System for recognizing and evaluating effectiveness of blood oxygen probe artifact signals based on deep learning

Abstract

This invention discloses a deep learning-based system for identifying and evaluating the effectiveness of pulse oximetry probe artifact signals. Relating to the field of deep learning, the system includes: a pulse oximetry probe signal acquisition module that acquires photoplethysmography (PPG) signals in real time and monitors the link status; combined with the synchronous acquisition and labeling of multi-source auxiliary signals; after denoising and correction by a preprocessing module; and an artifact signal feature extraction module that generates standardized feature vectors; a deep learning artifact identification module that analyzes features using a hybrid neural network model to identify artifact types and intensities; a pulse oximetry signal effectiveness evaluation module that performs a graded evaluation based on artifact information and signal quality parameters; a result output module that provides early warnings through visualization and voice prompts; and a data storage module that uses a distributed architecture to classify and manage data. The advantages of this invention are: accurate artifact identification and grading through deep learning, combined with multi-dimensional pulse oximetry signal effectiveness evaluation, significantly improves the accuracy of pulse oximetry monitoring, providing reliable support for clinical diagnosis and treatment.

Description

Technical Field

[0001] This invention relates to the field of deep learning, and in particular to a deep learning-based system for identifying and evaluating the effectiveness of pulse oximeter artifact signals. Background Technology

[0002] In intensive care, surgical anesthesia, and sleep apnea monitoring, blood oxygen saturation is a key indicator for assessing a patient's oxygenation status. However, factors such as motion artifacts, hypoperfusion, probe detachment, or electromagnetic interference often lead to signal distortion, potentially causing false alarms or missed diagnoses. With the development of artificial intelligence technology, it has become possible to analyze the morphological characteristics of photoplethysmography (PPG) waves through machine learning, such as waveform periodicity, amplitude variability, and spectral characteristics.

[0003] Current blood oxygen monitoring systems on the market generally suffer from significant disadvantages. Most only collect a single PPG signal, failing to simultaneously integrate multi-source auxiliary signals such as motion, ambient light, wearing pressure, and body temperature. Their anti-interference capabilities are weak, easily affected by limb movement, environmental interference, and improper wearing, resulting in artifacts and causing fluctuations and distortions in monitoring data. Artifact identification often employs simple algorithms such as fixed thresholds and traditional filtering, without utilizing deep learning hybrid models, making it impossible to accurately identify artifact types and interference intensity, resulting in insufficient generalization ability and recognition accuracy. Furthermore, they lack a multi-dimensional, standardized blood oxygen signal effectiveness grading and evaluation system, only able to simply determine the presence or absence of a signal, unable to objectively quantify signal usability. In addition, most systems lack a comprehensive real-time warning and device linkage mechanism, resulting in delayed anomaly responses. Their data storage architecture is simplistic, lacking classification management and end-to-end traceability capabilities, and their device adaptability is limited, failing to cover diverse clinical scenarios and providing stable, accurate, and traceable reliable support for clinical blood oxygen monitoring. Summary of the Invention

[0004] To improve the existing system, a deep learning-based system for identifying and evaluating the effectiveness of pulse oximeter probe artifacts is provided. This system identifies and classifies artifacts through deep learning, combines multi-dimensional pulse oximeter signal effectiveness evaluation, multi-form early warning, and full-process data traceability, adapts to multiple clinical scenarios, significantly improves the accuracy and stability of pulse oximeter monitoring, and provides reliable support for clinical diagnosis and treatment.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A deep learning-based system for identifying and evaluating the effectiveness of pulse oximeter probe artifacts includes:

[0007] Pulse Oxygen Probe Signal Acquisition Module: Acquires the raw photoplethysmography (PPG) signal output by the pulse oxygen probe in real time and detects the connectivity of the acquisition link. When an acquisition interruption or signal loss occurs, it generates an acquisition anomaly prompt signal.

[0008] Multi-source auxiliary signal acquisition module: Acquires multi-source auxiliary signals related to blood oxygen monitoring. All auxiliary signals are synchronized and marked with the same timestamp as the original PPG signal.

[0009] Signal preprocessing module: Receives the raw PPG signal and multi-source auxiliary signals, and performs synchronous preprocessing, including signal denoising, baseline correction, signal normalization and outlier removal;

[0010] Artifact signal feature extraction module: Based on the preprocessed signal, extract the time domain features, frequency domain features, and morphological features of the PPG signal, as well as the feature parameters of the multi-source auxiliary signals. After feature filtering and fusion processing, a standardized artifact recognition feature vector is formed.

[0011] Deep learning artifact recognition module: It has a built-in deep learning hybrid model that integrates convolutional neural networks and recurrent neural networks, trained and optimized by multi-scene data. It identifies artifact types by learning the mapping relationship between artifact features and artifact types, and classifies the interference intensity.

[0012] Blood oxygen signal effectiveness assessment module: Receives artifact recognition results and preprocessed PPG signals, and combines artifact type, artifact intensity and quality parameters of the PPG signal itself to establish a multi-dimensional effectiveness assessment system to grade and evaluate blood oxygen signals.

[0013] Results output and early warning module: Receives artifact recognition results, blood oxygen signal effectiveness assessment results and abnormal acquisition prompt signals, and provides early warnings in multiple forms, including visual display, voice prompts and signal output;

[0014] Data storage and traceability module: Stores all data during system operation. It adopts a distributed storage architecture and classifies and manages the stored data according to time, patient information, and monitoring scenario, supporting data traceability and review.

[0015] Preferably, the pulse oximeter signal acquisition module specifically includes:

[0016] Probe adapter unit: Establishes connection with various models and specifications of pulse oximeter probes, adapting to pulse oximeter monitoring equipment in different clinical scenarios;

[0017] Signal acquisition unit: Real-time acquisition of raw photoplethysmography (PPG) signals output by the pulse oximeter, with a focus on capturing signals in the red and infrared bands;

[0018] Acquisition rate adjustment unit: Based on clinical monitoring needs, the acquisition rate can be adjusted within a preset range to adapt to the monitoring needs of different patients;

[0019] Acquisition Status Monitoring Unit: Real-time detection of acquisition link connectivity. When acquisition is interrupted or signal is lost, an acquisition anomaly prompt signal is generated and quickly transmitted to the result output and early warning module.

[0020] Preferably, the multi-source auxiliary signal acquisition module specifically includes:

[0021] Multi-dimensional signal acquisition unit: synchronously acquires various auxiliary signals, captures human limb movement signals through a three-axis accelerometer, acquires ambient light interference signals through a light sensor, detects probe wearing pressure signals through a pressure sensor, and simultaneously acquires human body temperature signals;

[0022] Signal synchronization unit: Using the same timestamp as the original PPG signal of the blood oxygen probe, all acquired auxiliary signals are synchronized and aligned with the PPG signal in time;

[0023] Data Acquisition Status Monitoring Unit: Monitors the working status of each sensor in real time. When a sensor malfunction or signal acquisition abnormality occurs, it generates an abnormal prompt signal and transmits it to the result output and early warning module.

[0024] Preferably, the signal preprocessing module specifically includes:

[0025] Signal denoising unit: Real-time identification of environmental electromagnetic interference and sensor inherent noise in the signal, filtering out interference signals by adjusting filtering parameters, and retaining effective information in PPG signal and auxiliary signal;

[0026] Signal correction and normalization unit: Eliminates baseline drift of PPG signal through baseline correction, maps signals of different magnitudes to a preset range, and eliminates recognition deviation caused by differences in signal amplitude.

[0027] Outlier Removal Unit: Identifies and removes sudden abnormal signals that occur during the acquisition process, performs quality verification on the pre-processed signals, and synchronously transmits them to the artifact signal feature extraction module after passing the verification.

[0028] Preferably, the artifact signal feature extraction module specifically includes:

[0029] PPG signal feature extraction unit: Extracts multi-dimensional features related to artifacts in PPG signals, including time domain, frequency domain and morphological features, and captures signal peak, main frequency and waveform distortion degree;

[0030] Auxiliary signal feature extraction unit: Extracts feature parameters of multi-source auxiliary signals, including amplitude and frequency of limb motion signals, stability of probe wearing pressure signals, and intensity change rate of ambient light signals, to supplement the deficiencies of PPG signal features;

[0031] Feature filtering and fusion unit: Filters all extracted features, removes redundant and invalid features, and forms a standardized artifact recognition feature vector through fusion processing, which is then synchronously transmitted to the deep learning artifact recognition module.

[0032] Preferably, the deep learning artifact recognition module specifically includes:

[0033] Feature vector receiving and verification unit: Receives standardized artifact recognition feature vectors, performs quality verification on the feature vectors, and removes invalid and redundant feature data;

[0034] Deep learning model execution unit: Built-in hybrid model that integrates convolutional neural networks and recurrent neural networks trained and optimized with multi-scene data, which identifies artifact types by autonomously learning the mapping relationship between features and artifact types;

[0035] Intensity grading unit: The identified artifacts are classified into three levels according to the interference intensity: mild, moderate and severe, and the time nodes of artifact appearance are recorded simultaneously;

[0036] Recognition result output unit: synchronously transmits the recognition results of artifact type, intensity level, and occurrence time to the blood oxygen signal effectiveness assessment module and the result output and early warning module.

[0037] Preferably, the blood oxygen signal effectiveness assessment module specifically includes:

[0038] Multi-dimensional assessment system unit: Combining artifact type, artifact intensity and PPG signal quality parameters, a multi-dimensional assessment system is built, and the weights of each assessment indicator are adjusted according to the needs of clinical monitoring scenarios;

[0039] Assessment Level Determination Unit: The blood oxygen signal is graded and determined by the assessment system, and divided into three levels: effective, basically effective, and ineffective. The effective level corresponds to no artifacts or mild artifacts and excellent PPG signal quality. The basically effective level corresponds to moderate artifacts and PPG signal quality that meets the basic monitoring requirements. The ineffective level corresponds to severe artifacts and PPG signal that is severely distorted and cannot be used for blood oxygen calculation. The artifact intensity and PPG signal quality standards corresponding to each level are defined.

[0040] Evaluation result output unit: The final evaluation level result is synchronously transmitted to the result output and early warning module and the data storage and traceability module.

[0041] Preferably, the result output and early warning module specifically includes:

[0042] The result output unit uses three methods: visual display, voice prompt, and signal output. The visual display presents the artifact type, intensity, and blood oxygen signal validity level in real time. The voice prompt broadcasts key information synchronously. The signal output converts the results into standard control signals.

[0043] Anomaly warning triggering unit: Real-time monitoring of received data. When moderate or higher artifacts, invalid blood oxygen signals, or abnormal acquisition occur, an early warning mechanism is triggered, and timely handling is reminded through voice prompts.

[0044] Equipment linkage control unit: Transmits the output standard control signal to the blood oxygen monitoring equipment for equipment linkage control, triggers the suspension of invalid data recording, and prompts for adjustment of probe wearing status.

[0045] Preferably, the data storage and traceability module specifically includes:

[0046] Multi-source data storage unit: Stores all data during system operation, including raw PPG signals, multi-source auxiliary signals, preprocessed signals, artifact feature vectors, artifact recognition results, blood oxygen signal validity assessment results and abnormal records. It adopts a distributed storage architecture to store qualified data in real time.

[0047] Data classification and management unit: Classifies and organizes stored data according to time, patient information, and monitoring scenarios, establishes an index system, and sets data storage periods according to clinical needs;

[0048] Data traceability and output unit: Supports full-process traceability and retrospective analysis of stored data, and outputs data of specified dimensions according to query requirements.

[0049] Compared with the prior art, the advantages of the present invention are:

[0050] By synchronously acquiring multi-source auxiliary signals and raw PPG signals, uniformly preprocessing and fusing features, the system overcomes the limitations of single blood oxygenation signals being susceptible to interference from motion, ambient light, and wearing pressure, thus improving signal reliability from the source. Employing a deep learning hybrid model integrating convolutional neural networks and recurrent neural networks, and optimized through multi-scenario data training, it can accurately identify artifact types and classify interference intensity, with recognition accuracy and generalization ability far superior to traditional methods. A multi-dimensional blood oxygenation signal effectiveness evaluation system is established, achieving three-level signal classification and providing objective quantitative evidence for clinical monitoring. The system also features real-time monitoring of acquisition anomalies, multi-form early warning output, and device linkage functions, enabling rapid fault response and intervention prompts. Relying on a distributed storage and classification traceability architecture, it achieves full-process data management and review. Simultaneously, it is compatible with multiple types of blood oxygenation probes, flexibly adapting to different clinical scenarios. From signal acquisition, artifact identification, evaluation to early warning traceability, the entire chain is optimized, significantly improving the accuracy, stability, and practicality of blood oxygenation monitoring, effectively avoiding monitoring errors caused by artifacts, and providing reliable data support for clinical diagnosis and treatment. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the system proposed in this invention;

[0052] Figure 2 This is a diagram of the blood oxygen probe signal acquisition module proposed in this invention;

[0053] Figure 3 This is a diagram of the multi-source auxiliary signal acquisition module proposed in this invention;

[0054] Figure 4 This is a diagram of the signal preprocessing module proposed in this invention;

[0055] Figure 5 This is a diagram of the artifact signal feature extraction module proposed in this invention;

[0056] Figure 6 This is a diagram of the deep learning artifact recognition module proposed in this invention;

[0057] Figure 7 This is a diagram of the blood oxygen signal effectiveness assessment module proposed in this invention;

[0058] Figure 8 This is a diagram of the result output and early warning module proposed in this invention;

[0059] Figure 9 This is a diagram of the data storage and traceability module proposed in this invention. Detailed Implementation

[0060] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0061] See Figure 1 As shown, the deep learning-based system for identifying and evaluating the effectiveness of pulse oximeter artifact signals includes:

[0062] Pulse Oxygen Probe Signal Acquisition Module: Acquires the raw photoplethysmography (PPG) signal output by the pulse oxygen probe in real time and detects the connectivity of the acquisition link. When an acquisition interruption or signal loss occurs, it generates an acquisition anomaly prompt signal.

[0063] Multi-source auxiliary signal acquisition module: Acquires multi-source auxiliary signals related to blood oxygen monitoring. All auxiliary signals are synchronized and marked with the same timestamp as the original PPG signal.

[0064] Signal preprocessing module: Receives the raw PPG signal and multi-source auxiliary signals, and performs synchronous preprocessing, including signal denoising, baseline correction, signal normalization and outlier removal;

[0065] Artifact signal feature extraction module: Based on the preprocessed signal, extract the time domain features, frequency domain features, and morphological features of the PPG signal, as well as the feature parameters of the multi-source auxiliary signals. After feature filtering and fusion processing, a standardized artifact recognition feature vector is formed.

[0066] Deep learning artifact recognition module: It has a built-in deep learning hybrid model that integrates convolutional neural networks and recurrent neural networks, trained and optimized by multi-scene data. It identifies artifact types by learning the mapping relationship between artifact features and artifact types, and classifies the interference intensity.

[0067] Blood oxygen signal effectiveness assessment module: Receives artifact recognition results and preprocessed PPG signals, and combines artifact type, artifact intensity and quality parameters of the PPG signal itself to establish a multi-dimensional effectiveness assessment system to grade and evaluate blood oxygen signals.

[0068] Results output and early warning module: Receives artifact recognition results, blood oxygen signal effectiveness assessment results and abnormal acquisition prompt signals, and provides early warnings in multiple forms, including visual display, voice prompts and signal output;

[0069] Data storage and traceability module: Stores all data during system operation. It adopts a distributed storage architecture and classifies and manages the stored data according to time, patient information, and monitoring scenario, supporting data traceability and review.

[0070] See Figure 2 As shown, the pulse oximeter signal acquisition module specifically includes:

[0071] Probe adapter unit: Establishes connection with various models and specifications of pulse oximeter probes, adapting to pulse oximeter monitoring equipment in different clinical scenarios;

[0072] Signal acquisition unit: Real-time acquisition of raw photoplethysmography (PPG) signals output by the pulse oximeter, with a focus on capturing signals in the red and infrared bands;

[0073] Acquisition rate adjustment unit: Based on clinical monitoring needs, the acquisition rate can be adjusted within a preset range to adapt to the monitoring needs of different patients;

[0074] Acquisition Status Monitoring Unit: Real-time detection of acquisition link connectivity. When acquisition is interrupted or signal is lost, an acquisition anomaly prompt signal is generated and quickly transmitted to the result output and early warning module.

[0075] Specifically, using a continuous streaming acquisition mode, it simultaneously captures two original pulse wave signals, red light and infrared light, records waveform data frame by frame and binds a unique timestamp, performs real-time temporary storage and queue sorting of signals during the acquisition process, preserves the original fluctuation morphology, amplitude changes and periodic characteristics of the pulse wave, and the two signals are acquired independently and transmitted in parallel without interfering with each other, completely restoring the real photoelectric response information of human peripheral blood circulation.

[0076] Based on the differences in needs of clinical monitoring scenarios and patient types, the algorithmic adjustment of the acquisition rate is completed within the preset control range; for scenarios such as routine adult monitoring, continuous monitoring of critically ill patients, and refined monitoring of newborns, the corresponding acquisition frequency standard is automatically matched to optimize the data acquisition density while ensuring signal integrity.

[0077] Simultaneously, the module performs real-time detection of the acquisition link continuity. It continuously monitors the logical state of the acquisition link from three dimensions: signal transmission continuity, waveform existence, and data frame integrity. If multiple consecutive frames of signal are missing, data transmission is interrupted, or there is no valid waveform output, it is immediately determined that the acquisition is interrupted or the signal is lost. Once an anomaly is identified, a standardized acquisition anomaly prompt signal is immediately generated, clearly indicating the anomaly type, occurrence timestamp, and anomaly channel.

[0078] See Figure 3 As shown, the multi-source auxiliary signal acquisition module specifically includes:

[0079] Multi-dimensional signal acquisition unit: synchronously acquires various auxiliary signals, captures human limb movement signals through a three-axis accelerometer, acquires ambient light interference signals through a light sensor, detects probe wearing pressure signals through a pressure sensor, and simultaneously acquires human body temperature signals;

[0080] Signal synchronization unit: Using the same timestamp as the original PPG signal of the blood oxygen probe, all acquired auxiliary signals are synchronized and aligned with the PPG signal in time;

[0081] Data Acquisition Status Monitoring Unit: Monitors the working status of each sensor in real time. When a sensor malfunction or signal acquisition abnormality occurs, it generates an abnormal prompt signal and transmits it to the result output and early warning module.

[0082] Specifically, a timestamp system that is completely consistent with the original PPG signal of the blood oxygen probe is invoked to mark the timing of each auxiliary signal point by point, and all auxiliary signals are aligned with the PPG signal at the millisecond level. Through timing matching logic, it is ensured that the PPG signal, motion signal, light interference signal, pressure signal and body temperature signal at the same time point correspond one by one.

[0083] Each auxiliary signal is continuously monitored, and logical judgments are made from three dimensions: signal continuity, amplitude rationality, and data integrity to identify abnormal acquisition conditions such as continuous signal interruption, no effective data output, and abnormal amplitude exceeding limits.

[0084] See Figure 4 As shown, the signal preprocessing module specifically includes:

[0085] Signal denoising unit: Real-time identification of environmental electromagnetic interference and sensor inherent noise in the signal, filtering out interference signals by adjusting filtering parameters, and retaining effective information in PPG signal and auxiliary signal;

[0086] Signal correction and normalization unit: Eliminates baseline drift of PPG signal through baseline correction, maps signals of different magnitudes to a preset range, and eliminates recognition deviation caused by differences in signal amplitude.

[0087] Outlier Removal Unit: Identifies and removes sudden abnormal signals that occur during the acquisition process, performs quality verification on the pre-processed signals, and synchronously transmits them to the artifact signal feature extraction module after passing the verification.

[0088] Specifically, the interference types in the signal are distinguished through feature recognition logic: environmental electromagnetic interference manifests as high-frequency irregular fluctuations, while sensor inherent noise is low-amplitude steady-state clutter, both of which are significantly different from the physiological fluctuation characteristics of PPG signals and the physical change characteristics of auxiliary signals. Filtering parameters are dynamically adjusted for different interference types: targeted frequency band suppression is used for electromagnetic interference, preserving the physiologically effective frequency band of PPG signals (0.5-10Hz) and the characteristic frequency band of auxiliary signals; and smoothing logic is used for sensor inherent noise to filter out low-amplitude clutter without destroying the original form of the signal.

[0089] To address the baseline drift issue that is prone to occur in PPG signals, the module identifies the slow offset pattern of the signal baseline through dynamic trend tracking, eliminates the drift effect through real-time compensation, and restores the true fluctuation pattern of the PPG signal. Then, normalization processing is initiated: signals of different magnitudes are uniformly mapped to a preset 0-1 range, and the identification deviation caused by the difference in signal amplitude is eliminated by proportional scaling, while retaining the relative change characteristics within each signal.

[0090] See Figure 5 As shown, the artifact signal feature extraction module specifically includes:

[0091] PPG signal feature extraction unit: Extracts multi-dimensional features related to artifacts in PPG signals, including time domain, frequency domain and morphological features, and captures signal peak, main frequency and waveform distortion degree;

[0092] Auxiliary signal feature extraction unit: Extracts feature parameters of multi-source auxiliary signals, including amplitude and frequency of limb motion signals, stability of probe wearing pressure signals, and intensity change rate of ambient light signals, to supplement the deficiencies of PPG signal features;

[0093] Feature filtering and fusion unit: Filters all extracted features, removes redundant and invalid features, and forms a standardized artifact recognition feature vector through fusion processing, which is then synchronously transmitted to the deep learning artifact recognition module.

[0094] Specifically, for PPG signals, artifact-related features are comprehensively extracted from three dimensions: time domain, frequency domain, and morphology. Time domain feature extraction focuses on capturing the fluctuation patterns related to artifacts, including abnormal fluctuation amplitudes of signal peaks and troughs, stability of the pulse cycle, and rate of change of signal amplitude, accurately identifying features such as peak abrupt changes and period disorder caused by artifacts. Frequency domain feature extraction focuses on the signal frequency distribution, capturing abnormal frequency components caused by artifacts, distinguishing between the physiological signal's main frequency and artifact interference frequencies, and extracting features such as main frequency offset and clutter frequency proportion. Morphological feature extraction analyzes the waveform integrity of PPG signals, including the slope changes of rising and falling edges, waveform distortion degree, and waveform symmetry, identifying abnormal morphologies such as waveform distortion and overlap caused by artifacts.

[0095] By combining the correlation between multi-source auxiliary signals and artifacts, targeted feature parameters are extracted to supplement the deficiencies of PPG signal features; limb motion signals are used to extract motion amplitude, motion frequency, and motion duration to determine the presence of motion artifacts and motion intensity; probe wearing pressure signals are used to extract the stability, fluctuation amplitude, and number of abrupt changes in pressure values ​​to identify artifacts caused by probe loosening or excessive pressure; ambient light signals are used to extract the rate of change of light intensity and abnormal light intensity peaks to capture artifact features caused by ambient light interference; body temperature signals are used to extract the fluctuation range to help determine whether signal interference is indirectly caused by abnormal body temperature. All auxiliary features are bound to the PPG signal features of the corresponding timestamp.

[0096] By analyzing the correlation of features, redundant, repetitive, and invalid features that are irrelevant to artifacts are eliminated, while core features with high recognizability and strong correlation are retained. Then, through feature fusion logic, the time domain, frequency domain, and morphological features of the PPG signal are integrated with the features of multi-source auxiliary signals according to weights, eliminating redundant interference between features and forming a unified and dimensionally standardized artifact recognition feature vector.

[0097] See Figure 6 As shown, the deep learning artifact recognition module specifically includes:

[0098] Feature vector receiving and verification unit: Receives standardized artifact recognition feature vectors, performs quality verification on the feature vectors, and removes invalid and redundant feature data;

[0099] Deep learning model execution unit: Built-in hybrid model that integrates convolutional neural networks and recurrent neural networks trained and optimized with multi-scene data, which identifies artifact types by autonomously learning the mapping relationship between features and artifact types;

[0100] Intensity grading unit: The identified artifacts are classified into three levels according to the interference intensity: mild, moderate and severe, and the time nodes of artifact appearance are recorded simultaneously;

[0101] Recognition result output unit: synchronously transmits the recognition results of artifact type, intensity level, and occurrence time to the blood oxygen signal effectiveness assessment module and the result output and early warning module.

[0102] Specifically, the core of the hybrid model adopts a hybrid architecture of "CNN + LSTM + residual connection". The CNN part has 3 layers of convolution and pooling layers. The first layer of convolution kernel size is 3×3, which captures local spatial features such as PPG waveform distortion and sudden changes in motion signal amplitude. The second layer strengthens the coupling of multiple features. The third layer extracts global spatial features. The pooling layer reduces the dimensionality by 2 times, retaining key features while reducing the amount of computation. The LSTM part has 3 temporal layers, which receive the CNN output and capture the evolution of artifacts over time. Each layer has backpropagation logic to avoid the loss of long-term temporal features. The residual connection layer spans the CNN and LSTM layers, supplementing the gradient propagation path and adapting to the dual feature requirements of blood oxygenation artifacts of "spatial morphology + temporal change". The output layer is a hybrid structure of multi-classification and regression, corresponding to the artifact type and intensity classification dimensions.

[0103] The training process consists of four stages: Dataset construction: Collect artifact data from three scenarios: adults, newborns, and intensive care units. Divide the dataset into training / validation / test sets in an 8:1:1 ratio and label the artifact type, intensity, and occurrence scenario; Basic parameter settings: The batch size is fixed at 32, the initial learning rate is 0.001, the number of iterations is 200, the optimizer is adaptive moment estimation, and the loss function is adapted to the dual objectives of artifact multi-classification and intensity regression; Scenario-based fine-tuning: To address the high artifact burst rate in the intensive care unit scenario, the dropout rate of the LSTM layers is increased to 0.3 to enhance generalization; For the newborn scenario, the CNN convolution kernel size is adjusted to 2×2 to strengthen the recognition of subtle artifacts; Validation iteration: Validate with the test set every 10 rounds, retain the model version with artifact type recognition accuracy >95% and intensity classification accuracy >92%, and discard overfitted or underfitted versions;

[0104] After the input feature vector is processed by CNN to extract spatial features and LSTM to learn temporal patterns, the output layer is mapped to a structured result: dimension 1 corresponds to 5 types of artifacts, dimension 2 corresponds to 3 levels of intensity (light / moderate / severe), and dimension 3 is bound to timestamps and confidence thresholds. The output results are precisely correlated with the time period of the original PPG signal, for example, "10:25-10:30 time period, motion artifact, moderate intensity". At the same time, results below the threshold are marked as "to be reviewed", ensuring that the output is strongly correlated with the temporal features of the blood oxygen signal, providing an accurate basis for subsequent effectiveness evaluation.

[0105] See Figure 7 As shown, the blood oxygen signal effectiveness assessment module specifically includes:

[0106] Multi-dimensional assessment system unit: Combining artifact type, artifact intensity and PPG signal quality parameters, a multi-dimensional assessment system is built, and the weights of each assessment indicator are adjusted according to the needs of clinical monitoring scenarios;

[0107] Assessment Level Determination Unit: The blood oxygen signal is graded and determined by the assessment system, and divided into three levels: effective, basically effective, and ineffective. The effective level corresponds to no artifacts or mild artifacts and excellent PPG signal quality. The basically effective level corresponds to moderate artifacts and PPG signal quality that meets the basic monitoring requirements. The ineffective level corresponds to severe artifacts and PPG signal that is severely distorted and cannot be used for blood oxygen calculation. The artifact intensity and PPG signal quality standards corresponding to each level are defined.

[0108] Evaluation result output unit: The final evaluation level result is synchronously transmitted to the result output and early warning module and the data storage and traceability module.

[0109] Specifically, the multi-dimensional evaluation system is a comprehensive evaluation system encompassing three dimensions: artifact interference, signal quality, and acquisition status. The artifact interference dimension covers artifact type, intensity level, and duration; the signal quality dimension includes waveform integrity, periodicity, amplitude rationality, and effective pulsation ratio; and the acquisition status dimension relates to the stability of the acquisition link. The weights are automatically adjusted according to the clinical monitoring scenario: routine monitoring scenarios emphasize artifact interference and basic waveform quality, critical care monitoring scenarios increase the weight of signal stability and continuity, and neonatal scenarios strengthen the proportion of signal precision.

[0110] Under the premise of eliminating the influence of artifact interference, the quality of the PPG signal itself is judged to identify whether the signal has a clear pulse waveform, regular cardiac cycle, and stable amplitude change, and to determine whether there are non-artifact quality problems such as congenital waveform defects and period disorder. The analysis results are quantified into quality levels, distinguished into three levels: excellent, average, and poor, which serve as the core basis for effectiveness evaluation.

[0111] The three indicators—artifact interference results, PPG signal native quality, and acquisition status—are comprehensively integrated and classified into three levels according to preset rules: the effective level corresponds to no or mild artifacts, excellent signal quality, and stable acquisition status, which can be directly used for blood oxygen calculation; the basic effective level corresponds to moderate artifacts, average signal quality, and waveform not completely destroyed, which can only meet basic monitoring needs; the ineffective level corresponds to severe artifacts, poor signal quality, and severely distorted waveform, which cannot support blood oxygen value calculation at all. The influence of instantaneous interference is excluded in the judgment process, and the comprehensive performance over a continuous period is used as the final basis.

[0112] See Figure 8 As shown, the results output and early warning module specifically includes:

[0113] The result output unit uses three methods: visual display, voice prompt, and signal output. The visual display presents the artifact type, intensity, and blood oxygen signal validity level in real time. The voice prompt broadcasts key information synchronously. The signal output converts the results into standard control signals.

[0114] Anomaly warning triggering unit: Real-time monitoring of received data. When moderate or higher artifacts, invalid blood oxygen signals, or abnormal acquisition occur, an early warning mechanism is triggered, and timely handling is reminded through voice prompts.

[0115] Equipment linkage control unit: Transmits the output standard control signal to the blood oxygen monitoring equipment for equipment linkage control, triggers the suspension of invalid data recording, and prompts for adjustment of probe wearing status.

[0116] Specifically, the system determines in real time whether to trigger an alert based on preset rules: an emergency alert is triggered when there is an interruption in data acquisition, signal loss, severe artifacts, or invalid blood oxygen signals; a regular alert is triggered when there are moderate artifacts or the signal is basically valid; after an alert is triggered, the system automatically records the trigger time, abnormality type, level, and other information to form an alert log.

[0117] See Figure 9 As shown, the data storage and traceability module specifically includes:

[0118] Multi-source data storage unit: Stores all data during system operation, including raw PPG signals, multi-source auxiliary signals, preprocessed signals, artifact feature vectors, artifact recognition results, blood oxygen signal validity assessment results and abnormal records. It adopts a distributed storage architecture to store qualified data in real time.

[0119] Data classification and management unit: Classifies and organizes stored data according to time, patient information, and monitoring scenarios, establishes an index system, and sets data storage periods according to clinical needs;

[0120] Data traceability and output unit: Supports full-process traceability and retrospective analysis of stored data, and outputs data of specified dimensions according to query requirements.

[0121] Specifically, the verified multi-source data is standardized and packaged, a unique identifier is added to each type of data, and core metadata such as timestamp, patient basic information, and monitoring scenario type are bound. The data is then split according to data type and time period, and the original signal, feature data, identification and evaluation results, and abnormal records are split separately, while the temporal correlation and logical correspondence between the data are preserved.

[0122] It receives traceability query requests from clinical or system users, quickly locates the corresponding data through the index based on the query conditions, and retrieves the complete data flow chain, including the entire process data from signal acquisition, preprocessing, artifact identification, effectiveness assessment to outputting early warnings; it also supports retrospective analysis and can output data reports of specified dimensions as needed, clearly presenting the data change patterns and the entire process of abnormal occurrence, providing complete data support for clinical optimization of monitoring programs and investigation of system problems, and completing the closed loop of data storage and traceability.

[0123] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0124] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0125] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based system for identifying and evaluating the effectiveness of pulse oximeter artifact signals, characterized in that, include: Pulse Oxygen Probe Signal Acquisition Module: Acquires the raw photoplethysmography (PPG) signal output by the pulse oxygen probe in real time and detects the connectivity of the acquisition link. When an acquisition interruption or signal loss occurs, it generates an acquisition anomaly prompt signal. Multi-source auxiliary signal acquisition module: Acquires multi-source auxiliary signals related to blood oxygen monitoring. All auxiliary signals are synchronized and marked with the same timestamp as the original PPG signal. Signal preprocessing module: Receives the raw PPG signal and multi-source auxiliary signals, and performs synchronous preprocessing, including signal denoising, baseline correction, signal normalization and outlier removal; Artifact signal feature extraction module: Based on the preprocessed signal, extract the time domain features, frequency domain features, and morphological features of the PPG signal, as well as the feature parameters of the multi-source auxiliary signals. After feature filtering and fusion processing, a standardized artifact recognition feature vector is formed. Deep learning artifact recognition module: It has a built-in deep learning hybrid model that integrates convolutional neural networks and recurrent neural networks, trained and optimized by multi-scene data. It identifies artifact types by learning the mapping relationship between artifact features and artifact types, and classifies the interference intensity. Blood oxygen signal effectiveness assessment module: Receives artifact recognition results and preprocessed PPG signals, and combines artifact type, artifact intensity and quality parameters of the PPG signal itself to establish a multi-dimensional effectiveness assessment system to grade and evaluate blood oxygen signals. Results output and early warning module: Receives artifact recognition results, blood oxygen signal effectiveness assessment results and abnormal acquisition prompt signals, and provides early warnings in multiple forms, including visual display, voice prompts and signal output; Data storage and traceability module: Stores all data during system operation. It adopts a distributed storage architecture and classifies and manages the stored data according to time, patient information, and monitoring scenario, supporting data traceability and review.

2. The deep learning-based pulse oximeter artifact signal identification and effectiveness evaluation system according to claim 1, characterized in that, The pulse oximeter signal acquisition module specifically includes: Probe adapter unit: Establishes connection with various models and specifications of pulse oximeter probes, adapting to pulse oximeter monitoring equipment in different clinical scenarios; Signal acquisition unit: Real-time acquisition of raw photoplethysmography (PPG) signals output by the pulse oximeter, with a focus on capturing signals in the red and infrared bands; Acquisition rate adjustment unit: Based on clinical monitoring needs, the acquisition rate can be adjusted within a preset range to adapt to the monitoring needs of different patients; Acquisition Status Monitoring Unit: Real-time detection of acquisition link connectivity. When acquisition is interrupted or signal is lost, an acquisition anomaly prompt signal is generated and quickly transmitted to the result output and early warning module.

3. The deep learning-based pulse oximeter artifact signal identification and effectiveness evaluation system according to claim 1, characterized in that, The multi-source auxiliary signal acquisition module specifically includes: Multi-dimensional signal acquisition unit: synchronously acquires various auxiliary signals, captures human limb movement signals through a three-axis accelerometer, acquires ambient light interference signals through a light sensor, detects probe wearing pressure signals through a pressure sensor, and simultaneously acquires human body temperature signals; Signal synchronization unit: Using the same timestamp as the original PPG signal of the blood oxygen probe, all acquired auxiliary signals are synchronized and aligned with the PPG signal in time; Data Acquisition Status Monitoring Unit: Monitors the working status of each sensor in real time. When a sensor malfunction or signal acquisition abnormality occurs, it generates an abnormal prompt signal and transmits it to the result output and early warning module.

4. The deep learning-based pulse oximeter artifact signal identification and effectiveness evaluation system according to claim 1, characterized in that, The signal preprocessing module specifically includes: Signal denoising unit: Real-time identification of environmental electromagnetic interference and sensor inherent noise in the signal, filtering out interference signals by adjusting filtering parameters, and retaining effective information in PPG signal and auxiliary signal; Signal correction and normalization unit: Eliminates baseline drift of PPG signal through baseline correction, maps signals of different magnitudes to a preset range, and eliminates recognition deviation caused by differences in signal amplitude. Outlier Removal Unit: Identifies and removes sudden abnormal signals that occur during the acquisition process, performs quality verification on the pre-processed signals, and synchronously transmits them to the artifact signal feature extraction module after passing the verification.

5. The deep learning-based pulse oximeter artifact signal identification and effectiveness evaluation system according to claim 1, characterized in that, The artifact signal feature extraction module specifically includes: PPG signal feature extraction unit: Extracts multi-dimensional features related to artifacts in PPG signals, including time domain, frequency domain and morphological features, and captures signal peak, main frequency and waveform distortion degree; Auxiliary signal feature extraction unit: Extracts feature parameters of multi-source auxiliary signals, including amplitude and frequency of limb motion signals, stability of probe wearing pressure signals, and intensity change rate of ambient light signals, to supplement the deficiencies of PPG signal features; Feature filtering and fusion unit: Filters all extracted features, removes redundant and invalid features, and forms a standardized artifact recognition feature vector through fusion processing, which is then synchronously transmitted to the deep learning artifact recognition module.

6. The deep learning-based pulse oximeter artifact signal identification and effectiveness evaluation system according to claim 1, characterized in that, The deep learning artifact recognition module specifically includes: Feature vector receiving and verification unit: Receives standardized artifact recognition feature vectors, performs quality verification on the feature vectors, and removes invalid and redundant feature data; Deep learning model execution unit: Built-in hybrid model that integrates convolutional neural networks and recurrent neural networks trained and optimized with multi-scene data, which identifies artifact types by autonomously learning the mapping relationship between features and artifact types; Intensity grading unit: The identified artifacts are classified into three levels according to the interference intensity: mild, moderate and severe, and the time nodes of artifact appearance are recorded simultaneously; Recognition result output unit: synchronously transmits the recognition results of artifact type, intensity level, and occurrence time to the blood oxygen signal effectiveness assessment module and the result output and early warning module.

7. The deep learning-based pulse oximeter artifact signal identification and effectiveness evaluation system according to claim 1, characterized in that, The blood oxygen signal effectiveness assessment module specifically includes: Multi-dimensional assessment system unit: Combining artifact type, artifact intensity and PPG signal quality parameters, a multi-dimensional assessment system is built, and the weights of each assessment indicator are adjusted according to the needs of clinical monitoring scenarios; Assessment Level Determination Unit: The blood oxygen signal is graded and determined by the assessment system, and divided into three levels: effective, basically effective, and ineffective. The effective level corresponds to no artifacts or mild artifacts and excellent PPG signal quality. The basically effective level corresponds to moderate artifacts and PPG signal quality that meets the basic monitoring requirements. The ineffective level corresponds to severe artifacts and PPG signal that is severely distorted and cannot be used for blood oxygen calculation. The artifact intensity and PPG signal quality standards corresponding to each level are defined. Evaluation result output unit: The final evaluation level result is synchronously transmitted to the result output and early warning module and the data storage and traceability module.

8. The deep learning-based pulse oximeter artifact signal identification and effectiveness evaluation system according to claim 1, characterized in that, The result output and early warning module specifically includes: The result output unit uses three methods: visual display, voice prompt, and signal output. The visual display presents the artifact type, intensity, and blood oxygen signal validity level in real time. The voice prompt broadcasts key information synchronously. The signal output converts the results into standard control signals. Anomaly warning triggering unit: Real-time monitoring of received data. When moderate or higher artifacts, invalid blood oxygen signals, or abnormal acquisition occur, an early warning mechanism is triggered, and timely handling is reminded through voice prompts. Equipment linkage control unit: Transmits the output standard control signal to the blood oxygen monitoring equipment for equipment linkage control, triggers the suspension of invalid data recording, and prompts for adjustment of probe wearing status.

9. The deep learning-based pulse oximeter artifact signal identification and effectiveness evaluation system according to claim 1, characterized in that, The data storage and traceability module specifically includes: Multi-source data storage unit: Stores all data during system operation, including raw PPG signals, multi-source auxiliary signals, preprocessed signals, artifact feature vectors, artifact recognition results, blood oxygen signal validity assessment results and abnormal records. It adopts a distributed storage architecture to store qualified data in real time. Data classification and management unit: Classifies and organizes stored data according to time, patient information, and monitoring scenarios, establishes an index system, and sets data storage periods according to clinical needs; Data traceability and output unit: Supports full-process traceability and retrospective analysis of stored data, and outputs data of specified dimensions according to query requirements.

Images