Newborn body temperature abnormity prediction method and device and storage medium

Abstract

The invention relates to the technical field of abnormal body temperature prediction, in particular to a newborn abnormal body temperature prediction method and device and a storage medium. The method comprises the following steps: collecting the body temperature and physiological signals of a newborn in real time, and carrying out preprocessing; based on fast Fourier transform, time sequence features reflecting body temperature changes in the body temperature and physiological signals are extracted, feature vectors are generated, the body temperature and physiological signals are processed through phase coupling analysis and a frequency domain mask inhibition mechanism, and a low-frequency noise interference interval caused by neonatal breathing irregularity and skin conduction disturbance is dynamically recognized; on the basis of the feature vectors, the body temperature abnormal risk probability is predicted in real time through the long-short-term memory network, the gating weight of the long-short-term memory network is dynamically regulated and controlled through two factors of the body temperature change rate and the physiological state label, and a feature fusion mechanism is optimized. According to the invention, through multi-modal physiological signal fusion and a dynamic feature weighting mechanism, the accuracy and real-time performance of newborn body temperature anomaly prediction are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of body temperature abnormality prediction technology, and more specifically, to a method, device, and storage medium for predicting abnormal body temperature in newborns. Background Technology

[0002] Newborns, especially premature and low birth weight infants, are highly sensitive to changes in ambient temperature due to their underdeveloped thermoregulatory system, making them prone to abnormal body temperature (including hypothermia and hyperthermia). Traditional methods of temperature monitoring rely primarily on timed measurements, which lack the ability to monitor and analyze trends in newborn temperature changes in real time, making it difficult to detect potential risks and implement effective interventions in the early stages of abnormal temperature. Furthermore, changes in newborn body temperature are not only influenced by environmental factors but are also closely related to physiological parameters such as heart rate, respiratory rate, and blood oxygen saturation; single temperature data cannot comprehensively reflect the newborn's health status. In complex and variable situations, such as abnormal body temperature caused by infection, dehydration, or central nervous system diseases, other physiological signals often accompany abnormal temperatures, increasing the difficulty and uncertainty of accurate diagnosis. To improve the quality and efficiency of newborn care and reduce complications, there is an urgent need for a technological solution capable of automatically and accurately predicting abnormal body temperature. Therefore, this paper provides methods, devices, and storage media for predicting abnormal body temperature in newborns. Summary of the Invention

[0003] The purpose of this invention is to provide a method, device, and storage medium for predicting abnormal body temperature in newborns, in order to solve the problems mentioned in the background art that existing methods for monitoring body temperature in newborns cannot achieve real-time continuous monitoring, are difficult to accurately predict the risk of abnormal body temperature, and lack the ability to conduct collaborative analysis of multimodal physiological parameters.

[0004] To achieve the above objectives, the present invention aims to provide a method for predicting abnormal body temperature in newborns, comprising the following steps: S1. Real-time acquisition and preprocessing of newborns' body temperature and physiological signals; S2. Based on Fast Fourier Transform, extract the temporal features reflecting changes in body temperature from body temperature and physiological signals, generate feature vectors, and process body temperature and physiological signals through phase coupling analysis and frequency domain masking suppression mechanism to dynamically identify the low-frequency noise interference range caused by irregular breathing and skin conductivity disturbances in newborns. S3. Based on feature vectors, the Long Short-Term Memory (LSTM) network is used to predict the probability of abnormal body temperature in real time. The gating weights of the LSTM network are dynamically adjusted by two factors: the rate of change of body temperature and physiological state labels, thereby optimizing the feature fusion mechanism. S4. Risk classification judgment is made based on the probability of abnormal risks and combined with risk classification rules.

[0005] As a further improvement to this technical solution, step S2, extracting temporal features reflecting changes in body temperature from body temperature and physiological signals, includes the following steps: S2.1. A fixed time window is used to divide the continuous body temperature and physiological signal time series into multiple analysis segments, forming a sliding window structure; S2.2 Calculate the time-domain characteristics of body temperature and various physiological signals within each sliding window; S2.3. For the body temperature and physiological signal sequences within each sliding window, the frequency domain features are extracted using Fast Fourier Transform to be used for periodic fluctuations in body temperature and physiological signals. S2.4. Using the cross-modal analysis method, based on the body temperature and physiological signal sequences of the sliding window, calculate the dynamic coupling characteristics to reflect the coordinated fluctuation relationship between body temperature changes and other physiological activities. S2.5. Use a feature selection algorithm to filter out time-domain features, frequency-domain features, and dynamic coupling features that reflect changes in body temperature, and combine the filtered features into a unified feature vector. , In the formula, The first feature vector is the first feature vector. One characteristic, The first eigenvector in the original feature vector One characteristic, The final number of features selected. For feature index, The feature index is used for filtering.

[0006] As a further improvement to this technical solution, in S2.3, the frequency domain features of the body temperature and physiological signal sequences within each sliding window are extracted using Fast Fourier Transform, including the following steps: S2.31. By processing the body temperature and physiological signal sequences through phase coupling analysis and frequency domain masking suppression mechanism, the low-frequency noise interference range caused by irregular breathing and skin conductivity disturbance is dynamically identified, and an enhanced time domain signal is generated. S2.32. Apply Fast Fourier Transform to the enhanced time-domain signal to convert the enhanced time-domain signal into a frequency-domain complex spectrum. S2.33. Calculate the corresponding amplitude spectrum for each frequency domain complex spectrum, select the amplitude of the main frequency components in the frequency band, and construct the frequency domain feature vector. S2.34. Normalize each frequency domain feature vector; S2.35. All normalized frequency domain feature vectors are concatenated and fused to form a joint frequency domain feature vector.

[0007] As a further improvement to this technical solution, in S2.31, the body temperature and physiological signal sequences are processed through phase coupling analysis and frequency domain masking suppression mechanisms to dynamically identify low-frequency noise interference intervals caused by irregular breathing and skin conductivity disturbances, including the following steps: S2.311 Extract the RR interval sequence from the physiological signal, calculate the instantaneous heart rate variability, synchronously export the respiratory signal, use Hilbert transform to obtain the instantaneous phase of the instantaneous heart rate variability and the respiratory signal, and calculate the phase difference between the two. When the phase difference is greater than the upper limit threshold a, mark the x interval centered at the current time as the respiratory interference interval. S2.312. Calculate the second derivative of body temperature. If the absolute value of the second derivative is greater than the derivative threshold... When the current time point is used as the center, a fixed-length window is extended as the skin conductivity interference range; S2.313. Extract the reference noise signal from the breathing interference region and the skin conduction interference region, window the reference noise signal and perform a fast Fourier transform to obtain the complex spectrum, and take the modulus of the complex spectrum to obtain the interference spectrum template. S2.314. Divide the frequency bands, calculate the proportion of interference energy according to the frequency bands, and construct a frequency domain mask based on the interference spectrum template; S2.315. Apply a window function to the original body temperature and physiological signals, perform a fast Fourier transform, multiply the complex spectrum by the constructed frequency domain mask, and restore the processed complex spectrum to the enhanced time domain signal through an inverse Fourier transform.

[0008] As a further improvement to this technical solution, step S3, which uses a long short-term memory network based on feature vectors to predict the probability of abnormal body temperature in real time, includes the following steps: S3.1 Organize the feature vectors of each time period constructed in the sliding window method into a time-series input sequence in chronological order; S3.2 Input the temporal input sequence into a multi-layer stacked long short-term memory network, model the temporal dependency between features through its gating mechanism, extract deep temporal features representing the trend of body temperature change, and dynamically adjust the weights of each modality feature based on the body temperature trend and physiological state. S3.3, The feature vector output by the adjusted Long Short-Term Memory network The input is fed into a fully connected neural network to perform discriminative modeling of the risk of abnormal body temperature. S3.4. Generate the probability value of abnormal body temperature at the current moment through the output layer of the fully connected network.

[0009] As a further improvement to this technical solution, in step S3.2, the dynamic adjustment of the modal feature weights based on body temperature trends and physiological states includes the following steps: S3.21 Extract a scalar value representing the body temperature trend from the hidden states of the Long Short-Term Memory network. ; S3.22, the scalar value representing the trend of body temperature. Input a lightweight state classification network to classify physiological states and output the current physiological state label of the newborn. S3.23. Based on the physiological state label, query the corresponding pre-trained static gating vector, and dynamically adjust the weights in combination with the current body temperature change rate to form the final gating weight vector. S3.24. Feature vectors output by the Long Short-Term Memory network Element-wise scaling with a gated weight vector yields the adjusted feature vector output by the Long Short-Term Memory (LSTM) network. .

[0010] As a further improvement to this technical solution, in step S3.23, the weights are dynamically adjusted based on the current rate of body temperature change to form the final gating weight vector, including the following steps: S3.231, Use scalar values ​​that represent the trend of body temperature. As an indicator of the rate of change in body temperature per unit time, it is used to represent the direction and intensity of the current trend in body temperature. S3.232. Based on the direction and absolute value of the slope, construct the corresponding dynamic adjustment amount; S3.233. Combine the static gating vector with the dynamic adjustment amount to calculate the final gating weight vector.

[0011] As a further improvement to this technical solution, step S4, which involves risk classification based on the probability of abnormal risks and in conjunction with risk grading rules, includes the following steps: S4.1 Receive the probability value of abnormal body temperature output in step S3; S4.2 Classify body temperature risk levels based on the probability value of abnormal body temperature risk; S4.3. Determine the risk level of the current body temperature status based on the range of the probability value of abnormal body temperature. S4.4 Output the current body temperature risk level result.

[0012] On the other hand, the present invention provides a neonatal body temperature abnormality prediction device, comprising: The data acquisition module is used to collect and preprocess the newborn's body temperature and physiological signals in real time. The feature extraction module is used to extract the temporal features reflecting changes in body temperature from body temperature and physiological signals using fast Fourier transform, generate feature vectors, and process body temperature and physiological signals through phase coupling analysis and frequency domain masking suppression mechanism to dynamically identify the low-frequency noise interference range caused by irregular breathing and skin conductivity disturbances in newborns. The abnormal body temperature prediction module is used to predict the probability of abnormal body temperature in real time using a long short-term memory network. It also optimizes the feature fusion mechanism by dynamically adjusting the gating weights of the long short-term memory network through a dual factor of body temperature change rate and physiological state label. The risk classification module is used to classify risks by utilizing the probability of abnormal risks and combining them with risk grading rules.

[0013] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the neonatal body temperature abnormality prediction method described in any of the preceding claims.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The neonatal body temperature abnormality prediction method, device, and storage medium of this invention significantly improve the accuracy and real-time performance of neonatal body temperature abnormality prediction through multimodal physiological signal fusion and dynamic feature weighting mechanism. By employing a Long Short-Term Memory (LSTM) network combined with a dual-factor control of body temperature change rate and physiological state labels for gating weights, adaptive fusion of multi-source information such as heart rate, respiration, and blood oxygen saturation is achieved, effectively enhancing the model's adaptability to different clinical states. This allows for accurate identification of potential abnormal body temperature trends amidst complex physiological fluctuations, contributing to early intervention and clinical decision support.

[0015] 2. The neonatal body temperature abnormality prediction method, device, and storage medium of this invention introduce a noise suppression strategy based on phase coupling analysis and frequency domain masking suppression mechanism. This strategy can dynamically identify and suppress low-frequency interference caused by irregular breathing and skin conductivity disturbances, significantly improving the signal-to-noise ratio of body temperature and physiological signals. Combined with Fast Fourier Transform and temporal feature extraction, the model's ability to capture the periodicity and stability of body temperature changes is further enhanced, providing a more stable and reliable physiological data foundation for early warning of neonatal body temperature abnormalities, and improving the system's robustness and clinical applicability. Attached Figure Description

[0016] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: Please refer to Figure 1 As shown, this embodiment provides a method for predicting abnormal body temperature in newborns, including the following steps: S1. Real-time acquisition and preprocessing of newborns' body temperature and physiological signals; In this embodiment, body temperature includes core body temperature (rectal / ear temperature) and peripheral body temperature (axillary temperature, skin temperature); physiological signals include heart rate, respiratory rate, and blood oxygen saturation; and data are collected by sensors such as thermistors, infrared thermometers, PPG sensors, and chest impedance bands. Preprocessing includes time alignment and interpolation completion, and outlier detection and removal. S2. Based on Fast Fourier Transform, extract the temporal features reflecting changes in body temperature from body temperature and physiological signals, generate feature vectors, and process body temperature and physiological signals through phase coupling analysis and frequency domain masking suppression mechanism to dynamically identify the low-frequency noise interference range caused by irregular breathing and skin conductivity disturbances in newborns. In this embodiment, extracting temporal features reflecting changes in body temperature from body temperature and physiological signals includes the following steps: S2.1. A fixed time window (5 minutes or 10 minutes) is used to divide the continuous body temperature and physiological signal time series into multiple analysis segments to form a sliding window structure; S2.2 Within each sliding window, calculate the time-domain characteristics of body temperature and various physiological signals, including mean, variance, skewness, kurtosis, extreme value amplitude, fluctuation range, coefficient of variation, etc., to represent local change trends and stability. S2.3. For the body temperature and physiological signal sequence within each sliding window, the Fast Fourier Transform is used to extract frequency domain features, including spectral distribution, main frequency components, and energy concentration, for the periodic fluctuations in body temperature and physiological signals. The process involves extracting frequency domain features from the body temperature and physiological signal sequences within each sliding window using Fast Fourier Transform, including the following steps: S2.31. By processing the body temperature and physiological signal sequences through phase coupling analysis and frequency domain masking suppression mechanism, the low-frequency noise interference range caused by irregular breathing and skin conductivity disturbance is dynamically identified, and an enhanced time domain signal (body temperature and physiological signal sequence) is generated to improve the quality and stability of the real physiological oscillation components in the spectrum. Furthermore, phase coupling analysis and frequency domain masking suppression mechanisms are used to address low-frequency noise interference caused by irregular breathing and skin conductivity disturbances in neonatal body temperature monitoring. Irregular breathing (periodic breathing or apnea) leads to phase desynchronization between the heart rate variability (HRV) signal and the respiratory rhythm, manifesting as abnormal energy fluctuations in the low-frequency range (0.05–0.15 Hz) in the frequency domain. Skin conductivity disturbances (limb movement, changes in electrode contact) can cause abrupt changes in the skin temperature signal (sudden increase in the second derivative), forming artifacts in the time domain and leaking as broadband noise in the frequency domain (especially affecting 0.05–0.15 Hz). The study targets the 15–2 Hz physiological rhythm frequency band. Traditional methods (fixed-band filtering or time-domain smoothing) struggle to dynamically distinguish such noise from genuine thermoregulatory physiological oscillations (vasomotor fluctuations), leading to distorted feature extraction and reduced accuracy in predicting abnormal body temperature. This study utilizes phase coupling analysis (detecting the instantaneous phase difference between HRV and respiratory signals) to identify respiratory interference zones in real time, avoiding inaccuracies caused by individual differences in fixed-threshold methods. Simultaneously, by combining second-order derivative mutation detection of body temperature signals, the study accurately captures skin conductivity disturbance periods, overcoming the misjudgment problems inherent in traditional amplitude-threshold-based methods. Furthermore, a reference noise template is constructed and a frequency domain mask is generated. Interference bands are dynamically suppressed according to their energy proportions (low / medium / high frequency) to achieve targeted spectrum purification, maximizing the preservation of genuine physiological oscillation components such as heart rate variability rhythms. Further, signal reconstruction within the interference zone (FFT to mask to inverse FFT) enhances the stability and signal-to-noise ratio of the time-domain signal, improving the reliability of spectral characteristics such as dominant frequency and energy concentration. This mechanism effectively eliminates pseudo-physiological fluctuations caused by breathing and skin disturbances, avoids noise being misjudged as abnormal body temperature regulation, and ensures that the frequency domain features truly reflect the synergistic mechanism between body temperature and physiological signals. This provides solid support for the dynamic gating weight adjustment in the LSTM model and ultimately significantly improves the accuracy and robustness of abnormal body temperature risk classification. The body temperature and physiological signal sequences are processed using phase coupling analysis and frequency domain masking suppression mechanisms to dynamically identify low-frequency noise interference ranges caused by irregular breathing and skin conductivity disturbances. This includes the following steps: S2.311. Extract the RR interval sequence from the physiological signal (the RR interval sequence refers to the time interval sequence between the R waves (the most prominent peak in the electrocardiogram) between consecutive heartbeats, used to reflect heart rate variability; here, the physiological signal specifically refers to the heart rate signal), and calculate the instantaneous heart rate variability (the calculation of instantaneous heart rate variability specifically involves: interpolating and resampling the RR interval sequence to obtain heart rate variability signals with equal time intervals; then, using a sliding window, calculating the local standard deviation or frequency domain feature (LF / HF ratio) at each moment to obtain the instantaneous heart rate at that moment). The respiratory signal is simultaneously exported using the variability index (preferably using chest impedance band data; if unavailable, the respiratory signal is exported from the PPG (photoplethysmography) using respiratory-induced modulation (RIM) or PPG envelope waveform extraction). The instantaneous phase of the instantaneous heart rate variability and the respiratory signal is obtained using Hilbert transform, and the phase difference between the two is calculated. When the phase difference is greater than the upper limit threshold a (in this embodiment, the upper limit threshold a is 0.5π radians), the x interval centered at the current time is marked as the respiratory interference interval (in this embodiment, x is 200ms). Among them, the instantaneous phase of instantaneous heart rate variability for: ; Instantaneous phase of respiratory signal for: ; In the formula, For Hilbert transform operators, For time, For a moment Instantaneous heart rate variability signal, This represents the phase angle when taken as a complex number, i.e., the instantaneous phase of the analytic signal. For a moment respiratory signals; Phase difference between the two for: ; S2.312. Calculate the second derivative of body temperature. If the absolute value of the second derivative is greater than the derivative threshold... (In this embodiment, , When the temperature is measured in degrees Celsius per square second, a fixed-length window is extended with the current time point as the skin conductivity interference range. The second derivative of body temperature (skin temperature) is: ; In the formula, For skin temperature at all times The value, Indicates time The second derivative, i.e., acceleration, This represents the acceleration component of skin temperature changes, reflecting the trend of drastic temperature changes. S2.313. Extract reference noise signals from the breathing interference range and the skin conductivity interference range (breathing interference: separate abnormal segments from the breathing signal (ranges where the breathing amplitude is < 30% of the baseline); skin interference: extract abrupt change segments from the skin temperature signal (ranges where the absolute value of the second derivative of the skin temperature is greater than the skin temperature threshold e; in this embodiment, the empirical value of the skin temperature threshold e is 0.05°C / s). 2 (This reference noise signal only takes the abnormal time period within the interference window, not the entire sliding window). Window the reference noise signal (Hanning window) and perform a fast Fourier transform to obtain the complex spectrum. Take the magnitude (i.e. amplitude) of the complex spectrum to obtain the interference spectrum template. S2.314. Divide the frequency bands (low frequency 0.05-0.15Hz, mid frequency 0.15-0.5Hz, high frequency 0.5-2Hz), calculate the interference energy ratio according to the frequency bands, and construct a frequency domain mask based on the interference spectrum template (calculate the energy ratio according to the frequency band based on the spectrum of the reference noise signal, and then construct the mask) to suppress noise energy leakage. S2.315. Apply a window function (Hanning window) to the original body temperature and physiological signals, perform a fast Fourier transform, multiply the complex spectrum by the constructed frequency domain mask, suppress the energy of the interference frequency band, and restore the processed complex spectrum to the enhanced time domain signal through an inverse Fourier transform, thereby realizing the reconstruction of the sequence after interference suppression. S2.32. Apply Fast Fourier Transform (FFT) to the enhanced time-domain signal to convert the enhanced time-domain signal into a frequency-domain complex spectrum; S2.33. Calculate the corresponding amplitude spectrum for each frequency domain complex spectrum, select the amplitude of the main frequency components in the frequency band (the main frequency is 0.05–2Hz, covering physiological activity rhythms such as body temperature regulation and heart rate variability), and construct the frequency domain feature vector; S2.34. Normalize each frequency domain feature vector (normalize the total amplitude). S2.35. Concatenate and fuse all normalized frequency domain feature vectors (including body temperature, heart rate, respiratory rate, and blood oxygen saturation) to form a joint frequency domain feature vector. S2.4. Using cross-modal analysis, based on the body temperature and physiological signal sequences within a sliding window, calculate the dynamic coupling characteristics, including correlation coefficient, mutual information, phase synchronization index, and cross-power spectral density (the specific steps are: first, perform pairwise analysis on the body temperature sequence and various physiological signals (heart rate, respiratory rate, blood oxygen saturation) within each sliding window, and calculate the Pearson correlation coefficient, mutual information value, phase synchronization index, and cross-power spectral density between them), reflecting the coordinated fluctuation relationship between body temperature changes and other physiological activities; S2.5. Using a feature selection algorithm (recursive feature elimination), select time-domain features, frequency-domain features, and dynamically coupled features that reflect changes in body temperature (specifically: first, use the performance indicators of the abnormal body temperature prediction model (classification accuracy or loss function) as the evaluation standard, iteratively train the model and gradually eliminate features that contribute the least to the prediction until a subset of the most representative features that are sensitive to changes in body temperature is retained), and combine the selected features into a unified feature vector. , In the formula, The first feature vector is the first feature vector. The feature is a complete set of features pieced together from multiple sources, including body temperature, physiological signals, frequency domain features, and coupling features. The first eigenvector in the original feature vector One characteristic, , The final number of features selected. The feature index represents the first feature retained after feature selection. The index number of each feature in the original feature vector.

[0019] S3. Based on feature vectors, the Long Short-Term Memory (LSTM) network is used to predict the probability of abnormal body temperature in real time. The gating weights of the LSTM network are dynamically adjusted by two factors: the rate of change of body temperature and physiological state labels, thereby optimizing the feature fusion mechanism. In this embodiment, the probability of abnormal body temperature risk is predicted in real time using a long short-term memory network based on feature vectors, including the following steps: S3.1 Organize the feature vectors of each time period constructed in the sliding window method into a time-series input sequence in chronological order; S3.2. The temporal input sequence is fed into a multi-layered stacked Long Short-Term Memory (LSTM) network. Its gating mechanism models the temporal dependencies between features, extracting deep temporal features representing the trend of body temperature changes. The weights of each modality feature are dynamically adjusted based on the body temperature trend and physiological state. The basic architecture of the LSTM network includes an input layer, one or more LSTM hidden layers, and an output layer. The input layer receives a feature sequence organized by time steps, with each time step's input being a feature vector. The LSTM hidden layer consists of multiple units, each containing an input gate, a forget gate, and an output gate, used to control the writing, retention, and output of information, capturing long-term dependencies through internal memory units. The hidden state vector is propagated over time, extracting the temporal patterns of body temperature and physiological signals. Finally, the output of the LSTM is fed into a subsequent fully connected layer (multilayer perceptron) for predicting or classifying abnormal body temperature probabilities. In particular, during the temperature rise / fall phase (rapid fever or recovery period), the indicative weights of different physiological signals (heart rate, respiration, blood oxygen) to abnormal body temperature vary significantly (heart rate signals are more sensitive in the early stages of fever, while respiratory signals are more critical during high fever); the same body temperature value has different pathological significance under different physiological states (normal, low fever, high fever), requiring differentiated fusion of multimodal features (the weight of blood oxygen saturation needs to be higher in high fever than in normal states); traditional static weighted models (fixed weight fusion or single LSTM gating) cannot adapt to such dynamic correlations, leading to inaccurate feature fusion and reduced prediction specificity; based on body temperature trends and physiological states, the weights of each modality feature are dynamically adjusted, driving the basic gating vector through physiological state labels, solidifying the expert experience weights under different states (including automatically increasing the channel weights of respiratory features in high fever states), generating dynamic compensation quantities through the rate of body temperature change, and strengthening modal features strongly correlated with the current trend in real time (amplifying the sensitivity of heart rate variability features when body temperature rises rapidly); through Real-time sensitivity tuning under state constraints is achieved to avoid the overfitting risk of purely data-driven models; based on LSTM output... Channel-level reweighting is performed on each modal feature. This technology breaks through the limitations of traditional LSTM coarse-grained fusion, enabling the model to prioritize mutation indicators such as heart rate during periods of rapid changes in body temperature, and automatically focus on high-discriminative modalities (coordinated changes in respiration and blood oxygenation) under abnormal conditions such as high fever, significantly improving its generalization ability in complex clinical scenarios. Dynamically adjusting the weights of each modality feature based on body temperature trends and physiological states includes the following steps: S3.21 Extract a scalar value representing the body temperature trend from the hidden states of the Long Short-Term Memory (LSTM) network. (Specifically: input the hidden state vector into a fully connected layer and output a scalar value) As an estimate of the trend of body temperature change ( (The unit is temperature / time), and the hidden state vector is retained for subsequent steps. S3.22, the scalar value representing the trend of body temperature. Input the lightweight state classification network (the lightweight state classification network consists of 3 layers of convolutional neural network, with 16 / 32 / 4 channels per layer and 3 kernels per layer) to classify the physiological state and output the current physiological state label of the newborn (including normal, low fever, high fever). S3.23. Based on the physiological state label, query the corresponding pre-trained static gating vector (this static gating vector is generated by the statistical analysis of the multimodal contribution of historical samples in similar states, representing the importance distribution of each modal feature (including heart rate, respiratory rate, blood oxygen saturation, etc.) to body temperature prediction in the current state), and dynamically adjust the weights in combination with the current body temperature change rate to form the final gating weight vector. Furthermore, based on the current rate of change in body temperature, the weights are dynamically adjusted to form the final gating weight vector, including the following steps: S3.231, Use scalar values ​​that represent the trend of body temperature. As an indicator of the rate of change of body temperature per unit time, it is used to represent the direction and intensity of the current trend of body temperature change (including the increase, decrease or fluctuation range). S3.232. Based on the direction (positive / negative) and absolute value (rate strength) of the slope, construct the corresponding dynamic adjustment amount. This vector is used to reflect the sensitivity compensation of modal information under different rates of body temperature change, where, In the formula, For learnable scaling factor, A learnable weight vector (dimension 1) This means that through the training process of a neural network model, a large amount of newborn body temperature and physiological data is used to automatically learn and adjust through optimization algorithms (such as gradient descent); S3.233. Combine the static gating vector and the dynamic adjustment amount to calculate the final gating weight vector. (It is a weight vector that has been non-linearly normalized, with values ​​ranging from [0,1]). In the formula, It is a sigmoid nonlinear activation function. Current physiological state The generated basic gating vector Current physiological state; S3.24. Feature vectors output by Long Short-Term Memory (LSTM) networks Element-wise scaling with gated weight vectors is applied to achieve dynamic weighted fusion of features from different modalities, forming the adjusted feature vector output by the Long Short-Term Memory (LSTM) network. , In the formula, This involves Hadamard product (element-wise multiplication) to achieve adaptive reweighting of feature channels across modalities; specifically, a static gating vector and a dynamic adjustment vector. and the final gating weight vector All dimensions are The dimension d is consistent with that of the feature vector output by the Long Short-Term Memory network.

[0020] S3.3, The feature vector output by the adjusted Long Short-Term Memory (LSTM) network. The input is fed into a fully connected neural network (Multilayer Perceptron MLP) to perform discriminative modeling of abnormal body temperature risk. The basic architecture of the fully connected neural network (Multilayer Perceptron MLP) includes an input layer, one or more hidden layers, and an output layer. The input layer receives feature vectors from the output of LSTM or other modules. The hidden layer consists of several fully connected neurons, equipped with an activation function (ReLU) to introduce nonlinearity. The number of layers and the number of neurons per layer can be adjusted according to the task complexity. The final output layer uses the Sigmoid activation function to compress the result to the 0~1 range, which is used to output the probability of abnormal body temperature risk at the current time. S3.4. Generate the probability value of abnormal body temperature at the current moment through the output layer of the fully connected network; The probability value of abnormal body temperature is: ; In the formula, For the current moment The probability value of abnormal body temperature (ranging from 0 to 1, representing the probability of abnormal body temperature). This is the Sigmoid function, used to compress a linear output to the range of 0 to 1. This is the weight matrix of the output layer of a fully connected neural network (i.e., a multilayer perceptron, MLP), with dimension 1. ,in It is the dimension of the input features. This is a bias term, a scalar.

[0021] S4. Risk classification judgment is made based on the probability of abnormal risks and combined with risk classification rules; In this embodiment, risk classification judgment is performed based on the probability of abnormal risks and combined with risk grading rules, including the following steps: S4.1 Receive the probability value of abnormal body temperature output in step S3; S4.2. Classify body temperature risk levels based on the probability value of abnormal body temperature risk. Body temperature risk levels include low risk (normal), medium risk (suspicious), and high risk (abnormal). Low risk (normal): ; Medium risk (suspicious): ; High risk (abnormal): ; in, The body temperature risk threshold, This is the upper limit threshold for body temperature risk. S4.3. Determine the risk level of the current body temperature status based on the range of the probability value of abnormal body temperature. S4.4 Output the current body temperature risk level result.

[0022] Example 2: This example provides a neonatal body temperature abnormality prediction device, including: The data acquisition module is used to collect and preprocess the newborn's body temperature and physiological signals in real time. The feature extraction module is used to extract the temporal features reflecting changes in body temperature from body temperature and physiological signals using fast Fourier transform, generate feature vectors, and process body temperature and physiological signals through phase coupling analysis and frequency domain masking suppression mechanism to dynamically identify the low-frequency noise interference range caused by irregular breathing and skin conductivity disturbances in newborns. The abnormal body temperature prediction module is used to predict the probability of abnormal body temperature in real time using a long short-term memory network. It also optimizes the feature fusion mechanism by dynamically adjusting the gating weights of the long short-term memory network through a dual factor of body temperature change rate and physiological state label. The risk classification module is used to classify risks by utilizing the probability of abnormal risks and combining them with risk grading rules.

[0023] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the neonatal body temperature abnormality prediction method described above.

[0024] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method of predicting neonatal abnormal body temperature, characterized by, The method comprises the following steps: S1, real-time acquisition of the body temperature and physiological signals of the newborn and preprocessing; S2, based on the fast Fourier transform, extracting the time sequence characteristics reflecting the body temperature changes in the body temperature and physiological signals, generating a characteristic vector, and processing the body temperature and physiological signals through phase coupling analysis and frequency domain mask suppression mechanism, dynamically identifying the low-frequency noise interference interval caused by irregular breathing and skin conductive disturbance of the newborn; S3, based on the characteristic vector, using the long short-term memory network to predict the body temperature abnormal risk probability in real time, and through the body temperature change rate and physiological state label double factors, dynamically regulating the gating weight of the long short-term memory network, optimizing the feature fusion mechanism; S4, based on the abnormal risk probability and combining the risk classification rules to judge the risk classification.

2. The method of claim 1, wherein: In the S2, the time sequence characteristics reflecting the body temperature changes are extracted from the body temperature and physiological signals, comprising the following steps: S2.1, using a fixed time window to divide the continuous body temperature and physiological signal time sequence into multiple analysis segments, forming a sliding window structure; S2.2, in each sliding window, calculating the time domain characteristics of the body temperature and various physiological signals; S2.3, for the body temperature and physiological signal sequence in each sliding window, using the fast Fourier transform to extract the frequency domain characteristics for the periodic fluctuations in the body temperature and physiological signals; S2.4, using a cross-modal analysis method, based on the body temperature and physiological signal sequence of the sliding window, calculating the dynamic coupling characteristics, reflecting the cooperative fluctuation relationship between the body temperature changes and other physiological activities; S2.5, filtering out time domain features, frequency domain features and dynamic coupling features reflecting body temperature changes by a feature selection algorithm, and combining the filtered features into a unified feature vector , , wherein, is the i-th feature of the original feature vector, is the i-th feature of the original feature vector, is the i-th feature of the original feature vector, is the i-th feature of the original feature vector, is the number of final selected features, is the feature index, is the filtered feature index.

3. The method of claim 2, wherein: In S2.3, for the body temperature and physiological signal sequence in each sliding window, using the fast Fourier transform to extract the frequency domain characteristics, comprising the following steps: S2.31, processing the body temperature and physiological signal sequence through phase coupling analysis and frequency domain mask suppression mechanism, dynamically identifying the low-frequency noise interference interval caused by irregular breathing and skin conductive disturbance, and generating an enhanced time domain signal; S2.32, applying the fast Fourier transform to the enhanced time domain signal, converting the enhanced time domain signal into a frequency domain complex spectrum; S2.33, calculating the amplitude spectrum corresponding to each frequency domain complex spectrum, selecting the main frequency component amplitude in the frequency band, and constructing a frequency domain characteristic vector; S2.34, normalizing each frequency domain characteristic vector; S2.35, splicing and fusing all the normalized frequency domain characteristic vectors to construct a joint frequency domain characteristic vector.

4. The method of claim 3, wherein: In the S2.31, the body temperature and physiological signal sequence are processed through the phase coupling analysis and frequency domain mask suppression mechanism to dynamically identify the low-frequency noise interference interval caused by irregular breathing and skin conductive disturbance, comprising the following steps: S2.311, extracting the RR interval sequence from the physiological signals, calculating the instantaneous heart rate variability, synchronously deriving the respiratory signal, using Hilbert transform to obtain the instantaneous phase of the instantaneous heart rate variability and the respiratory signal, and calculating the phase difference therebetween, when the phase difference is greater than an upper threshold a, marking an x interval centered at the current time as a respiratory interference interval; S2.312, calculating a second derivative of the body temperature, and if the absolute value of the second derivative is greater than a derivative threshold, then expanding a fixed length window centered on the current time point as the skin conduction interference interval then expanding a fixed length window centered on the current time point as the skin conduction interference interval; S2.313, extract the reference noise signal of the respiratory disturbance interval and the skin conductive disturbance interval, window the reference noise signal and perform fast Fourier transform to obtain a complex spectrum, and take the module length of the complex spectrum to obtain a disturbance spectrum template; S2.314, divide the frequency band, calculate the disturbance energy proportion according to the frequency band, and construct a frequency domain mask based on the disturbance spectrum template; S2.315, apply a window function to the original body temperature and physiological signals, perform fast Fourier transform, multiply the complex spectrum by the constructed frequency domain mask, and restore the processed complex spectrum to an enhanced time domain signal through inverse Fourier transform.

5. The neonatal abnormal body temperature prediction method of claim 1, wherein: In the S3, the long short-term memory network is used to predict the body temperature abnormality risk probability in real time based on the feature vector, including the following steps: S3.1, organize the feature vectors of each time period constructed in a sliding window manner into a time sequence input sequence in time sequence; S3.2, input the time sequence input sequence into a multi-layer stacked long short-term memory network, model the time dependence between features through the gating mechanism thereof, extract deep time sequence features representing body temperature change trends, and dynamically adjust the feature weight of each mode based on the body temperature trend and physiological state; S3.3, the feature vector output by the adjusted long short-term memory network to the subsequent fully connected neural network, and discriminant modeling of the body temperature abnormality risk is performed; S3.4, generate the body temperature abnormality risk probability value at the current time through the output layer of the fully connected network.

6. The method of claim 5, wherein: In the S3.2, the feature weight of each mode is dynamically adjusted based on the body temperature trend and physiological state, including the following steps: S3.21, extracting a scalar value representative of the temperature trend from the long short-term memory network hidden state ; S3.22, the scalar value representing the body temperature trend The input lightweight state classification network is used to classify the physiological state, and output the physiological state label of the newborn at the current time. S3.23, according to the physiological state label, query the corresponding pre-trained static gating vector, and dynamically adjust the weight by combining the current body temperature change rate to form a final gating weight vector; S3.24, the feature vector output by the long short-term memory network The gating weight vector is applied element-wise to scale the feature vector output by the long short-term memory network .

7. The method of claim 6, wherein: In the S3.23, the weight is dynamically adjusted by combining the current body temperature change rate to form a final gating weight vector, including the following steps: S3.231, using a scalar value representative of the body temperature trend As a speed indicator of the body temperature change per unit time, used to represent the change direction and intensity of the current body temperature trend; S3.232, construct a corresponding dynamic adjustment amount according to the direction and absolute value of the slope; S3.233, calculate the final gating weight vector by combining the static gating vector and the dynamic adjustment amount.

8. The method of claim 1, wherein: In the S4, the risk classification is determined based on the abnormality risk probability and combined with the risk classification rule, including the following steps: S4.1, receive the body temperature abnormality risk probability value output by step S3; S4.2, divide the body temperature risk level based on the body temperature abnormality risk probability value; S4.3, determine the risk level of the current body temperature state according to the interval in which the body temperature abnormality risk probability value is located; S4.4, output the body temperature risk level result at the current time.

9. A neonatal abnormal body temperature prediction device characterized by comprising: It includes: A collection module for collecting and preprocessing the body temperature and physiological signals of the newborn in real time; A feature extraction module for extracting time sequence features reflecting body temperature changes in the body temperature and physiological signals using fast Fourier transform, generating a feature vector, and dynamically identifying low-frequency noise disturbance intervals caused by irregular breathing and skin conductive disturbance of the newborn by processing the body temperature and physiological signals through phase coupling analysis and frequency domain mask suppression mechanism; A body temperature abnormality prediction module for predicting the body temperature abnormality risk probability in real time using a long short-term memory network, and dynamically regulating the gating weight of the long short-term memory network through the body temperature change rate and the physiological state label to optimize the feature fusion mechanism; A risk classification module is configured to perform risk classification by using the abnormal risk probability and in combination with a risk grading rule.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program, when executed by a processor, implements the steps of the new-born abnormal body temperature prediction method according to any one of claims 1 to 8.

Images