IoT-based Picc catheter safety monitoring and alarm system and method

Abstract

This invention relates to the field of medical device monitoring technology, specifically to a PICC catheter safety monitoring and alarm system and method based on the Internet of Things (IoT). Specifically, it involves: real-time acquisition of multimodal physiological parameters via intelligent sensors, followed by multi-level fusion, denoising, and normalization processing via an edge computing module to generate anti-loss seals; cloud-based analysis of time-series data and patient group correlation characteristics using an adaptive hybrid model, achieving unsupervised anomaly detection through contrastive learning and variational autoencoders, and lightweight deployment of the complex model using knowledge distillation technology; dynamic calculation of risk levels based on Bayesian anomaly scoring and weighted time-series indices, employing a tiered early warning strategy to match differentiated alarm notification methods, simultaneously generating personalized nursing suggestions and optimizing monitoring frequency; and integrating data, risk alerts, and decision support through a visual interactive interface. This invention achieves accurate identification, real-time early warning, and proactive intervention of PICC catheter abnormalities, effectively reducing false alarm rates and medical risks.

Description

Technical Field

[0001] This invention relates to the field of medical device monitoring technology, specifically to a Picc catheter safety monitoring and alarm system and method based on the Internet of Things. Background Technology

[0002] PICC (Peripherally Inserted Central Catheter) is a widely used intravenous infusion tool in clinical practice, providing long-term intravenous therapy support and suitable for patients requiring continuous infusion, parenteral nutrition, or chemotherapy. However, due to the long-term indwelling of PICC catheters, a series of safety risks may arise during their use, including but not limited to catheter displacement, blockage, infection, and venous thrombosis. These risks not only affect the patient's treatment outcome but may also lead to serious complications and even endanger life. Current PICC catheter monitoring methods mainly rely on regular manual examinations and ultrasound assessments. This method has problems such as poor real-time performance, dependence on medical experience, limited detection frequency, and inability to provide continuous monitoring data, resulting in some abnormalities not being detected and handled in a timely manner.

[0003] Chinese invention patent application CN118762838A discloses an image-based PICC intravenous therapy information monitoring system and method. By performing time-series analysis on high-definition images of the patient's puncture site at different time points within a target monitoring period, the system can determine whether there are signs of infection at the puncture site. This can improve the accuracy and timeliness of infection monitoring, help detect signs of infection early, and thus take timely measures to reduce the incidence and severity of infection.

[0004] In recent years, with the development of IoT technology, intelligent medical monitoring systems have been gradually applied in clinical settings. These systems collect patient physiological data through sensors and combine wireless communication, cloud computing, and artificial intelligence technologies for data analysis and anomaly detection, which is expected to improve the safety monitoring capabilities of PICC catheters. However, current intelligent monitoring systems still have shortcomings in terms of sensor stability, data processing accuracy, anomaly prediction capabilities, and alarm timeliness, and cannot fully meet clinical needs. Therefore, there is an urgent need for an IoT-based PICC catheter safety monitoring and alarm system to improve the real-time performance, accuracy, and intelligence of catheter monitoring, thereby effectively reducing the incidence of PICC catheter-related complications and improving patient safety and medical quality. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the background technology by proposing an Internet of Things-based Picc catheter safety monitoring and alarm system and method.

[0006] The technical solution of this invention: a method for safety monitoring and alarm of Picc catheters based on the Internet of Things, comprising the following specific implementation steps:

[0007] S1. Periodically collect Picc catheter environment data to generate physiological parameter data stream;

[0008] S2. Based on the multi-level fusion denoising method, the original signal is decomposed into multiple scales and denoised by adaptive threshold through wavelet transform and empirical mode decomposition. Combined with Kalman filtering, the smoothness of the time series signal is optimized. Z-score normalization is used to generate standardized data, and anti-lost seals are generated based on hash function and finite field operation.

[0009] S3. Based on the reliability of the anti-lost seal test data, an adaptive hybrid model is constructed, integrating graph neural network and Transformer+LSTM model. Combined with patient relationship graph, group association analysis and long and short-term time series feature capture are performed. Comparative learning is used to construct positive and negative samples to distinguish abnormalities. Error detection abnormality distribution is reconstructed through variational autoencoder. Knowledge distillation is used to deploy the complex model in a lightweight manner and output the detection results.

[0010] S4. Calculate the comprehensive risk value by combining the weighted time series anomaly index, dynamically adjust the risk threshold, determine the warning level, divide the risk into low, medium and high levels, and adopt a graded alarm strategy.

[0011] S5. Based on the risk grading results, patients are divided into low, medium and high risk levels, and differentiated nursing measures are formulated according to the level. The monitoring frequency and nursing plan are dynamically adjusted based on the patient's historical data, intelligent operation reminders are pushed, nursing logs are automatically generated and risk strategies are optimized.

[0012] S6. Visualizes standardized physiological data, test results, risk levels, corresponding alarm strategies, personalized maintenance recommendations, and optimal care plans.

[0013] The preferred implementation process of the multi-level fusion denoising method is as follows:

[0014] S21. The original signal is decomposed into multiple scales using wavelet transform to extract different frequency components. Threshold denoising is then performed on the high-frequency components using an adaptive soft thresholding method, which dynamically adjusts the threshold based on the noise level.

[0015]

[0016] In the formula, W f (a,b) represent the signal coefficients after wavelet transform; f(t) represents the original sensor signal at time t; ψ(t) represents the mother wavelet function; a represents the scale parameter; b represents the translation parameter; Represents the wavelet coefficients after denoising; λ represents the adaptive threshold; σ represents the noise standard deviation; N represents the data length; sign() represents the sign function;

[0017] S22. Decompose the signal f(t) into intrinsic mode functions, and dynamically remove high-frequency noise and low-frequency trend noise based on spectrum analysis and IMF energy threshold screening:

[0018] EMD decomposition:

[0019] Calculate the energy ratio of the IMF components:

[0020] Set energy threshold E th If Ei <E th If it is a noise component, it will be removed.

[0021] In the formula, IMF i (t) represents the i-th intrinsic mode component; r n (t) represents the residual component; n represents the total number of intrinsic mode components;

[0022] S23. Introduce Kalman filtering, model the relationship between signal and noise through state equations, predict the state and covariance, calculate and update the Kalman gain, realize signal smoothing and anomaly prediction, and optimize the denoising effect.

[0023] The preferred optimization process for improving noise reduction is as follows:

[0024] S31. State-space modeling: Define the state equation as: X t =Ax t-1 +BU t +W t Z t =HX t +V t ;

[0025] In the formula, X t Z represents the actual signal at the current time t; t A represents the measured signal; B represents the state transition matrix; and U represents the measured signal. t W represents the control input matrix and control input, respectively; t V t Represents process noise and measurement noise; H represents the observation matrix;

[0026] S32. Predicting State Estimation:

[0027] Predicting the covariance matrix:

[0028] in, Indicates the predicted state at the current moment; This represents the optimal state estimate at the previous moment; P represents the prediction error covariance matrix;t-1 Let represent the error covariance matrix at the previous time step; Q represents the process noise covariance matrix.

[0029] S33. Calculate the Kalman gain:

[0030] Updated state estimate:

[0031] Update the covariance matrix:

[0032] In the formula, K t R represents the Kalman gain; T represents the measurement noise covariance matrix; T represents the matrix transpose operation. P represents the optimal state estimate at the current moment, i.e., the final denoised signal; t I represents the updated error covariance matrix; I represents the identity matrix.

[0033] S34. Output optimized, denoised data.

[0034] Preferably, the process for generating the anti-loss seal is as follows:

[0035] S41. Calculate the information packet Inp = H(data);

[0036] Where H is a predefined hash function; data is the binary data form of the normalized data;

[0037] S42. Select a random number k∈F p Calculate the basic encapsulation parameter PB = r + k·Inp mod p;

[0038] Where p is a predefined large prime number satisfying p = 3 mod 4; F p Let F be a predefined finite field; r is a predefined random number drawn from the finite field F. p The elements selected from;

[0039] S43. Calculate the derived encapsulation parameter PD = s + k·Inp mod p;

[0040] Where s is a predefined random number drawn from a finite field F p The elements selected from;

[0041] S44. Calculate the comprehensive packaging parameter PS = k·(2r+τ)+k 2 ·Inp+ρk mod p;

[0042] Where τ and ρ are predefined encapsulation factors, τ and ρ ∈ F p ;

[0043] S45. Generate anti-loss seal AS = {basic package parameter PB, derived package parameter PD, comprehensive package parameter PS}.

[0044] Preferably, the inspection process based on the reliability of the anti-lost seal inspection data is as follows:

[0045] S51. Calculate the information packet Inp′ = H(data′);

[0046] Where data′ is the binary data of the received standardized physiological data data;

[0047] S52. Construct the following equation: Eq1 = (PB) 2 +τ·PD; Eq2=U+PS·Inp′ mod p;

[0048] Eq3 = (PD) 2 +ρ·PB; Eq4=V+PS·Inp′ mod p;

[0049] Where U and V are predefined unsealing factors, U = r 2 +τs mod p,V=s 2 +ρr mod p;

[0050] S53. If Eq1 = Eq2 and Eq3 = Eq4, it means that there is no packet loss in the standardized data, which ensures the reliability of the data.

[0051] The preferred detection process of the adaptive hybrid model is as follows:

[0052] S61. Construct a patient relationship graph, treating each patient as a node in the graph, and defining edges as those using the same type of PICC catheter and having similar treatment plans. The graph neural network (GNN) captures the potential relationships between different patients by training the nodes and edges, and learns the feature representation of each patient to identify normal and abnormal patterns.

[0053] S62. Transformer processes input data through a self-attention mechanism, captures long-term dependent features in time series, focuses on the correlation between time steps in PICC catheter data, and identifies long-term trends.

[0054] S63 and LSTM are used to capture short-term time dependence and extract short-term information from short-term fluctuations in the monitoring data of PICC catheters.

[0055] Preferably, the detection process for anomaly distributions by reconstructing errors using a variational autoencoder is as follows:

[0056] S71. Through comparative learning, identify abnormal data in the absence of labels, construct positive and negative samples, and learn how to distinguish between them:

[0057] Positive samples: Selected samples from normal PICC catheter monitoring data;

[0058] Negative samples: Negative samples are constructed through artificial generation and through abnormal catheter monitoring data;

[0059] S72. By using a variational autoencoder (VAE), the distribution of normal PICC catheter data is learned, and the reconstruction error is calculated to determine whether the data is abnormal.

[0060] Preferably, the process for classifying warning levels is as follows:

[0061] S81. Calculate the overall risk score using Bayesian anomaly scoring combined with a weighted time-series anomaly index:

[0062] Bayesian outlier score:

[0063] Weighted Time Series Anomaly Index:

[0064] w t =e -λ(T-t) ;

[0065] In the formula, S BRS Let P(A) represent the Bayesian anomaly score, i.e., the probability of anomaly A occurring given a detection result X; P(A) represents the prior probability of anomaly based on historical data; P(X|A) represents the probability of observing current data X under anomaly conditions; P(X) represents the total probability density of current data X; S WTAI Indicates the weighted time series anomaly index; w t Indicates the weighting coefficient; A t The anomalous fraction is represented at time t; T represents the observation time window; λ represents the attenuation coefficient.

[0066] S82, Based on risk score S BRS Risk classification is performed using an adaptive threshold setting method to determine the warning level:

[0067] Grading strategy: Adaptive dynamic grading is used, based on the mean-variance method of historical outlier data (adjusting alert thresholds).

[0068]

[0069]

[0070] Where μ AS represents the mean of the anomaly scores, reflecting the average level of Bayesian anomaly scores across all samples in the historical data; N represents the number of historical data samples; S BRS,i σ represents the Bayesian outlier score of the i-th sample; A The standard deviation of outlier scores reflects the volatility or dispersion of Bayesian outlier scores in historical data.

[0071] S83. Setting Risk Levels:

[0072] Level 1: Low risk, S BRS <μ A ;

[0073] Level 2: Medium risk, μ A ≤S BRS <μ A +(1+α·S WTAI )σ A ;

[0074] Level 3: High Risk, S BRS ≥μ A +(1+α·S WTAI )σ A ;

[0075] Where α is the weighting factor, and α>0.

[0076] The technical solution of this invention: A method for monitoring and alarming the safety of Picc catheters based on the Internet of Things (IoT), which is used to execute the above-mentioned method for monitoring and alarming the safety of Picc catheters based on the Internet of Things (IoT), comprising:

[0077] Intelligent PICC catheter sensor module for monitoring physiological parameters;

[0078] The edge data processing module is used to preprocess the collected physiological parameter data;

[0079] The cloud-based data analysis module is used to intelligently assess the usage of PICC catheters and identify catheter abnormalities by employing big data analysis and deep learning algorithms.

[0080] The intelligent alarm and notification module is used to predict potential risks and send alarm information through smart terminals;

[0081] The intelligent decision support module is used to provide medical advice by combining historical data with personalized patient information;

[0082] The user interaction module provides a data visualization interface.

[0083] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects:

[0084] This invention designs an IoT-based Picc catheter safety monitoring and alarm system and method. It utilizes flexible microsensors to collect multimodal physiological parameters in real time, combining a multi-level fusion denoising strategy (wavelet transform, EMD decomposition, and Kalman filtering) with a seal-loss prevention mechanism to significantly improve signal quality and ensure data integrity. This effectively solves the problems of high false alarm rate and poor real-time performance caused by noise interference in traditional monitoring systems. In the cloud, an adaptive hybrid model (GNN+Transformer+LSTM) is used to achieve group data association analysis and temporal dependency modeling, combined with contrastive learning and variational autoencoder for unsupervised anomaly detection. The system overcomes the limitations of single-patient data, improving the accuracy of identifying catheter displacement, blockage, and infection. A dynamic risk grading strategy based on Bayesian scoring and weighted time-series indices optimizes medical resource allocation through tiered alerts (low, medium, and high risk) and personalized care recommendations, reducing the risks of over-warning and delayed intervention. Knowledge distillation technology enables lightweight deployment of complex models, supporting real-time inference and low-latency response. Simultaneously, the user interaction module integrates data visualization and decision support, forming a closed-loop management system from monitoring and analysis to intervention. Ultimately, this achieves precise early warning, proactive intervention, and a comprehensive improvement in clinical nursing efficiency for PICC catheter abnormalities. Attached Figure Description

[0085] Figure 1 This is a system architecture diagram of a Picc catheter safety monitoring and alarm system based on the Internet of Things proposed in this invention;

[0086] Figure 2 This is a flowchart of a method for monitoring and alarming the safety of Picc catheters based on the Internet of Things proposed in this invention. Detailed Implementation

[0087] Example 1, as Figure 1 As shown, the PICC catheter safety monitoring and alarm system based on the Internet of Things proposed in this invention includes: an intelligent PICC catheter sensor module, an edge data processing module, a cloud data analysis module, an intelligent alarm and notification module, an intelligent decision support module, and a user interaction module.

[0088] The intelligent PICC catheter sensor module uses flexible microsensors embedded in the outer wall of the catheter (without affecting catheter operation) to monitor physiological parameters, including but not limited to catheter position, blood flow resistance, temperature, and pH value.

[0089] The edge data processing module preprocesses the collected physiological parameter data, including but not limited to denoising and normalization;

[0090] The cloud-based data analysis module uses big data analysis and deep learning algorithms to intelligently assess the usage of PICC catheters and identify catheter abnormalities, including but not limited to blockage, displacement, and infection.

[0091] The intelligent alarm and notification module uses AI analysis to predict potential risks and sends alarm information through smart terminals.

[0092] The intelligent decision support module combines historical data with personalized patient information to provide medical advice;

[0093] The user interaction module provides a data visualization interface.

[0094] Example 2, as Figure 2 As shown, the IoT-based Picc catheter safety monitoring and alarm method proposed in this invention is applied to the IoT-based Picc catheter safety monitoring and alarm system proposed in Example 1. The specific implementation steps are as follows:

[0095] S1. The intelligent PICC catheter sensor module completes device initialization and reference value calibration through a flexible micro-sensor array (including but not limited to pressure sensors, flow sensors, temperature sensors, and pH sensors).

[0096] The system then enters the implementation phase, where the intelligent PICC catheter sensor module periodically collects PICC catheter environmental data (including but not limited to catheter internal pressure, blood flow rate, temperature, and pH value) to form a physiological parameter data stream.

[0097] Then, the physiological parameter data stream is transmitted to the edge data processing module in real time via the NB-IoT unit;

[0098] S2. The edge data processing module constructs a multi-level fusion denoising strategy, which removes high-frequency noise through wavelet transform, removes trend noise by combining EMD decomposition, and smooths the signal using Kalman filtering. The specific implementation process is as follows:

[0099] S21. Preliminary noise suppression based on wavelet transform: extracting signal features at different scales and dispersing noise into high-frequency components, specifically:

[0100] S2101, Wavelet Decomposition: Wavelet transform is used to decompose the original signal at multiple scales, extracting components of different frequencies.

[0101] In the formula, W f (a,b) represent the signal coefficients after wavelet transform; f(t) represents the original sensor signal at time t; ψ(t) represents the mother wavelet function; a represents the scaling parameter (controlling frequency resolution); b represents the translation parameter (controlling time position);

[0102] S2102, Threshold Denoising: High-frequency components are subjected to threshold denoising processing using an adaptive soft thresholding method, specifically:

[0103]

[0104] In the formula, Represents the wavelet coefficients after denoising; λ represents the adaptive threshold; σ represents the noise standard deviation; N represents the data length; sign() represents the sign function;

[0105] Therefore, an adaptive thresholding method is adopted to dynamically adjust the threshold according to the noise level, thereby improving the adaptability of denoising and reducing information loss.

[0106] S22. Empirical Mode Decomposition (EMD) Denoising: Combining IMF energy threshold screening, noise components are dynamically removed to avoid information loss, while enhancing the fidelity of low-frequency signals. Specifically:

[0107] S2201, EMD decomposition: Decomposes the signal f(t) into a series of intrinsic mode functions (IMFs):

[0108]

[0109] In the formula, IMF i (t) represents the i-th intrinsic mode component, i.e., signal components at different scales; r n (t) represents the residual component, i.e., the trend noise; n represents the total number of intrinsic mode components;

[0110] S2202. Based on spectrum analysis, remove high-frequency noise and low-frequency trend noise, specifically as follows:

[0111] Calculate the energy ratio of the IMF components:

[0112] Set energy threshold E th If Ei <E th If it is a noise component, it will be removed.

[0113] Accordingly: By combining IMF energy threshold screening, noise components are dynamically removed to avoid information loss, while enhancing the fidelity of low-frequency signals.

[0114] S23. Time Series Optimization Based on Kalman Filtering: Wavelet transform and EMD have significantly reduced noise, but slight signal fluctuations may still exist. To further improve signal smoothness and predictability, Kalman filtering is introduced to smooth the signal under low noise conditions and provide outlier prediction capabilities, thereby further improving the denoising effect. Specifically:

[0115] S2301, State-space modeling: Let signal X... t Let Z be the state variable and the measured value. t Due to the influence of noise, the state equation is: X t =Ax t-1 +BUt +W t Z t =HX t +V t ;

[0116] In the formula, X t Z represents the actual signal at the current time t; t A represents the measured signal (containing noise); B represents the state transition matrix (set as the identity matrix); C and D represent the measured signal (containing noise); E represents the measured signal (containing noise); F represents the measured signal (containing noise); U represents ... A represents the state transition matrix (set as the identity matrix); B and U represent the measured signal (containing noise); U represents the measured signal (containing noise); A represents the state transition matrix ( t These represent the control input matrix and the control input (which can be ignored in this embodiment); W t V t H represents process noise and measurement noise; H represents the observation matrix, i.e., the relationship between measurement and state, which is set as the identity matrix H = I in this embodiment;

[0117] S2302, Recursive Filtering Calculation:

[0118] (1) Prediction steps:

[0119] Predictive state estimation:

[0120] in, It represents the predicted state at the current moment, that is, the inference of the actual signal value at the current moment based on the information of the previous moment; This represents the optimal state estimate at the previous moment;

[0121] Predicting the covariance matrix:

[0122] in, P represents the prediction error covariance matrix, i.e., the degree of uncertainty of the predicted signal; t-1 represents the error covariance matrix of the previous time step; Q represents the process noise covariance matrix, which describes the degree of random variation of the signal itself;

[0123] (2) Update steps:

[0124] Calculate the Kalman gain:

[0125] In the formula, K t The Kalman gain represents the weight of the measured and predicted information, which determines the degree of confidence in the measured value during the update; R represents the measurement noise covariance matrix, which is the magnitude of the sensor measurement noise; T represents the matrix transpose operation.

[0126] Updated state estimate:

[0127] In the formula, This represents the optimal state estimate at the current moment, i.e., the final denoised signal;

[0128] Update the covariance matrix:

[0129] In the formula, P t This represents the updated error covariance matrix, i.e., the magnitude of the estimation error at the current time step; I represents the identity matrix, ensuring dimension matching.

[0130] Therefore, Kalman filtering can smooth signals under low noise conditions and provide outlier prediction capabilities, further improving the denoising effect.

[0131] S24. Since PICC catheter safety monitoring involves multiple sensor data (including but not limited to catheter internal pressure, blood flow rate, temperature, and pH value), and different data have different dimensions and distributions, normalization processing is required for comprehensive analysis. Therefore, the Z-score normalization method is used to standardize the data of different modalities after noise reduction to obtain standardized physiological data.

[0132] S25. Generate anti-loss seals for standardized data. The process for generating anti-loss seals is as follows:

[0133] S2501, Calculate the information packet Inp = H(data);

[0134] Where H is a predefined hash function; data is the binary data form of the normalized data;

[0135] S2502, Select a random number k∈F p Calculate the basic encapsulation parameter PB = r + k·Inp mod p;

[0136] Where p is a predefined large prime number satisfying p = 3 mod 4; F p Let F be a predefined finite field; r is a predefined random number drawn from the finite field F. p The elements selected from;

[0137] S2503, Calculate the derived package parameter PD = s + k·Inp mod p;

[0138] Where s is a predefined random number drawn from a finite field F p The elements selected from;

[0139] S2504, Calculate the comprehensive packaging parameter PS = k·(2r+τ)+k 2 ·Inp+ρk mod p;

[0140] Where τ and ρ are predefined encapsulation factors, τ and ρ ∈ F p ;

[0141] S2505, Generate anti-lost seal AS = {basic package parameter PB, derived package parameter PD, comprehensive package parameter PS};

[0142] S26. Transmit {Anti-loss seal AS = {Basic encapsulation parameter PB, Derivative encapsulation parameter PD, Comprehensive encapsulation parameter PS}, and standardized physiological data} to the cloud data analysis module.

[0143] S3, the cloud-based data analysis module uses an adaptive hybrid model (GNN+Transformer+LSTM), combining self-supervised learning and knowledge distillation. Through multi-layer feature fusion of deep learning, it achieves accurate analysis, anomaly identification, and risk assessment of PICC catheter monitoring data. The specific implementation process is as follows:

[0144] S31. Receive {Anti-loss seal AS = {Basic encapsulation parameter PB, Derived encapsulation parameter PD, Integrated encapsulation parameter PS}, standardized physiological data}, extract the anti-loss seal AS = {Basic encapsulation parameter PB, Derived encapsulation parameter PD, Integrated encapsulation parameter PS} and the standardized physiological data data, and perform packet loss detection on the standardized physiological data data. The detection process is as follows:

[0145] S3101, Calculate the information packet Inp′ = H(data′);

[0146] Where data′ is the binary data of the received standardized physiological data data;

[0147] S3102. Construct the following equation: Eq1=(PB) 2 +τ·PD; Eq2=U+PS·Inp′mod p; Eq3=(PD) 2 +ρ·PB; Eq4=V+PS·Inp′ mod p;

[0148] Where U and V are predefined unsealing factors, U = r 2 +τs mod p,V=s 2 +ρr mod p;

[0149] S3103. If Eq1 = Eq2 and Eq3 = Eq4, it means that no packet loss occurred during the transmission of standardized physiological data from the edge data processing module to the cloud data analysis module, thus ensuring the reliability of the data; otherwise, an alarm will be issued immediately.

[0150] S32. Construct an adaptive hybrid model by integrating graph neural networks (GNN) and temporal modeling methods (Transformer+LSTM) to achieve a global perspective analysis of single PICC catheter monitoring data and group data (including but not limited to monitoring data of other patients in the hospital). By combining PICC catheter sensor data and data from other patient groups in the hospital, the system's anomaly detection capability is improved. Specifically:

[0151] S3201. Construct a patient relationship diagram:

[0152] (1) Graph structure construction: Each patient is treated as a node in the graph. The node features include, but are not limited to, the monitoring data of the PICC catheter sensor and the patient's basic clinical information (including but not limited to treatment history and disease progression);

[0153] (2) Definition of edge: If two patients use the same type of PICC catheter and have similar treatment plans, there is an edge connecting them;

[0154] (3) The role of GNN: GNN captures the potential relationships between different patients by training nodes and edges, and learns the feature representation of each patient to identify normal and abnormal patterns at the group level.

[0155] S3202, Temporal Modeling: Using Transformer and LSTM to capture the long-term and short-term dependencies of time series data, specifically:

[0156] Transformer: Transformer processes input data through a self-attention mechanism, captures long-term dependent features in time series, focuses on the correlation between time steps in PICC catheter data, and identifies long-term trends;

[0157] LSTM: LSTM is used to capture short-term time dependence, effectively extracting short-term information from the short-term fluctuations in the monitoring data of PICC catheters, and strengthening the continuous modeling between local time steps;

[0158] Accordingly: GNN first constructs a patient relationship graph to provide global information at the group level. After receiving this information, Transformer and LSTM model the time series data of PICC catheters, respectively capturing long-term and short-term dependencies, improving the model's recognition ability, generating structured association features between patients, and obtaining potential abnormal trends predicted by the time series model.

[0159] S33. Construct an anomaly detection model, and perform unsupervised anomaly detection on the data through contrastive learning and variational autoencoder (VAE), learning the anomaly distribution, specifically as follows:

[0160] S3301. Through contrastive learning, abnormal data is identified in the absence of labels. Positive and negative samples are constructed, and the model learns how to distinguish between them.

[0161] Positive samples: Selected samples from normal PICC catheter monitoring data;

[0162] Negative samples: Negative samples are constructed through artificial generation (including but not limited to adding noise) or through abnormal catheter monitoring data;

[0163] S3302, Variational Autoencoder (VAE) for reconstruction error detection: Through VAE, the distribution of normal PICC catheter data is learned, and the reconstruction error is calculated to determine whether the data is abnormal;

[0164] It should be noted that the working principle of VAE is as follows: VAE compresses the input data into the latent space through the encoder, and then reconstructs it back to the original data through the decoder. If the reconstruction error is large, it means that the distribution of the input data and the training data is significantly different, which may be abnormal data.

[0165] Therefore: both contrastive learning and VAE are used for unsupervised learning. The former learns similarity by constructing positive and negative samples, while the latter determines anomalies by reconstructing errors, thus enhancing the model's anomaly detection capability and reducing its dependence on labeled data.

[0166] S34. The anomaly detection model constructed in step S33 is lightweighted through knowledge distillation, thereby enabling real-time inference, specifically as follows:

[0167] S3401, Cloud Model Training: Construct an anomaly detection model in the cloud and train it to detect anomaly patterns in PICC conduit data;

[0168] S3402, Knowledge Distillation: Using the knowledge distillation method, complex cloud models are "compressed" into lighter models, and the student model (lightweight model) is trained through the output of the teacher model (cloud model);

[0169] Distillation process: The output of the teacher model is provided as soft targets to the student model to help the student model learn and reason efficiently on edge devices;

[0170] S3403. Deploy the distilled student model (lightweight model). The lightweight model performs the inference task and outputs the detection result {the probability P(X|A) of observing the current data X under anomalous conditions, and the anomalous score A at time t}. t};

[0171] S35. The probability P(X|A) of observing the current data X under abnormal conditions in the detection result {the abnormality score A at time t}. tThe data is transmitted to the intelligent alarm and notification module.

[0172] S4. The intelligent alarm and notification module constructs an intelligent alarm and risk-level early warning method to achieve accurate early warning of abnormal PICC catheter states, promptly notify medical staff and patients, and provide reasonable risk level assessments, specifically:

[0173] S41. Calculate the overall risk score using the Bayesian Risk Score (BRS) combined with the Weighted Temporal Anomaly Index (WTAI), specifically as follows:

[0174] Bayesian outlier score:

[0175] Weighted Time Series Anomaly Index:

[0176] w t =e -λ(T-t) ;

[0177] In the formula, S BRS Let P(A) represent the Bayesian anomaly score, i.e., the probability of anomaly A occurring given a detection result X; P(A) represents the prior probability of anomaly based on historical data; P(X|A) represents the probability of observing current data X under anomaly conditions; P(X) represents the total probability density of current data X; S WTAI This represents the weighted time-series anomaly index, i.e., the degree of cumulative anomalies in recent data; w t This represents the weighting coefficient; the closer to the current time point, the higher the weight. A t λ represents the outlier score at time t; T represents the observation time window; λ represents the decay coefficient, which determines that the data further away from the current time point T has a smaller impact on the total score.

[0178] S42. Risk score S calculated based on step S41 BRS Risk classification is performed using an adaptive threshold setting method to determine the warning level, specifically as follows:

[0179] Tiering Strategy: Adaptive Dynamic Risk Stratification (ARS) is used, adjusting alert thresholds based on the mean-variance method of historical anomaly data.

[0180]

[0181]

[0182] Where μ AS represents the mean of the anomaly scores, reflecting the average level of Bayesian anomaly scores across all samples in the historical data; N represents the number of historical data samples; S BRS,i σ represents the Bayesian outlier score of the i-th sample; A The standard deviation of outlier scores reflects the volatility or dispersion of Bayesian outlier scores in historical data.

[0183] Risk level setting:

[0184] Level 1: Low risk, S BRS <μ A ;

[0185] Level 2: Medium risk, μ A ≤S BRS <μ A +(1+α·S WTAI )σ A ;

[0186] Level 3: High Risk, S BRS ≥μ A +(1+α·S WTAI )σ A ;

[0187] Here, α is a weighting factor, and α>0 means that when there are many historical anomalies, the system will increase the sensitivity of risk alerts.

[0188] S43. Based on the risk level, a graded alert strategy is adopted to ensure timely and accurate notification without affecting the patient's normal life. The alert strategy graded table is shown in Table 1.

[0189] Table 1 Alarm Strategy Classification Table

[0190]

[0191] S44. Transmit the risk level and corresponding alarm strategy to the intelligent decision support module;

[0192] S5, the intelligent decision support module, after obtaining the risk level, makes a decision based on the patient's current health status and provides targeted treatment suggestions, specifically:

[0193] S51. Based on the risk classification results in step S4, each monitored object will be assigned a risk level L (low risk, medium risk, high risk) and corresponding specific operational recommendations. L∈{1,2,3} is defined to represent {low, medium, and high risk} respectively.

[0194] S52. Based on risk level L, provide personalized maintenance recommendations:

[0195] (1) Low risk (L=1): Normal use, regular maintenance:

[0196] Recommendation: Follow the routine care plan and check the catheter status daily, including for any leakage, signs of infection, etc.

[0197] Medical staff remind patients to pay attention to the catheter's fixation to prevent it from dislodging or becoming twisted.

[0198] Monitoring recommendations: Maintain normal monitoring frequency, collect data once every 24 hours, and record the condition of the skin around the catheter;

[0199] (2) Medium risk (L=2): Increase monitoring frequency and provide focused care:

[0200] Potential risks include, but are not limited to, loosening of catheter fixation, local redness and swelling, slight leakage, and tendency to become blocked.

[0201] Recommendation: Increase the monitoring frequency to every 6 to 12 hours, focusing on observing whether there is further redness, swelling, or increased pain around the catheter.

[0202] Additional nursing interventions include, but are not limited to: re-secure the catheter to ensure stability; change dressings and perform aseptic disinfection when there is slight leakage; and adjust the patient's arm movements appropriately to avoid excessive pulling on the catheter.

[0203] (3) High risk (L=3): Immediate intervention and emergency treatment:

[0204] Potential risks include, but are not limited to, severe effusion, extensive redness and swelling, suspected infection, catheter blockage, and catheter displacement.

[0205] Management measures include, but are not limited to: immediately notifying healthcare professionals and assessing whether catheter removal or replacement is necessary; using thrombolytic agents to treat catheter blockage; if signs of infection appear (including but not limited to local redness, swelling, and fever), sending the sample for culture and administering antibiotics; monitoring the patient's vital signs and preventing serious complications.

[0206] S53. Generate a personalized nursing optimization plan, specifically as follows:

[0207] Based on the patient's previous nursing records, the nursing plan is adjusted individually (including but not limited to the early intervention of nursing plan for high-risk patients);

[0208] Adjust the frequency of monitoring data collection; for high-risk patients, the data collection interval can be shortened.

[0209] Intelligent push nursing instructions remind nurses when to perform specific nursing procedures, including but not limited to catheter flushing and dressing changes;

[0210] S54. Automatically record all nursing operations and generate nursing logs for doctors and nurses to review. Based on historical data, optimize risk grading strategies, combine machine learning to summarize the handling methods of similar cases, and push the optimal nursing plan.

[0211] S6. The user interaction module visually displays standardized physiological data, test results, risk levels, corresponding alarm strategies, personalized maintenance suggestions, and optimal care plans.

[0212] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for safety monitoring and alarm of Picc catheters based on the Internet of Things, characterized in that, The specific implementation steps include the following: S1. Periodically collect Picc catheter environment data to generate physiological parameter data stream; S2. Based on the multi-level fusion denoising method, the original signal is decomposed into multiple scales and denoised by adaptive threshold through wavelet transform and empirical mode decomposition. Combined with Kalman filtering, the smoothness of the time series signal is optimized. Z-score normalization is used to generate standardized data, and anti-lost seals are generated based on hash function and finite field operation. The implementation process of the multi-level fusion denoising method is as follows: S21. The original signal is decomposed into multiple scales using wavelet transform to extract different frequency components. Threshold denoising is then performed on the high-frequency components using an adaptive soft thresholding method, which dynamically adjusts the threshold based on the noise level. ； ； ； In the formula, W f (a,b) represent the signal coefficients after wavelet transform; f(t) represents the original sensor signal at time t; denoted by ; a represents the mother wavelet function; represents the scaling parameter; b represents the translation parameter; The denoised wavelet coefficients are represented by λ; the adaptive threshold is represented by σ; the noise standard deviation is represented by N; and the data length is represented by N. Represents a symbolic function; S22. Decompose the signal f(t) into intrinsic mode functions, and dynamically remove high-frequency noise and low-frequency trend noise based on spectrum analysis and IMF energy threshold screening: EMD decomposition: ; Calculate the energy ratio of the IMF components: ; Set energy threshold E th If Ei <E th If it is a noise component, it will be removed. In the formula, IMF i (t) represents the i-th intrinsic mode component; r n (t) represents the residual component; n represents the total number of intrinsic mode components; S23. Introduce Kalman filtering, model the relationship between signal and noise through state equations, predict the state and covariance, calculate and update the Kalman gain, realize signal smoothing and anomaly prediction, and optimize the denoising effect. S3. Based on the reliability of the anti-lost seal test data, an adaptive hybrid model is constructed, integrating graph neural network and Transformer+LSTM model. Combined with patient relationship graph, group association analysis and long and short-term time series feature capture are performed. Comparative learning is used to construct positive and negative samples to distinguish abnormalities. Error detection abnormality distribution is reconstructed through variational autoencoder. Knowledge distillation is used to deploy the complex model in a lightweight manner and output the detection results. S4. Calculate the comprehensive risk value by combining the weighted time series anomaly index, dynamically adjust the risk threshold, determine the warning level, divide the risk into low, medium and high levels, and adopt a graded alarm strategy. S5. Based on the risk grading results, patients are divided into low, medium and high risk levels, and differentiated nursing measures are formulated according to the level. The monitoring frequency and nursing plan are dynamically adjusted based on the patient's historical data, intelligent operation reminders are pushed, nursing logs are automatically generated and risk strategies are optimized. S6. Visualizes standardized physiological data, test results, risk levels, corresponding alarm strategies, personalized maintenance recommendations, and optimal care plans.

2. The method for monitoring and alarming the safety of a Picc catheter based on the Internet of Things according to claim 1, characterized in that, The optimization process for improving the noise reduction effect is as follows: S31. State-space modeling: Define the state equation as: X t =Ax t-1 +BU t +W t Z t =HX t +V t ; In the formula, X t Z represents the actual signal at the current time t; t A represents the measured signal; B represents the state transition matrix; and U represents the measured signal. t W represents the control input matrix and control input, respectively; t V t Represents process noise and measurement noise; H represents the observation matrix; S32. Predicting State Estimation: ; Predicting the covariance matrix: ; in, Indicates the predicted state at the current moment; This represents the optimal state estimate at the previous moment; P represents the prediction error covariance matrix; t-1 Let represent the error covariance matrix at the previous time step; Q represents the process noise covariance matrix. S33. Calculate the Kalman gain: ; Updated state estimate: ; Update the covariance matrix: ; In the formula, K t R represents the Kalman gain; T represents the measurement noise covariance matrix; T represents the matrix transpose operation. P represents the optimal state estimate at the current moment, i.e., the final denoised signal; t I represents the updated error covariance matrix; I represents the identity matrix. S34. Output optimized, denoised data. .

3. The method for monitoring and alarming the safety of a Picc catheter based on the Internet of Things according to claim 1, characterized in that, The process of generating the anti-loss seal is as follows: S41. Calculate the information packet Inp = H(data); Where H is a predefined hash function; data is the binary data form of the normalized data; S42. Select a random number k∈F p Calculate the basic package parameter PB = r + k × Inp mod p; Where p is a predefined large prime number satisfying p = 3 mod 4; F p Let F be a predefined finite field; r is a predefined random number drawn from the finite field F. p The elements selected from; S43. Calculate the derived package parameter PD = s + k × Inp mod p; Where s is a predefined random number drawn from a finite field F p The elements selected from; S44. Calculate the integrated packaging parameters PS = k × (2r + τ) + k 2 ×Inp+ρk mod p; Where τ and ρ are predefined encapsulation factors, τ and ρ ∈ F p ; S45. Generate anti-loss seal AS={basic packaging parameter PB, derived packaging parameter PD, comprehensive packaging parameter PS}.

4. The method for monitoring and alarming the safety of a Picc catheter based on the Internet of Things according to claim 3, characterized in that, The verification process based on the reliability of the anti-loosening seal inspection data is as follows: S51, Computational Information Packet ; in, The received standardized physiological data is in binary format; S52. Construct the following equation: Eq1 = (PB) 2 +τ×PD;Eq2=U+PS× mod p; Eq3 = (PD) 2 +ρ×PB;Eq4=V+PS× mod p; Where U and V are predefined unsealing factors, U=r 2 +τs mod p,V=s 2 +ρr mod p; S53. If Eq1=Eq2 and Eq3=Eq4, it means that there is no packet loss in the standardized data, which ensures the reliability of the data.

5. The method for monitoring and alarming the safety of a Picc catheter based on the Internet of Things according to claim 1, characterized in that, The detection process of the adaptive hybrid model is as follows: S61. Construct a patient relationship graph, treating each patient as a node in the graph, and defining edges as those using the same type of PICC catheter and having similar treatment plans. The graph neural network (GNN) captures the potential relationships between different patients by training the nodes and edges, and learns the feature representation of each patient to identify normal and abnormal patterns. S62. Transformer processes input data through a self-attention mechanism, captures long-term dependent features in time series, focuses on the correlation between time steps in PICC catheter data, and identifies long-term trends. S63 and LSTM are used to capture short-term time dependence and extract short-term information from the monitoring data of PICC catheters that have short-term fluctuations.

6. The method for monitoring and alarming the safety of a Picc catheter based on the Internet of Things according to claim 1, characterized in that, The detection process of anomaly distribution by reconstructing error using a variational autoencoder is as follows: S71. Through comparative learning, identify abnormal data in the absence of labels, construct positive and negative samples, and learn how to distinguish between them: Positive samples: Selected samples from normal PICC catheter monitoring data; Negative samples: Negative samples are constructed through artificial generation and through abnormal catheter monitoring data; S72. By using a variational autoencoder (VAE), the distribution of normal PICC catheter data is learned, and the reconstruction error is calculated to determine whether the data is abnormal.

7. A method for monitoring and alarming the safety of a Picc catheter based on the Internet of Things according to claim 1, characterized in that, The process for classifying warning levels is as follows: S81. Calculate the overall risk score using Bayesian anomaly scoring combined with a weighted time-series anomaly index: Bayesian outlier score: ; Weighted Time Series Anomaly Index: ; ； In the formula, S BRS Let P(A) represent the Bayesian anomaly score, i.e., the probability of anomaly A occurring given a detection result X; P(A) represents the prior probability of anomaly based on historical data; P(X|A) represents the probability of observing current data X under anomaly conditions; P(X) represents the total probability density of current data X; S WTAI Indicates the weighted time series anomaly index; w t Indicates the weighting coefficient; A t This represents the anomaly score at time t; T represents the observation time window; λ represents the attenuation coefficient. S82, Based on risk score S BRS Risk classification is performed using an adaptive threshold setting method to determine the warning level: Tiering Strategy: Adaptive dynamic tiering is used, adjusting alert thresholds based on the mean-variance method of historical outlier data. ； ； In the formula, μ A S represents the mean of the anomaly scores, reflecting the average level of Bayesian anomaly scores across all samples in the historical data; N represents the number of historical data samples; S BRS,i σ represents the Bayesian outlier score of the i-th sample; A The standard deviation of outlier scores reflects the volatility or dispersion of Bayesian outlier scores in historical data. S83. Setting Risk Levels: Level 1: Low risk, S BRS <μ A ; Level 2: Medium risk, μ A ≤S BRS <μ A +(1+α×S WTAI )σ A ; Level 3: High Risk, S BRS ≥μ A +(1+α×S WTAI )σ A ; Where α is the weighting factor, and α>0.

8. An Internet of Things (IoT) based Picc catheter safety monitoring and alarm system, used to execute the IoT-based Picc catheter safety monitoring and alarm method according to any one of claims 1 to 7, characterized in that, include: Intelligent PICC catheter sensor module for monitoring physiological parameters; The edge data processing module is used to preprocess the collected physiological parameter data; The cloud-based data analysis module is used to intelligently assess the usage of PICC catheters and identify catheter abnormalities by employing big data analysis and deep learning algorithms. The intelligent alarm and notification module is used to predict potential risks and send alarm information through smart terminals; The intelligent decision support module is used to provide medical advice by combining historical data with personalized patient information; The user interaction module provides a data visualization interface.

Images