features, and standardize behavioral data. S23. Time alignment: Based on timestamps, linear interpolation is used to align multimodal data from different acquisition frequencies to the same time scale; S24. Feature Extraction: Different feature extraction methods are used for different modal data to obtain a standardized single-modal feature set.

4. The method according to claim 1, characterized in that: In step S3, the deep fusion of multimodal data includes: S31. Single-modal feature enhancement: Employ an autoencoder model to extract high-level features from each modality's data, separating specific features from their corresponding features; S32. Intermediate layer feature fusion: A cross-attention mechanism is used to extract intermediate layer features of each single modality's high-level features, achieve fine-grained alignment, and then input them into the Transformer model for fusion. S33. Dynamic Loss Optimization: A triple loss function is constructed by combining the output layer loss, the intermediate layer fine-grained loss, and the fusion feature loss, and the matching label value is dynamically adjusted; S34. Feature fusion screening: Principal component analysis algorithm is used to reduce the dimensionality of the fused feature vector and remove redundant features.

5. The method according to claim 1, characterized in that: In step S4, the real-time assessment of nursing risk includes: S41. Construct a risk assessment indicator system: covering physiological risk indicators, behavioral risk indicators, environmental risk indicators, and text-related risk indicators, with each indicator assigned a corresponding weight; S42. Construct a dynamic risk assessment model: Use an improved LSTM model, introduce an attention mechanism, and input fused feature vectors to calculate the comprehensive risk value and individual risk values; S43. Risk Level Classification: Based on the comprehensive risk value, nursing risks are classified into four levels: no risk, low risk, medium risk, and high risk. S44. False Alarm Analysis and Correction: Construct a false alarm analysis model, input patient information, risk type and confidence vector of each indicator, and filter false alarm signals.

6. The method according to claim 1, characterized in that: In step S5, the generation of the personalized intervention plan includes: S51. Construct an intervention knowledge base: Store intervention measures according to risk type, risk level, and individual patient characteristics to form a standardized intervention template library; S52. Personalized matching and optimization: Based on risk level, risk type and individual patient characteristics, match and optimize basic intervention plans from the intervention knowledge base; S53. Plan Detailing: Clarify the specific content of the intervention measures, the timing of implementation, the person responsible for implementation, the frequency of implementation, and the expected results.

7. The method according to claim 1, characterized in that: In step S6, the intervention execution and feedback iteration includes: S61. Intervention Push and Execution: Push the intervention plan to the nursing staff's terminal, and the nursing staff will record the execution status after execution; S62. Intervention effect evaluation: Collect multimodal data after intervention, calculate the risk value after intervention, and compare it with the pre-intervention value to evaluate the intervention effect; S63. Model and Program Iteration: Update the intervention knowledge base and adjust the risk assessment model parameters and intervention program based on the intervention effect; S64. Closed-loop maintenance: Regularly update the risk assessment model, risk assessment indicator system, and intervention knowledge base.

8. The method according to any one of claims 1-7, characterized in that: The method can be applied to hospital clinical nursing, elderly care, and home care scenarios, and is compatible with existing wearable devices, monitors, sensors, and electronic medical record systems. The method is executed on servers or edge nodes with a certain computing power, which must meet the computing power requirements for synchronous operation and real-time fusion of Transformer and LSTM models. During the collection and transmission of multimodal data, data desensitization and encrypted transmission processing are performed to ensure patient data privacy and security and meet medical data compliance requirements.

Images