Multi-modal vital signs ai early warning system and method for critical patients

Abstract

This invention relates to the field of health monitoring technology, and more particularly to an AI-based early warning system and method for multimodal vital signs in acute and critical illnesses. The system includes: a data acquisition module for collecting multimodal vital sign data from patients; a data evaluation module for selectively removing or triggering modal complementation mechanisms on unqualified data; a feature extraction module for determining an initial set of feature parameters and implementing an adaptive adjustment strategy for feature extraction based on the maximum parameter change rate and clinical risk acceleration; a feature screening module for determining effective features based on feature discrimination coefficients and generating a screened feature set; a time-series prediction module for inputting the screened feature set into a time-series prediction network and outputting risk probability values ​​and early warning time windows for each preset critical event; and an early warning module for determining the early warning level based on the risk probability values ​​and early warning time windows and generating an early warning signal. This invention improves the speed and accuracy of early warning response.

Description

Technical Field

[0001] This invention relates to the field of health monitoring technology, and in particular to an AI-based early warning system and method for multimodal vital signs in acute and critical illnesses. Background Technology

[0002] Monitoring vital signs in critically ill patients is a core aspect of clinical treatment. Organ dysfunction in critically ill patients often occurs at the microscopic pathophysiological level. The body's strong compensatory abilities can mask early changes in the condition. Once the condition becomes severe, it progresses extremely rapidly, requiring medical staff to provide precise and timely treatment. Currently, widely used clinical methods for monitoring vital signs are mainly divided into two categories: contact and non-contact. Traditional contact monitoring relies on equipment such as electrocardiograms, pulse oximeters, and blood pressure cuffs. Although considered the clinical "gold standard," patient movement is restricted due to lead wires, and sensor displacement can easily introduce measurement artifacts, making it particularly difficult for special populations such as newborns and burn patients. In recent years, non-contact monitoring technologies such as video-based remote photoplethysmography have developed rapidly. These technologies extract pulse signals by detecting subtle color changes in facial videos. However, they are extremely sensitive to changes in lighting, skin color differences, and subtle patient movements. In real critical care environments, patients are often obstructed by medical equipment, leading to a significant decrease in signal quality. To overcome the limitations of single-modality monitoring, multimodal fusion technology has become a research hotspot. By integrating multi-source data such as video, audio, radar, wearable sensors, and electronic medical records, and utilizing artificial intelligence for feature-level fusion, the robustness and accuracy of monitoring can be significantly improved. However, existing technologies still have the following drawbacks: single data modality leads to insufficient information completeness; traditional threshold alarm mechanisms have poor warning timeliness, triggering alarms only after the patient enters the decompensated stage; high false positive rates lead to "alarm fatigue," reducing the willingness of medical staff to respond; and non-contact monitoring is not adaptable to complex clinical environments. Therefore, developing an AI-based early warning system and method for acute and critical illnesses that can integrate multimodal data, possess high-precision real-time warning capabilities, and effectively reduce false positive rates is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0003] To address this, the present invention provides an AI-based early warning system and method for multimodal vital signs in acute and critical illnesses, which overcomes the problems of poor non-contact monitoring signal quality, delayed early warning response, and high false positive rate in existing technologies for special populations that cannot be effectively monitored by conventional monitoring methods.

[0004] To achieve the above objectives, the present invention provides an AI-based early warning system for multimodal vital signs in acute and critical illnesses, comprising: The data acquisition module is used to collect multimodal vital sign data of patients in a non-contact or micro-contact manner; The data evaluation module, which is connected to the data acquisition module, is used to calculate the activity disturbance index and signal quality index of each modality data in real time, and determine the qualification of the data based on the activity disturbance index and signal quality index, and select to perform data rejection or trigger the modal complementation mechanism for unqualified data. The feature extraction module, which is connected to the data evaluation module, is used to determine an initial set of feature parameters based on the time-domain and waveform features of qualified data, and to execute an adaptive adjustment strategy for feature extraction based on the maximum parameter change rate and clinical risk acceleration, including dynamically adjusting the sliding window length and feature update frequency of feature extraction. The feature filtering module, which is connected to the feature extraction module, is used to determine the feature discrimination coefficient based on the inter-class separability and intra-class compactness of each feature in the initial feature parameter set, mark features with a discrimination coefficient greater than a preset discrimination threshold as effective features, perform clustering based on the Pearson correlation coefficient between effective features, and select the feature with the highest discrimination coefficient from each cluster as the representative feature of that cluster to generate a filtered feature set. The time series prediction module is connected to the feature filtering module to input the filtered feature set into the time series prediction network and output the risk probability value and warning time window of each preset critical event. The early warning module, which is connected to the time-series prediction module, is used to determine the early warning level based on the risk probability value and the early warning time window, and generate an early warning signal.

[0005] Furthermore, the data evaluation module determines the data's validity based on the activity disturbance index and the signal quality index, wherein... When the signal quality index is greater than or equal to the first preset quality threshold and the activity disturbance index is less than or equal to the first preset motion threshold, the data is deemed qualified. If the signal quality index is less than the first preset quality threshold or the activity disturbance index is greater than the first preset motion threshold, the data is deemed unqualified.

[0006] Furthermore, the activity disturbance index is determined based on the short-time amplitude fluctuation variance of the signal and the signal mutation rate between adjacent sampling points; The signal quality index is determined based on the signal-to-noise ratio and the effective waveform percentage, where the effective waveform percentage is the proportion of sampling points whose signal amplitude is within a preset effective amplitude range to the total number of sampling points.

[0007] Furthermore, in response to a signal quality index being less than a first preset quality threshold and greater than or equal to a second preset quality threshold, or an activity disturbance index being greater than a first preset motion threshold and less than or equal to a second preset motion threshold, the data evaluation module triggers a modal complementation mechanism, using other qualified modal data to reconstruct the feature parameters of the current modality through a pre-trained intermodal mapping model to replace the current unqualified data.

[0008] Furthermore, in response to a signal quality index being less than a second preset quality threshold and an activity disturbance index being greater than a second preset motion threshold, the data evaluation module performs data rejection, discards the current unqualified data, and generates a data source check alarm signal.

[0009] Furthermore, the feature extraction module executes the following adaptive adjustment strategy, including: When the maximum parameter change rate is greater than the first preset rate threshold or the clinical risk acceleration is greater than the first preset acceleration threshold, the system switches to fast response mode, shortens the sliding window length of feature extraction to the first preset window length, and increases the feature update frequency to the first preset update frequency. When the maximum parameter change rate is less than the second preset rate threshold and the clinical risk acceleration is less than or equal to the second preset acceleration threshold, the system switches to long-term monitoring mode, extends the sliding window length of feature extraction to the second preset window length, and reduces the feature update frequency to the second preset update frequency. The clinical risk acceleration is the second derivative of the risk probability value output by the time-series prediction module.

[0010] Furthermore, the inter-class separability is determined based on the distance between cluster centers of different class samples in the feature space; The intra-class compactness is determined based on the average Euclidean distance between samples of the same class in the feature space; The feature discrimination coefficient is the ratio of inter-class separability to intra-class compactness.

[0011] Furthermore, the specific process of clustering based on the Pearson correlation coefficient between effective features by the feature selection module is as follows: Calculate the Pearson correlation coefficient between any two valid features; When the absolute value of the Pearson correlation coefficient is greater than or equal to the preset correlation threshold, the two corresponding features will be classified into the same cluster.

[0012] Furthermore, the early warning module determines the early warning level based on the risk probability value and the early warning time window, wherein, When the risk probability value is less than the first preset probability threshold and the warning time window is greater than or equal to the first preset time window threshold, it is determined to be a level three warning. When the risk probability value is greater than or equal to the first preset probability threshold and the warning time window is greater than or equal to the first preset time window threshold, or when the risk probability value is less than the first preset probability threshold and the warning time window is less than the first preset time window threshold, it is determined to be a level two warning. When the risk probability value is greater than or equal to the first preset probability threshold and the warning time window is less than the first preset time window threshold, it is determined to be a Level 1 warning.

[0013] This invention also provides an AI-based early warning method for multimodal vital signs in acute and critical illnesses, comprising: Step S1: Collect non-contact or micro-contact multimodal vital sign data from the patient; Step S2: Calculate the activity disturbance index and signal quality index of each modality data in real time, determine the qualification of the data based on the activity disturbance index and signal quality index, and select to perform data removal or trigger the modal complementation mechanism for unqualified data. Step S3: Determine the initial feature parameter set based on the time domain features and waveform features of qualified data, and execute an adaptive adjustment strategy for feature extraction based on the maximum parameter change rate and clinical risk acceleration, including dynamically adjusting the sliding window length and feature update frequency of feature extraction. Step S4: Determine the feature discrimination coefficient based on the inter-class separability and intra-class compactness of each feature in the initial feature parameter set. Mark the features with a discrimination coefficient greater than the preset discrimination threshold as effective features. Perform clustering based on the Pearson correlation coefficient between effective features. Select the feature with the highest discrimination coefficient from each cluster as the representative feature of that cluster to generate the filtered feature set. Step S5: Input the filtered feature set into the time series prediction network and output the risk probability value and warning time window for each preset critical event; Step S6: Determine the warning level based on the risk probability value and the warning time window, and generate a warning signal.

[0014] Compared with existing technologies, the advantages of this invention lie in its introduction of two independent evaluation dimensions: the activity disturbance index and the signal quality index. This allows for real-time assessment of the interference level and signal quality of each modality. When the signal quality index falls below a threshold or the activity disturbance index exceeds a threshold, the system automatically determines the data to be unqualified and selectively performs data removal or triggers a modal complementation mechanism based on the severity. This mechanism effectively solves the problem of poor signal quality caused by patient movement and environmental interference in non-contact monitoring, preventing low-quality data from directly entering subsequent analysis processes, thereby reducing the false positive rate of the early warning system.

[0015] Furthermore, this invention addresses data that is unqualified but does not meet the rejection criteria by employing a modal complementation mechanism. This mechanism utilizes other qualified modal data and reconstructs the feature parameters of the current modality through a pre-trained inter-modal mapping model, effectively replacing the unqualified data. This mechanism is particularly suitable for scenarios involving sudden agitation or changes in body position in critically ill patients. It can maintain the continuity of vital sign monitoring even when a particular modality temporarily fails, avoiding early warning blind spots or delayed responses caused by missing data.

[0016] Furthermore, this invention dynamically adjusts the window length and update frequency of feature extraction based on the maximum parameter change rate and the acceleration of clinical risk. When a patient's condition deteriorates rapidly, the system automatically switches to a rapid response mode, employing a short-window, high-frequency feature extraction strategy to ensure timely warnings; when the patient's condition is relatively stable, the system switches to a long-term monitoring mode, employing a long-window, low-frequency strategy to reduce computational load and minimize spurious fluctuations. This adaptive mechanism effectively balances the dual needs of timely warnings and long-term stable monitoring in acute and critical care scenarios.

[0017] Furthermore, this invention first calculates the feature discrimination coefficient of each feature by the ratio of inter-class separability to intra-class compactness, and then selects the most discriminative features for classifying critical events. Next, it clusters the effective features using the Pearson correlation coefficient and retains only the representative features with the highest discrimination coefficient from each cluster. This selection mechanism, while ensuring the amount of feature information, eliminates feature redundancy to the greatest extent, reduces the input dimensionality of the time-series prediction network, thereby reducing the risk of model overfitting and improving the prediction's generalization ability and computational efficiency.

[0018] Furthermore, this invention employs a multi-task temporal prediction network, using the filtered feature set as input to simultaneously output the risk probability value and expected warning time window for each preset critical event. Based on the combined judgment of these two factors, the system classifies warnings into three levels. This grading mechanism provides clear decision-making references for clinical medical staff: Level 1 warnings indicate the need for immediate intervention, Level 2 warnings indicate the need for enhanced monitoring, and Level 3 warnings indicate that routine observation is acceptable, effectively avoiding the false alarms and missed alarms caused by single-threshold warnings. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the module connections of the AI ​​early warning system for multimodal vital signs in acute and critical illnesses according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the process of determining the passability of data based on the activity disturbance index and signal quality index, as described in an embodiment of the present invention. Figure 3 This is a flowchart illustrating how the warning level is determined based on the risk probability value and the warning time window, as described in an embodiment of the present invention. Figure 4This is a flowchart of the AI-based early warning method for multimodal vital signs in acute and critical illnesses according to an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0021] Those skilled in the art will understand that the method for determining the above-mentioned parameters for a single item in this invention can be as follows: selecting the value with the highest proportion based on the data distribution as the preset standard parameter; using weighted summation to use the obtained value as the preset standard parameter; substituting each historical data into a specific formula and using the value obtained by the formula as the preset standard parameter; or other selection methods, as long as the present invention can clearly define different specific situations in the single item determination process through the obtained values.

[0022] Please see Figure 1 As shown in the figure, the AI-based early warning system for multimodal vital signs in acute and critical illnesses according to an embodiment of the present invention includes: The data acquisition module is used to collect multimodal vital sign data of patients in a non-contact or micro-contact manner; The data evaluation module, which is connected to the data acquisition module, is used to calculate the activity disturbance index and signal quality index of each modality data in real time, and determine the qualification of the data based on the activity disturbance index and signal quality index, and select to perform data rejection or trigger the modal complementation mechanism for unqualified data. The feature extraction module, which is connected to the data evaluation module, is used to determine an initial set of feature parameters based on the time-domain and waveform features of qualified data, and to execute an adaptive adjustment strategy for feature extraction based on the maximum parameter change rate and clinical risk acceleration, including dynamically adjusting the sliding window length and feature update frequency of feature extraction. The feature filtering module, which is connected to the feature extraction module, is used to determine the feature discrimination coefficient based on the inter-class separability and intra-class compactness of each feature in the initial feature parameter set, mark features with a discrimination coefficient greater than a preset discrimination threshold as effective features, perform clustering based on the Pearson correlation coefficient between effective features, and select the feature with the highest discrimination coefficient from each cluster as the representative feature of that cluster to generate a filtered feature set. The time series prediction module is connected to the feature filtering module to input the filtered feature set into the time series prediction network and output the risk probability value and warning time window of each preset critical event. The early warning module, which is connected to the time-series prediction module, is used to determine the early warning level based on the risk probability value and the early warning time window, and generate an early warning signal.

[0023] Specifically, the data acquisition module acquires multimodal data in the following ways, including but not limited to: using millimeter-wave radar to transmit continuous frequency-modulated waves to the patient's chest and receiving the echo signals; extracting intermediate frequency signals after mixing and filtering; demodulating the patient's cardiac impact map signal, respiratory waveform signal, and body motion signal from the signal; using the cardiac impact map signal to calculate heart rate and heart rate variability parameters; and using the respiratory waveform signal to calculate respiratory rate and chest and abdominal movement amplitude parameters. It also continuously acquires a video stream of the patient's face using an optical camera, detecting and tracking the facial region in each frame; extracting the spatiotemporal variation signal of skin pixels within the region of interest; and obtaining a video-based photoplethysmography signal after independent component analysis or chromaticity model decomposition; from which heart rate and heart rate variability parameters are calculated. Heterogeneous and relative blood oxygen saturation parameters; using mattress-type piezoelectric sensors or textile electrodes, micro-vibration signals and electrocardiogram (ECG) signals are collected under conditions where the patient is in contact with the mattress or clothing. After pre-amplification, analog-to-digital conversion, and filtering, ECG and respiratory waveforms are obtained for the extraction and cross-validation of heart rate, arrhythmia events, and respiratory rate parameters; an accelerometer is fixed to the patient's chest wall or integrated into a wearable patch to synchronously acquire acceleration signals along three axes, which are used to calculate the patient's body motion intensity, positional changes, and activity disturbance index, providing motion reference information for the data evaluation module; all the above-mentioned sensors and acquisition devices are started synchronously according to a preset sampling rate. The raw data of each modality is timestamped to form a multimodal synchronous data stream, which is uniformly transmitted to the data evaluation module for subsequent processing.

[0024] It is understandable that this invention primarily addresses common technical challenges faced by the following specific populations in clinical monitoring: For patients with extensive burns, traditional contact electrodes or sensors cannot be adhered to the damaged skin surface; forced contact not only causes severe pain but may also lead to secondary infections; For newborns, especially premature infants, whose skin barrier function is not yet fully developed, repeated application and removal of adhesive patches or electrodes can easily cause skin damage, and the added weight of the sensors and the binding of wires can interfere with the infant's independent activities and growth and development; For patients with highly infectious diseases (such as severe respiratory infectious diseases, multidrug-resistant bacterial infections, etc.), reducing or avoiding contact with the monitored individuals is crucial. Direct physical contact with medical devices can significantly reduce the occupational exposure risk for healthcare workers during the wearing, adjustment, and removal of devices. For patients with impaired consciousness or agitation, sensors are easily displaced due to involuntary limb movements. Traditional contact monitoring not only has poor data continuity, but frequent sensor resetting operations also increase the nursing burden and stimulate the patient's emotions. For elderly or debilitated patients who require long-term bed rest monitoring, continuous pressure contact (such as finger clip pulse oximeters or chest electrode patches) may lead to local skin pressure sores or circulatory disorders. Micro-contact mattress sensors or non-contact radar monitoring can achieve continuous monitoring without disturbing the patient.

[0025] Specifically, the time-series prediction network employs a sequence modeling structure (such as LSTM, GRU, or Transformer), taking a filtered feature sequence within a continuous time window as input and memorizing the historical physiological parameter change trends through its internal state memory. The network output has two prediction heads for each preset critical event: one is a risk probability prediction head, outputting the probability of the event occurring between 0 and 1; the other is a time window prediction head, outputting the remaining time expected for the event to occur.

[0026] Please see Figure 2 As shown, it is a flowchart of an embodiment of the present invention for determining the passability of data based on the activity disturbance index and the signal quality index; When the signal quality index is greater than or equal to the first preset quality threshold and the activity disturbance index is less than or equal to the first preset motion threshold, the data is deemed qualified. When the signal quality index is less than the first preset quality threshold or the activity disturbance index is greater than the first preset motion threshold, the data is deemed unqualified. In this embodiment of the invention, the first preset mass threshold is 0.7 and the first preset motion threshold is 0.3.

[0027] Specifically, the activity disturbance index is determined based on the short-time amplitude fluctuation variance of the signal and the signal abrupt change rate between adjacent sampling points. The activity disturbance index = first weighting coefficient × normalized short-time amplitude fluctuation variance + second weighting coefficient × normalized signal abrupt change rate between adjacent sampling points, where... The normalized short-time amplitude fluctuation variance is determined as follows: calculate the variance of the original signal amplitude within the current sliding window to obtain the short-time amplitude fluctuation variance, and normalize its range to the interval between 0 and 1; The normalized signal mutation rate between adjacent sampling points is determined as follows: calculate the absolute value of the difference between adjacent sampling points and take the mean value to obtain the signal mutation rate between adjacent sampling points, and normalize its range to the interval of 0 to 1. In this embodiment of the invention, the first weighting coefficient is 0.4 and the second weighting coefficient is 0.6.

[0028] The signal quality index is determined based on the signal-to-noise ratio (SNR) and the effective proportion of the waveform. The signal quality index = third weighting coefficient × normalized SNR + fourth weighting coefficient × effective proportion of the waveform, where... The normalized signal-to-noise ratio (SNR) is determined as follows: a fast Fourier transform is performed on the original signal within the current sliding window. The sum of the power spectrum energies corresponding to the physiological frequency bands in the spectrum is taken as the signal power, and the sum of the power spectrum energies of the remaining frequency bands in the spectrum is taken as the noise power. The ratio of signal power to noise power is calculated, and the original SNR in decibels is obtained by logarithmic transformation. Then, the normalized signal-to-noise ratio in the interval from 0 to 1 is obtained by range normalization. The effective percentage of the waveform is determined as follows: count the number of sampling points within the current sliding window whose amplitude falls within the preset effective amplitude range, and calculate the ratio of this number to the total number of sampling points in the window; the preset effective amplitude range is the interval of ±1.5 times the standard deviation of the signal amplitude mean. In this embodiment of the invention, the third weighting coefficient is 0.5 and the fourth weighting coefficient is 0.5.

[0029] Specifically, the data evaluation module responds to the signal quality index being less than a first preset quality threshold and greater than or equal to a second preset quality threshold, or the activity disturbance index being greater than a first preset motion threshold and less than or equal to a second preset motion threshold, by triggering a modal complementation mechanism, and using other qualified modal data to reconstruct the feature parameters of the current modality through a pre-trained intermodal mapping model to replace the current unqualified data; The data evaluation module responds to the signal quality index being less than the second preset quality threshold and the activity disturbance index being greater than the second preset motion threshold by performing data removal, discarding the current unqualified data and generating a data source check alarm signal. In this embodiment of the invention, the second preset mass threshold is 0.4 and the second preset motion threshold is 0.6.

[0030] Specifically, the pre-trained intermodal mapping model refers to a cross-modal mapping network pre-trained using offline collected multimodal synchronous data as training samples and deep learning algorithms. This model takes qualified source modal data as input and unqualified target modal feature parameters as output, and is used to achieve feature-level reconstruction and substitution between modalities. This invention does not limit the specific network architecture of the intermodal mapping model; those skilled in the art can choose any one or a combination of fully connected neural networks, convolutional neural networks, recurrent neural networks, long short-term memory networks, gated recurrent units, Transformer networks, or conditional generative adversarial networks according to the actual application scenario. During the training phase, synchronous data containing multiple modalities is collected as a training set. The temporal domain features, frequency domain features, and waveform morphology features of the source modalities are extracted as input features, and the corresponding feature parameters of the target modalities are used as supervision labels. The network parameters are optimized by minimizing the loss function between the predicted output and the true label, enabling the model to learn the nonlinear mapping relationship between different modalities. In practical deployment, when a certain modality data is judged to be partially unqualified, the system automatically calls the pre-trained intermodal mapping model, inputs the other qualified modal data into the model, and outputs the estimated feature parameters of the currently unqualified modality after forward calculation. This replaces the original unqualified data and is sent to the feature extraction module for subsequent processing, thereby ensuring the continuity of vital sign monitoring and the stability of the early warning system when a single modality fails temporarily.

[0031] It is understandable that when the signal quality index is less than the first preset quality threshold and greater than or equal to the second preset quality threshold, the signal quality index is at a medium level. The current modal signal still has some usable information, but the signal-to-noise ratio is insufficient or the waveform distortion has affected the reliability of independent feature extraction. At this time, it is not advisable to use the original data directly, nor is it serious enough to need to be completely discarded. Therefore, by using the modal complementarity mechanism to reconstruct the data with the help of other qualified modes, the data quality can be improved while retaining the original information. The activity disturbance index is greater than the first preset motion threshold and less than or equal to the second preset motion threshold. The logic behind this classification is as follows: when the activity disturbance index is at a moderate level, the patient exhibits mild to moderate body movement or agitation; the signal is affected by motion interference to some extent but has not completely collapsed. At this point, some physiological components in the original data are still discernible. Directly removing them would waste information, while direct use might introduce errors. Therefore, correction and replacement through modal complementarity mechanisms is the optimal processing strategy. When the signal quality index is less than the second preset quality threshold and the activity disturbance index is greater than the second preset motion threshold, the signal quality index has dropped to an extremely low level, indicating severe signal degradation. Simultaneously, the high activity disturbance index indicates intense patient activity or agitation. Both factors combined completely overwhelm the physiological components of the current modality signal, making effective recovery through complementary mechanisms impossible. Continuing to retain or attempt to reconstruct this modality data is not only unhelpful but may also introduce erroneous information that interferes with subsequent predictions; therefore, data removal is performed.

[0032] Specifically, the feature extraction module performs the following adaptive adjustment strategy, including: When the maximum parameter change rate is greater than the first preset rate threshold or the clinical risk acceleration is greater than the first preset acceleration threshold, the system switches to fast response mode, shortens the sliding window length of feature extraction to the first preset window length, and increases the feature update frequency to the first preset update frequency. When the maximum parameter change rate is less than the second preset rate threshold and the clinical risk acceleration is less than or equal to the second preset acceleration threshold, the system switches to long-term monitoring mode, extends the sliding window length of feature extraction to the second preset window length, and reduces the feature update frequency to the second preset update frequency. Otherwise, the system switches to the normal monitoring mode. In the normal monitoring mode, the sliding window length for feature extraction is set to 10 seconds, and the feature update frequency is set to 10 seconds / time. The maximum parameter change rate refers to the maximum value among all physiological parameter change rates that can be extracted from the current qualified data. The clinical risk acceleration is the second derivative of the risk probability value output by the time series prediction module. When the system is in the initial operation stage and the time series prediction module has not yet output the risk probability value, the value of the clinical risk acceleration is zero. In this embodiment of the invention, the first preset rate threshold is 0.15 / second, and the second preset rate threshold is 0.05 / second; the first preset acceleration threshold is 0.05 / second², and the second preset acceleration threshold is 0; the first preset window length is 2 seconds, and the second preset window length is 60 seconds; the first preset update frequency is 1 second / time, and the second preset update frequency is 30 seconds / time.

[0033] Understandably, when the maximum parameter change rate exceeds the first preset rate threshold or the clinical risk acceleration exceeds the first preset acceleration threshold, it indicates that the patient's physiological state is rapidly deteriorating, and the rate of deterioration may be accelerating. At this time, the system switches to a rapid response mode, adopting a strategy of shortening the sliding window length and increasing the feature update frequency to ensure timely capture of rapid changes in the condition, thus gaining valuable preparation time for clinical intervention. Conversely, when the maximum parameter change rate is lower than the second preset rate threshold and the clinical risk acceleration is less than or equal to the second preset acceleration threshold, it indicates that the patient's physiological state is relatively stable and there are no signs of accelerated deterioration. At this time, the system switches to a long-term monitoring mode, adopting a strategy of extending the sliding window length and reducing the feature update frequency to smooth normal fluctuations in physiological signals, reduce computational load, and reduce false alarms caused by noise. For the normal state between the two, the system adopts a normal monitoring mode, achieving a balance between response speed and stability.

[0034] Specifically, the inter-class separability is determined based on the distance between cluster centers of different categories of samples in the feature space. The feature parameters collected by different monitoring devices are regarded as different categories. For a certain feature, the feature space center point of all samples of the feature under each category (i.e. the average value of the samples of that category) is calculated first. Then, the average Euclidean distance between the center points of different categories is calculated. The larger the average distance, the more significant the difference in the numerical distribution of the feature among different monitoring devices. The intra-class compactness is determined based on the average Euclidean distance between samples of the same class in the feature space. Specifically, for the same feature extracted by the same monitoring device, the average Euclidean distance between samples of the feature at multiple sampling time points is calculated. The smaller the average distance, the more stable the value of the feature is within the same modality, and the better the consistency. The feature discrimination coefficient is the ratio of inter-class separability to intra-class compactness. The larger the ratio, the more separated the samples of different classes are while keeping the samples of the same class compact, and the stronger the discrimination ability. In this embodiment of the invention, the preset discrimination threshold is set to 2.0.

[0035] Specifically, the clustering process performed by the feature selection module based on the Pearson correlation coefficient between effective features is as follows: Calculate the Pearson correlation coefficient between any two valid features; When the absolute value of the Pearson correlation coefficient is greater than or equal to the preset correlation threshold, the two corresponding features will be classified into the same cluster. In this embodiment of the invention, the preset relevant threshold value ranges from 0.70 to 0.95, with a preferred value of 0.85.

[0036] Understandably, inter-class separability reflects the spatial separation of features from different modalities, while intra-class compactness reflects the temporal stability of features within the same modality. A higher ratio indicates that the feature exhibits significant differences between different modalities while maintaining stability within the same modality, thus possessing good discriminative ability. By setting a preset discriminative threshold, features with low discriminative power or high noise levels below this threshold are removed, thereby selecting effective features. The Pearson correlation coefficient is used to measure the linear correlation between pairs of features. When the absolute value of the correlation coefficient is greater than or equal to the preset correlation threshold, the features are considered to carry similar discriminative information and are grouped into the same cluster. Since features within the same cluster have high information overlap, not all need to be retained. Therefore, the feature with the highest discriminative coefficient in each cluster is selected as the representative feature of that cluster, preserving the strongest discriminative power of that category while minimizing feature dimensionality.

[0037] Please see Figure 3 As shown, it is a flowchart of an embodiment of the present invention for determining the warning level based on the risk probability value and the warning time window; When the risk probability value is less than the first preset probability threshold and the warning time window is greater than or equal to the first preset time window threshold, it is determined to be a level three warning. It is recommended to observe normally and no special intervention is needed for the time being. When the risk probability value is greater than or equal to the first preset probability threshold and the warning time window is greater than or equal to the first preset time window threshold, or when the risk probability value is less than the first preset probability threshold and the warning time window is less than the first preset time window threshold, it is determined to be a level two warning, prompting medical staff to strengthen monitoring, increase attention, and prepare for intervention. When the risk probability value is greater than or equal to the first preset probability threshold and the warning time window is less than the first preset time window threshold, it is determined to be a Level 1 warning, and the emergency procedure must be initiated immediately and clinical intervention must be carried out. In this embodiment of the invention, the first preset probability threshold is set to 0.6, and the first preset time window threshold is set to 30 minutes. It should be noted that the first preset probability threshold and the first preset time window threshold can be adaptively adjusted according to the specific requirements of the clinical department, the severity of the patient's condition, and the results of historical data statistics. For example, in the intensive care unit (ICU) scenario, the threshold can be appropriately lowered to improve sensitivity, and in the general ward scenario, the threshold can be appropriately increased to reduce false alarms.

[0038] Please see Figure 4 As shown, the AI-based early warning method for multimodal vital signs in acute and critical illnesses according to an embodiment of the present invention includes: Step S1: Collect non-contact or micro-contact multimodal vital sign data from the patient; Step S2: Calculate the activity disturbance index and signal quality index of each modality data in real time, determine the qualification of the data based on the activity disturbance index and signal quality index, and select to perform data removal or trigger the modal complementation mechanism for unqualified data. Step S3: Determine the initial feature parameter set based on the time domain features and waveform features of qualified data, and execute an adaptive adjustment strategy for feature extraction based on the maximum parameter change rate and clinical risk acceleration, including dynamically adjusting the sliding window length and feature update frequency of feature extraction. Step S4: Determine the feature discrimination coefficient based on the inter-class separability and intra-class compactness of each feature in the initial feature parameter set. Mark features with a discrimination coefficient greater than a preset discrimination threshold as effective features. Perform clustering based on the Pearson correlation coefficient between effective features, and select the feature with the highest discrimination coefficient from each cluster as the representative feature of that cluster to generate a filtered feature set. The calculation of the feature discrimination coefficient and the selection of effective features are completed during the offline training phase of the system, generating a fixed filtered feature set. This feature set is directly used in the online prediction phase, and the filtering calculation is not repeated. Step S5: Input the filtered feature set into the time series prediction network and output the risk probability value and warning time window for each preset critical event; Step S6: Determine the warning level based on the risk probability value and the warning time window, and generate a warning signal.

[0039] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in the details for the sake of brevity.

[0040] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A critical illness multi-modal vital signs AI early warning system, characterized in that, include: The data acquisition module is used to collect multimodal vital sign data of patients in a non-contact or micro-contact manner; The data evaluation module, which is connected to the data acquisition module, is used to calculate the activity disturbance index and signal quality index of each modality data in real time, and determine the qualification of the data based on the activity disturbance index and signal quality index, and select to perform data rejection or trigger the modal complementation mechanism for unqualified data. The feature extraction module, which is connected to the data evaluation module, is used to determine an initial set of feature parameters based on the time-domain and waveform features of qualified data, and to execute an adaptive adjustment strategy for feature extraction based on the maximum parameter change rate and clinical risk acceleration, including dynamically adjusting the sliding window length and feature update frequency of feature extraction. The feature filtering module, which is connected to the feature extraction module, is used to determine the feature discrimination coefficient based on the inter-class separability and intra-class compactness of each feature in the initial feature parameter set, mark features with a discrimination coefficient greater than a preset discrimination threshold as effective features, perform clustering based on the Pearson correlation coefficient between effective features, and select the feature with the highest discrimination coefficient from each cluster as the representative feature of that cluster to generate a filtered feature set. The time series prediction module is connected to the feature filtering module and is used to input the filtered feature set into the time series prediction network and output the risk probability value and warning time window of each preset critical event. The early warning module, which is connected to the time-series prediction module, is used to determine the early warning level based on the risk probability value and the early warning time window, and generate an early warning signal.

2. The AI-based early warning system for multimodal vital signs in acute and critical illnesses according to claim 1, characterized in that, The data evaluation module determines the data's validity based on the activity disturbance index and signal quality index. When the signal quality index is greater than or equal to the first preset quality threshold and the activity disturbance index is less than or equal to the first preset motion threshold, the data is deemed qualified. If the signal quality index is less than the first preset quality threshold or the activity disturbance index is greater than the first preset motion threshold, the data is deemed unqualified.

3. The AI-based early warning system for multimodal vital signs in acute and critical illnesses according to claim 2, characterized in that, The activity disturbance index is determined based on the short-time amplitude fluctuation variance of the signal and the signal mutation rate between adjacent sampling points; The signal quality index is determined based on the signal-to-noise ratio and the effective waveform percentage, where the effective waveform percentage is the proportion of sampling points whose signal amplitude is within a preset effective amplitude range to the total number of sampling points.

4. The AI-based early warning system for multimodal vital signs in acute and critical illnesses according to claim 3, characterized in that, The data evaluation module responds to the signal quality index being less than a first preset quality threshold and greater than or equal to a second preset quality threshold, or the activity disturbance index being greater than a first preset motion threshold and less than or equal to a second preset motion threshold, by triggering a modal complementation mechanism. The module then uses other qualified modal data to reconstruct the feature parameters of the current modality through a pre-trained intermodal mapping model to replace the current unqualified data.

5. The AI-based early warning system for multimodal vital signs in acute and critical illnesses according to claim 4, characterized in that, The data evaluation module responds to the signal quality index being less than the second preset quality threshold and the activity disturbance index being greater than the second preset motion threshold by performing data rejection, discarding the current unqualified data and generating a data source check alarm signal.

6. The AI-based early warning system for multimodal vital signs in acute and critical illnesses according to claim 5, characterized in that, The feature extraction module executes the following adaptive adjustment strategy, including: When the maximum parameter change rate is greater than the first preset rate threshold or the clinical risk acceleration is greater than the first preset acceleration threshold, the system switches to fast response mode, shortens the sliding window length of feature extraction to the first preset window length, and increases the feature update frequency to the first preset update frequency. When the maximum parameter change rate is less than the second preset rate threshold and the clinical risk acceleration is less than or equal to the second preset acceleration threshold, the system switches to long-term monitoring mode, extends the sliding window length of feature extraction to the second preset window length, and reduces the feature update frequency to the second preset update frequency. The clinical risk acceleration is the second derivative of the risk probability value output by the time-series prediction module.

7. The AI-based early warning system for multimodal vital signs in acute and critical illnesses according to claim 6, characterized in that, The inter-class separability is determined based on the distance between cluster centers of different categories of samples in the feature space; The intra-class compactness is determined based on the average Euclidean distance between samples of the same class in the feature space; The feature discrimination coefficient is the ratio of inter-class separability to intra-class compactness.

8. The AI-based early warning system for multimodal vital signs in acute and critical illnesses according to claim 7, characterized in that, The specific process of clustering based on the Pearson correlation coefficient between effective features in the feature selection module is as follows: Calculate the Pearson correlation coefficient between any two valid features; When the absolute value of the Pearson correlation coefficient is greater than or equal to the preset correlation threshold, the two corresponding features will be classified into the same cluster.

9. The AI-based early warning system for multimodal vital signs in acute and critical illnesses according to claim 8, characterized in that, The early warning module determines the early warning level based on the risk probability value and the early warning time window, wherein, When the risk probability value is less than the first preset probability threshold and the warning time window is greater than or equal to the first preset time window threshold, it is determined to be a level three warning. When the risk probability value is greater than or equal to the first preset probability threshold and the warning time window is greater than or equal to the first preset time window threshold, or when the risk probability value is less than the first preset probability threshold and the warning time window is less than the first preset time window threshold, it is determined to be a level two warning. When the risk probability value is greater than or equal to the first preset probability threshold and the warning time window is less than the first preset time window threshold, it is determined to be a Level 1 warning.

10. A method for AI-based early warning of multimodal vital signs in acute and critical illnesses, applied to the AI-based early warning system for multimodal vital signs in acute and critical illnesses as described in any one of claims 1-9, characterized in that, include: Step S1: Collect non-contact or micro-contact multimodal vital sign data from the patient; Step S2: Calculate the activity disturbance index and signal quality index of each modality data in real time, determine the qualification of the data based on the activity disturbance index and signal quality index, and select to perform data removal or trigger the modal complementation mechanism for unqualified data. Step S3: Determine the initial feature parameter set based on the time domain features and waveform features of qualified data, and execute an adaptive adjustment strategy for feature extraction based on the maximum parameter change rate and clinical risk acceleration, including dynamically adjusting the sliding window length and feature update frequency of feature extraction. Step S4: Determine the feature discrimination coefficient based on the inter-class separability and intra-class compactness of each feature in the initial feature parameter set. Mark the features with a discrimination coefficient greater than the preset discrimination threshold as effective features. Perform clustering based on the Pearson correlation coefficient between effective features. Select the feature with the highest discrimination coefficient from each cluster as the representative feature of that cluster to generate the filtered feature set. Step S5: Input the filtered feature set into the time series prediction network and output the risk probability value and warning time window for each preset critical event; Step S6: Determine the warning level based on the risk probability value and the warning time window, and generate a warning signal.

Images