Intelligent early warning system and method for paralytic nursing based on Internet of Things technology

Abstract

The invention discloses a paralytic nursing intelligent early warning system and method based on the Internet of Things technology. The system comprises a multi-modal physiological parameter acquisition module, an edge calculation preprocessing unit, an intelligent data transmission module, a multi-scale time sequence feature extraction module, a space-time diagram convolutional network module, a cross-modal attention fusion module, a multi-task risk prediction module and a model training and optimization module. According to the system, multi-mode data such as electrocardio, blood pressure, blood oxygen, eye movement tracks and voice are collected, preprocessed at an edge end and then transmitted to a cloud end; a multi-scale convolutional network is adopted to extract time sequence features, parameter association is modeled through space-time diagram convolution, cross-modal data fusion is realized by using an attention mechanism, and finally risk classification, anomaly detection and trend prediction are completed through a multi-task network. The early-stage, accurate and explainable early warning of the stroke risk is realized, and the early warning accuracy and clinical practicability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent early warning technology, specifically to an intelligent early warning system and method for stroke patient care based on Internet of Things (IoT) technology. Background Technology

[0002] Stroke is characterized by high incidence and high disability rates, and its prognosis is closely related to the availability of early warning and timely intervention. Current clinical monitoring systems suffer from the following shortcomings: First, they have limited monitoring dimensions, primarily focusing on basic vital signs such as electrocardiogram and blood pressure, lacking monitoring of specific neurological functions such as eye movements and speech. Second, their early warning mechanisms are simplistic, relying on single-parameter threshold alarms and failing to identify complex correlations between multiple parameters, leading to high false alarm rates and delayed warnings. Third, their data analysis is superficial, making it difficult to extract deep features from multimodal data. Fourth, balancing data privacy and computational efficiency is challenging. Summary of the Invention

[0003] To address the aforementioned technical shortcomings, the purpose of this invention is to provide an intelligent early warning system and method for stroke patient care based on Internet of Things (IoT) technology, thereby resolving the problems mentioned in the background section.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides an intelligent early warning system for stroke patient care based on Internet of Things technology, including a multimodal physiological parameter acquisition module for real-time acquisition of the patient's physiological parameters; An edge computing preprocessing unit, connected to the acquisition module, is used to preprocess the acquired raw data; The intelligent data transmission module, connected to the preprocessing unit, is used to encrypt and transmit the preprocessed data to the cloud. Cloud-based data processing centers include: A multi-scale temporal feature extraction module is used to extract multi-scale temporal features from physiological signals; The spatiotemporal graph convolutional network module is used to model the spatiotemporal correlations between multiple physiological parameters; A cross-modal attention fusion module is used to fuse heterogeneous multimodal data; The multi-task risk prediction module is used for stroke risk classification, anomaly detection, and trend prediction. The early warning and visualization module is used to trigger tiered early warnings based on prediction results and provide clinical decision support.

[0005] Preferably, the multimodal physiological parameter acquisition module includes: The core vital signs monitoring unit includes an electrocardiogram monitoring subsystem, a blood pressure monitoring subsystem, and a blood oxygen monitoring subsystem; The nervous system function monitoring unit includes an eye movement tracking monitoring module and a speech feature analysis module.

[0006] Preferably, the multi-scale temporal feature extraction module specifically includes: Three parallel convolutional branches use small, medium, and large convolutional kernels respectively; A feature recalibration mechanism is used to adaptively calibrate the weights of each feature channel; Temporal attention mechanism is used to weight feature sequences in the time dimension.

[0007] Preferably, the spatiotemporal graph convolutional network module specifically includes: The dynamic graph structure building unit learns the edge weights between physiological parameter nodes through an attention mechanism; The graph convolution operation unit is used to implement multi-hop neighbor information propagation; The temporal dynamic modeling unit integrates a gated loop unit to capture the temporal evolution of parameters.

[0008] Preferably, the cross-modal attention fusion module specifically includes: Feature projection networks are used to map features from various modalities to a unified semantic space; A cross-modal multi-head attention mechanism is used to achieve bidirectional information exchange between modalities; Context-aware gating fusion networks are used to dynamically adjust the fusion weights of each modality based on the clinical context.

[0009] Preferably, the multi-task risk prediction module specifically includes: A shared encoder is used to learn a general high-level feature representation; The risk classification branch uses temperature-regulated Softmax to output the risk probability; The anomaly detection branch is used to output the independent anomaly probability of each physiological parameter; The trend prediction branch integrates temporal convolutional networks and self-attention mechanisms for trend prediction.

[0010] Preferably, the signal preprocessing of the edge computing preprocessing unit includes: Wavelet transform is used for signal denoising; Outlier detection is performed using machine learning-based methods. Data standardization is performed using sliding window normalization.

[0011] Preferably, it further includes a model training and optimization module, which is configured as follows: A federated learning framework is used to train the model while protecting data privacy. A dynamically weighted multi-task loss function is used for collaborative optimization. The model is compressed and accelerated using knowledge distillation, quantization, and pruning techniques.

[0012] This invention also provides an intelligent early warning method for stroke patient care based on Internet of Things (IoT) technology, comprising the following steps: The multimodal physiological parameter acquisition module collects the patient's physiological parameters, eye movement trajectory, and voice signals in real time. The raw data collected is preprocessed in the edge computing preprocessing unit; The pre-processed data is encrypted and transmitted to the cloud server via an intelligent data transmission module. On the cloud server, feature extraction and fusion are performed through a multi-scale temporal feature extraction module, a spatiotemporal graph convolutional network module, and a cross-modal attention fusion module. Stroke risk prediction is performed using a multi-task risk prediction module. The prediction results trigger a tiered early warning mechanism and provide decision support to medical staff.

[0013] Preferably, the multi-scale temporal feature extraction employs a parallel multi-branch convolutional network combined with feature recalibration and temporal attention mechanisms; the spatiotemporal graph convolutional network models spatiotemporal correlations by dynamically learning adjacency matrices and gated recurrent units; and the cross-modal attention fusion achieves heterogeneous data fusion through multi-head attention and context-aware gating.

[0014] The beneficial effects of the present invention are as follows: Compared with the prior art, the present invention has the following significant advantages: More comprehensive monitoring: By integrating vital signs and neurological function monitoring, it has achieved three-dimensional perception of stroke-related signs for the first time, and can capture early specific symptoms such as nystagmus and dysarthria.

[0015] More accurate early warning: By employing multi-scale feature extraction and spatiotemporal graph convolutional networks, it can identify weak precursors of anomalies and changes in multiple parameters, significantly improving the early warning time while greatly reducing the false alarm rate.

[0016] Smarter Decision Making: Through a cross-modal attention mechanism, physiological data from different sources are intelligently integrated and decision weights are adjusted in a personalized manner based on clinical context, making risk assessment more targeted.

[0017] More efficient architecture: It adopts a layered architecture of "edge preprocessing + cloud deep analysis" to ensure real-time performance while ensuring data privacy and security through federated learning.

[0018] More credible results: Provides interpretable decision-making evidence, clearly demonstrates key signs that trigger warnings, enhances clinicians' trust, and facilitates precise intervention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The intelligent early warning system for stroke patient care based on Internet of Things technology of the present invention adopts a distributed intelligent monitoring architecture and achieves accurate early warning of stroke risk through multi-level data processing and analysis.

[0021] The intelligent early warning system for stroke patient care based on Internet of Things technology of the present invention includes the following modules: Multimodal physiological parameter acquisition module: used to comprehensively collect various physiological data of patients and form a three-dimensional sensing network.

[0022] Edge computing preprocessing unit: Deployed on edge devices close to the data source, used to perform preliminary cleaning, compression and quality assessment of raw data; Intelligent data transmission module: used to securely and reliably upload pre-processed data to the cloud server via an encrypted channel; Multi-scale temporal feature extraction module: running in the cloud, used to extract deep features at different time scales from physiological signals; Spatiotemporal graph convolutional network module: used to model the complex relationships between multiple physiological parameters and their spatiotemporal evolution laws; Cross-modal attention fusion module: used to fuse heterogeneous physiological signals and neural system functional signals; Multi-task risk prediction module: used to perform comprehensive risk assessment, anomaly detection, and trend prediction; Model training and optimization module: used for continuous learning, model optimization and performance improvement of the system; The above modules work together to form a complete system from data collection to risk warning.

[0023] The specific tasks of each module are as follows: 2.1 Multimodal physiological parameter acquisition module It consists of two sub-units: a core vital signs monitoring unit and a nervous system function monitoring unit, aiming to achieve a three-dimensional perception of the patient's physiological state.

[0024] The core vital signs monitoring unit includes: ECG monitoring subsystem: Employs a medical-grade analog front-end chip to acquire ECG signals at a sampling frequency of 250Hz. This subsystem features a high common-mode rejection ratio (CMRR>110dB), effectively suppressing environmental interference, and integrates lead dropout detection functionality.

[0025] Blood pressure monitoring subsystem: Based on the oscillometric principle, it uses a high-precision pressure sensor to automatically perform non-invasive blood pressure measurements every 30 minutes. The system precisely controls the inflation pump and deflation valve, and integrates accelerometer data to correct for changes in body position, enabling it to identify key patterns such as orthostatic hypotension.

[0026] Blood oxygen monitoring subsystem: Employs an integrated chip to acquire photoplethysmography (PPG) signals at a sampling rate of 100Hz. This system utilizes an adaptive ambient light cancellation algorithm, maintaining high measurement accuracy even during periods of slight patient movement.

[0027] The nervous system function monitoring unit includes: Eye Movement Tracking Monitoring Module: This module is an innovative design for neurological assessment of stroke. It uses a miniature infrared camera to record eye movements at a frame rate of 60Hz. Through computer vision algorithms, it accurately calculates the eye position and can effectively identify characteristic changes common in stroke patients, such as nystagmus, slowed saccades, and abnormal smooth tracking. These are often early signs of brainstem or cerebellar lesions.

[0028] Speech Feature Analysis Module: This module is used to capture stroke-related speech dysfunction. It employs a microphone array to acquire speech signals at a 16kHz sampling rate and uses beamforming technology to suppress ambient noise. The system extracts multiple acoustic features, including fundamental frequency, formants, speech rate, and articulation (such as jitter and shimmer), and can sensitively reflect articulation disorders and decreased speech fluency.

[0029] 2.2 Edge Computing Preprocessing Unit This module is deployed on edge computing devices close to the data source and is used for signal denoising, outlier detection, and data standardization.

[0030] Specifically, signal denoising involves using a wavelet transform algorithm, selecting the db4 wavelet as the mother wavelet, performing N-level decomposition on the signal, and processing the detail coefficients using an adaptive threshold function. Outlier detection: Using machine learning-based methods (such as the Isolation Forest algorithm), abnormal data points caused by sensor malfunction or motion interference can be effectively identified. Data standardization: A sliding window normalization method was used. The window size was 5 minutes. The mean (μ) and standard deviation (σ) of the data within the window were calculated, and Z-score standardization was performed: X_norm = (X - μ) / σ.

[0031] 2.3 Multi-scale temporal feature extraction module Employing a parallel multi-scale convolutional architecture, this module comprehensively captures the dynamic characteristics of physiological signals and is implemented based on the following algorithm: Let the input physiological signal be X ∈ R^(T×D), where T is the time step and D is the feature dimension.

[0032] Parallel convolution branches: First branch (high frequency): Uses a smaller convolution kernel K1 to output feature F1 ∈ R^(T×C1); the calculation formula is: F1 = ReLU(BN(X ∗ W1 + b1)); Second branch (mid-frequency): Using a medium-sized convolution kernel K2, output feature F2 ∈ R^(T×C2); The third branch (low frequency): uses a larger convolution kernel K3 to output features F3 ∈ R^(T×C3).

[0033] Feature fusion and enhancement: Concatenation: Concatenate the outputs of the three branches along the channel dimension to form F_cat = [F1, F2, F3] ∈ R^(T×(C1+C2+C3)).

[0034] 1. Feature Recalibration (SENet): Squeeze: Calculates channel statistics using global average pooling: z_c = (1 / T) ∑ F_cat(t, c); Excitation: Learn the relationship between channels through a fully connected layer to generate a weight vector: s = σ(W2 δ(W1 z)), where δ is ReLU and σ is Sigmoid; 2. Recalibration: Weight the original features: F_se = s · F_cat; 3. Temporal attention mechanism: Calculate the attention weight A = softmax(V_a tanh(W_a F_se^T)), weight the feature sequence in the time dimension, and finally output F_out = A · F_se.

[0035] The symbols in the above algorithm formula have the following meanings: X: Input physiological signal matrix; T: Time step; D: Input feature dimension; K1, K2, K3: Convolution kernel size; F1, F2, F3: Feature maps output by each branch; W1, W2, W3: Convolution weight matrices; b1, b2, b3: Convolution bias terms; F_cat: The concatenated feature map; z: Channel statistics vector; s: Channel weight vector; F_se: The recalibrated feature map; A: Temporal attention weight vector; F_out: Final output features.

[0036] 2.4 Spatiotemporal Graph Convolutional Network Module This module models the complex relationships between multiple physiological parameters using graph structures and captures their spatiotemporal dynamics.

[0037] Specifically, it is implemented through the following algorithm: 1. Dynamic Graph Construction: Each physiological parameter (such as heart rate, systolic blood pressure, and blood oxygen saturation) is treated as a node in a graph. The edge weight α_ij between nodes i and j is dynamically learned through an attention mechanism.

[0038] e_ij = LeakyReLU(a^T [W h_i || W h_j]) α_ij = exp(e_ij) / ∑ exp(e_ik) 2. Spatial Graph Convolution: Aggregates neighbor node information to update the current node's features. h_i' = σ(∑ α_ij W h_j) 3. Temporal dynamic modeling: The updated node feature sequence is input into the gated recurrent unit (GRU), and its reset gate r_t and update gate z_t mechanism are used to adaptively capture the temporal evolution of parameters.

[0039] The symbols in the above algorithm formula have the following meanings: h_i, h_j: Node feature vectors; e_ij: Attention coefficient; α_ij: Normalized attention weights; W: Learnable weight matrix; a: Attention vector; h_i': Updated node characteristics; σ: Activation function; r_t: Reset gate; z_t: Update gate.

[0040] 2.5 Cross-modal attention fusion module This module is used to achieve intelligent fusion of heterogeneous data such as ECG, blood pressure, eye movement, and voice.

[0041] Specifically, the implementation algorithm is as follows: Feature projection: Map the features X_m of each modality to a unified feature space through an independent projection network: H_m = LayerNorm(X_m W_m + b_m).

[0042] Cross-modal multi-head attention: For each modality m, a query vector Q_m = H_m W_q is generated, while the key K and value V are generated by concatenating features from all modalities. Features after cross-modal interaction are computed using scaled dot product attention.

[0043] Attention(Q_m, K, V) = softmax( (Q_m K^T) / √D_k ) V Context-aware gated fusion: Based on the patient's clinical context features c (such as age, underlying diseases), the fusion weights g_m = σ(U_g [Z_m; c] + b_g) of each modality are calculated, and the final fusion feature is Z = ∑ g_m · Z_m.

[0044] In the above algorithm, the meanings of each symbol are as follows: X_m; the input features of the m-th modality; H_m; Modal characteristics after projection; W_m, b_m; Projection weights and biases; Q_m; query matrix; K, V; key and value matrices; W_q, W_k, W_v; Attention weight matrix; D_k; Attention dimension; Z_m; Attention output; c; Clinical context feature vector; g_m; gating weight; Z; final fusion feature.

[0045] 2.6 Multi-task risk prediction module This module achieves multi-objective collaborative optimization through shared representation learning and task-specific networks.

[0046] It is implemented through the following algorithm: Shared encoder: Employs a residual network structure to extract high-level general feature representation H_shared from the fused features Z.

[0047] Task-specific branches: Risk classification branch: Outputs the classification probability of stroke risk p = softmax(W_c H_shared / τ), where τ is a temperature parameter used to smooth the probability distribution and improve model calibration.

[0048] Anomaly detection branch: Designed as a multi-label classification network, outputting the independent anomaly probability p_k^anom = σ(W_a^k H_shared) for each physiological parameter.

[0049] Trend prediction branch: Integrating Temporal Convolutional Network (TCN) and self-attention mechanism, it predicts the risk trend p_trend over a future period of time.

[0050] Interpretability analysis: Based on attribution methods (such as Integrated Gradients), it provides feature-level contribution analysis for each prediction.

[0051] 2.7 The model training and optimization module includes: Multi-task loss function: The dynamically weighted multi-task loss is L_total = λ1 L_class + λ2 L_anom + λ3 L_trend.

[0052] Federated Learning Framework: To protect data privacy, federated learning is adopted. Each hospital client trains the model locally and only uploads the gradients of the model with differential privacy noise added to the cloud for aggregation.

[0053] Model compression: Generate lightweight models suitable for edge deployment through knowledge distillation, quantization (such as FP32 to INT8), and structured pruning techniques.

[0054] Adaptive inference: Dynamically selects model computation paths of different complexities based on real-time risk scores and system load.

[0055] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A smart early warning system for stroke patient care based on Internet of Things (IoT) technology, characterized in that, It includes a multimodal physiological parameter acquisition module for real-time acquisition of patients' physiological parameters; An edge computing preprocessing unit, connected to the acquisition module, is used to preprocess the acquired raw data; The intelligent data transmission module, connected to the preprocessing unit, is used to encrypt and transmit the preprocessed data to the cloud. Cloud-based data processing centers include: A multi-scale temporal feature extraction module is used to extract multi-scale temporal features from physiological signals; The spatiotemporal graph convolutional network module is used to model the spatiotemporal correlations between multiple physiological parameters; A cross-modal attention fusion module is used to fuse heterogeneous multimodal data; The multi-task risk prediction module is used for stroke risk classification, anomaly detection, and trend prediction. The early warning and visualization module is used to trigger tiered early warnings based on prediction results and provide clinical decision support.

2. The intelligent early warning system for stroke patient care based on Internet of Things technology as described in claim 1, characterized in that, The multimodal physiological parameter acquisition module includes: The core vital signs monitoring unit includes an electrocardiogram monitoring subsystem, a blood pressure monitoring subsystem, and a blood oxygen monitoring subsystem; The nervous system function monitoring unit includes an eye movement tracking monitoring module and a speech feature analysis module.

3. The intelligent early warning system for stroke patient care based on Internet of Things technology as described in claim 1, characterized in that, The multi-scale temporal feature extraction module specifically includes: Three parallel convolutional branches use small, medium, and large convolutional kernels respectively; A feature recalibration mechanism is used to adaptively calibrate the weights of each feature channel; Temporal attention mechanism is used to weight feature sequences in the time dimension.

4. The intelligent early warning system for stroke patient care based on Internet of Things technology as described in claim 1, characterized in that, The spatiotemporal graph convolutional network module specifically includes: The dynamic graph structure building unit learns the edge weights between physiological parameter nodes through an attention mechanism; The graph convolution operation unit is used to implement multi-hop neighbor information propagation; The temporal dynamic modeling unit integrates a gated loop unit to capture the temporal evolution of parameters.

5. The intelligent early warning system for stroke patient care based on Internet of Things technology as described in claim 1, characterized in that, The cross-modal attention fusion module specifically includes: Feature projection networks are used to map features from various modalities to a unified semantic space; A cross-modal multi-head attention mechanism is used to achieve bidirectional information exchange between modalities; Context-aware gating fusion networks are used to dynamically adjust the fusion weights of each modality based on the clinical context.

6. The intelligent early warning system for stroke patient care based on Internet of Things technology as described in claim 1, characterized in that, The multi-task risk prediction module specifically includes: A shared encoder is used to learn a general high-level feature representation; The risk classification branch uses temperature-regulated Softmax to output the risk probability; The anomaly detection branch is used to output the independent anomaly probability of each physiological parameter; The trend prediction branch integrates temporal convolutional networks and self-attention mechanisms for trend prediction.

7. The intelligent early warning system for stroke patient care based on Internet of Things technology as described in claim 1, characterized in that, The signal preprocessing of the edge computing preprocessing unit includes: Wavelet transform is used for signal denoising; Outlier detection is performed using machine learning-based methods. Data standardization is performed using sliding window normalization.

8. The intelligent early warning system for stroke patient care based on Internet of Things technology as described in claim 1, characterized in that, It also includes a model training and optimization module, which is configured as follows: A federated learning framework is used to train the model while protecting data privacy. A dynamically weighted multi-task loss function is used for collaborative optimization. The model is compressed and accelerated using knowledge distillation, quantization, and pruning techniques.

9. The intelligent early warning method for stroke patient care based on Internet of Things technology as described in claim 1, characterized in that, Includes the following steps: The multimodal physiological parameter acquisition module collects the patient's physiological parameters, eye movement trajectory, and voice signals in real time. The raw data collected is preprocessed in the edge computing preprocessing unit; The pre-processed data is encrypted and transmitted to the cloud server via an intelligent data transmission module. On the cloud server, feature extraction and fusion are performed through a multi-scale temporal feature extraction module, a spatiotemporal graph convolutional network module, and a cross-modal attention fusion module. Stroke risk prediction is performed using a multi-task risk prediction module. The prediction results trigger a tiered early warning mechanism and provide decision support to medical staff.

10. The intelligent early warning method for stroke patient care based on Internet of Things technology as described in claim 9, characterized in that, The multi-scale temporal feature extraction employs a parallel multi-branch convolutional network combined with feature recalibration and temporal attention mechanisms; the spatiotemporal graph convolutional network models spatiotemporal correlations by dynamically learning the adjacency matrix and gated recurrent units. The cross-modal attention fusion achieves heterogeneous data fusion through multi-head attention and context-aware gating.