Auxiliary identification method and device for monophasic depression and biphasic depression

Abstract

The invention provides an auxiliary identification method and device for single-phase depression and double-phase depression. The method comprises the steps that a millimeter wave frequency modulation continuous wave radar is used for collecting thoracic cavity micro-motion signals of a subject in a non-contact mode; filtering and phase demodulation are carried out on the thoracic cavity micro-motion signal, respiration components and heartbeat components are separated, and an RR interval sequence is extracted from the heartbeat components; calculating heart rate variability characteristics according to the RR interval sequence, wherein the heart rate variability characteristics comprise a time domain index, a frequency domain index and a nonlinear index; and inputting the heart rate variability characteristics into a pre-trained auxiliary identification model to obtain a depression identification result of the subject. According to the invention, a non-contact measurement method is adopted, so that stress reaction possibly caused by a contact sensor is avoided, and interference to physiological signals is reduced; and meanwhile, complex features in the data can be automatically extracted by using a deep learning model, so that the identification accuracy and generalization ability are improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to an auxiliary method and apparatus for identifying unipolar depression and bipolar depression. Background Technology

[0002] Depression is a common mental disorder that severely impacts patients' quality of life. Differentiating between unipolar and bipolar depression is crucial for developing appropriate treatment plans, but clinical diagnosis presents challenges.

[0003] Currently, the diagnosis of depression mainly relies on subjective assessments by clinicians, such as structured interviews, questionnaires, and medical history reviews. These methods are time-consuming and easily influenced by the doctor's experience and the patient's subjective statements, resulting in limited accuracy. In addition, brain imaging techniques, such as fMRI and PET, are also used for diagnosis. While these can provide some physiological information, they are costly, complex to operate, and difficult to popularize. Summary of the Invention

[0004] This invention provides an auxiliary method and apparatus for identifying unipolar depression and bipolar depression, in order to overcome the deficiencies in the prior art.

[0005] This invention provides an auxiliary method for differentiating between unipolar depression and bipolar depression, comprising the following steps: Millimeter-wave frequency-modulated continuous wave radar was used to collect the micro-motion signals of the subject's chest cavity in a non-contact manner. The thoracic cavity micromotion signal is filtered and phase demodulated to separate the respiratory and cardiac components, and the RR interval sequence is extracted from the cardiac component. Heart rate variability characteristics are calculated based on the RR interval sequence, and the heart rate variability characteristics include time-domain indicators, frequency-domain indicators, and nonlinear indicators; The heart rate variability features are input into a pre-trained auxiliary identification model to obtain the depression identification result of the subject; the auxiliary identification model is trained based on historical heart rate variability features and corresponding depression labels.

[0006] According to the present invention, an auxiliary method for differentiating unipolar and bipolar depression includes filtering and demodulating the thoracic cavity micromotion signal, separating the respiratory and cardiac components, and extracting the RR interval sequence from the cardiac component, comprising: The thoracic cavity micro-motion signal is bandpass filtered to remove noise and power frequency interference; Phase demodulation is performed on the filtered signal to obtain a baseband signal containing respiratory and heartbeat information; The respiratory and cardiac components were separated using independent component analysis or empirical mode decomposition. Peak detection was performed on the isolated heartbeat components to obtain the heartbeat peak sequence; The time interval between adjacent heartbeat peaks is calculated to obtain the RR interval sequence.

[0007] According to the present invention, an auxiliary method for differentiating between unipolar and bipolar depression is provided, wherein calculating heart rate variability characteristics based on the RR interval sequence includes: Based on the statistical characteristics of the RR interval sequence, time-domain analysis is performed to obtain the time-domain index, which includes at least one of the following: standard deviation of NN interval, root mean square of the difference between adjacent NN intervals, percentage of adjacent NN intervals with a difference greater than 50ms, mean NN interval, heart rate variability trigonometric index, and trigonometric interpolation of the NN interval histogram. The power spectral density is estimated by performing power spectral density estimation on the RR interval sequence to obtain the power spectral density, and the power spectral density is integrated within a specified frequency band to obtain the frequency domain index. The frequency domain index includes at least one of the following: extremely low frequency power, low frequency power, high frequency power, low frequency to high frequency power ratio, total power, and normalized power. The nonlinear index is obtained by performing graph analysis, entropy analysis, fractal analysis, and information theory analysis on the RR interval series. The nonlinear index includes at least one of graph parameters, sample entropy, approximate entropy, multi-scale entropy, scaling index of detrended fluctuation analysis, correlation dimension, transition entropy, and mutual information.

[0008] According to the present invention, an auxiliary identification method for unipolar and bipolar depression is provided, wherein the heart rate variability features are input into a pre-trained auxiliary identification model to obtain the depression identification result of the subject, comprising: The heart rate variability features are subjected to explicit feature screening or implicit feature importance analysis to determine important features from the heart rate variability features; The key features are input into the auxiliary identification model to obtain the depression identification results of the subject.

[0009] According to the present invention, an auxiliary identification method for unipolar and bipolar depression is provided, wherein inputting the important features into the auxiliary identification model to obtain the depression identification result of the subject includes: The important features are input into the input layer of the auxiliary discrimination model, and the input layer standardizes the important features to obtain the original features; The original features are input into the embedding layer of the auxiliary discrimination model. The embedding layer maps the original features to the latent space and then performs position encoding on the mapped features to obtain the encoded features. The encoded features are input into the attention layer of the auxiliary discrimination model, and the attention layer performs multi-head attention aggregation on the encoded features to obtain attention features; The attention features are input into the feedforward layer of the auxiliary discrimination model, where the feedforward layer enhances the attention features, and then performs residual connection and layer normalization on the enhanced features to obtain the normalized enhanced representation features. The normalized enhanced representation features are input into the integration layer of the auxiliary discrimination model. The integration layer stacks multiple encoder layers and integrates features from different levels using skip connections or residual connections to obtain multi-scale fusion features. The multi-scale fused features are input into the aggregation layer of the auxiliary discrimination model. The aggregation layer uses attention pooling or global average pooling to aggregate the features in the time dimension, generating a fixed-length feature vector, and thus obtaining the global feature vector. The global feature vector is input into the classification layer of the auxiliary identification model, and the classification layer performs classification through a fully connected layer and an activation function to obtain the depression identification result.

[0010] According to the auxiliary identification method for unipolar depression and bipolar depression provided by the present invention, the step of performing implicit feature importance analysis on the heart rate variability features to determine important features from the heart rate variability features includes: The attention weights of each heart rate variability feature are calculated using the attention mechanism in the auxiliary discrimination model. Based on the attention weights of each heart rate variability feature, the importance score of each heart rate variability feature is determined. Heart rate variability features with importance scores greater than a threshold are considered as important features.

[0011] According to the present invention, an auxiliary identification method for unipolar depression and bipolar depression is provided, wherein the training steps of the auxiliary identification model include: The training dataset is preprocessed, and the training dataset includes historical subjects' heart rate variability characteristics and corresponding depression labels; The training dataset is divided into a training set, a validation set, and a test set; Training the auxiliary discrimination model using the training set includes: Define the loss function; Select an optimizer, which includes Adam, SGD, or other gradient descent optimization algorithms; Set the batch size and the number of training rounds, and divide the training set into multiple batches of training data according to the batch size; In each training epoch, a batch of training data is input into the model, the loss value is calculated, and the model parameters are updated using the optimizer based on the loss value.

[0012] The present invention also provides an auxiliary device for differentiating unipolar depression and bipolar depression, comprising the following modules: The acquisition unit is used to non-contactly acquire the chest cavity micro-motion signals of the subject using millimeter-wave frequency-modulated continuous wave radar; The separation unit is used to filter and demodulate the thoracic cavity micro-motion signal, separate the respiratory component and the heartbeat component, and extract the RR interval sequence from the heartbeat component; The calculation unit is used to calculate heart rate variability characteristics based on the RR interval sequence, wherein the heart rate variability characteristics include time-domain indicators, frequency-domain indicators, and nonlinear indicators; The identification unit is used to input the heart rate variability features into a pre-trained auxiliary identification model to obtain the depression identification result of the subject; the auxiliary identification model is trained based on historical heart rate variability features and corresponding depression labels.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the auxiliary identification method for unipolar depression and bipolar depression as described above.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the auxiliary identification method for unipolar depression and bipolar depression as described above.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the auxiliary identification method for unipolar depression and bipolar depression as described above.

[0016] This invention provides a method and apparatus for auxiliary identification of unipolar and bipolar depression. It acquires chest cavity micro-motion signals using non-contact millimeter-wave radar, extracts heart rate variability features, and utilizes a pre-trained auxiliary identification model to predict depression type, achieving objective, efficient, and accurate auxiliary identification of depression. Because it employs a non-contact measurement method, it avoids the stress response that may be caused by contact sensors, reducing interference with physiological signals. Simultaneously, the use of a deep learning model automatically extracts complex features from the data, improving identification accuracy and generalization ability. This provides clinicians with more reliable diagnostic evidence, assists in developing more appropriate treatment plans, avoids the subjectivity and limitations of traditional methods, and improves diagnostic efficiency and accuracy. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is one of the flowcharts illustrating the auxiliary identification method for unipolar depression and bipolar depression provided by the present invention.

[0019] Figure 2 This is the second flowchart of the auxiliary identification method for unipolar depression and bipolar depression provided by the present invention.

[0020] Figure 3 This is a schematic diagram of the ROC curve for model performance evaluation provided by the present invention.

[0021] Figure 4 This is a schematic diagram of the auxiliary identification device for unipolar depression and bipolar depression provided by the present invention.

[0022] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] Unipolar depression is a type of depression characterized primarily by persistent low mood. Patients mainly experience symptoms such as depressed mood, loss of interest, and lack of energy, and have no history of manic or hypomanic episodes. Bipolar depression refers to the depressive episodes of bipolar disorder, in which patients experience both depressive and manic or hypomanic episodes, with alternating mood states. Currently, the clinical methods for differentiating between unipolar and bipolar depression include the following: Firstly, clinical interviews and questionnaires are used. Doctors conduct structured or unstructured interviews with patients, combined with questionnaires such as the Self-Rating Depression Scale and the Hamilton Depression Rating Scale (HAMD), to assess the patient's symptoms, medical history, and family history. This method is highly dependent on the doctor's subjective judgment and experience, and assessment results may vary between different doctors. Patients' subjective statements may be influenced by factors such as social expectations and memory biases, leading to inaccurate information. In addition, interviews and questionnaires are usually time-consuming and inefficient.

[0025] Secondly, brain imaging techniques, such as functional magnetic resonance imaging (fMRI) and positron emission tomography (PET), can detect changes in brain activity, providing some physiological evidence for the diagnosis of depression. However, brain imaging examinations are expensive and require specialized equipment and operators. The scanning process is relatively complex, and patient cooperation may be limited. Furthermore, the interpretation of brain imaging results requires specialized knowledge and experience, and inconsistencies may exist between different studies.

[0026] Third, machine learning algorithms, such as Support Vector Machines (SVM) and Random Forests, can be used to identify depression based on clinical data (e.g., demographic information, symptom scores) or physiological signals (e.g., electrocardiograms, electroencephalograms). Traditional machine learning algorithms typically require manually designed features, which rely on the knowledge and experience of domain experts. Manually extracted features may not fully capture the complex information in the data. Furthermore, the performance of traditional machine learning algorithms largely depends on feature selection and parameter tuning, potentially limiting their generalization ability. They are particularly robust to mental illnesses with significant individual variability.

[0027] In summary, traditional methods for identifying depression generally suffer from problems such as high subjectivity, high cost, complex operation, limited accuracy, and poor robustness, making them difficult to meet clinical needs. Therefore, there is an urgent need to develop an objective, efficient, and accurate method for identifying depression.

[0028] To address this, the present invention provides an auxiliary method for differentiating between unipolar and bipolar depression. Specifically, Figure 1 This is one of the flowcharts illustrating the auxiliary identification method for unipolar and bipolar depression provided by this invention. Figure 2 This is the second flowchart of the auxiliary identification method for unipolar and bipolar depression provided by the present invention, as shown below. Figure 1 and Figure 2 As shown, the method includes steps 110, 120, 130 and 140.

[0029] Step 110: Use millimeter-wave frequency-modulated continuous wave radar to non-contactly collect the subject's chest cavity micro-motion signals.

[0030] Here, millimeter-wave frequency-modulated continuous wave radar (FMCW radar) refers to a type of radar operating in the millimeter-wave frequency band (30GHz-300GHz). It measures target distance and velocity by transmitting a continuous wave signal whose frequency varies linearly with time and analyzing the frequency difference between the received and transmitted signals. "Non-contact" means that the measurement is performed without direct contact with the subject's body, avoiding the discomfort and interference that contact sensors may cause. Chest micromotion signals refer to the minute displacement changes on the surface of the chest cavity caused by heartbeats and respiratory movements, containing rich cardiovascular and respiratory physiological information. Non-contact acquisition of chest micromotion signals from subjects can avoid the stress response that may be caused by traditional contact measurement methods, reduce interference with physiological signals, and improve the comfort and repeatability of data acquisition.

[0031] For example, a 77-81 GHz band FMCW millimeter-wave radar can be used, placed 1-2 meters away from the subject's chest at an angle within ±15°. The subject remains at rest, and non-contact monitoring is performed for 5-10 minutes. IQ demodulation is used to acquire complex signals containing micron-level chest cavity movement information, enabling precise capture of micro-displacements caused by heartbeat and respiration. The subject can be someone initially diagnosed with depression, requiring further differentiation between unipolar and bipolar depression.

[0032] After the radar is placed, it can be powered on and self-calibrated to ensure stable transmission frequency and power, detect the environmental noise baseline, automatically adjust gain parameters, and confirm the data transmission link is unobstructed. After self-checking, a rapid scan is performed to determine the subject's position. A fine scan is performed within the determined distance range, and a dynamic tracking area is set to automatically compensate for minor body position changes. After tracking and positioning are completed, the received reflected signal is converted into I (in-phase) and Q (quadrature) components by a mixer, and the I / Q signals are sampled by an ADC at a sampling rate of up to 2000Hz. After sampling, the sampled data undergoes preliminary filtering to eliminate high-frequency noise, and the raw I / Q data is transmitted to the computing platform for further processing. A range-velocity map (RDM) is constructed based on the collected data. This involves performing a Fast Fourier Transform (FFT) on the data from each sweep cycle to obtain the range dimension information, performing a second FFT on the data from multiple sweep cycles to obtain the velocity (Doppler) dimension information, and generating a two-dimensional RDM matrix representing the velocity distribution of the target at different distances. After obtaining the RDM, the improved CFAR (Constant False Alarm Rate) algorithm can be applied to detect thoracic motion signals: in, Indicates position Adaptive threshold at the location, Represents a sliding window Size, Indicates position The test results from the relevant department Indicates the signal strength in RDM. It's a sliding window. =1.5, =3dB.

[0033] It should be noted that, compared to the traditional CFAR algorithm, this section introduces... A fixed offset of 3dB effectively suppresses background noise, preventing it from being misidentified as a target signal and thus reducing the false alarm rate. Simultaneously, the sliding window is dynamically adjusted. The size and position of the device can better adapt to changes in the range and amplitude of chest cavity movement, avoiding the inclusion of the movement of other objects or environmental interference, thereby improving the detection accuracy of the target signal. In other words, compared to traditional methods, the embodiments of this invention can more accurately detect weak chest cavity movement signals in complex environments, reducing the impact of noise and interference, and improving the sensitivity and accuracy of signal detection.

[0034] Furthermore, after detecting chest movement signals, a Kalman filter can be applied for target tracking to improve the stability and accuracy of chest movement signal detection, reduce the influence of noise and interference, and compensate for the subject's subtle body movements. The state equation and measurement equation for target tracking are as follows: in, It is a state vector. These are measured values. and These are process noise and measurement noise, respectively. It is the state transition matrix. It is the observation matrix.

[0035] Step 120: Filter and demodulate the thoracic cavity micromotion signal to separate the respiratory and cardiac components, and extract the RR interval sequence from the cardiac component.

[0036] Considering that the raw chest cavity micromotion signals acquired by radar contain various noises and interferences, such as high-frequency noise, power frequency interference, and motion artifacts, which can affect subsequent signal analysis and feature extraction, filtering and phase demodulation are performed on the signal after acquisition to improve signal quality and highlight effective information. Filtering refers to removing unwanted frequency components from the signal while retaining the target frequency components; for example, a bandpass filter can be used to filter the chest cavity micromotion signals. Phase demodulation refers to extracting phase information from the received radar signal. Phase information reflects the displacement changes of the target object; for example, an arctangent function demodulation can be used to demodulate the phase of the chest cavity micromotion signals. Given that filtering can effectively remove noise and improve the accuracy of phase demodulation, this embodiment of the invention preferably filters the chest cavity micromotion signals first, and then performs phase demodulation on the filtered signal.

[0037] Furthermore, the filtered and phase-demodulated thoracic micromotion signal contains respiratory and cardiac components. Typically, in depression identification, the RR interval sequence within the cardiac component is closely related to the activity of the autonomic nervous system and can reflect an individual's emotional state and mental health level. Therefore, this embodiment of the invention separates the cardiac component from the signal and extracts the RR interval sequence from it. The RR interval sequence refers to the time interval sequence between consecutive heartbeats, i.e., the time interval sequence between adjacent R-wave peaks.

[0038] The RR interval sequence can be obtained based on the following formula: The radar echo signal is converted into thoracic displacement information through phase demodulation: in, This indicates information about chest cavity displacement. Indicates the radar wavelength. Indicates phase change, Indicates the sampling time The I signal, Indicates the sampling time Q signal, Indicates the sampling time The I signal, Indicates the sampling time The Q signal.

[0039] Meanwhile, before separating the heartbeat component, signal enhancement techniques can be applied to improve the signal-to-noise ratio of the weak heartbeat component: in, This indicates the enhanced signal. Represents the original signal. Indicates enhancement filter, This represents variational mode decomposition. This represents the mixing coefficient, with a value of 0.3.

[0040] Next, a bandpass filter is applied to separate the respiratory components. (0.1-0.5Hz) and heartbeat components (0.8-2Hz): in, This represents the impulse response of a bandpass filter.

[0041] Furthermore, an adaptive wavelet transform is applied to the heartbeat component to improve the signal-to-noise ratio: in, This represents the heartbeat component after adaptive wavelet transform. Represents discrete wavelet transform. This represents the inverse discrete wavelet transform. This represents the threshold function, and a soft threshold can be used.

[0042] The transformed heartbeat component is then bandpass filtered to confine the signal within the effective frequency band, filtering out baseline drift and high-frequency noise. The specific formula is as follows: in, Indicates the output signal. Represents the forward coefficient, Indicates the input signal and its past The delay value at each moment, Indicates the feedback coefficient. Indicates the past The output feedback value at each moment Indicates the discrete-time sequence number. This indicates the delay order index.

[0043] The transformed heartbeat components are then subjected to a differential operation to enhance the high-frequency components and highlight the rapid changes in the R wave. The specific formula is as follows: in, This represents the heartbeat signal after the differential operation. This represents the heartbeat components after the above transformation.

[0044] The heartbeat component after the above differential operation is squared to eliminate negative values ​​and enhance large amplitude components. The specific formula is as follows: in, This represents the squared heart rate component.

[0045] The heart rate component after squaring the above signal is then subjected to a moving average to smooth the signal and form an envelope. The specific formula is as follows: ( =30) in, This represents the envelope after a moving average.

[0046] Finally, adaptive threshold detection is performed on the envelope after moving average. This allows for dynamic adjustment of the detection threshold based on signal characteristics, leading to more accurate detection of heartbeat peaks and reduced false positives and false negatives. The detection threshold... It can be adjusted based on the following formula: in, This represents the mean of the envelope. The variance of the envelope is represented. =0.35, =1.5.

[0047] The above-mentioned bandpass filtering, differential operation, signal squaring, moving average, and adaptive threshold detection of heartbeat components can effectively extract the features of heartbeat signals, reduce the influence of noise and interference, and adaptively adjust the detection threshold to improve the accuracy and robustness of heartbeat detection. This avoids missed heartbeats due to weak signals or noise interference, as well as threshold incompatibility due to individual differences or environmental changes.

[0048] After applying bandpass filtering, differential operation, signal squaring, moving average, and adaptive threshold detection to the heartbeat components, artifact removal and verification are performed on the detected heartbeat peaks, as detailed below: First, exclude intervals that do not fall within the physiologically reasonable range: in, Indicates the first RR intervals, This indicates the effective RR interval, which falls within a physiologically reasonable range.

[0049] Next, outliers with excessively large variations from adjacent intervals are identified and removed: in, Indicates the current RR interval The moving average of the previous 5 RR intervals.

[0050] After artifact removal and verification, the RR interval sequence is determined, and anomaly repair, resampling, and interpolation are performed on the RR interval sequence, as detailed below: Perform difference correction on the detected true or abnormal RR intervals: in, Indicates the repaired RR interval. Indicates the target time point. Indicates the first A valid heartbeat time point Indicates the first A valid RR interval value This represents the spline interpolation function. This represents a valid set of data.

[0051] In addition, the unequal-interval RR interval sequences were converted into equally-interval sampling sequences: in, Indicates in RR interval after time point resampling Indicates the original time point The RR interval value after repair, Indicates the first The actual time of each heartbeat This represents the defined uniform time grid points. This represents cubic spline interpolation. It is the time point for uniform sampling at 4Hz.

[0052] Step 130: Calculate heart rate variability characteristics based on the RR interval sequence. Heart rate variability characteristics include time-domain indicators, frequency-domain indicators, and nonlinear indicators.

[0053] Specifically, heart rate variability (HRV) refers to the degree to which the time interval between heartbeats (i.e., the RR interval) changes over time. It reflects the autonomic nervous system's ability to regulate the heart and is used to characterize the activity and balance of the autonomic nervous system.

[0054] Heart rate variability characteristics include time-domain indices, frequency-domain indices, and nonlinear indices. Time-domain indices are calculated based on the statistical characteristics of the RR interval sequence, describing the overall variation and dispersion of the RR interval sequence. Frequency-domain indices are obtained through spectral analysis of the RR interval sequence, reflecting heart rate fluctuations at different frequency components and the activity of the sympathetic and parasympathetic nervous systems. Nonlinear indices are calculated based on nonlinear dynamics theory, describing the complexity and chaos of the RR interval sequence.

[0055] Step 140: Input the heart rate variability features into the pre-trained auxiliary identification model to obtain the depression identification results of the subjects; the auxiliary identification model is trained based on historical heart rate variability features and corresponding depression labels.

[0056] Specifically, since heart rate variability (HRV) characteristics can reflect the functional state of the autonomic nervous system, and the autonomic nervous system is closely related to emotion regulation and mental disorders, combining HRV characteristics can objectively assess an individual's emotional state and assist in the diagnosis and differential diagnosis of depression. In this embodiment of the invention, HRV characteristics are input into a pre-trained auxiliary identification model. This model learns the complex relationship between HRV characteristics and depression type, and predicts the subject's depression type based on the input HRV characteristics, obtaining the subject's depression identification result. This result is used to distinguish whether the subject has unipolar or bipolar depression. For example, this result can be the probability that the subject has unipolar or bipolar depression, or it can be the depression type classification result directly given by the model.

[0057] The auxiliary identification model is trained based on historical heart rate variability features and corresponding depression labels. It can be a deep learning model, such as a convolutional neural network (CNN), a recurrent neural network (RNN), a long short-term memory network (LSTM), a Transformer network, or a combination thereof, or a traditional machine learning model, such as a support vector machine (SVM) or a random forest.

[0058] The auxiliary identification method for unipolar and bipolar depression provided in this invention uses non-contact millimeter-wave radar to collect chest cavity micro-motion signals, extracts heart rate variability features, and uses a pre-trained auxiliary identification model to predict depression type, achieving objective, efficient, and accurate auxiliary identification of depression. Because a non-contact measurement method is used, it avoids the stress response that may be caused by contact sensors, reducing interference with physiological signals. Simultaneously, the use of a deep learning model can automatically extract complex features from the data, improving identification accuracy and generalization ability. This provides clinicians with more reliable diagnostic evidence, assists in developing more appropriate treatment plans, avoids the subjectivity and limitations of traditional methods, and improves diagnostic efficiency and accuracy.

[0059] Based on the above embodiments, the thoracic cavity micromotion signal is filtered and phase demodulated to separate the respiratory and cardiac components, and the RR interval sequence is extracted from the cardiac component, including: Bandpass filtering is applied to the thoracic cavity micro-motion signal to remove noise and power frequency interference; Phase demodulation is performed on the filtered signal to obtain a baseband signal containing respiratory and heartbeat information; The respiratory and cardiac components were separated using independent component analysis or empirical mode decomposition. Peak detection was performed on the isolated heartbeat components to obtain the heartbeat peak sequence; The time interval between adjacent heartbeat peaks is calculated to obtain the RR interval sequence.

[0060] Specifically, bandpass filtering of thoracic cavity micromotion signals refers to selecting an appropriate frequency range, allowing only signals within that range to pass through, while suppressing signals in other frequency ranges. This can filter out noise and power frequency interference in thoracic cavity micromotion signals, thereby improving the signal-to-noise ratio and making subsequent signal processing more accurate.

[0061] After filtering, considering that the signal acquired by the radar is a modulated signal, it needs to be demodulated to obtain the original baseband signal. The filtered signal is phase demodulated to obtain the baseband signal containing breathing and heartbeat information. This signal contains information on the changes in chest cavity displacement caused by breathing and heartbeat.

[0062] Since the respiratory and heartbeat components are mixed together in the baseband signal, it is difficult to directly extract the RR interval sequence. Therefore, it is necessary to separate the respiratory and heartbeat components. In this embodiment of the invention, independent component analysis or empirical mode decomposition is used to separate the respiratory and heartbeat components, thereby obtaining an independent heartbeat component, which facilitates subsequent peak detection and RR interval sequence extraction.

[0063] After separating the heartbeat components, peak detection is performed on these components to determine the time point of each heartbeat, thus obtaining a heartbeat peak sequence. For example, the Pan-Tompkins algorithm or other peak detection algorithms can be used to find the peaks in the heartbeat signal.

[0064] After obtaining the heart rate peak sequence, the time interval between adjacent heart rate peaks is calculated to obtain the RR interval sequence. This sequence reflects the rhythmic changes of the heartbeat and is the basis for heart rate variability analysis.

[0065] Based on any of the above embodiments, calculating heart rate variability characteristics according to the RR interval sequence includes: Time-domain analysis is performed based on the statistical characteristics of the RR interval sequence to obtain time-domain indicators, which include at least one of the following: standard deviation of NN interval, root mean square of the difference between adjacent NN intervals, percentage of adjacent NN intervals with a difference greater than 50ms, mean NN interval, heart rate variability trigonometric index, and trigonometric interpolation of the NN interval histogram. The power spectral density is estimated for the RR interval sequence to obtain the power spectral density. The power spectral density is then integrated over a specified frequency band to obtain frequency domain indices. The frequency domain indices include at least one of the following: extremely low frequency power, low frequency power, high frequency power, low frequency to high frequency power ratio, total power, and normalized power. Graph analysis, entropy analysis, fractal analysis, and information theory analysis are performed on the RR interval series to obtain nonlinear indices. The nonlinear indices include at least one of graph parameters, sample entropy, approximate entropy, multiscale entropy, scaling exponent for detrended fluctuation analysis, correlation dimension, transfer entropy, and mutual information.

[0066] The standard deviation of NN intervals (SDNN) is used to reflect the overall HRV level, and its calculation formula is as follows: in, Indicates the first RR intervals, This represents the average RR interval. This represents the total number of RR intervals.

[0067] The root mean square of successive differences (RMSSD) between adjacent neural inter-neuron intervals primarily reflects parasympathetic nerve activity, and its calculation formula is as follows: in, Indicates the first +1 RR interval.

[0068] The percentage of adjacent neural intervals differing by more than 50 ms (pNN50) reflects the parasympathetic nervous system's regulatory capacity. The calculation formula is as follows: in, This indicates the number of consecutive RR intervals with a difference of more than 50 ms.

[0069] The standard deviation of successive differences (SDSD) is used to reflect short-term variability and is calculated using the following formula: in, This represents the average difference between adjacent RR periods.

[0070] The Heart Rate Variance Triangular Index (HRV Triangular Index) is used for overall HRV assessment, and the calculation formula is as follows: Construct a histogram of RR intervals with a bin width of 20ms, and calculate the ratio of the total number of samples to the highest point of the histogram.

[0071] The overall HRV assessment using the triangular interpolation of the NN interval histogram (TINN) is calculated using the following formula: TINN=MN M and N are baseline widths obtained by fitting triangles onto the RR interval histogram.

[0072] In addition, the following methods can be used to estimate the power spectral density of RR interval sequences: (1) Fast Fourier Transform (FFT): Applying the Hamming window function, the FFT has 1024 points: in, It is a window function. It is the resampled RR interval sequence. For sampling points, For the target frequency, This is the signal length.

[0073] (2) Autoregressive (AR) models are used to improve spectral resolution: in, It is the model order (usually taken as 16-20). It is white noise, and the AR coefficients are estimated using the Burg algorithm. .

[0074] (3) Lomb-Scargle periodogram, suitable for data sampled at unequal intervals: in, Is to make time offset, It is the j-th RR interval. It is the timestamp corresponding to the j-th RR interval. It is angular frequency.

[0075] The power spectral density is estimated by performing power spectral density estimation on the RR interval sequence. The power spectral density is then integrated over a specified frequency band to obtain the extremely low frequency power, low frequency power, high frequency power, low frequency to high frequency power ratio, total power, and normalized power.

[0076] Among them, very low frequency power (VLF), with a specified frequency band of 0.003-0.04Hz, reflects long-term regulatory mechanisms such as body temperature regulation and vasomotor activity. Here, This represents the power spectral density.

[0077] Low frequency (LF), within the specified frequency band of 0.04-0.15Hz, reflects the combined action of the sympathetic and parasympathetic nervous systems, primarily indicating baroreflex activity. High-frequency power (HF), within a specified frequency band of 0.15-0.4Hz, primarily reflects parasympathetic nerve activity and is related to respiration. Total Power (TP), at 0.4 Hz, reflects the overall HRV level: The ratio of low-frequency to high-frequency power (LF / HF ratio) reflects the balance between the sympathetic and non-sympathetic inductance systems. Standardize power to eliminate the effects of total power variations: Wherein, LF norm represents low-frequency normalized power, and HF norm represents high-frequency normalized power.

[0078] In addition, graphical analysis of RR interval sequences includes: Construct a Poincaré scatter plot with the x-axis representing RR. i The vertical axis is RR i+1 And calculate SD1 (the minor axis of the ellipse) and SD2 (the major axis of the ellipse): Based on this, the area and shape parameters (SD1 / SD2) of the ellipse are calculated using the following formula: Entropy analysis was performed on the RR interval series to obtain sample entropy, approximate entropy, and multi-scale entropy. Among these, sample entropy (SampEn) is used to measure the complexity and regularity of the time series. Where m is the template length (value 2), r is the similarity tolerance (value is 0.2 times the standard deviation), A is the number of matches with template length m+1, and B is the number of matches with template length m.

[0079] Approximate entropy (ApEn) is similar to SampEn, but includes self-matching: in, This represents a measure of the regularity of a time series under the embedding dimension m and the tolerance threshold r.

[0080] Multiscale entropy (MSE) analyzes the complexity of signals across multiple time scales: in, Represents the first of the original time series There are 10 data points, where τ is the scaling factor. It is the first under scale τ Each coarse-grained data point represents a coarse-grained time series. Calculate SampEn to generate a curve showing the relationship between entropy and scale.

[0081] Fractal analysis was performed on the RR interval series to obtain the scaling exponent and related dimension of the detrended volatility analysis. The scaling exponent for detrended volatility analysis can be used to evaluate the self-similarity of the signal's entropy at different time scales by performing detrended volatility analysis (DFA) on the RR interval series. (a) Integrating the RR interval series: (b) Divide the integral sequence into windows of length n; (c) Fit a linear trend within each window ; (d) Calculate the root mean square fluctuation: (e) Repeat the calculation for different window lengths n to obtain relation; (f) Short-term variability index (4n6) and long-term variability index (16n64).

[0082] Furthermore, the correlation dimension (CD) estimates the fractal dimension of the system: in, It's the Heaviside function. and It is to reconstruct vectors in phase space. It is the radius threshold. It is the total number of vectors in the reconstructed phase space. For the relevant integrals, The relevant dimension.

[0083] Information theory analysis is performed on the RR interval series to obtain the transition entropy and mutual information. The transition entropy (TE) quantifies the directionality of information flow. in, It is the target sequence Historical embeddings, including One past value; It is the source sequence Historical embeddings, including A past value.

[0084] Mutual information (MI) assesses the interdependence of two time series: Based on any of the above embodiments, heart rate variability features are input into a pre-trained auxiliary identification model to obtain the depression identification results of the subject, including: Perform explicit feature screening or implicit feature importance analysis on heart rate variability features to identify important features from the heart rate variability features; Key features are input into the auxiliary identification model to obtain the depression identification results of the subjects.

[0085] Optionally, L1 regularized regression pre-screening can be used to achieve explicit feature screening. The L1 regularized regression pre-screening can be performed using the following formula: in, It is the objective function. These are the model coefficients to be learned. It is the sample size. It is the number of features. It is the first The true label of each sample It is the first The first sample 1 eigenvalue, It is the regularization intensity hyperparameter.

[0086] Optionally, an internal attention mechanism can be used to perform implicit feature importance analysis, calculate the contribution of each feature to the decision, and determine important features based on their contribution. The contribution of each feature to the decision... It can be determined based on the following formula: Based on any of the above embodiments, key features are input into the auxiliary identification model to obtain the depression identification results of the subject, including: Important features are input into the input layer of the auxiliary discrimination model, and the input layer standardizes the important features to obtain the original features; The original features are input into the embedding layer of the auxiliary discrimination model. The embedding layer maps the original features to the latent space and then performs position encoding on the mapped features to obtain the encoded features. The encoded features are input into the attention layer of the auxiliary discrimination model, and the attention layer performs multi-head attention aggregation on the encoded features to obtain the attention features; Attention features are input into the feedforward layer of the auxiliary discrimination model. The feedforward layer enhances the attention features and performs residual connection and layer normalization on the enhanced features to obtain the normalized enhanced representation features. The normalized enhanced representation features are input into the integration layer of the auxiliary discriminator model. The integration layer stacks multiple encoder layers and integrates features from different levels using skip connections or residual connections to obtain multi-scale fused features. Multi-scale fused features are input into the aggregation layer of the auxiliary discrimination model. The aggregation layer uses attention pooling or global average pooling to aggregate the features in the time dimension, generating a fixed-length feature vector, and thus obtaining the global feature vector. The global feature vector is input into the classification layer of the auxiliary identification model. The classification layer classifies the data through a fully connected layer and an activation function to obtain the depression identification result.

[0087] Specifically, the input data of the input layer is in the form of , It is the time step. This refers to the feature dimension. For a 5-minute window, with a sampling rate of 4Hz, T=1200; for the complete HRV feature set, F is approximately 20-30. Z-score standardization is applied to all features, making the mean 0 and the standard deviation 1. The standardization formula is as follows: in, and These are the mean and standard deviation of feature j, respectively.

[0088] For the embedding layer, the original features are mapped to a D-dimensional latent space, and the mapping formula is as follows: in, It is a learnable projection matrix. It is a bias term, D=128.

[0089] After mapping, position encoding is performed, that is, sine and cosine position information is added to capture the relative positional relationships in the sequence: Where pos is the position index and i is the dimension index, the final embedding is represented as: .

[0090] For the attention layer, a multi-head self-attention mechanism is used, where the query matrix Q, key matrix K, and value matrix V can be calculated based on the following formula: in, It is a learnable parameter matrix.

[0091] After obtaining the above matrix, attention is calculated using the following formula: in, It is a scaling factor. It refers to the number of heads that need attention.

[0092] Furthermore, multi-head attention aggregation is performed on the above attention calculation results to obtain attention features: in, and It is a learnable parameter matrix.

[0093] The feedforward layer consists of a feedforward network module, a residual module, and a normalization module. The feedforward network module is used to enhance the representation of attention features, and the enhancement formula is as follows: in, It is an internal dimension.

[0094] Next, the output of the feedforward network module is fed into the residual module to alleviate the gradient problem in deep networks: Next, the output of the residual module is fed into the normalization module to enhance the stability of training. in, and They are the mean and standard deviation. and These are learnable scaling and offset parameters.

[0095] The integration layer stacks multiple encoder layers, each combining an attention mechanism and a feedforward network. The computational mechanism of the encoder layer is as follows: Optionally, embodiments of the present invention construct a 4-layer Transformer encoder layer to enhance the model's expressive power: For the aggregation layer, attention pooling is used to aggregate information along the time dimension, focusing on key time points: in, It is the output of the final encoder layer at time step t. These are learnable parameters. =64.

[0096] For the classification layer, the global feature vector is first mapped to the classification space through a fully connected layer: in, .

[0097] Next, the mapped vector is input into the output layer to generate the final classification result: in, It is the Sigmoid activation function.

[0098] Based on any of the above embodiments, implicit feature importance analysis is performed on heart rate variability features to determine important features from the heart rate variability features, including: The attention weights of each heart rate variability feature are calculated using the attention mechanism in the auxiliary discrimination model. Based on the attention weights of each heart rate variability feature, the importance score of each heart rate variability feature is determined. Heart rate variability features with importance scores greater than a threshold are considered important features.

[0099] Specifically, the weight matrix of the attention mechanism in the auxiliary discrimination model is extracted, and the model's attention to different features and time points is analyzed. : in, It is the column corresponding to feature f in the embedding matrix. It is the attention weight at time step t.

[0100] Next, according to Generate a feature importance hotspot to visually display the contribution of each feature to classification decisions, analyze the attention distribution over time, and identify the time window with the greatest diagnostic value.

[0101] Next, the marginal contribution of each feature to the prediction result is calculated. : Where N is the geometry of all features, S is the subset excluding feature i, and f is the model prediction function.

[0102] Randomly shuffle the values ​​of individual features and observe the impact on model performance: Among them, I f It is the importance score. original This refers to the model's original performance, Score. permuted(f) It refers to the performance of the model after the feature values ​​are shuffled.

[0103] Based on any of the above embodiments, the training steps of the auxiliary identification model include: The training dataset was preprocessed, and it included historical subjects' heart rate variability characteristics and corresponding depression labels. The training dataset is divided into a training set, a validation set, and a test set; Train the auxiliary discrimination model using the training set, including: Define the loss function; Choose an optimizer, which may include Adam, SGD, or other gradient descent optimization algorithms; Set the batch size and number of training rounds, and divide the training set into multiple batches of training data according to the batch size; In each training epoch, a batch of training data is input into the model, the loss value is calculated, and the model parameters are updated using the optimizer based on the loss value.

[0104] Specifically, a total of 200 patient data points can be collected as the training dataset, including 120 UPD patients and 80 BDP patients. Age, gender, disease severity, and other factors should be matched between the two groups as closely as possible. The training dataset is divided into training, validation, and test sets in a 7:1.5:1.5 ratio to ensure consistent class distribution.

[0105] In addition, after obtaining the training dataset, data augmentation can be performed on the training dataset, as follows: First, a sliding window is created by sliding the original 5-minute window in 2.5-minute increments to generate an overlapping window: Where s=600 is the step size (2.5 minutes × 4 Hz) and w=1200 is the window size (5 minutes × 4 Hz).

[0106] Next, noise is injected by adding small-amplitude Gaussian noise to enhance the model's robustness: in, It is the standard deviation of each feature. =0.05 is the noise intensity parameter.

[0107] Next, time scaling is applied to slightly stretch and deform the time series: in, It is a random scaling factor.

[0108] Next, a loss function is defined. This embodiment of the invention employs a weighted binary cross-entropy loss function and introduces class weights to achieve class balance. This allows the model to better handle imbalanced sample sizes in unipolar and bipolar depression, preventing the model from favoring the class with more samples and improving the ability to identify the minority class, thereby enhancing overall classification performance and generalization ability. Specifically, the class weights can be set inversely proportional to the sample size of each class, causing the model to pay more attention to the class with fewer samples, thus balancing the influence of different classes during training.

[0109] The loss function formula is as follows: Class weights in the loss function and The calculation formula is as follows: in, and These are the sample numbers for the positive class (UPD) and the negative class (BPD), respectively. .

[0110] Next, select an optimizer, which can include the Adam optimizer, which combines momentum and adaptive learning rate for optimization: in, It is the gradient exponential moving average at the current moment. It is a moment gradient, It is the exponential moving average of the squared gradient. These are the updated model parameters. =0.9, =0.999, =10 -8 Initial learning rate =0.001.

[0111] Furthermore, embodiments of the present invention employ cosine return learning rate scheduling: in, T=100 is the total number of epochs.

[0112] Gradually increase the learning rate during the initial training phase: in, It is the number of warm-up epochs.

[0113] To prevent overfitting, improve generalization ability, and reduce the model's dependence on training data, this embodiment of the invention also employs a regularization strategy to constrain the model's parameters during training, thereby reducing the model's complexity. Specifically: First, Dropout is used to prevent overfitting and improve generalization ability: Dropout can be applied after a fully connected layer, with a dropout rate of p=0.3.

[0114] Next, weight decay (L2 regularization) is used to control model complexity: in, =0.0001 is the regularization strength.

[0115] Next, gradient clipping is used to prevent gradient explosion: in, =1.0 is the gradient clipping threshold.

[0116] After setting the relevant parameters, set the batch size and the number of training epochs, and divide the training set into multiple batches of training data according to the batch size. You can set the batch size to 32 and the number of training epochs to 100.

[0117] Furthermore, training is stopped when the validation set performance shows no improvement for 10 consecutive epochs. Where, patience is the number of consecutive epochs in which the performance of the validation set failed to improve.

[0118] Finally, only the parameters of the model that perform best on the validation set are saved: When evaluating model performance, metrics such as true positive (TP), false positive (FP), true negative (TN), false negative (FN), accuracy, precision, recall / sensitivity, specificity, F1 score, ROC curve and AUC value, precision-recall curve and mean precision (AP) can be used.

[0119] The validation methods include k-fold cross-validation (k=5), leave-one-out validation, independent test set validation, Bootstrap resampling, etc.

[0120] Figure 3 This is a schematic diagram of the ROC curve for model performance evaluation provided by the present invention, as shown below. Figure 3 As shown, the auxiliary identification model provided in this embodiment of the invention has been verified, with an AUC of 0.93, indicating that the auxiliary identification model provided in this embodiment of the invention has good classification performance and can effectively distinguish between unipolar and bipolar patients. An AUC value of 0.93 means that when a unipolar patient and a bipolar patient are randomly selected, the model has a 93% probability of correctly distinguishing them, indicating that the model has high identification ability and clinical application value.

[0121] The auxiliary identification device for unipolar depression and bipolar depression provided by the present invention will be described below. The auxiliary identification device for unipolar depression and bipolar depression described below can be referred to in correspondence with the auxiliary identification method for unipolar depression and bipolar depression described above.

[0122] Based on any of the above embodiments Figure 4This is a schematic diagram of the auxiliary identification device for unipolar and bipolar depression provided by the present invention, as shown below. Figure 4 As shown, the device includes: Acquisition unit 410 is used to non-contactly acquire the chest cavity micromotion signals of the subject using millimeter-wave frequency-modulated continuous wave radar; Separation unit 420 is used to filter and demodulate the thoracic cavity micromotion signal, separate the respiratory component and the heartbeat component, and extract the RR interval sequence from the heartbeat component; The calculation unit 430 is used to calculate heart rate variability characteristics based on the RR interval sequence. The heart rate variability characteristics include time-domain indicators, frequency-domain indicators, and nonlinear indicators. The discrimination unit 440 is used to input heart rate variability features into a pre-trained auxiliary discrimination model to obtain the depression discrimination result of the subject; the auxiliary discrimination model is trained based on historical heart rate variability features and corresponding depression labels.

[0123] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 5 As shown, the electronic device may include a processor 510, a communications interface 520, a memory 530, and a communication bus 540. The processor 510, communications interface 520, and memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions from the memory 530 to execute auxiliary identification methods for unipolar and bipolar depression.

[0124] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0125] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the auxiliary identification methods for unipolar depression and bipolar depression provided by the above methods.

[0126] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the auxiliary identification methods for unipolar depression and bipolar depression provided by the methods described above.

[0127] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0128] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for auxiliary differentiation between unipolar depression and bipolar depression, characterized in that, include: Millimeter-wave frequency-modulated continuous wave radar was used to collect the micro-motion signals of the subject's chest cavity in a non-contact manner. The thoracic cavity micromotion signal is filtered and phase demodulated to separate the respiratory and cardiac components, and the RR interval sequence is extracted from the cardiac component. Heart rate variability characteristics are calculated based on the RR interval sequence, and the heart rate variability characteristics include time-domain indicators, frequency-domain indicators, and nonlinear indicators; The heart rate variability features are input into a pre-trained auxiliary identification model to obtain the depression identification result of the subject; the auxiliary identification model is trained based on historical heart rate variability features and corresponding depression labels.

2. The auxiliary identification method for unipolar depression and bipolar depression according to claim 1, characterized in that, The process of filtering and demodulating the thoracic cavity micromotion signal, separating the respiratory and cardiac components, and extracting the RR interval sequence from the cardiac component includes: The thoracic cavity micro-motion signal is bandpass filtered to remove noise and power frequency interference; Phase demodulation is performed on the filtered signal to obtain a baseband signal containing respiratory and heartbeat information; The respiratory and cardiac components were separated using independent component analysis or empirical mode decomposition. Peak detection was performed on the isolated heartbeat components to obtain the heartbeat peak sequence; The time interval between adjacent heartbeat peaks is calculated to obtain the RR interval sequence.

3. The auxiliary identification method for unipolar depression and bipolar depression according to claim 1, characterized in that, The calculation of heart rate variability characteristics based on the RR interval sequence includes: Based on the statistical characteristics of the RR interval sequence, time-domain analysis is performed to obtain the time-domain index, which includes at least one of the following: standard deviation of NN interval, root mean square of the difference between adjacent NN intervals, percentage of adjacent NN intervals with a difference greater than 50ms, mean NN interval, heart rate variability trigonometric index, and trigonometric interpolation of the NN interval histogram. The power spectral density is estimated by performing power spectral density estimation on the RR interval sequence to obtain the power spectral density, and the power spectral density is integrated within a specified frequency band to obtain the frequency domain index. The frequency domain index includes at least one of the following: extremely low frequency power, low frequency power, high frequency power, low frequency to high frequency power ratio, total power, and normalized power. The nonlinear index is obtained by performing graph analysis, entropy analysis, fractal analysis, and information theory analysis on the RR interval series. The nonlinear index includes at least one of graph parameters, sample entropy, approximate entropy, multi-scale entropy, scaling index of detrended fluctuation analysis, correlation dimension, transition entropy, and mutual information.

4. The auxiliary identification method for unipolar depression and bipolar depression according to claim 1, characterized in that, The step of inputting the heart rate variability features into a pre-trained auxiliary identification model to obtain the depression identification result of the subject includes: The heart rate variability features are subjected to explicit feature screening or implicit feature importance analysis to determine important features from the heart rate variability features; The key features are input into the auxiliary identification model to obtain the depression identification results of the subject.

5. The auxiliary identification method for unipolar depression and bipolar depression according to claim 4, characterized in that, The step of inputting the key features into the auxiliary identification model to obtain the depression identification result of the subject includes: The important features are input into the input layer of the auxiliary discrimination model, and the input layer standardizes the important features to obtain the original features; The original features are input into the embedding layer of the auxiliary discrimination model. The embedding layer maps the original features to the latent space and then performs position encoding on the mapped features to obtain the encoded features. The encoded features are input into the attention layer of the auxiliary discrimination model, and the attention layer performs multi-head attention aggregation on the encoded features to obtain attention features; The attention features are input into the feedforward layer of the auxiliary discrimination model, where the feedforward layer enhances the attention features, and then performs residual connection and layer normalization on the enhanced features to obtain the normalized enhanced representation features. The normalized enhanced representation features are input into the integration layer of the auxiliary discrimination model. The integration layer stacks multiple encoder layers and integrates features from different levels using skip connections or residual connections to obtain multi-scale fusion features. The multi-scale fused features are input into the aggregation layer of the auxiliary discrimination model. The aggregation layer uses attention pooling or global average pooling to aggregate the features in the time dimension, generating a fixed-length feature vector, and thus obtaining the global feature vector. The global feature vector is input into the classification layer of the auxiliary identification model, and the classification layer performs classification through a fully connected layer and an activation function to obtain the depression identification result.

6. The auxiliary identification method for unipolar depression and bipolar depression according to claim 4, characterized in that, The implicit feature importance analysis of the heart rate variability features, and the determination of important features from the heart rate variability features, includes: The attention weights of each heart rate variability feature are calculated using the attention mechanism in the auxiliary discrimination model. Based on the attention weights of each heart rate variability feature, the importance score of each heart rate variability feature is determined. Heart rate variability features with importance scores greater than a threshold are considered as important features.

7. The auxiliary identification method for unipolar depression and bipolar depression according to claim 1, characterized in that, The training steps of the auxiliary identification model include: The training dataset is preprocessed, and the training dataset includes historical subjects' heart rate variability characteristics and corresponding depression labels; The training dataset is divided into a training set, a validation set, and a test set; Training the auxiliary discrimination model using the training set includes: Define the loss function; Select an optimizer, which includes Adam, SGD, or other gradient descent optimization algorithms; Set the batch size and the number of training rounds, and divide the training set into multiple batches of training data according to the batch size; In each training epoch, a batch of training data is input into the model, the loss value is calculated, and the model parameters are updated using the optimizer based on the loss value.

8. A device for assisting in the differentiation of unipolar depression and bipolar depression, characterized in that, include: The acquisition unit is used to non-contactly acquire the chest cavity micro-motion signals of the subject using millimeter-wave frequency-modulated continuous wave radar; The separation unit is used to filter and demodulate the thoracic cavity micro-motion signal, separate the respiratory component and the heartbeat component, and extract the RR interval sequence from the heartbeat component; The calculation unit is used to calculate heart rate variability characteristics based on the RR interval sequence, wherein the heart rate variability characteristics include time-domain indicators, frequency-domain indicators, and nonlinear indicators; The identification unit is used to input the heart rate variability features into a pre-trained auxiliary identification model to obtain the depression identification result of the subject; the auxiliary identification model is trained based on historical heart rate variability features and corresponding depression labels.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the auxiliary identification method for unipolar depression and bipolar depression as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the auxiliary identification method for unipolar depression and bipolar depression as described in any one of claims 1 to 7.

Images