A machine learning based sentiment profile prediction system

Abstract

The present application relates to the technical field of mental health data analysis, in particular to an emotion profile prediction system based on machine learning, which comprises an interval identification module, a physiological characteristic analysis module, an index construction module, an emotion recognition module and a crisis warning module.In the present application, the normalized physiological deviation vector in the sensitive interval is extracted, and the continuous time fluctuation dispersion is calculated, so as to accurately quantify the instability characteristics of the physiological signal under specific conditions.Combining with the psychological score, a psychological load accumulation index reflecting the physical and mental double stress is constructed, the load increment rate is used to linearly extrapolate the future numerical change trend, the prediction trajectory is geometrically compared with the disease evolution threshold curve, so as to accurately locate the emotion trajectory jump point and the crisis outbreak moment, realize the transition from static state evaluation to dynamic crisis warning, and significantly improve the prospective monitoring of the complex emotional pathology evolution of malignant tumor patients.

Description

Technical Field

[0001] This invention relates to the field of mental health data analysis technology, and in particular to a machine learning-based emotion profile prediction system. Background Technology

[0002] The field of mental health data analytics involves assessing an individual's psychological state and emotional responses through various data collection, processing, and analysis methods, providing a basis for mental health intervention, diagnosis, and treatment. This field includes emotion monitoring, psychological state assessment, physiological signal analysis, and mental illness prediction.

[0003] Among them, the emotion profile prediction system refers to predicting a patient's emotional state in a specified situation based on the patient's physiological and psychological data. In particular, it is designed to address emotional problems such as anxiety, depression and PTSD that may occur in patients with malignant tumors after they are transferred out of the ICU. The system usually establishes an individual's emotion profile through multidimensional data collection, including physiological signals (such as heart rate, blood pressure, respiratory rate, etc.), psychological questionnaires, behavioral analysis and other methods.

[0004] Existing technologies rely solely on static physiological and psychological data points to assess current emotional states, ignoring the dynamic temporal fluctuations of physiological signals within specific sensitive intervals. This results in an inability to effectively quantify the cumulative effects of psychological load over time, making it difficult to capture the instability trends of physiological indicators under load changes. Consequently, there is a lack of dynamic basis for predicting the future evolution of emotional states, making it impossible to determine the specific timing of potential emotional crises. This leads to significant delays in clinical interventions and severely weakens the ability to manage and prevent acute emotional deterioration risks in patients with malignant tumors at an early stage. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a machine learning-based sentiment profile prediction system.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a sentiment profile prediction system based on machine learning includes:

[0007] The interval identification module collects electrocardiogram signals, chest and abdominal respiratory movement signals, and fingertip pulse wave signals from malignant tumor patients transferred from the ICU, marks multiple types of sensitive intervals, and generates a set of physiological signal sensitive intervals.

[0008] The physiological feature analysis module performs normalization processing on each type of sensitive interval in the set of physiological signal sensitive intervals, constructs a normalized physiological offset vector, and calculates the physiological signal fluctuation dispersion of the normalized physiological offset vector over continuous time.

[0009] The index construction module obtains the patient's current hospital anxiety and depression scale score and post-traumatic stress disorder checklist score, performs normalization processing on each, and combines them with the normalized physiological offset vector to calculate the psychological load accumulation index, and extracts the load increment rate of the index for multiple monitoring periods.

[0010] The emotion recognition module constructs a feature space based on the psychological load accumulation index and the physiological signal fluctuation dispersion, and filters the patient's potential emotional profile category within the feature space.

[0011] The crisis early warning module uses the load increment rate to extrapolate the future psychological load accumulation index, predicts the evolution direction of the potential emotional profile category, and obtains the emotional crisis early warning result.

[0012] As a further aspect of the present invention, the set of physiological signal sensitive intervals includes an anxiety sensitive interval, a depression sensitive interval, and a PTSD sensitive interval; the physiological signal fluctuation dispersion includes the variance of the normalized heart rate variability parameter, the variance of the normalized respiratory rate trough, and the variance of the normalized systolic blood pressure-heart rate linkage amplitude; the psychological load accumulation index includes the normalized hospital anxiety and depression scale score, the normalized post-traumatic stress disorder checklist score, and the norm of the normalized physiological offset vector; the load increment rate is specifically the difference change value of the psychological load accumulation index in adjacent monitoring periods; the emotional potential profile categories include an anxiety-dominant profile, a depression-dominant profile, and a PTSD comorbid profile; and the emotional crisis early warning result includes the predicted time of the emotional trajectory jump point, the extrapolated value of the psychological load accumulation index, and the evolution direction of the emotional potential profile category.

[0013] As a further aspect of the present invention, the interval identification module includes:

[0014] The feature sequence generation submodule collects electrocardiogram (ECG), chest and abdominal respiratory motion signals, and fingertip pulse wave signals from patients with malignant tumors transferred from the ICU. It calculates the time-domain characteristics of the RR interval for the ECG signals to obtain the heart rate variability sequence, calculates the peak interval data for the chest and abdominal respiratory motion signals to obtain the respiratory rate sequence, and calculates the amplitude changes of feature points for the fingertip pulse wave signals to obtain the systolic blood pressure fluctuation sequence, thus establishing a multidimensional physiological feature sequence set.

[0015] The single-source state labeling submodule, based on the multidimensional physiological feature sequence set, sets a sliding window for the heart rate variability sequence and calculates the variance value of the heart rate variability data in the window, filters the time period when the variance value exceeds the preset resting baseline threshold, and marks and generates anxiety sensitive intervals. For the respiratory rate sequence, it calculates the first-order difference value of adjacent respiratory rate data, marks the time period when the difference value is continuously negative and generates depression sensitive intervals, and establishes a single-source pathological state interval table.

[0016] The comprehensive interval summarization submodule calculates the sliding covariance values ​​of the systolic blood pressure fluctuation sequence and the heart rate variability sequence based on the multidimensional physiological feature sequence set. It filters out time periods where the covariance values ​​exceed a preset linkage threshold and marks them to generate PTSD sensitive intervals. These intervals are then summarized with the anxiety sensitive intervals and depression sensitive intervals in the single-source pathological state interval table to generate a set of physiological signal sensitive intervals.

[0017] As a further aspect of the present invention, the physiological characteristic analysis module includes:

[0018] The differential value extraction submodule calls the set of physiological signal sensitive intervals, extracts heart rate variability data for the anxiety sensitive interval and calculates the difference between the heart rate variability data and the resting baseline mean, extracts respiratory rate data for the depression sensitive interval and calculates the difference between the respiratory rate trough and the resting baseline mean, and extracts systolic blood pressure and heart rate data for the PTSD sensitive interval and calculates the difference between the systolic blood pressure and heart rate linkage amplitude and the resting baseline mean, generating the original physiological differential value set;

[0019] The offset vector construction submodule performs normalization operations on the original physiological difference value set for the heart rate variability difference, respiratory rate difference, and linkage amplitude difference, respectively, to construct a normalized physiological offset vector.

[0020] The dispersion analysis submodule sets a continuous time statistical window, calculates the variance of the normalized physiological offset vector in the time dimension, obtains the fluctuation data of the physiological signal in each sensitive interval, and generates the physiological signal fluctuation dispersion.

[0021] As a further aspect of the present invention, the index construction module includes:

[0022] The psychological rating vector construction submodule obtains the patient's current hospital anxiety and depression scale score and post-traumatic stress disorder list score, performs normalization operations on them respectively, and maps them to the same numerical range as the physiological signals to construct a normalized psychological rating vector.

[0023] The load index calculation submodule concatenates and combines the normalized physiological offset vector and the normalized psychological score vector in the feature dimension to establish a combined load feature vector with both physiological and psychological features. The vector magnitude of the combined load feature vector is calculated and defined as the psychological load cumulative index.

[0024] The incremental rate analysis submodule, based on the cumulative psychological load index of the current monitoring period, retrieves the cumulative psychological load index of the previous monitoring period as a benchmark, performs a difference operation on the cumulative psychological load index of the current period and the previous period, calculates the difference in index change within adjacent monitoring periods, and generates the load incremental rate.

[0025] As a further aspect of the present invention, the emotion recognition module includes:

[0026] The feature space construction submodule sets the psychological load accumulation index and the physiological signal fluctuation dispersion as spatial coordinate axes, positions the psychological load accumulation index to the load intensity coordinate axis, and positions the physiological signal fluctuation dispersion to the physiological instability coordinate axis to generate feature space sample data.

[0027] The probability calculation and analysis submodule inputs the feature space sample data into the Gaussian mixture model and maps it to the preset distribution area of ​​anxiety, depression and PTSD comorbidity patterns. For each sample point, it calculates the conditional probability value belonging to each specified pathological emotional pattern and generates the posterior probability set of the pathological pattern.

[0028] The profile category determination submodule iterates through the conditional probability values ​​for each pathological emotional pattern in the posterior probability set of the pathological patterns, filters the target item with the highest probability value, locks the corresponding specified pathological pattern, defines the target pathological pattern as a classification label representing the patient's current psychopathological characteristics, and outputs the potential emotional profile category.

[0029] As a further aspect of the present invention, the crisis early warning module includes:

[0030] The evolution threshold matching submodule compares the current potential emotional profile category with the preset emotional pathological evolution standard data, filters the specified evolution threshold value that matches the current category feature, maps the evolution threshold value to the time axis and connects them to form a dynamic judgment boundary, and generates the disease evolution threshold curve.

[0031] The index extrapolation calculation submodule obtains the psychological load accumulation index as the current baseline value, calls the load increment rate as the numerical change rate parameter, takes the current monitoring time as the calculation starting point, performs linear extrapolation on the numerical change trend of the psychological load accumulation index in the future time series, and generates the future load extrapolation trajectory.

[0032] The early warning result generation submodule calls the future load extrapolation trajectory and the disease evolution threshold curve, calculates the time coordinates when the trajectory and the curve coincide, marks the time corresponding to the coordinates as the emotional trajectory jump point, analyzes the trajectory data trend after the jump point, predicts the evolution direction of the potential emotional profile category in the next stage, and generates an emotional crisis early warning result.

[0033] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0034] In this invention, by extracting the normalized physiological offset vector within the sensitive interval and calculating the continuous-time fluctuation dispersion, the instability characteristics of physiological signals under specific circumstances are accurately quantified. Combined with psychological scoring, a psychological load accumulation index reflecting the dual pressure of mind and body is constructed. The load increment rate is used to linearly extrapolate the future numerical change trend, and the predicted trajectory is geometrically compared with the disease evolution threshold curve. This allows for the precise location of emotional trajectory jumps and crisis outbreak moments, realizing the transformation from static state assessment to dynamic crisis early warning, and significantly improving the prospective monitoring of the complex emotional pathological evolution of malignant tumor patients. Attached Figure Description

[0035] Figure 1 This is a system flowchart of the present invention;

[0036] Figure 2 This is a flowchart of the interval identification module of the present invention;

[0037] Figure 3 This is a flowchart of the physiological characteristic analysis module of the present invention;

[0038] Figure 4 This is a flowchart of the index construction module of the present invention;

[0039] Figure 5 This is a flowchart of the emotion recognition module of the present invention;

[0040] Figure 6 This is a flowchart of the crisis early warning module of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0042] Please see Figure 1 A machine learning-based sentiment profile prediction system includes:

[0043] The interval identification module collects electrocardiogram signals, chest and abdominal respiratory movement signals, and fingertip pulse wave signals from malignant tumor patients transferred from the ICU, marks multiple types of sensitive intervals, and generates a set of physiological signal sensitive intervals.

[0044] The physiological feature analysis module performs normalization processing on each type of sensitive interval in the set of sensitive intervals of physiological signals, constructs a normalized physiological offset vector, and calculates the dispersion of physiological signal fluctuations in the normalized physiological offset vector over continuous time.

[0045] The index construction module obtains the patient's current hospital anxiety and depression scale score and post-traumatic stress disorder checklist score, performs normalization processing on each, and combines them with the normalized physiological offset vector to calculate the psychological load accumulation index, and extracts the load increment rate of the index for multiple monitoring periods.

[0046] The emotion identification module constructs a feature space based on the psychological load accumulation index and the physiological signal fluctuation dispersion, and filters the patient's potential emotional profile category within the feature space.

[0047] The crisis early warning module uses the load increment rate to extrapolate the future psychological load accumulation index, predicts the evolution direction of potential emotional profile categories, and obtains emotional crisis early warning results.

[0048] The set of physiological signal sensitive intervals includes the anxiety sensitive interval, the depression sensitive interval, and the PTSD sensitive interval. The physiological signal fluctuation dispersion includes the variance of the normalized heart rate variability parameter, the variance of the normalized respiratory rate trough, and the variance of the normalized systolic blood pressure-heart rate linkage amplitude. The psychological load accumulation index includes the normalized hospital anxiety and depression scale score, the normalized post-traumatic stress disorder checklist score, and the normalized physiological offset vector norm. The load increment rate is specifically the difference change value of the psychological load accumulation index in adjacent monitoring periods. The emotional potential profile categories include the anxiety-dominant profile, the depression-dominant profile, and the PTSD comorbid profile. The emotional crisis early warning results include the predicted time of the emotional trajectory jump point, the extrapolated value of the psychological load accumulation index, and the evolution direction of the emotional potential profile category.

[0049] Please see Figure 2 The interval recognition module includes:

[0050] The feature sequence generation submodule collects electrocardiogram (ECG), chest and abdominal respiratory motion signals, and fingertip pulse wave signals from patients with malignant tumors transferred from the ICU. It calculates the time-domain characteristics of the RR interval for the ECG signals to obtain the heart rate variability sequence, calculates the peak interval data for the chest and abdominal respiratory motion signals to obtain the respiratory rate sequence, and calculates the amplitude changes of feature points for the fingertip pulse wave signals to obtain the systolic blood pressure fluctuation sequence, thus establishing a multidimensional physiological feature sequence set.

[0051] The specific process of calculating the time-domain characteristics of the RR interval from the electrocardiogram signal to obtain the heart rate variability sequence is as follows:

[0052] Morphological filtering algorithms are applied to ECG signals to remove baseline drift and power frequency interference noise;

[0053] The position of the R wave peak in the QRS complex of an electrocardiogram signal was located using the differential threshold detection method.

[0054] Calculate the time difference between the peak positions of two consecutive R waves to obtain the original RR interval value;

[0055] Abnormal data points exceeding the preset physiological limit in the original RR interval values ​​are removed, and the processed values ​​are arranged according to the time of heartbeat to generate a heart rate variability sequence.

[0056] The process of calculating the RR interval time-domain features of an electrocardiogram (ECG) signal to obtain the heart rate variability sequence is as follows: The feature sequence generation submodule first calls the morphological filter processor, sets the structuring element to a linear structure, and calculates the width parameter of the structuring element based on the sampling frequency divided by a preset baseline drift frequency lower limit. For example, if the sampling frequency is... Hertz, setting the lower limit of the drift frequency to Hertz, then the width parameter is set to The processor performs a sampling point; morphological opening operation (erosion followed by dilation) on the raw ECG signal to remove low-frequency baseline drift caused by breathing or body movement; then morphological closing operation (dilation followed by erosion) is performed to smooth high-frequency spikes and power line interference noise, outputting clean ECG waveform data; next, the processor applies differential threshold detection logic to calculate the first-order difference sequence of the clean ECG waveform data, amplifies the high-frequency slope characteristics of the QRS complex through first-order differential amplification, and sets the amplitude threshold to [value missing]. millivolts, this threshold is calculated based on the product of the mean amplitude of pure electrocardiogram waveform data and a preset coefficient, for example, the mean is... millivolts, coefficient set to Mark the peaks in the difference sequence that exceed the amplitude threshold as candidate R-wave locations; before and after the candidate R-wave locations The processor searches for a maximum point within a millisecond time window and locks the time coordinate of this maximum point as the final R-wave vertex position. Then, the processor extracts the timestamps of two adjacent R-wave vertex positions in chronological order and performs a subtraction operation; for example, if the current R-wave time is... seconds, the previous R-wave time was seconds, the original RR interval value is calculated to be seconds; for a set of raw RR interval values ​​obtained, the physiological limit range is set to... Instant The range is determined by statistically analyzing the RR interval data of healthy individuals at rest, calculating the mean and standard deviation of the dataset, setting the lower limit as the mean minus three times the standard deviation, and the upper limit as the mean plus three times the standard deviation; iterating through all raw RR interval values, those less than... seconds or greater Data points in seconds are identified as abnormal artifacts and removed, while values ​​that conform to physiological logic are retained. Finally, the retained RR interval values ​​are sorted on the time axis according to the corresponding heartbeat time to construct a heart rate variability sequence that changes over time.

[0057] The single-source state labeling submodule, based on a multidimensional physiological feature sequence set, sets a sliding window for the heart rate variability sequence and calculates the variance value of the heart rate variability data within the window. It filters out time periods where the variance value exceeds the preset resting baseline threshold and marks them to generate anxiety-sensitive intervals. For the respiratory rate sequence, it calculates the first-order difference value of adjacent respiratory rate data, marks the time periods where the difference value is continuously negative and generates depression-sensitive intervals, and establishes a single-source pathological state interval table.

[0058] First, retrieve the patient's historical heart rate variability data at rest, extracting data for a duration of [duration missing]. For a data segment of hours, calculate the variance of that segment as the resting baseline variance. For example, the calculated resting baseline variance is... Milliseconds squared, set resting baseline threshold as The millisecond squared threshold is calculated by taking the mean and standard deviation of the historical resting state variance dataset, setting the threshold to the mean plus twice the standard deviation; subsequently, for the currently acquired heart rate variability sequence, the length is set to... A sliding window of seconds, with a step size of 1. Calculate the variance of all RR intervals within the sliding window, one by one, for each second; if the variance calculated for a certain window is... Milliseconds squared, exceeding the preset value. If a millisecond squared threshold is applied, the time period corresponding to that window is marked as an anxiety-sensitive candidate segment, and consecutively overlapping candidate segments are merged to generate an anxiety-sensitive interval. Simultaneously, for the respiratory rate sequence, the difference between the current respiratory rate and the previous respiratory rate is calculated to obtain the first-order difference value; if the current respiratory rate is... times per minute, the previous moment was If the first difference is 10 times per minute, then the first difference value is 100 times per minute. Set the duration threshold to The threshold is set at seconds, based on statistical distribution data of respiratory depression response time induced by depressive mood, and is selected from the distribution's [number]th [second]. The percentile value is determined; if the first difference value is in a continuous sequence If the respiratory rate remains negative or zero for a continuous period of seconds, i.e., the respiratory rate shows a monotonous decrease or maintains a low trend, then this continuous period of time is marked and a depression sensitive interval is generated. Finally, the determined anxiety sensitive interval and depression sensitive interval, along with their corresponding start and end points, are filled into the corresponding category columns of the single-source pathological state interval table.

[0059] The comprehensive interval summary submodule calculates the sliding covariance values ​​of the systolic blood pressure fluctuation sequence and the heart rate variability sequence based on the multidimensional physiological feature sequence set. It filters out time periods where the covariance values ​​exceed the preset linkage threshold and marks them to generate PTSD sensitive intervals. It then summarizes these intervals with the anxiety sensitive intervals and depression sensitive intervals in the single-source pathological state interval table to generate a set of physiological signal sensitive intervals.

[0060] Use the systolic blood pressure fluctuation sequence and heart rate variability sequence, set a synchronized sliding window, and set the window length to [value missing]. Seconds; within each sliding window, calculate the covariance between systolic blood pressure data and heart rate variability data, which reflects the direction and intensity of the linkage between the two sets of signals; set the linkage threshold to... This threshold is calculated by statistically analyzing physiological signal datasets of patients with post-traumatic stress disorder under stress-triggered scenarios, calculating a set of correlation coefficients between systolic blood pressure and heart rate variability, and selecting the first value from this set. Percentile values ​​are determined; if the absolute value of the calculated covariance is greater than... This indicates a strong, abnormal synchronous fluctuation or reverse antagonism between the two, and this time period is marked as a sensitive interval for post-traumatic stress disorder; subsequently, the module performs interval union operation to retrieve the anxiety-sensitive interval from the single-source pathological state interval table, such as the time period. point Instant point Seconds, and the sensitive interval for depression, such as time periods. point Instant point Seconds, and newly formed sensitive intervals for post-traumatic stress disorder, such as time periods. point Instant point Seconds; for intervals with overlapping times, for example point Instant point If a second falls within the range of both anxiety and post-traumatic stress disorder, then it is combined into a single continuous time period. point Instant point Seconds; All independent time periods after merging are aggregated to generate the final set of physiological signal sensitive intervals.

[0061] Please see Figure 3 The physiological characteristic analysis module includes:

[0062] The differential value extraction submodule calls the set of physiological signal sensitive intervals, extracts heart rate variability data for the anxiety sensitive interval and calculates the difference between the heart rate variability data and the resting baseline mean, extracts respiratory rate data for the depression sensitive interval and calculates the difference between the respiratory rate trough and the resting baseline mean, and extracts systolic blood pressure and heart rate data for the PTSD sensitive interval and calculates the difference between the systolic blood pressure and heart rate linkage amplitude and the resting baseline mean, generating the original physiological differential value set;

[0063] First, the system accesses a pre-defined database and retrieves the mean physiological parameters of the patient at rest over the past week as the resting baseline mean. This mean is calculated by arithmetically averaging historical monitoring data; for example, the mean resting heart rate variability is calculated as follows: Milliseconds, mean resting respiratory rate is The mean amplitude of the correlation between resting systolic blood pressure and heart rate is [number] times per minute. Unit; subsequently, the module performs differentiated extraction logic for different classification intervals within the set of physiological signal sensitive intervals:

[0064] For the anxiety-sensitive zone, extract all heart rate variability data points within that zone and identify the maximum fluctuation peak, for example, a value of [value missing]. Milliseconds, perform subtraction, and use the peak value. Subtract the resting baseline value The difference value was obtained. millisecond;

[0065] For the depression-sensitive region, the respiratory rate sequence within this region is extracted, and the minimum point in the sequence is searched as the trough of respiratory inhibition. For example, the trough value is... Perform subtraction operations every minute, using the valley value. Subtract the resting baseline value The difference value was obtained. times per minute;

[0066] For the sensitive intervals of post-traumatic stress disorder, the systolic blood pressure and heart rate sequences within that time period are first extracted. The average systolic blood pressure and average heart rate within that interval are then calculated. Assuming the calculated average systolic blood pressure for the interval is... mmHg, mean heart rate over intervals Once per minute; then the instantaneous coupling strength is calculated point by point, selecting a sampling time within the interval, if the systolic pressure at that time is mmHg, heart rate If the systolic blood pressure deviates from the mean by [number] times per minute, then the deviation is [percentage]. mmHg, heart rate deviation from mean is Every minute, multiplying the two yields the instantaneous coupling product at that moment. Sum the instantaneous coupling products at all sampling times within the interval and take the arithmetic mean. Assume the calculated average coupling product is... The normalization coefficient is set to... This coefficient is used to adjust the large value of the product of blood pressure and heart rate to the standard scoring level. The larger the coefficient, the higher the final amplitude amplification factor. Here, the average coupling product is used. Multiply by a coefficient The calculated amplitude of the correlation between systolic blood pressure and heart rate is... Finally, perform the subtraction operation and use the calculated linkage amplitude. Subtract the resting baseline value The difference value was obtained. ;

[0067] Finally, the heart rate variability difference values ​​calculated above will be used to... Difference in milliseconds and respiratory rate Frequency per minute, amplitude difference of linkage These are then compiled into a unified set of original physiological difference values.

[0068] The offset vector construction submodule performs normalization operations on the differences in heart rate variability, respiratory rate, and linkage amplitude based on the original set of physiological differences, and constructs normalized physiological offset vectors.

[0069] Normalization was performed on the difference data of different dimensions in the original physiological difference value set to eliminate the influence of dimensions; the theoretical maximum fluctuation range of heart rate variability difference was set as follows. to The theoretical range of the difference in respiratory rate in milliseconds is to The theoretical range of the amplitude difference in the linkage is [number] times per minute. to These ranges are calculated by statistically analyzing the distribution characteristics of physiological monitoring big data of malignant tumor patients, and then calculating the distribution of each parameter. percentile to the 1st Percentile intervals determined; regarding the aforementioned The difference in heart rate variability in milliseconds is calculated by performing a minimum-maximum normalization operation. Divide by get Regarding the aforementioned The difference in respiratory rate per minute is first mapped to an absolute deviation. Divide by get Regarding the aforementioned The difference in amplitude of the linkage is calculated. Divide by get The above three normalized values , , Arranged in a fixed dimensional order, a three-dimensional normalized physiological offset vector is constructed.

[0070] The dispersion analysis submodule sets a continuous time statistical window, calculates the variance of the normalized physiological offset vector in the time dimension, obtains the fluctuation data of physiological signals in each sensitive interval, and generates physiological signal fluctuation dispersion.

[0071] Set the continuous time statistics window length to Seconds; within this window, a series of normalized physiological offset vectors are collected. For the value of each dimension in the vector, the statistical variance calculation logic is executed: First, extract the value of the first dimension within the current window. For all normalized sample points of dimension (heart rate variability), calculate their mean, then calculate the sum of squares of the differences between each sample point and the mean, divide by the number of samples minus one, and obtain the variance of that dimension. For example, the calculation result is... Similarly, calculate the first... The variance of the dimension (respiratory rate) is , No. The variance of the dimension (linkage amplitude) is Then, the variance values ​​of all dimensions are summed. Finally, calculate the arithmetic mean. Divide by Get about ; this value Defined as the dispersion of physiological signal fluctuations, it quantitatively characterizes the overall instability of a patient's physiological signals within the sensitive range.

[0072] Please see Figure 4 The index building module includes:

[0073] The psychological rating vector construction submodule obtains the patient's current hospital anxiety and depression scale score and post-traumatic stress disorder list score, performs normalization operations on them respectively, and maps them to the same numerical range as the physiological signals to construct a normalized psychological rating vector.

[0074] Read the patient's completed Hospital Anxiety and Depression Scale. The total score of this scale is... If the patient's current score is The scale is scored as follows: [Score missing]; Simultaneously, a list of post-traumatic stress disorder (PTSD) cases is retrieved; the total score for this scale is [Score missing]. If the patient's current score is Score; perform normalization operation on the hospital anxiety and depression scale score. Divide the score by the total score The normalized values ​​were obtained. ;Assess post-traumatic stress disorder (PTSD) using a checklist. Divide the score by the total score The normalized values ​​were obtained. Map these two values ​​to arrive Within the closed interval, and by combining them in sequence, a two-dimensional normalized psychological rating vector is constructed. .

[0075] The load index calculation submodule concatenates and combines the normalized physiological offset vector and the normalized psychological score vector in the feature dimension to establish a combined load feature vector with both physiological and psychological features. The vector magnitude of the combined load feature vector is calculated and defined as the psychological load cumulative index.

[0076] First, the previously constructed three-dimensional normalized physiological offset vector is called. With two-dimensional normalized psychological rating vector The two are then combined in series along the characteristic dimensions, and the formula for calculating the cumulative psychological load index is applied: ,in, The calculated cumulative psychological load index represents the magnitude of the overall physical and mental load. The dimension of the physiological feature vector is determined based on the number of physiological signal types, and its value is [value missing]. ; The dimension of the psychological feature vector is determined based on the number of types of psychological scales, and its value is [value missing]. ; The first in the physiological feature vector The values ​​of each component are derived from the normalized physiological offset vector, index. Value to ; The first in the psychological feature vector The values ​​of each component are derived from the normalized psychological rating vector, indexed... Value to ; This is the mind-body coupling weighting coefficient, which reflects the degree of influence of psychological scores on overall workload. It is set based on the Pearson correlation coefficient between psychological intervention and improvement in physiological indicators in the patient's past medical records. The larger the correlation coefficient (i.e., the more significant the impact of psychological state on physiological recovery), the greater the weighting. In this example, it is set to [value missing]. .

[0077] The first step is to calculate the sum of squares of the physiological characteristics: substitute the values... , , The calculation process is as follows The summation result is ;

[0078] The second step is to calculate the sum of squares of the psychological characteristics: substitute the values... , The calculation process is as follows The summation result is approximately ;

[0079] The third step is to calculate the sum by substituting the weighting coefficients: The calculation process is as follows The result is ;

[0080] Step 4: Calculate the final exponent by taking the square root: The result is approximately ; this value The output is the current cumulative psychological load index.

[0081] The incremental rate analysis submodule, based on the cumulative psychological load index of the current monitoring period, retrieves the cumulative psychological load index of the previous monitoring period as a benchmark, performs a difference operation on the cumulative psychological load index of the current period and the previous period, calculates the difference in index change within adjacent monitoring periods, and generates the load incremental rate.

[0082] Retrieve the previous monitoring period, for example The cumulative psychological load index was calculated minutes ago; let's assume the value is... Obtain the cumulative psychological load index for the current monitoring period. Perform a difference operation, using the current value. Subtract the value from the previous period The difference is obtained. ; this difference Defined as the rate of increase in workload, this positive value indicates that the patient's psychological workload shows an upward trend during the monitoring period, and the rate of increase is per cycle. .

[0083] Please see Figure 5 The emotion recognition module includes:

[0084] The feature space construction submodule sets the psychological load accumulation index and the physiological signal fluctuation dispersion as spatial coordinate axes, positions the psychological load accumulation index to the load intensity coordinate axis, and positions the physiological signal fluctuation dispersion to the physiological instability coordinate axis to generate feature space sample data.

[0085] Establish a two-dimensional Cartesian coordinate system, with the horizontal axis defined as the load intensity axis and the vertical axis defined as the physiological instability axis; read the currently calculated cumulative psychological load index. Map it to the horizontal axis coordinate; read the currently calculated physiological signal fluctuation dispersion. This is then mapped to the vertical axis coordinates; thus determining a feature space sample point in the coordinate system. The location of this sample point in geometric space comprehensively reflects the patient's current psychological stress intensity and physiological stability.

[0086] The probability calculation and analysis submodule inputs the feature space sample data into the Gaussian mixture model and maps it to the preset distribution area of ​​anxiety, depression and PTSD comorbidity patterns. For each sample point, it calculates the conditional probability value belonging to each specified pathological emotional pattern and generates the posterior probability set of the pathological pattern.

[0087] A pre-trained Gaussian mixture model is preloaded, which includes preset Gaussian distribution parameters for three comorbidity patterns: anxiety, depression, and post-traumatic stress disorder; the current feature space sample points are then loaded. As the input vector, the simplified two-dimensional Gaussian probability density function formula is applied for calculation: ,in, This represents the conditional probability value of the target pattern. Pi, with a value of approximately ; For the first The determinant of the pattern covariance matrix represents the breadth of the distribution; For the first The inverse of the pattern covariance matrix; The current input feature space sample column vector ; For the first The mean column vector of the class pattern represents the center position of the pathological pattern; Represents the matrix transpose operation; This represents the difference between the sample vector and the mean vector.

[0088] The complete calculation and classification process is illustrated by calculating the probability density of the "Post-Traumatic Stress Disorder (PTSD) pattern" (the calculation logic for anxiety and depression patterns is the same, only the parameters differ):

[0089] Set preset parameters: PTSD mode mean vector PTSD mode covariance matrix ;

[0090] The first step is to calculate the difference vector: ;

[0091] The second step is to calculate the determinant and inverse of the covariance matrix: Determinant ;

[0092] Inverse matrix ;

[0093] The third step is to calculate the squared Mahalanobis distance term of the exponential part. : Calculate first ;

[0094] Recalculate ;

[0095] Fourth step, calculate the coefficient terms mentioned earlier. Calculate the denominator ; Calculate coefficients ;

[0096] Fifth step, calculate the final probability density value. Calculate the exponential term ;

[0097] Final result ;

[0098] Similarly, the anxiety pattern density is calculated. and depression pattern density ;

[0099] Step 6: Perform normalization to obtain the posterior probability: calculate the total density and sum. ;

[0100] Calculate the probability of pattern attribution for post-traumatic stress disorder. ;

[0101] Calculate the probability of belonging to the anxiety pattern ;

[0102] Calculate the probability of belonging to the depression pattern ;

[0103] Finally, a posterior probability set of the pathological pattern is generated, which includes the above three probability values.

[0104] The profile category determination submodule iterates through the conditional probability values ​​for each pathological emotional pattern in the posterior probability set of the pathological pattern, filters the target item with the highest probability value, locks the corresponding specified pathological pattern, defines the target pathological pattern as a classification label that represents the patient's current psychopathological characteristics, and outputs the potential emotional profile category.

[0105] Traversing the values ​​in the posterior probability set of pathological patterns: , , ;Execute numerical comparison logic to identify the maximum value as ;because This corresponds to the probability of a post-traumatic stress disorder (PTSD) pattern. Therefore, the pathological emotional pattern corresponding to this maximum value is identified as the "PTSD pattern." The "PTSD pattern" is used as the target pathological pattern and converted into a textual classification label: "PTSD-dominant psychopathological features." The classification results show that, at the current monitoring time, the patient's psychophysiological characteristics data have the highest matching degree with the typical pathological model of PTSD in the multidimensional space, belonging to the PTSD-dominant pathological state, which is different from a simple anxiety or depression state. Finally, the potential emotional profile category is output through the interface.

[0106] Please see Figure 6 The crisis early warning module includes:

[0107] The evolution threshold matching submodule compares the current potential emotional profile category with the preset emotional pathological evolution standard data, filters the specified evolution threshold value that matches the current category feature, maps the evolution threshold value to the time axis and connects them to form a dynamic judgment boundary, and generates the disease evolution threshold curve.

[0108] First, the current output emotional potential profile category is identified as "post-traumatic stress disorder-dominant psychopathological features". Then, pathological evolution standard data corresponding to this category is retrieved from a pre-defined database. This data is a set of critical values ​​that change over time. This set is determined by longitudinally tracking patients with similar symptoms, using survival analysis to calculate hazard ratios at different stress levels, and identifying the risk critical values ​​at each time point. For example, the risk critical value at the [missing information - likely a specific time point]. The critical value for critical illness in hours , No. Hours , No. Hours These critical values ​​are mapped to a coordinate system with time as the horizontal axis and load index as the vertical axis according to the corresponding time points. Spline interpolation is used to smoothly connect these discrete critical points to form a dynamic judgment boundary line, namely the disease evolution threshold curve. This curve intuitively divides the safe area and the crisis area.

[0109] The index extrapolation calculation submodule obtains the psychological load cumulative index as the current baseline value, calls the load increment rate as the numerical change rate parameter, takes the current monitoring time as the calculation starting point, performs linear extrapolation on the numerical change trend of the psychological load cumulative index in the future time series, and generates the future load extrapolation trajectory.

[0110] The process of performing linear extrapolation on the numerical trend of the psychological load accumulation index over future time series is as follows:

[0111] Set the time resolution parameters for the future time series that define the extrapolation accuracy, and the prediction termination time point that defines the prediction range;

[0112] Using the current monitoring time as the origin of the time axis, a series of continuous and equally spaced discrete prediction time points are generated in the positive direction of the time axis based on the time resolution parameter.

[0113] The load increment rate is locked as a numerical evolution gradient that remains constant across all discrete forecast time points.

[0114] Set the cumulative psychological load index as the initial intercept value of the linear extrapolation model;

[0115] Iterate through each discrete prediction time point and calculate the time difference between the current discrete prediction time point and the current monitoring time point;

[0116] Calculate the product of the load increment rate and the time difference to obtain the total load drift at the discrete prediction time point relative to the current monitoring time point;

[0117] The total load drift is directly added to the psychological load accumulation index to calculate the expected psychological load value corresponding to the discrete prediction time point.

[0118] Each discrete prediction time point is combined with the corresponding expected psychological load value to form a two-dimensional coordinate point of time coordinates and load coordinates;

[0119] Connect all the two-dimensional coordinate points in chronological order to form a continuous path in which the psychological load increases or decreases linearly over time, and construct the trajectory of future load extrapolation.

[0120] The process of performing linear extrapolation on the numerical trend of the cumulative psychological load index over future time series is as follows: The time resolution parameter is set to... Minutes, setting the prediction termination time to after the current moment. Minutes; based on the current monitoring time (the [number]th minute); Generating a timeline with the origin at (minutes) along the positive direction of the time axis. , , Until A series of integer discrete prediction time points; locking the load increment rate calculated above. Assuming a constant evolution gradient, the monitoring period is also... minutes, then the gradient is Every minute; accumulating the psychological load index at the current moment. Set as the initial intercept; iterate through the first discrete prediction time point (the... (minutes), calculate the time difference as Minutes, calculate the total load drift as The total drift is added to the initial intercept. , obtained the The expected psychological load value per minute is Similarly, regarding the first At the minute point, calculate the time difference. The total drift is The expected value is obtained after superposition. ; calculate all the coordinate points , Combine them sequentially; connect these coordinate points with straight line segments in chronological order to construct a line starting from the current value with a slope of The ray, i.e., the trajectory extrapolated from the future load.

[0121] The early warning result generation submodule calls the future load extrapolation trajectory and the disease evolution threshold curve, calculates the time coordinates when the trajectory and the curve coincide, marks the time corresponding to the coordinates as the emotional trajectory jump point, analyzes the trajectory data trend after the jump point, and predicts the evolution direction of the potential emotional profile category in the next stage, generating an emotional crisis early warning result.

[0122] The constructed future load extrapolation trajectory and the generated disease evolution threshold curve are placed in the same coordinate system; a geometric intersection operation is performed to detect whether the two lines intersect; assuming the disease evolution threshold curve is at the [missing information]th [missing information]th [missing information]th [missing information]. The value at the minute is The extrapolated trajectory is in the first... The calculation of the value at the minute is as follows: ;because Greater than And the first Trajectory value per minute Less than the value corresponding to the curve (assuming it is) If the two lines are in the first position, then determine the position of the two lines. Minutes to the The intersection occurs between minutes; the overlapping time coordinates are accurately calculated using linear interpolation to represent the future... Minutes; this moment is marked as the emotional trajectory jump point; analysis of the trajectory data after the jump point reveals that the extrapolated trajectory remains above the threshold curve, indicating that the condition is about to spiral out of control; based on this, the potential emotional profile category will be predicted to be... Minutes later, it will evolve from the current "post-traumatic stress disorder-dominant type" to the "post-traumatic stress disorder critical outbreak type"; ultimately generating a model containing "expected..." The emotional crisis warning result is "entering a crisis state in minutes".

[0123] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A sentiment profile prediction system based on machine learning, characterized in that, The system includes: The interval identification module collects electrocardiogram signals, chest and abdominal respiratory movement signals, and fingertip pulse wave signals from malignant tumor patients transferred from the ICU, marks multiple types of sensitive intervals, and generates a set of physiological signal sensitive intervals. The physiological feature analysis module performs normalization processing on each type of sensitive interval in the set of physiological signal sensitive intervals, constructs a normalized physiological offset vector, and calculates the physiological signal fluctuation dispersion of the normalized physiological offset vector over continuous time. The index construction module obtains the patient's current hospital anxiety and depression scale score and post-traumatic stress disorder checklist score, performs normalization processing on each, and combines them with the normalized physiological offset vector to calculate the psychological load accumulation index, and extracts the load increment rate of the index for multiple monitoring periods. The emotion recognition module constructs a feature space based on the psychological load accumulation index and the physiological signal fluctuation dispersion, and filters the patient's potential emotional profile category within the feature space. The crisis early warning module uses the load increment rate to extrapolate the future psychological load accumulation index, predicts the evolution direction of the potential emotional profile category, and obtains the emotional crisis early warning result.

2. The machine learning-based sentiment profile prediction system according to claim 1, characterized in that, The set of sensitive physiological signal intervals includes anxiety-sensitive intervals, depression-sensitive intervals, and PTSD-sensitive intervals. The dispersion of physiological signal fluctuations includes the variance of normalized heart rate variability parameters, the variance of normalized respiratory rate troughs, and the variance of normalized systolic blood pressure-heart rate linkage amplitudes. The cumulative psychological load index includes the normalized hospital anxiety and depression scale score, the normalized post-traumatic stress disorder checklist score, and the norm of the normalized physiological offset vector. The load increment rate is specifically the difference change value of the cumulative psychological load index in adjacent monitoring periods. The potential emotional profile categories include anxiety-dominant profiles, depression-dominant profiles, and PTSD comorbid profiles. The emotional crisis early warning results include the predicted time of emotional trajectory jump points, the extrapolated value of the cumulative psychological load index, and the evolution direction of the potential emotional profile categories.

3. The machine learning-based sentiment profile prediction system according to claim 1, characterized in that, The interval identification module includes: The feature sequence generation submodule collects electrocardiogram (ECG), chest and abdominal respiratory motion signals, and fingertip pulse wave signals from patients with malignant tumors transferred from the ICU. It calculates the time-domain characteristics of the RR interval for the ECG signals to obtain the heart rate variability sequence, calculates the peak interval data for the chest and abdominal respiratory motion signals to obtain the respiratory rate sequence, and calculates the amplitude changes of feature points for the fingertip pulse wave signals to obtain the systolic blood pressure fluctuation sequence, thus establishing a multidimensional physiological feature sequence set. The single-source state labeling submodule, based on the multidimensional physiological feature sequence set, sets a sliding window for the heart rate variability sequence and calculates the variance value of the heart rate variability data in the window, filters the time period when the variance value exceeds the preset resting baseline threshold, and marks and generates anxiety sensitive intervals. For the respiratory rate sequence, it calculates the first-order difference value of adjacent respiratory rate data, marks the time period when the difference value is continuously negative, generates depression sensitive intervals, and establishes a single-source pathological state interval table. The comprehensive interval summarization submodule calculates the sliding covariance values ​​of the systolic blood pressure fluctuation sequence and the heart rate variability sequence based on the multidimensional physiological feature sequence set. It filters out time periods where the covariance values ​​exceed a preset linkage threshold and marks them to generate PTSD sensitive intervals. These intervals are then summarized with the anxiety sensitive intervals and depression sensitive intervals in the single-source pathological state interval table to generate a set of physiological signal sensitive intervals.

4. The machine learning-based sentiment profile prediction system according to claim 3, characterized in that, The specific process of calculating the time-domain characteristics of the RR interval from the electrocardiogram signal to obtain the heart rate variability sequence is as follows: Morphological filtering algorithms are applied to ECG signals to remove baseline drift and power frequency interference noise; The position of the R wave peak in the QRS complex of an electrocardiogram signal was located using the differential threshold detection method. Calculate the time difference between the peak positions of two consecutive R waves to obtain the original RR interval value; Abnormal data points that exceed the preset physiological limit in the original RR interval values ​​are removed, and the processed values ​​are arranged according to the time of heartbeat to generate a heart rate variability sequence.

5. The machine learning-based sentiment profile prediction system according to claim 3, characterized in that, The physiological characteristic analysis module includes: The differential value extraction submodule calls the set of physiological signal sensitive intervals, extracts heart rate variability data for the anxiety sensitive interval and calculates the difference between the heart rate variability data and the resting baseline mean, extracts respiratory rate data for the depression sensitive interval and calculates the difference between the respiratory rate trough and the resting baseline mean, and extracts systolic blood pressure and heart rate data for the PTSD sensitive interval and calculates the difference between the systolic blood pressure and heart rate linkage amplitude and the resting baseline mean, generating the original physiological differential value set; The offset vector construction submodule performs normalization operations on the original physiological difference value set for the heart rate variability difference, respiratory rate difference, and linkage amplitude difference, respectively, to construct a normalized physiological offset vector. The dispersion analysis submodule sets a continuous time statistical window, calculates the variance of the normalized physiological offset vector in the time dimension, obtains the fluctuation data of the physiological signal in each sensitive interval, and generates the physiological signal fluctuation dispersion.

6. The machine learning-based sentiment profile prediction system according to claim 5, characterized in that, The index construction module includes: The psychological rating vector construction submodule obtains the patient's current hospital anxiety and depression scale score and post-traumatic stress disorder list score, performs normalization operations on them respectively, and maps them to the same numerical range as the physiological signals to construct a normalized psychological rating vector. The load index calculation submodule concatenates and combines the normalized physiological offset vector and the normalized psychological score vector in the feature dimension to establish a combined load feature vector with both physiological and psychological features. The vector magnitude of the combined load feature vector is calculated and defined as the psychological load cumulative index. The incremental rate analysis submodule, based on the cumulative psychological load index of the current monitoring period, retrieves the cumulative psychological load index of the previous monitoring period as a benchmark, performs a difference operation on the cumulative psychological load index of the current period and the previous period, calculates the difference in index change within adjacent monitoring periods, and generates the load incremental rate.

7. The machine learning-based sentiment profile prediction system according to claim 6, characterized in that, The emotion recognition module includes: The feature space construction submodule sets the psychological load accumulation index and the physiological signal fluctuation dispersion as spatial coordinate axes, positions the psychological load accumulation index to the load intensity coordinate axis, and positions the physiological signal fluctuation dispersion to the physiological instability coordinate axis to generate feature space sample data. The probability calculation and analysis submodule inputs the feature space sample data into the Gaussian mixture model and maps it to the preset distribution area of ​​anxiety, depression and PTSD comorbidity patterns. For each sample point, it calculates the conditional probability value belonging to each specified pathological emotional pattern and generates the posterior probability set of the pathological pattern. The profile category determination submodule iterates through the conditional probability values ​​for each pathological emotional pattern in the posterior probability set of the pathological patterns, filters the target item with the highest probability value, locks the corresponding specified pathological pattern, defines the target pathological pattern as a classification label representing the patient's current psychopathological characteristics, and outputs the potential emotional profile category.

8. The machine learning-based sentiment profile prediction system according to claim 7, characterized in that, The formula for calculating the conditional probability of belonging to each specified pathological emotional pattern is as follows: ； in, The conditional probability value for the target pathological emotional pattern; Pi For the first The determinant of the pattern covariance matrix, For the first The inverse of the pattern covariance matrix, The current input feature space sample column vector, For the first The mean column vector of the class pattern, This represents the matrix transpose operation. This represents the difference between the sample vector and the mean vector.

9. The machine learning-based sentiment profile prediction system according to claim 7, characterized in that, The crisis early warning module includes: The evolution threshold matching submodule compares the current potential emotional profile category with the preset emotional pathological evolution standard data, filters the specified evolution threshold value that matches the current category feature, maps the evolution threshold value to the time axis and connects them to form a dynamic judgment boundary, and generates the disease evolution threshold curve. The index extrapolation calculation submodule obtains the psychological load accumulation index as the current baseline value, calls the load increment rate as the numerical change rate parameter, takes the current monitoring time as the calculation starting point, performs linear extrapolation on the numerical change trend of the psychological load accumulation index in the future time series, and generates the future load extrapolation trajectory. The early warning result generation submodule calls the future load extrapolation trajectory and the disease evolution threshold curve, calculates the time coordinates when the trajectory and the curve coincide, marks the time corresponding to the coordinates as the emotional trajectory jump point, analyzes the trajectory data trend after the jump point, predicts the evolution direction of the potential emotional profile category in the next stage, and generates an emotional crisis early warning result.

10. The machine learning-based sentiment profile prediction system according to claim 9, characterized in that, The process of performing linear extrapolation on the numerical trend of the psychological load accumulation index over future time series is as follows: Set the time resolution parameters for the future time series that define the extrapolation accuracy, and the prediction termination time point that defines the prediction range; Using the current monitoring time as the origin of the time axis, continuous and equally spaced discrete prediction time points are generated in the positive direction of the time axis based on the time resolution parameter; The load increment rate is locked as a numerical evolution gradient that remains constant across all discrete forecast time points. Set the cumulative psychological load index as the initial intercept value of the linear extrapolation model; Iterate through each discrete prediction time point and calculate the time difference between the current discrete prediction time point and the current monitoring time point; Calculate the product of the load increment rate and the time difference to obtain the total load drift at the discrete prediction time point relative to the current monitoring time point; The total load drift is directly added to the psychological load accumulation index to calculate the expected psychological load value corresponding to the discrete prediction time point. Each discrete prediction time point is combined with the corresponding expected psychological load value to form a two-dimensional coordinate point of time coordinates and load coordinates; Connect all two-dimensional coordinate points sequentially according to the logical order of time to form a continuous path in which the psychological load increases or decreases linearly over time, and construct the trajectory of future load extrapolation.

Images