Signal data cleaning method and system based on ICU monitoring

Abstract

The application relates to the technical field of ICU signal processing, and discloses a signal data cleaning method and system based on ICU monitoring. The method comprises the following steps: receiving original data streams of different vital sign monitoring devices through a multi-channel physiological signal acquisition interface, and converting the original data streams into standardized data with unified time stamps, sampling frequencies and amplitude dimensions; evaluating the signal smoothness of the standardized data in real time, triggering multi-dimensional feature extraction when the signal smoothness is lower than a preset threshold, and obtaining time domain, frequency domain and nonlinear feature vectors; inputting the feature vectors into an anomaly detection model, and outputting an anomaly confidence score and a type label; calling a dynamic cleaning strategy according to the anomaly confidence score and the type label, performing interpolation repair, noise suppression or abnormal segment elimination, and generating cleaned data; and finally, performing adaptive reconstruction and storage optimization on the cleaned data, and generating final storage signal data, so that the quality and utilization efficiency of ICU monitoring data can be improved, and the requirements of clinical diagnosis and treatment and medical research can be met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ICU signal processing, in particular to a signal data cleaning method and system based on ICU monitoring. BACKGROUND

[0002] In the intensive care unit (ICU) clinical diagnosis and treatment scene, vital sign monitoring data is an important basis for medical staff to judge the patient's condition change and formulate diagnosis and treatment plan, and its data quality is directly related to the accuracy of diagnosis and treatment decision and the patient's treatment effect. The types of devices used to collect vital signs in the current ICU are diverse, including electrocardiogram monitors, blood pressure monitors, blood oxygen saturation monitors, etc. The signal acquisition principles and data transmission protocols of different devices are different, resulting in significant inconsistency in the timestamp format, sampling frequency and amplitude dimension of the collected raw physiological signal data stream. For example, some electrocardiogram monitors use millisecond-level timestamps to record data, while some blood pressure monitors use second-level timestamps to store information; the sampling frequency of electrocardiogram signals may reach 500Hz, while the sampling frequency of respiratory signals is only 10Hz; at the same time, different devices may use different units to annotate the amplitude dimension of the same physiological parameter (such as blood pressure), some in mmHg, and some in kPa. This diversity of data format brings great obstacles to subsequent signal analysis and clinical application.

[0003] The ICU environment is complex, and factors such as electromagnetic interference generated by various medical devices in operation, poor electrode contact caused by patient body movement, line loss in signal transmission process, etc. will all cause a large amount of noise to be mixed into the collected physiological signals, and even abnormal situations such as signal mutation and loss may occur. The existing data cleaning methods mostly use single-dimensional processing methods, such as noise suppression through time domain filtering only, or simply removing signal loss segments, which are difficult to comprehensively deal with the complex and diverse quality problems in ICU physiological signals. At the same time, these methods often lack real-time evaluation mechanism of signal quality, and cannot dynamically adjust the cleaning strategy according to the actual quality status of the signal, resulting in unstable cleaning effect. When there are multiple problems such as noise interference, baseline drift, data loss, etc. in the signal, the traditional method is prone to over-cleaning or incomplete cleaning: over-cleaning may filter out the effective physiological information in the signal, affecting the accuracy of subsequent condition judgment; incomplete cleaning may cause abnormal data to remain, interfering with the data analysis results, and thus may mislead medical staff to make wrong diagnosis and treatment decisions.

[0004] The existing method directly stores the data in the original data format after data cleaning, without considering the characteristics of large amount of ICU monitoring data, and lacks reconstruction and storage optimization processing of the cleaned data. With the extension of the monitoring time, the amount of data will increase exponentially, not only occupying a large amount of storage resources, but also reducing the efficiency of subsequent data retrieval and analysis, which is not conducive to the long-term management and secondary use of clinical data. The existence of these problems makes it difficult for ICU monitoring signal data to fully play its value in clinical diagnosis and treatment, disease warning, medical research and other aspects, and an efficient and comprehensive data cleaning method is needed to solve the above problems. SUMMARY

[0005] The purpose of the present application is to provide a signal data cleaning method and system based on ICU monitoring, to solve the problems raised in the background art.

[0006] To achieve the above purpose, the present application provides a signal data cleaning method based on ICU monitoring, which comprises:

[0007] Through a multi-channel physiological signal acquisition interface, the original physiological signal data stream from different vital sign monitoring devices is received;

[0008] The original physiological signal data stream is converted into standardized physiological signal data with unified timestamp, unified sampling frequency and unified amplitude dimension;

[0009] The standardized physiological signal data is subjected to real-time signal smoothness evaluation to generate a quality evaluation result;

[0010] When the quality evaluation result indicates that the signal smoothness is lower than a preset threshold, a multi-dimensional feature extraction operation is triggered;

[0011] The multi-dimensional feature extraction operation is performed to extract time domain feature vectors, frequency domain feature vectors and nonlinear feature vectors from the standardized physiological signal data;

[0012] The time domain feature vectors, frequency domain feature vectors and nonlinear feature vectors are input into an anomaly detection model to output an anomaly confidence score and an anomaly type label;

[0013] Based on the anomaly confidence score and the anomaly type label, a corresponding dynamic cleaning strategy is called;

[0014] According to the dynamic cleaning strategy, interpolation repair, noise suppression or abnormal segment rejection operation is performed on the standardized physiological signal data to generate cleaned physiological signal data;

[0015] The cleaned physiological signal data is subjected to adaptive data reconstruction and storage optimization processing to generate final storage signal data.

[0016] Preferably, the real-time signal stationarity evaluation on the standardized physiological signal data is performed to generate a quality evaluation result, specifically including:

[0017] calculating a local variance entropy value of the standardized physiological signal data in a sliding time window;

[0018] detecting a gradient amplitude of the local variance entropy value over time;

[0019] when the gradient amplitude continuously exceeds a dynamic fluctuation threshold, determining that the signal stationarity is below a preset threshold and generating the quality evaluation result.

[0020] Preferably, the multi-dimensional feature extraction operation is performed, specifically including:

[0021] extracting a signal zero-crossing rate and a peak-valley interval sequence from the standardized physiological signal data as the time-domain feature vector;

[0022] generating a frequency domain energy distribution matrix as the frequency domain feature vector through spectral entropy analysis;

[0023] calculating a multi-scale sample entropy value to form the nonlinear feature vector.

[0024] Preferably, the time-domain feature vector, the frequency domain feature vector, and the nonlinear feature vector are input into an anomaly detection model, specifically including:

[0025] performing feature fusion on the time-domain feature vector, the frequency domain feature vector, and the nonlinear feature vector to generate a fusion feature tensor;

[0026] performing spatial mapping on the fusion feature tensor through a multi-layer perception network to generate a high-dimensional feature representation;

[0027] based on a similarity comparison between the high-dimensional feature representation and a preset anomaly pattern library, outputting the anomaly confidence score and the anomaly type label.

[0028] Preferably, based on the anomaly confidence score and the anomaly type label, a corresponding dynamic cleaning strategy is called, specifically including:

[0029] when the anomaly type label is transient interference and the anomaly confidence score is below a first threshold, activating a noise suppression strategy;

[0030] when the anomaly type label is data missing and the anomaly confidence score is above a second threshold, activating an interpolation repair strategy;

[0031] when the anomaly type label is persistent distortion, activating an abnormal segment rejection strategy.

[0032] Preferably, the interpolation repair, noise suppression or abnormal segment rejection operation on the standardized physiological signal data according to the dynamic cleaning strategy specifically includes:

[0033] Identify the missing segment boundary in the standardized physiological signal data;

[0034] Generate an optimal fitting curve based on the morphological features of the preceding and subsequent valid data segments;

[0035] Fill the missing segment with a cubic spline interpolation using the optimal fitting curve.

[0036] Preferably, the interpolation repair, noise suppression or abnormal segment rejection operation on the standardized physiological signal data according to the dynamic cleaning strategy specifically includes:

[0037] Locate the continuous abnormal points in the standardized physiological signal data that meet the abnormal type label;

[0038] Calculate the duration of the continuous abnormal points and the statistical difference degree of the adjacent normal segments;

[0039] When the statistical difference degree exceeds the morphological tolerance threshold, mark the continuous abnormal points as the segment to be rejected;

[0040] Replace with the mean sequence of the normal segments before and after the segment to be rejected.

[0041] Preferably, the adaptive data reconstruction and storage optimization processing on the cleaned physiological signal data specifically includes:

[0042] Analyze the local complexity index of the cleaned physiological signal data;

[0043] Dynamically adjust the data compression rate according to the local complexity index;

[0044] Use variable order difference encoding for lossless compression of high complexity segments;

[0045] Perform feature point-based lossy down-sampling on low complexity segments.

[0046] Preferably, after generating the final storage signal data, it further includes:

[0047] Establish an association index between the final storage signal data and the original device metadata;

[0048] Partition storage of the final storage signal data based on time series;

[0049] Generate data version snapshots and operation log records.

[0050] Preferably, the present application also includes an ICU monitoring-based signal data cleaning system, which comprises a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of an ICU monitoring-based signal data cleaning method as described above.

[0051] Compared with the prior art, the present application has the following advantages:

[0052] The raw data of different vital sign monitoring devices is received through a multi-channel physiological signal acquisition interface, and then converted into standardized data with unified time stamps, unified sampling frequencies and unified amplitude dimensions, effectively eliminating the differences in data formats collected by different devices. This standardization process can form a unified data specification for the originally disordered multi-source physiological signals, enabling subsequent signal analysis, quality assessment and cleaning operations to be carried out on a consistent data basis, avoiding analysis errors caused by inconsistent data formats, and also providing the possibility for correlation analysis between different types of physiological signals, allowing medical staff to more conveniently judge the patient's condition by integrating multi-dimensional physiological information.

[0053] In terms of signal quality processing, the method first performs real-time stationarity evaluation on the standardized physiological signals and generates a quality assessment result, and only triggers the multi-dimensional feature extraction operation when the signal stationarity is below a preset threshold. This on-demand triggering mechanism can reduce unnecessary computational consumption and improve overall data processing efficiency. During the multi-dimensional feature extraction process, time-domain, frequency-domain and nonlinear feature vectors are extracted from the standardized data simultaneously. Compared with traditional single-dimensional feature extraction methods, this method can more comprehensively characterize the feature information of the signal, covering the amplitude variation in the time dimension, the energy distribution in the frequency dimension, and the complex dynamic characteristics in the nonlinear dimension, thereby more accurately capturing the abnormal features in the signal and providing more abundant and reliable basis for subsequent anomaly detection.

[0054] After inputting the multi-dimensional feature vector into the anomaly detection model, the abnormal confidence score and the abnormal type label can be output, and the corresponding dynamic cleaning strategy can be called according to these information, realizing accurate processing for different abnormal types. For data missing, interpolation repair operation can be used to supplement the missing data points, and the complete form of the signal is restored as much as possible; for noise interference, noise suppression operation can effectively filter out irrelevant interference signals and retain useful physiological information; for severely distorted abnormal signal segments, abnormal segment rejection operation can avoid the influence of abnormal data on subsequent analysis. This dynamic and targeted cleaning method can flexibly adjust the processing method according to the specific situation of signal abnormalities, avoiding the problems of over-cleaning or incomplete cleaning caused by traditional single cleaning strategy, ensuring that the cleaned data can not only remove invalid interference information, but also maximize the retention of effective physiological signal features.

[0055] After generating the cleaned physiological signal data, the method further performs adaptive data reconstruction and storage optimization processing. Adaptive data reconstruction can organize and integrate the cleaned data according to the actual characteristics of the data and the clinical application requirements, forming a data structure that is easier to analyze and use. Storage optimization processing can reduce the space occupied by data storage and reduce storage resource consumption while ensuring data integrity and availability, improving the efficiency of subsequent data retrieval, query and analysis, facilitating long-term management and clinical secondary use of ICU monitoring data, and enabling the cleaned data to more efficiently serve patient condition monitoring, diagnosis and treatment plan optimization, and medical research and other fields. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 A working principle diagram of the signal data cleaning method based on ICU monitoring according to the present application is shown in the figure.

[0057] Figure 2 A flowchart of multi-dimensional feature extraction operation.

[0058] Figure 3 A flowchart of abnormality detection model input and output.

[0059] Figure 4 A flowchart of interpolation repair operation.

[0060] Figure 5 A flowchart of adaptive data reconstruction and storage optimization processing. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0062] Please refer to Figure 1 The present application provides a signal data cleaning method based on ICU monitoring, which comprises:

[0063] The raw physiological signal data streams from different vital sign monitoring devices are received through a multi-channel physiological signal acquisition interface. The raw physiological signal data streams are converted into standardized physiological signal data with unified time stamps, unified sampling frequencies, and unified amplitude dimensions. Real-time signal stationarity evaluation is performed on the standardized physiological signal data to generate quality evaluation results. When the quality evaluation results indicate that the signal stationarity is below a preset threshold, a multi-dimensional feature extraction operation is triggered. The multi-dimensional feature extraction operation is performed to extract time-domain feature vectors, frequency-domain feature vectors, and nonlinear feature vectors from the standardized physiological signal data. The time-domain feature vectors, the frequency-domain feature vectors, and the nonlinear feature vectors are input into an anomaly detection model to output an anomaly confidence score and an anomaly type label. Based on the anomaly confidence score and the anomaly type label, a corresponding dynamic cleaning strategy is called. According to the dynamic cleaning strategy, interpolation repair, noise suppression, or abnormal segment rejection operations are performed on the standardized physiological signal data to generate cleaned physiological signal data. Adaptive data reconstruction and storage optimization processing is performed on the cleaned physiological signal data to generate final storage signal data.

[0064] Embodiment 1: refer to Figure 2 The raw physiological signal data streams from different vital sign monitoring devices are received through a multi-channel physiological signal acquisition interface. These devices can include electrocardiogram monitors, blood oxygen saturation monitors, arterial blood pressure monitors, etc. Since these devices come from different manufacturers, their output data differs in time stamp alignment, sampling frequency, and amplitude dimension. Therefore, first, for time stamp alignment, the reference and format of each device data stream time stamp need to be identified. For devices with different reference times, a selected universal time standard (such as Coordinated Universal Time UTC) is used as a reference to convert and adjust the time stamps of each device data to this universal time system. For inconsistent time stamp formats, such as some devices using milliseconds and others using seconds, conversion is performed according to a unified time precision. In terms of unified sampling frequency, the target sampling frequency is first determined, which is usually based on clinical application requirements or relevant industry standards. For device data with a sampling frequency higher than the target frequency, a downsampling algorithm is used, such as the commonly used decimation algorithm, to select data points from the original data at a certain ratio to reduce the sampling frequency of the data to the target value. For device data with a sampling frequency lower than the target frequency, an interpolation algorithm such as linear interpolation or spline interpolation is used to insert new data points between existing data points based on their characteristics, thereby increasing the sampling frequency of the data to the target value.

[0065] For amplitude dimension unification, the amplitude dimensions of the output data of each device need to be determined, such as some devices using millivolts and others using microvolts for electrocardiogram signals. According to the standard specifications of physiological signal amplitude dimensions, the amplitude dimensions of all device data are converted to standard dimensions. By multiplying the corresponding conversion coefficients, the different dimension data is unified.

[0066] The heterogeneous data streams are converted into standardized physiological signal data with unified time stamp, sampling frequency and amplitude dimension through the above steps. The time stamp is unified in international standard time format, the sampling frequency is adjusted to 250 Hz according to the clinical monitoring requirements, and the amplitude dimension is converted into standard physical units according to the signal type, such as converting the electrocardiogram into millivolts, the blood oxygen saturation into percentage, and the blood pressure into millimeters of mercury.

[0067] After obtaining the standardized physiological signal data, the system immediately starts the real-time signal stationarity evaluation process. This process uses a sliding time window mechanism, with a window length of 5 seconds to balance real-time and evaluation accuracy. In each sliding window, the local variance entropy value of the signal is calculated. The calculation of the local variance entropy value is based on the variance distribution characteristics of the data points in the window, and the discrete degree of the variance sequence is quantified by the concept of information entropy. Specifically, the system first calculates the instantaneous variance of each sampling point in the window to form a variance sequence, then estimates the probability distribution of the sequence, and finally applies the information entropy formula to obtain the entropy value representing the complexity of signal fluctuation. At the same time, the system continuously monitors the gradient amplitude of the local variance entropy value over time, which reflects the mutation of the signal fluctuation rate. The gradient amplitude is calculated using the first-order difference method, and the random fluctuations are smoothed by moving average filtering. The dynamic fluctuation threshold is self-adaptively adjusted according to the historical signal fluctuation characteristics, and its initial value is set based on the signal type, such as the threshold of electrocardiogram signal is usually higher than that of blood oxygen signal. When the gradient amplitude continuously exceeds the dynamic fluctuation threshold, it indicates that the signal has non-stationary changes, and the system determines that the signal stationarity is below the preset threshold and generates the corresponding quality evaluation result.

[0068] When the quality assessment result indicates insufficient signal stationarity, the system automatically triggers a multi-dimensional feature extraction operation. This operation conducts in-depth analysis of the signal from three dimensions: time domain, frequency domain, and nonlinearity. Time domain feature extraction includes calculating the signal zero-crossing rate and the peak-valley interval sequence. The signal zero-crossing rate is obtained by counting the number of times the signal crosses zero in a unit of time, which reflects the rough estimation of signal frequency and amplitude variation. The peak-valley interval sequence is extracted by first identifying the local maximum and minimum points of the signal waveform, then calculating the time interval between adjacent peak and valley points to form a sequence describing the periodicity of the signal. Frequency domain feature extraction is achieved through spectral entropy analysis, generating a frequency domain energy distribution matrix. This process first performs a fast Fourier transform on the signal to obtain the power spectrum, then calculates the energy proportion of each frequency band, and finally applies the concept of information entropy to quantify the uniformity of the frequency spectrum energy distribution. The higher the spectral entropy value, the more dispersed the frequency spectrum energy distribution, and the more complex the signal. Nonlinear feature extraction calculates the multiscale sample entropy value, which involves multiscale coarse-grained processing of the original signal, i.e., resampling the signal at multiple time scales, and then calculating the sample entropy value at each scale. Sample entropy reflects the probability of new patterns in the signal, and multiscale analysis reveals the characteristics of signal complexity changes with scale. The extraction of these three types of feature vectors provides comprehensive signal representation for subsequent anomaly detection.

[0069] Taking the electrocardiogram signal as an example, when the signal baseline drift occurs due to poor electrode contact, the local variance entropy value will significantly increase, and its gradient amplitude may continue to exceed the dynamic fluctuation threshold. At this time, the system triggers the feature extraction operation, the time domain signal zero-crossing rate may abnormally decrease, and the periodicity of the peak-valley interval sequence is destroyed; the frequency domain spectral entropy value abnormally increases, indicating that the energy distribution becomes dispersed; the multiscale sample entropy value at different scales presents an abnormal change pattern. These characteristic changes collectively indicate a decrease in signal quality, providing a basis for subsequent cleaning decisions. The entire implementation process is fully automated and does not require human intervention, achieving real-time quality assessment and feature extraction of ICU monitoring signals.

[0070] Embodiment 2: see Figure 3 , which involves the specific process of inputting multi-dimensional feature vectors into an anomaly detection model and processing. After completing the time domain, frequency domain, and nonlinear feature extraction of physiological signal data, this method enters the feature fusion and anomaly recognition stage. The time domain feature vector from this embodiment contains the signal zero-crossing rate and the peak-valley interval sequence, the frequency domain feature vector contains the frequency domain energy distribution matrix generated by spectral entropy analysis, and the nonlinear feature vector contains the multiscale sample entropy value. Although these feature vectors describe the signal characteristics from different angles, they need to be integrated into a unified representation form for subsequent analysis.

[0071] The feature fusion process adopts a tensor concatenation operation to connect the three types of feature vectors in the feature dimension. The time-domain feature vector usually contains dozens of feature points, describing the statistical characteristics of the signal in the time dimension; the frequency-domain feature vector exists in the form of a matrix, containing the energy distribution information of multiple frequency bands; the nonlinear feature vector contains a series of entropy values at multiple scales. The system first converts these features into the same dimension vector representation, and then concatenates along the feature axis to form a three-dimensional fusion feature tensor. This tensor structure preserves the integrity of the original features while establishing the correlation between different feature dimensions. During the fusion process, the system normalizes each type of feature to eliminate the influence caused by different dimensions.

[0072] The fusion feature tensor is spatially mapped by a multi-layer perception network. The network contains three fully connected layers, each using a ReLU activation function for nonlinear transformation. The first layer of the network receives the fusion feature tensor as input and maps it to a hidden feature space; the second layer of the network further extracts high-level abstract features; the third layer of the network outputs the final high-dimensional feature representation. This high-dimensional feature representation captures the complex interactions and hidden patterns between the original features, and its dimension is usually much higher than that of the original features, thereby enhancing the expressiveness of the features. During network training, a large number of labeled abnormal signal samples are used for supervised learning, enabling the network to learn the feature representation that distinguishes normal and abnormal signals.

[0073] Based on the similarity comparison between the high-dimensional feature representation and the preset abnormal pattern library, the abnormal confidence score and the abnormal type label are output. The preset abnormal pattern library is established before system deployment and contains feature templates of various typical abnormal patterns, which are derived from historical clinical data analysis and expert knowledge. Each abnormal type in the abnormal pattern library has a corresponding feature signature, such as electrode detachment, motion artifact, device failure, etc. The similarity comparison uses the cosine similarity algorithm to calculate the angle cosine value between the high-dimensional feature representation and each abnormal template. The closer this value is to 1, the higher the similarity. After calculating the similarity with all abnormal templates, the system selects the abnormal pattern with the highest similarity as the candidate type.

[0074] The abnormal confidence score is obtained by normalization processing using the softmax function. The system inputs all similarity values into the softmax function to convert them into a probability distribution form, where the maximum probability value corresponds to the abnormal confidence score. This score is between 0 and 1, and the higher the value, the higher the confidence of the abnormal detection. The abnormal type label is determined according to the abnormal pattern category corresponding to the highest similarity, and the system maintains an abnormal type coding table to map the similarity matching result to a specific abnormal type description.

[0075] Taking the common electrode contact failure in ECG signals as an example, when the feature extraction stage detects abnormal features, the time domain features may show abnormal reduction of zero-crossing rate and disorder of peak-valley interval, the frequency domain features show abnormal dispersion of energy distribution, and the nonlinear features show abnormal multi-scale entropy values. These features form a feature tensor after fusion, which is input into a multi-layer perception network to obtain a high-dimensional feature representation. This representation has a high cosine similarity with the "electrode contact failure" template in the abnormal pattern library, and a high abnormal confidence score is calculated through the softmax function, and finally the abnormal type label of "electrode contact failure" is output. The whole process is completed within milliseconds, meeting the real-time requirements of ICU monitoring.

[0076] The abnormal pattern library of the system supports an online updating mechanism. When a new abnormal pattern appears, it can be added to the pattern library after expert labeling, so that the system has the ability of continuous learning. At the same time, the parameters of the multi-layer perception network can also be fine-tuned according to new data to improve the recognition ability of new types of abnormalities. This design enables the system to adapt to the monitoring needs of different ICU environments and cope with various possible signal abnormal situations. The whole implementation process is fully automated and can complete the whole process from feature fusion to abnormality recognition without human intervention, providing accurate decision basis for subsequent signal cleaning operations.

[0077] Embodiment 3: Decision-making process involving calling dynamic cleaning strategies based on abnormal detection results. After completing abnormal detection and obtaining abnormal confidence score and abnormal type label, appropriate signal cleaning strategies need to be selected according to these output results. The abnormal confidence score is a continuous value between 0 and 1, reflecting the credibility of the current signal segment being judged as abnormal, while the abnormal type label indicates the specific nature of the abnormality, such as transient interference, data loss or persistent distortion, etc.

[0078] The system maintains a strategy decision matrix, which defines the cleaning strategies corresponding to different abnormal types and confidence ranges. In the decision-making process, the system first analyzes the abnormal type label to determine the basic nature of the abnormality, and then makes a comprehensive judgment combined with the abnormal confidence score. For transient interference type of abnormality, its characteristics are usually short duration, large amplitude change but not leading to complete signal loss, which is usually caused by random noise or transient artifacts. When the abnormal type label is transient interference and the abnormal confidence score is lower than the first threshold value, the system activates the noise suppression strategy. The first threshold value is a value set after a large amount of clinical data verification, and its purpose is to distinguish between low confidence transient noise and high confidence significant abnormal events. Confidence scores below this threshold indicate that abnormal detection may be uncertain, so a relatively conservative noise suppression method is used to avoid over-processing of the signal.

[0079] For the data missing type of anomaly, its characteristic is the complete or partial loss of signal segments, which can be caused by sensor falling off, data transmission interruption, etc. When the anomaly type label is data missing and the anomaly confidence score is higher than the second threshold, the system activates the interpolation repair strategy. The second threshold is set significantly higher than the first threshold, and this design is based on clinical considerations: the judgment of data missing requires higher confidence, because interpolation repair is an active modification of the signal, which must be based on a high degree of certainty. The confidence score higher than the second threshold provides this certainty guarantee, ensuring that only in the case of clear confirmation of data missing will the signal reconstruction operation be performed.

[0080] For the persistent distortion type of anomaly, its characteristic is that the signal appears long-term, structural distortion, which can be caused by device failure, electrode aging or patient's violent movement, etc. When the anomaly type label is persistent distortion, no matter how high or low the anomaly confidence score is, the system will activate the abnormal segment rejection strategy. This design is based on the following considerations: persistent distortion often means a fundamental deterioration of signal quality, even if the confidence score is not high, it cannot be risky to retain the signal segment that may contain false information. In this case, the primary task of the system is to ensure the credibility of the stored signal, so the most thorough cleaning method is adopted.

[0081] In the strategy selection process, the system adopts a comprehensive evaluation method based on a weighted decision function. This decision function considers the anomaly type, confidence score and signal context information:

[0082]

[0083] Wherein: represents the strategy selection score, which is used to assist decision-making in boundary cases; is the anomaly type weight, which is set according to the severity of the anomaly type; is the basic score corresponding to the anomaly type; is the confidence weight, which controls the influence of the confidence score in decision-making; is the normalized anomaly confidence score; is the signal context weight; is the context score based on the historical quality of the signal and the clinical importance. This weighted decision function provides additional decision-making basis when there is a conflict between the suggestions given by the anomaly type and the confidence score.

[0084] Taking electrocardiogram signal processing as an example, when a short-term high-frequency noise is detected and classified as transient interference, if the confidence score is 0.6 (lower than the first threshold 0.7), the system selects the noise suppression strategy and uses a digital filtering method to smooth the signal. When there is an obvious data interruption and it is classified as data missing, if the confidence score reaches 0.9 (higher than the second threshold 0.85), the system starts interpolation repair and reconstructs based on the normal signal segments before and after. When persistent baseline drift and morphological distortion of the signal are detected, even if the confidence score is only 0.75, the system will directly mark the abnormal segment for rejection, because such persistent distortion may seriously affect clinical diagnosis.

[0085] The system also sets a priority mechanism for strategy execution. When multiple abnormal types are detected in the same signal segment, the system determines the processing priority according to the severity of the abnormality: persistent distortion is prior to data missing, and data missing is prior to transient interference. This priority setting ensures the timeliness of processing the most serious abnormality. At the same time, all strategy decisions and execution processes are recorded in the operation log, including abnormal type, confidence score, selected cleaning strategy, and execution time, etc. information, providing complete data support for subsequent quality audit and algorithm optimization.

[0086] The whole dynamic cleaning strategy calling process adopts modular design, and each strategy corresponds to an independent execution module. This design enables the system to flexibly extend new cleaning strategies. Strategy decision parameters (such as the first threshold, the second threshold, and various weight coefficients) can be adjusted through configuration files, so that the system can adapt to the specific needs of different ICU environments. Through this fine strategy scheduling mechanism, the system realizes targeted processing of abnormal signals, while ensuring signal quality and maximizing the retention of useful physiological information.

[0087] Embodiment 4: see Figure 4, involves specific procedures for performing signal repair and rejection operations according to dynamic cleaning strategies. After determining the cleaning strategy, the method implements corresponding processing techniques for different types of signal abnormalities. When the interpolation repair strategy is activated, the system first identifies the boundary of the missing segment in the standardized physiological signal data. The missing segment boundary is determined by the start and end timestamps of consecutive invalid data points, and the system scans the signal sequence for zero or null value intervals exceeding the preset duration. Based on the morphological features of the preceding and following valid data segments, an optimal fitting curve is generated, including the mean level, variance fluctuation range, and first derivative trend of the valid data segment. The system analyzes the valid signal segments 2 seconds before and after the missing segment, extracting their statistical features and waveform characteristics. The missing segment is filled with cubic spline interpolation using the optimal fitting curve, ensuring that the repaired curve satisfies the first derivative continuity condition at the boundary points and maintains the natural transition of the signal waveform. The interpolation node spacing is dynamically adjusted according to the signal type, for example, 5ms for ECG signals and 10ms for blood pressure signals.

[0088] When the abnormal segment rejection strategy is activated, the system locates the continuous abnormal points in the standardized physiological signal data that meet the abnormal type label. Continuous abnormal points are defined as a set of adjacent abnormal data points marked by the abnormal detection model. Calculate the duration of continuous abnormal points and the statistical difference of adjacent normal segments, using the KL divergence algorithm to quantify the probability distribution difference between the abnormal segment and the 1-second normal segments before and after it. When the statistical difference exceeds the morphological tolerance threshold, the continuous abnormal points are marked as segments to be rejected. The morphological tolerance threshold is dynamically set according to the signal type, for example, the threshold for ECG signals is set to 0.35, and for blood oxygen signals, it is set to 0.28. Replace the mean sequence of the normal segments before and after the segment to be rejected, using a weighted average method: normal points close to the abnormal segment boundary are given higher weights, and points far from the boundary are given lower weights, maintaining timestamp continuity. The system refers to dynamic parameter configurations when performing cleaning operations, as shown in Table 1.

[0089] Table 1: Dynamic cleaning strategy parameter configuration table.

[0090] Signal type Boundary identification threshold (ms) Morphology feature analysis window (s) Spline node spacing (ms) Statistical difference algorithm Morphology tolerance threshold Replacement segment weight distribution Electrocardiogram signal 120 Front 2.0 / back 2.0 5 KL divergence 0.35 Exponential decay Blood pressure signal 200 Front 1.5 / back 1.5 10 KL divergence 0.40 Linear decrease Blood oxygen signal 150 Front 1.0 / back 1.0 8 KL divergence 0.28 Uniform distribution Respiratory waveform 180 Front 2.5 / back 2.5 15 KL divergence 0.30 Gaussian distribution

[0091] Taking the data missing of arterial blood pressure signal as an example, the system detects a zero-value segment of 150 ms, and confirms the missing segment boundary as t = 123.45 s to t = 123.60 s. The morphological features of the effective segments of 1.5 seconds before (t = 121.95 s-123.45 s) and after (t = 123.60 s-125.10 s) are analyzed: the mean value of the former segment is 85 mmHg, the variance is 3.2, and the rising slope is 0.5 mmHg / ms; the mean value of the latter segment is 87 mmHg, the variance is 2.8, and the falling slope is -0.6 mmHg / ms. The generated optimal fitting curve matches the rising trend of the former segment at t = 123.45 s, matches the falling trend of the latter segment at t = 123.60 s, and fills 15 data points in the middle with cubic spline interpolation. When the blood oxygen saturation signal appears a continuous motion artifact, the system locates the continuous abnormal points at t = 456.70 s-457.20 s, with a duration of 500 ms. The KL divergence values of this segment and the normal segments before and after are calculated: the difference degree with the former segment (t = 456.20 s-456.70 s) is 0.42, and the difference degree with the latter segment (t = 457.20 s-457.70 s) is 0.39. Since the morphological tolerance threshold of the blood oxygen signal is 0.28, the system determines that this segment is a segment to be removed. An exponential decay weight distribution is used: the last 100 ms data points of the former normal segment are assigned a weight of 0.7, and the first 100 ms data points of the latter normal segment are assigned a weight of 0.3, to generate a replacement sequence to cover the abnormal period.

[0092] During the execution of the cleaning operation, the system monitors the processing effect in real time, checks the derivative continuity of the repaired segment and the adjacent signal after interpolation repair; verifies the statistical property consistency of the replacement sequence and the original signal after abnormal segment removal. All operations are completed in memory, the original data remains unchanged, and the processing result generates a new signal copy. The operation log records detailed processing parameters, including boundary position, analysis window range, fitting curve parameters, difference degree calculation results, and the final operation type.

[0093] Example 5: see Figure 5, and the specific process of adaptive reconstruction and storage optimization of the cleaned physiological signal data. After the signal cleaning operation, the processed data needs to be intelligently compressed and stored to balance the relationship between data preservation quality and storage efficiency. The system first analyzes the local complexity index of the cleaned physiological signal data, which is obtained by calculating the approximate entropy value of the signal in the sliding window. The approximate entropy value reflects the probability of new patterns in the signal sequence, and the higher the entropy value, the greater the signal complexity. The system uses an overlapping sliding window mechanism, and the window length is dynamically adjusted according to the signal type. The electrocardiogram uses a 256-sample window, and the blood oxygen signal uses a 128-sample window, with a window overlap rate of 50%. The approximate entropy value calculated for each window constitutes the local complexity profile of the signal, clearly identifying the distribution of smooth and complex segments in the signal.

[0094] The dynamic adjustment of the data compression rate according to the local complexity index is the core link of storage optimization. The system establishes a mapping relationship between the complexity value and the compression rate, with lower compression rates allocated to high complexity segments and higher compression rates allocated to low complexity segments. For electrocardiogram signals, when the local complexity index exceeds 0.8, the compression rate is set to 20%; when the complexity index is less than 0.3, the compression rate can be increased to 60%. This differential compression strategy ensures that important signal features are retained to the maximum extent, while achieving higher storage efficiency in smooth signal segments. The compression rate adjustment process uses a smooth transition mechanism to avoid artificial artifacts caused by sudden changes in compression rate between adjacent data segments.

[0095] Variable-order difference encoding is used for lossless compression of high complexity segments, and the order of difference encoding is automatically selected according to the signal change gradient. The system calculates the second derivative amplitude of the signal in real time, and when the derivative amplitude is large, high-order difference (such as fourth or fifth order) is used, and when the signal changes smoothly, low-order difference (such as first or second order) is used. During encoding, the system first records the initial sample value, then stores the difference sequence instead of the original sample value, which requires less storage space, thus achieving compression effect. For low complexity segments, feature point-based lossy downsampling is performed, retaining only the extreme points and inflection points of the signal waveform. The system locates the feature points by finding local maxima and minima, and also retains key inflection points with curvature changes exceeding a threshold. In the downsampling process, the sampling rate of the smooth segment can be reduced to one-fourth of the original sampling rate, while the linear interpolation segment between feature points is further approximated by a straight line.

[0096] After generating the final storage signal data, an association index is established between the final storage signal data and the original device metadata. The association index adopts a three-layer structure: the first layer records the device identifier and the signal type; the second layer stores the time range information; and the third layer saves the data quality label and the cleaning operation summary. The index data is stored in a lightweight JSON format, facilitating quick retrieval and parsing. The final storage signal data is stored in partitions based on time series, and each time partition corresponds to an independent data file. The partition granularity is flexibly set according to the data volume, and is usually set to one hour or six hours as a storage unit. This partition strategy supports quick data positioning by time range, avoiding the performance overhead caused by full scan.

[0097] Generating data version snapshots and operation log records is an important measure to ensure data traceability. Before each data cleaning operation, the system automatically creates a read-only snapshot of the original data. The snapshot adopts an incremental storage method and only records the differential data from the previous version. The operation log record details the processing pipeline, including the data reading timestamp, the applied cleaning strategy type, the executed specific operation parameters, the processing duration, and the output data characteristics. The log entry adopts a structured key-value pair format, facilitating machine parsing and manual review. The system maintains metadata headers for each data file, including the data acquisition start time, signal type, sampling rate, compression algorithm version, and associated snapshot and log identifiers.

[0098] The entire adaptive data reconstruction and storage optimization process adopts a pipeline architecture, with each processing stage independent and organically linked. The output of the complexity analysis module directly drives the compression strategy selection module, and the output of the compression execution module feeds into the storage management module, forming a complete processing chain. All modules share a unified memory data bus, avoiding unnecessary data copying and conversion overhead. The system monitors the storage space usage in real time during operation, and automatically starts the data archiving program when the remaining space is below the threshold, migrating historical data to secondary storage devices. At the same time, the system provides a data reconstruction interface, supporting complete reconstruction of the original signal waveform based on the stored differential encoding and feature point information, ensuring the data availability for clinical retrospective analysis.

[0099] It should be noted that, in this text, relational terms such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device.

[0100] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A method for cleaning signal data based on ICU monitoring, characterized in that, include: The system receives raw physiological signal data streams from different vital sign monitoring devices via a multi-channel physiological signal acquisition interface. The original physiological signal data stream is converted into standardized physiological signal data with a unified timestamp, a unified sampling frequency, and a unified amplitude dimension. Real-time signal stationarity assessment is performed on the standardized physiological signal data to generate quality assessment results; When the stability of the quality assessment result indicator signal is lower than a preset threshold, a multi-dimensional feature extraction operation is triggered. The multidimensional feature extraction operation is performed to extract time-domain feature vectors, frequency-domain feature vectors, and nonlinear feature vectors from the standardized physiological signal data. The time-domain feature vector, frequency-domain feature vector, and nonlinear feature vector are input into the anomaly detection model, and the anomaly confidence score and anomaly type label are output. Based on the anomaly confidence score and anomaly type label, the corresponding dynamic cleaning strategy is invoked. According to the dynamic cleaning strategy, interpolation repair, noise suppression or abnormal segment removal operations are performed on the standardized physiological signal data to generate cleaned physiological signal data. The physiological signal data after cleaning is subjected to adaptive data reconstruction and storage optimization processing to generate the final stored signal data.

2. The signal data cleaning method based on ICU monitoring according to claim 1, characterized in that, The process of performing real-time signal stationarity assessment on the standardized physiological signal data and generating quality assessment results specifically includes: Calculate the local variance entropy value of the standardized physiological signal data within the sliding time window; Detect the gradient magnitude of the local variance entropy value as a function of time; When the gradient amplitude continuously exceeds the dynamic fluctuation threshold, the signal stability is determined to be lower than the preset threshold, and the quality assessment result is generated.

3. The signal data cleaning method based on ICU monitoring according to claim 2, characterized in that, The execution of the multidimensional feature extraction operation specifically includes: The zero-crossing rate and peak-valley interval sequence of the standardized physiological signal data are extracted as the time-domain feature vector. The frequency domain energy distribution matrix is ​​generated by spectral entropy analysis and used as the frequency domain feature vector. The nonlinear feature vector is formed by calculating the entropy values ​​of the multi-scale samples.

4. The signal data cleaning method based on ICU monitoring according to claim 3, characterized in that, The step of inputting the time-domain feature vector, frequency-domain feature vector, and nonlinear feature vector into the anomaly detection model specifically includes: The time-domain feature vector, frequency-domain feature vector, and nonlinear feature vector are fused to generate a fused feature tensor. The fused feature tensor is spatially mapped using a multilayer perceptron to generate a high-dimensional feature representation. Based on the similarity comparison between the high-dimensional feature representation and the preset anomaly pattern library, the anomaly confidence score and anomaly type label are output.

5. The signal data cleaning method based on ICU monitoring according to claim 4, characterized in that, The step of invoking the corresponding dynamic cleaning strategy based on the anomaly confidence score and anomaly type label specifically includes: When the anomaly type label is transient interference and the anomaly confidence score is lower than the first threshold, the noise suppression strategy is activated. When the anomaly type label is missing data and the anomaly confidence score is higher than the second threshold, the interpolation repair strategy is activated. When the anomaly type label is persistent distortion, the anomaly segment removal strategy is activated.

6. The signal data cleaning method based on ICU monitoring according to claim 5, characterized in that, The step of performing interpolation repair, noise suppression, or outlier removal operations on the standardized physiological signal data according to the dynamic cleaning strategy specifically includes: Identify the boundaries of missing segments in the standardized physiological signal data; Based on the morphological characteristics of the preceding and following effective data segments, the optimal fitting curve is generated; The missing segment is filled using cubic spline interpolation using the optimal fitting curve.

7. The signal data cleaning method based on ICU monitoring according to claim 5, characterized in that, The step of performing interpolation repair, noise suppression, or outlier removal operations on the standardized physiological signal data according to the dynamic cleaning strategy specifically includes: Locate consecutive abnormal points in the standardized physiological signal data that match the abnormality type label; Calculate the statistical difference between the duration of the consecutive abnormal points and the adjacent normal segments; When the statistical difference exceeds the morphological tolerance threshold, the consecutive outliers are marked as segments to be removed. Replace it with the mean sequence of the normal segments before and after the segment to be removed.

8. The signal data cleaning method based on ICU monitoring according to claim 1, characterized in that, The adaptive data reconstruction and storage optimization processing of the cleaned physiological signal data specifically includes: Analyze the local complexity index of the physiological signal data after cleaning; The data compression rate is dynamically adjusted based on the local complexity index. Lossless compression of high-complexity segments is achieved using variable-order differential coding. Perform lossy downsampling based on feature points on low-complexity segments.

9. The signal data cleaning method based on ICU monitoring according to claim 8, characterized in that, After generating the final stored signal data, the process also includes: Establish an association index between the final stored signal data and the original device metadata; The final stored signal data is partitioned and stored based on the time series. Generate data version snapshots and operation logs.

10. A signal data cleaning system based on ICU monitoring, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the signal data cleaning method based on ICU monitoring as described in any one of claims 1 to 9.

Images