A cerebral hemorrhage condition dynamic early warning method and device

Abstract

The application discloses a cerebral hemorrhage condition dynamic early warning method and device, constructs a dynamic monitoring data set by integrating bedside physiological parameters and head CT review images, measures the reliability of data from the interruption distribution of physiological parameters, and extracts abnormal signals related to autonomic nervous regulation from the conventional fine fluctuations; with the aid of image quantitative indicators and cross comparison of review text conclusions, the possible hidden progress of the lesion is identified, and a comprehensive risk judgment is formed in combination with intracranial pressure related performance in physiological signs; risk signals from different sources are limited within the same time range for verification, the risk level is divided according to the duration of continuous deviation, and the corresponding disposal instruction is matched; further, the terminal delivery of early warning, the personnel response mode and the change of indicators after disposal are tracked, the situation of insufficient response or ineffective disposal is identified, and a review prompt is sent accordingly, forming a closed loop from disease identification, graded early warning to execution quality management.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, and in particular to a method and device for dynamic early warning of cerebral hemorrhage. Background Technology

[0002] Patients with cerebral hemorrhage experience rapid changes in their condition during the acute phase. Clinically, bedside monitoring devices are typically used to continuously collect physiological parameters such as heart rate, blood pressure, respiration, and blood oxygenation. A follow-up head CT scan is scheduled according to the clinical protocol to track changes in the hematoma and surrounding edema. These two monitoring methods operate independently. Physiological parameter monitoring is primarily triggered by whether a single indicator exceeds a preset range, while the assessment of the progression of the follow-up imaging relies mainly on the physician's visual comparison of the two images and the formation of a written conclusion.

[0003] This fragmented monitoring approach has significant limitations. On the one hand, the intervals between imaging follow-ups are often several days or even longer, meaning that a physician's conclusion of lesion stability based on visual comparison may not capture slow progression that is not yet visually noticeable but has already occurred. On the other hand, threshold judgments for individual physiological parameters are insufficient to reflect more subtle signs of deterioration, such as autonomic nervous system dysregulation, and they do not combine the progression trend on imaging with abnormal fluctuations in physiological parameters for comprehensive analysis. Furthermore, after existing warnings are issued, there is usually a lack of systematic tracking and verification mechanisms to determine whether effective treatment is received and whether the condition truly improves after treatment. Consequently, the operational quality of the warning mechanism itself is difficult to continuously evaluate and improve. Summary of the Invention

[0004] This invention discloses a dynamic early warning method and device for cerebral hemorrhage, which aims to integrate cranial CT images with bedside physiological parameter monitoring data to identify the progression trend of lesions that are not reflected in written review conclusions. It also combines abnormal fluctuations in physiological signals to comprehensively determine the risk level and generate graded early warning instructions. At the same time, it tracks and evaluates the delivery and handling effects of the early warning, making up for the shortcomings of existing methods that separate image interpretation and physiological monitoring and lack feedback on the quality of early warning execution. This forms a dynamic early warning method system that takes into account both the assessment of the condition and the quality management of the early warning.

[0005] The first aspect of this invention proposes a method for dynamic early warning of cerebral hemorrhage, comprising: Acquire bedside vital sign data and hematoma and edema images from head CT scans; determine the hematoma and edema area expansion boundary from the hematoma and edema images to delineate the hematoma and edema range; and construct a dynamic monitoring dataset by marking the interruption period of the vital sign data according to a preset time window. The monitoring interruption load is obtained by statistically analyzing the distribution of interruption periods in the dynamic monitoring dataset. The set of fluctuation segments is obtained by screening the abnormal fluctuations that are routinely filtered out from the vital sign data. The rhythm is analyzed based on the set of fluctuation segments to generate a fluctuation rhythm sequence. The fluctuation rhythm sequence is corrected by combining the monitoring interruption load to form an autonomic nervous system abnormality index. Based on the hematoma and edema range, the expansion rate of the space-occupying lesion is obtained by statistically examining the expansion amount during the negative period. The Cushing's reaction is identified through the vital signs data to determine the signs of intracranial hypertension. The signs of intracranial hypertension are compared with the space-occupying lesion expansion rate to form signs of deterioration risk. The intersection of the aforementioned signs of deterioration risk and the autonomic nervous system abnormality index is used to obtain a synchronous over-threshold window. The risk level is determined by statistically analyzing the critical zone dwell time based on the synchronous over-threshold window. An early warning instruction is generated by matching the treatment path according to the risk level. The system uses the warning command to push warning messages and establishes a terminal delivery record. It summarizes the silent and unresponsive items and the confirmation delay in the terminal delivery record to obtain a warning failure list. The system evaluates the delivery status through the warning failure list and outputs a warning review prompt.

[0006] A second aspect of the present invention provides a dynamic early warning device for cerebral hemorrhage, comprising: The data acquisition unit is used to acquire vital sign data from bedside monitoring and hematoma and edema images from head CT scans. It determines the hematoma and edema area expansion boundary from the hematoma and edema images to define the range of hematoma and edema. It constructs a dynamic monitoring dataset by marking the interruption period of the vital sign data according to a preset time window. An anomaly identification unit is used to statistically analyze the distribution of interruption periods in the dynamic monitoring dataset to obtain the monitoring interruption load, screen out the pseudo-variables that are routinely filtered out from the vital signs data to obtain a set of fluctuation segments, parse the rhythm based on the set of fluctuation segments to generate a fluctuation rhythm sequence, and combine the monitoring interruption load to correct the fluctuation rhythm sequence to form an autonomic nervous system abnormality index. The sign comparison unit is used to obtain the lesion expansion rate by statistically analyzing the expansion amount during the negative period of the hematoma and edema range, identify the Cushing's reaction through the vital sign data to determine the signs of intracranial hypertension, and compare the signs of intracranial hypertension with the lesion expansion rate to form signs of deterioration risk. The risk grading unit is used to obtain a synchronous over-threshold window by taking the intersection of the worsening risk signs and the autonomic nervous system abnormality index, determine the risk level based on the statistical critical zone dwell time of the synchronous over-threshold window, and generate an early warning instruction by matching the treatment path according to the risk level. The early warning verification unit is used to push early warning messages using the early warning command to establish a terminal delivery record, summarize the silent non-response items and confirmation delays in the terminal delivery record to obtain an early warning failure list, evaluate the delivery status through the early warning failure list, and output an early warning review prompt.

[0007] The beneficial effects of this invention are reflected in the following points: 1. Unlike conventional monitoring, this invention does not treat subtle fluctuations in vital signs as interference to be filtered out. Instead, it retains these fluctuations and incorporates them into the same monitoring dataset along with quantitative lesion data from imaging. This creates conditions for identifying abnormal signals related to autonomic nervous system regulation from seemingly meaningless fluctuations, giving previously discarded information reference value again. 2. Addressing the issue that follow-up reports rely on manual visual comparison and are prone to recording actual disease progression as stable, this invention introduces cross-checking of quantitative imaging values ​​and textual conclusions. It identifies the time periods where the two differ and uses this to calculate the true extent of lesion expansion. This judgment is combined with characteristic manifestations in vital signs that reflect changes in intracranial pressure, allowing the two originally separate assessment criteria—imaging and physiological—to corroborate each other. 3. This invention does not simply issue alerts separately upon receiving multiple risk signals. Instead, it requires that abnormal signals from different sources truly overlap in time before determining a state requiring attention, and classifies the urgency of handling based on the duration of the alert. After the alert is issued, the delivery status of the terminal, the response method of the personnel, and whether the condition has truly improved after handling will continue to be tracked. Situations where handling is ineffective or ineffective will be collected separately for managers to understand the actual performance of the alert mechanism in the execution process. Attached Figure Description

[0008] Figure 1 This is a flowchart of a dynamic early warning method for cerebral hemorrhage according to the present invention.

[0009] Figure 2 This is a schematic diagram of the hematoma false negative re-examination segment identification mechanism of the present invention.

[0010] Figure 3 This is a structural block diagram of a dynamic early warning device for cerebral hemorrhage according to the present invention.

[0011] Among them: 1-Re-examination time sequence; 2-Hollow circle; 3-Solid circle; 4-Measured value curve of hematoma and edema range; 5-Inconsistent negative conclusion; 6-False negative re-examination segment; 7-Negative period expansion amount; 8-Landslide expansion rate; 300-Dynamic early warning device for cerebral hemorrhage; 301-Data acquisition unit; 302-Abnormal identification unit; 303-Sign comparison unit; 304-Risk grading unit; 305-Early warning verification unit. Detailed Implementation

[0012] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0013] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0014] The technical solutions of the embodiments of this application will be described below.

[0015] like Figure 1 As shown, this embodiment of the invention provides a method for dynamic early warning of cerebral hemorrhage, including the following steps S101-S105: Step S101: Obtain vital sign data from bedside monitoring and hematoma and edema images from head CT scans. Determine the hematoma and edema extent from the hematoma and edema images and define the expansion boundary of the hematoma and edema area. Construct a dynamic monitoring dataset by marking the interruption period of vital sign data according to a preset time window.

[0016] Specifically, bedside vital sign data and head CT images of hematoma and edema were acquired. Vital sign data were simultaneously collected from five channels of the bedside monitor: heart rate, systolic blood pressure, diastolic blood pressure, respiratory rate, and oxygen saturation. Heart rate, respiratory rate, and oxygen saturation were continuously recorded at 2-second sampling points. Systolic and diastolic blood pressure were measured every 15 minutes using a non-invasive cuff. All readings for the five channels were based on the monitor's internal timing, thus the vital sign data were collected in parallel on the same timeline. Head CT images of hematoma and edema were obtained from the patient's baseline head CT scan upon admission and subsequent follow-up scans scheduled according to the clinical protocol. Most patients underwent 3 to 5 follow-up scans during their hospitalization. The scans were continuous and uninterrupted, with a slice thickness and interslice spacing of 5 mm, using a brain parenchyma window to show the contrast between the high-density areas of hematoma and the low-density areas of edema. Each scan was typically completed within minutes, with intervals often lasting several days. In the acute phase, hematomas typically appear as well-defined, high-density lesions with a CT value higher than 60 HU in hematoma and edema imaging, while the surrounding edema zone has a CT value lower than 20 HU. The boundary between these two density intervals is pre-defined accordingly. Vital sign data are continuously generated at intervals ranging from seconds to minutes. Hematoma and edema images are only generated at the time of clinically scheduled follow-up examinations. The former has a much higher temporal density than the latter. The timestamps of both are unified using the same clinical clock of the monitoring system. The acquisition time of the hematoma and edema image corresponding to a CT follow-up examination can be correlated with the vital sign data records around the same time. If a sampling point in the heart rate and blood oxygen saturation channels suddenly deviates from the adjacent points due to electrode detachment, it is discarded and linearly interpolated from the effective points before and after it. This processing is not applied to the respiratory rate, systolic blood pressure, and diastolic blood pressure channels for the time being, and will be refined in subsequent implementation examples.

[0017] The extent of hematoma and edema was determined by measuring the boundaries of the hematoma and edema areas in the images. Two density thresholds, 60 HU and 20 HU, were used to divide the hematoma and edema images into three categories: connected regions with a CT value of at least 60 HU were marked as hematoma candidate regions; connected regions with a CT value between 20 HU and 60 HU and adjacent to the hematoma candidate regions were marked as edema candidate regions; and the remaining regions were marked as background areas and not included in subsequent calculations. The two thresholds were derived from a statistical comparison of CT images of patients with cerebral hemorrhage in our hospital over the past three years with the results of manual delineation by radiologists. The largest connected region under three-dimensional 26-connectivity conditions was used to determine the hematoma and edema areas for each category of candidate regions. Hematoma and edema images acquired from the same patient at different times were segmented according to these two thresholds. Adjacent scans were first registered using skull landmarks to eliminate head position differences, and then segmented separately. Local low-density cavities within the hematoma area due to uneven blood clot distribution were filled using morphological closure operations in the segmentation results of the two hematoma and edema images. The expansion boundary of the hematoma area is calculated by taking the maximum distance from each point on the outer contour of the hematoma area to the centroid of the hematoma. The hematoma expansion radius of that layer is obtained by calculating it layer by layer. The expansion boundary of the edema area is calculated in the same way to obtain the edema expansion radius of that layer. The difference between the two radii is the width of the edema expansion relative to the hematoma in that layer. The areas of the connected regions of the hematoma and edema areas in each scanning layer are accumulated along the layer thickness direction and converted into hematoma volume and edema volume, respectively. The hematoma and edema range corresponding to this scan is composed of the hematoma volume, edema volume, and the time of acquisition in this follow-up examination. The hematoma and edema ranges of the same patient in two adjacent follow-up examinations are compared after the aforementioned registration: the increase or decrease in hematoma volume reflects whether the bleeding is still progressing within this time period, the change in the ratio of edema volume to hematoma volume reflects the growth rate of the edema zone relative to the hematoma, and the difference between the two hematoma and edema ranges is the basis for subsequent judgment.

[0018] In some embodiments, constructing a dynamic monitoring dataset by marking interruption periods of vital sign data according to a preset time window includes: reading the timestamps of each collection time from the vital sign data to form a vital sign time series; identifying interruption periods in the vital sign time series where the collection interval exceeds a preset threshold to obtain an interruption period list; classifying and labeling the interruption period list according to the direction of vital sign changes before and after the interruption to obtain an interruption marker file; aligning the interruption marker file with the vital sign time series according to the preset time window and merging them to construct a dynamic monitoring dataset.

[0019] The vital signs time series is formed by reading the timestamps of each acquisition moment from the vital signs data. The vital signs data for heart rate, respiratory rate, and blood oxygen saturation are each sampled at 2-second intervals and written to the corresponding acquisition moment, with timestamp accuracy down to the second, accumulating approximately 43,000 records per day for each channel. The vital signs data for systolic and diastolic blood pressure are measured at 15-minute intervals and written to the acquisition moment, accumulating 96 records per day for each channel. The record density of the two types of channels differs by hundreds of times. The vital signs time series is organized by channel, with all historical readings and corresponding timestamps arranged sequentially for each channel. Adjacent records from the same channel are not allowed to have the same timestamp. If the monitor generates two identical timestamps due to internal timing errors, the timestamp of the later record is delayed by 1 second. Sampling points where the original readings of the vital signs data exceed the physiological range of the channel are marked as suspected abnormal readings before being written into the vital signs time series. The marking information, along with the reading and timestamp, is written into the corresponding channel sequence. The specific physiological range boundaries will be provided in subsequent embodiments. Heart rate data is obtained from the monitor's ECG module, and blood oxygen saturation data is obtained from the pulse oximeter probe. The sensors for these two channels are installed at different locations on the patient's body, resulting in a physiological conduction delay of several seconds between readings. This delay is included in the original timestamp of each channel and is not separately corrected when writing the vital signs timeline. Data transmission between the monitor and the data acquisition terminal is real-time via the hospital's wired network. The number of readings transmitted per transmission is variable. Upon arrival at the acquisition terminal, the readings are written into the vital signs timeline according to their original acquisition times for each channel, without being reordered based on arrival order. Brief network congestion does not affect the final arrangement results.

[0020] The system identifies interruption periods in the vital signs time-series data collection when the sampling interval exceeds a preset threshold, generating a list of interruption periods. In the vital signs time-series data collection, the interruption threshold for the heart rate, respiratory rate, and blood oxygen saturation channels is set to three times their respective normal sampling intervals; that is, if no new sampling point is recorded for 6 seconds, the channel is considered interrupted. The interruption threshold for the systolic and diastolic blood pressure channels is set to twice the normal measurement interval; that is, if no new reading is recorded for 30 minutes, it is considered interrupted. Channels with continuous sampling have dense data, and occasionally missing one or two points is usually just normal fluctuation. Channels with low-frequency measurements have longer intervals, requiring several consecutive misses to rule out simply not being at the measurement time. The thresholds for these two types are set as integer multiples of this difference, with smaller multiples indicating greater sensitivity to interruptions. When the interval between two adjacent timestamps in the same channel exceeds the corresponding threshold in the vital signs time-series data collection, the interruption start time is the previous timestamp, and the interruption end time is the first newly added timestamp after the interval returns to normal. Based on an estimate of approximately 300,000 sampling points for continuous 7-day monitoring of the heart rate channel, interruptions triggered by electrode adjustments and other reasons typically do not exceed 20 times. The interruption period list summarizes all interruptions identified by each of the five channels. Each record is directly taken from the start and end readings in the vital signs time stream and includes four items: the channel it belongs to, the start time, the end time, and the valid reading. If the heart rate, respiratory rate, and blood oxygen saturation channels simultaneously enter an interruption within the same time range, each channel still generates a separate interruption period list record. If multiple consecutive intervals exceeding the threshold occur within the same channel and are consecutive, they are merged into one interruption period entry in the interruption period list, with the start and end times being the earliest start time and the latest end time within the merged range, respectively.

[0021] The interruption period list is categorized and labeled according to the direction of vital sign changes before and after the interruption to obtain an interruption marker file. The first valid reading after the interruption ends is recorded as a1, the last valid reading before the interruption begins as a0, and the corresponding channel's normal fluctuation amplitude is σ. The change factor K for each record in the interruption period list is K=(a1-a0) / σ: K>1 indicates an upward interruption, K<-1 indicates a downward interruption, and K in between indicates a stable interruption. The input in this step is always the start and end readings of each record in the interruption period list. The normal fluctuation amplitude of the heart rate channel is taken as the standard deviation of the readings in the first 2 hours after admission. The standard deviation of the heart rate for most patients falls between 5 and 10 beats / min. The standard deviations of the blood pressure channels are taken as the standard deviations of the readings in the first 2 hours after admission. The respiratory rate and blood oxygen saturation channels are calculated in the same way. The standard deviations of all five channels are recalculated every 24 hours using the latest 2-hour data. The interruption marker file consists of the classification results recorded in each interruption period list, along with the original channel, start time, and end time. When |K|>5, the judgment result is still recorded as either rising or falling, and simultaneously marked as high-power abnormality for manual verification of readings. This type of abnormality is recorded separately in the interruption marker file and does not affect the original three-category classification. If heart rate, respiratory rate, and blood oxygen saturation simultaneously enter an interruption within the same time range, each of the three channels obtains its own directional judgment based on its preceding and following readings. If one channel is judged as rising, the others can be classified as falling or stable. The judgments of the three channels are independent of each other. In the interruption marker file, the proportion of stable interruptions is significantly higher than the sum of rising and falling interruptions, indicating that most interruptions in this period occur during a phase with relatively small fluctuations in vital signs. When fluctuations are close to baseline, the readings often return to a similar level after being interrupted.

[0022] A dynamic monitoring dataset is constructed by aligning and merging the interruption marker files and vital sign time series according to a preset time window. The preset time window is 1 hour. The vital sign time series and interruption marker files are each aligned and merged according to this set of window divisions. The patient's admission time is used as the starting point of the first window, and subsequent windows are extended sequentially. If the monitoring in the last window ends in the middle of the window, the window is counted according to the actual monitoring duration rather than a fixed 1 hour, and the actual duration of the window is also marked in the window record. If the admission time of the first window is earlier than the hour but less than 1 hour, it is also treated as the actual duration. All readings in the vital sign time series that fall within the same 1-hour window are included in that window. The number of readings of each of the five channels within the window is also recorded in the corresponding record of that window, and stored side by side with the interruption classification count of the corresponding window for subsequent verification and traceability. Interruption records in the interruption marker file whose start times fall within a certain window range are assigned to that window. They are counted separately for rising, falling, and stationary interruptions and then recorded in the corresponding window. If an interruption crosses the boundary of two adjacent windows, it is assigned to the window where its start time is located. In these boundary cases, the vital signs timeline and the interruption marker file are always aligned to the same window and will not be counted twice. The dynamic monitoring dataset is organized by window. Each window's records contain the counts of each of the three types of interruptions within that window, as well as the start and end times of that window. A patient monitored for 7 days corresponds to 168 dynamic monitoring dataset window records. The first and last two windows, each less than one hour, are each counted as a separate record and are not merged with adjacent windows. Therefore, the total number of windows may be slightly more than the theoretical value for a full 7 days by one or two. The total number of window records in the dynamic monitoring dataset increases daily without any upper limit truncation.

[0023] Step S102: Statistically analyze the distribution of interruption periods in the dynamic monitoring dataset to obtain the monitoring interruption load. Screen the vital signs data to obtain a set of fluctuation segments by routinely filtering out artifact fluctuations. Analyze the rhythm based on the set of fluctuation segments to generate a fluctuation rhythm sequence. Combine the monitoring interruption load to correct the fluctuation rhythm sequence and form an autonomic nervous system abnormality index.

[0024] In some embodiments, the monitoring interruption load is obtained by statistically analyzing the distribution of interruption periods in the dynamic monitoring dataset. This includes: retrieving interruption periods from the dynamic monitoring dataset to obtain an interruption period sequence; generating an interruption frequency curve by statistically analyzing the number of interruptions per unit time based on the interruption period sequence; filtering out periodic interruptions based on the interruption frequency curve to form an unplanned interruption distribution; and accumulating the frequency of the unplanned interruption distribution in each preset time window to obtain the monitoring interruption load.

[0025] The interruption period sequence is obtained by retrieving interruption periods from the dynamic monitoring dataset. Each window record in the dynamic monitoring dataset already stores counts for rising, falling, and stable categories. Each window record is accessed sequentially, directly reading the pre-written counts for each category. These three counts are then added together to obtain the total number of interruptions for that window, without re-accessing other data sources outside the dynamic monitoring dataset. Windows are accessed sequentially according to the patient's admission time, starting from the first window and proceeding to the window at the current monitoring time. If all three counts for a particular window are zero, the process jumps directly to the next window. Each record in the interruption period sequence retains the start and end times of the window and the total number of interruptions for that window. These two pieces of information correspond completely to the original fields in the window records of the dynamic monitoring dataset. The sequence length continuously increases with the monitoring duration, and historical records are not overwritten or truncated by scrolling. The total number of interruptions for each window in the interrupted time series is not equal. The original sampling interval for the two blood pressure channels is relatively long, and the interruption threshold is correspondingly lenient. Therefore, their contribution to the total number of interruptions is usually significantly less than that of the three continuously acquired channels: heart rate, respiratory rate, and blood oxygen saturation. This difference is retained window-by-window in the interrupted time series without normalization. The interrupted time series only retains the total number of interruptions already summarized at the window level, without backtracking to any specific reading or individual interruption record for a particular channel. Subsequent statistics based on the interrupted time series are all built upon this level, eliminating the need for repeated cross-level verification of lower-level details. This approach ensures that the data volume of the interrupted time series itself is always proportional to the number of windows, and does not expand with the number of interruptions.

[0026] An interruption frequency curve is generated based on the number of interruptions per unit time in the interruption time sequence. The unit time is 1 hour, consistent with the window length of the dynamic monitoring dataset. Each record in the interruption time sequence corresponds to a 1-hour interval; the total number of interruptions for that record is directly taken as the interruption count for that interval, without needing to re-accumulate for each record. The interruption frequency curve uses the patient's admission time as 0, and each interval corresponds to an incrementing sequence number starting from 0. Multiplying the sequence number by 1 hour gives the starting time of that interval relative to the admission time. The first interval corresponds to the first hour after admission. The records from the five channels in the interruption time sequence have already been merged and counted in the previous step. The value of an interval on the interruption frequency curve is directly equal to the total number of interruptions corresponding to that interval, without distinguishing specific channels or requiring verification of the source for each record. The interruption time sequence is naturally arranged in window order, and the interruption frequency curve extends sequentially from one interval to the next. Adjacent intervals are connected end-to-end without any time gaps. If a certain interval has no vital sign data available for statistics due to reasons such as a complete power outage of the monitoring equipment, the number of interruptions in that interval is recorded as the same value as the previous interval, not as zero, to avoid misinterpreting power outages as sudden drops in interruption. The total number of intervals in the interruption frequency curve is always equal to the number of records in the interruption time sequence. The two correspond one-to-one using the same set of serial numbers. There is no situation where a certain interval cannot find a source record in the interruption time sequence. Subsequent statistics can therefore directly use the serial number to retrieve data without repeatedly checking the underlying detail of timestamps. Every update of the interruption frequency curve can be traced back to a specific source record of the interruption time sequence.

[0027] The unplanned interruption distribution is formed by filtering out periodic interruptions based on the interruption frequency curve. Nurses typically turn patients every 4 hours as ordered by doctors. The brief detachment of monitoring electrodes before and after turning is the most common periodic component on the interruption frequency curve, with a slight increase in frequency at shift changes. The identification method involves performing autocorrelation analysis on the interruption frequency curve to find the time intervals that cause the autocorrelation coefficient to reach a local maximum. Intervals falling within the range of 3.5 to 4.5 hours or 7.5 to 8.5 hours are considered periodic components corresponding to turning or shift changes. The amplitude of the component corresponding to turning is usually between 0.5 and 1.5 times, while the amplitude of the component corresponding to shift changes is smaller, mostly within 0.3 times. If a patient has never experienced the 4-hour cycle corresponding to the order to turn, the corresponding periodic component is treated as 0. The interruption frequency curve is subtracted by the estimated values ​​of these two periodic components in each interval. A negative difference is truncated to zero. The resulting difference sequence is the unplanned interruption distribution. The unplanned interruption distribution uses the same interval numbers and start time markers as the interruption frequency curve, only replacing the values ​​corresponding to each interval. The specific amplitude of the periodic component is continuously re-estimated based on the autocorrelation strength of the interruption frequency curve over the most recent 48 hours, instead of using a fixed amplitude determined at the time of admission. The values ​​of different time periods on the same day in the unplanned interruption distribution can be directly compared because the periodic component has been deducted based on the actual turning and shift change times of the day. There will be no systematic deviation due to different shift schedules in different wards. Therefore, the unplanned interruption distributions of two different wards are also comparable. This is why the unplanned interruption distribution is more suitable for subsequent load assessment than the original values ​​of the interruption frequency curve.

[0028] The monitoring interruption load is calculated by accumulating the frequency of unplanned interruptions across preset time windows. The monitoring interruption load is accumulated over a 6-hour rolling window. This window length is significantly longer than the 1-hour granularity of the unplanned interruption distribution. The 6-hour window is chosen to prevent drastic jumps in load due to isolated omissions or brief equipment failures. For example, if a certain interval experiences an isolated interruption due to a brief equipment failure, after being factored into the 6-hour rolling window, this interruption only increases the sum of values ​​within that window by 1, preventing a spike in the monitoring interruption load significantly higher than the preceding and following intervals within that hour. The value of each 1-hour interval in the unplanned interruption distribution is included in the subsequent 6 consecutive rolling windows. Therefore, an interval contributes to the monitoring interruption load for a continuous 6 hours, rather than just affecting its own hour. The rolling window moves every hour. After moving, the value of the interval that first moves out of the window is subtracted from the accumulated sum, and the value of the most recently entered interval is added to the accumulated sum. The monitoring interruption load thus obtains a new value hourly, without needing to re-accumulate all intervals within the window each time. The monitoring interruption load value with serial number N is the sum of the unplanned interruption distribution of interval N and the five intervals before it. When the interval number is less than 5, the summation is directly applied to the intervals from 0 to N, and it is no longer required to make up 6 intervals. Therefore, the monitoring interruption load value of these early windows naturally does not cover 6 hours, and the low value is a normal phenomenon and no additional compensation is made. After monitoring for 6 hours, the monitoring interruption load value of each new window truly covers the complete 6-hour interval. Thereafter, each update corresponds to a rolling window of fixed length.

[0029] In some embodiments, the process of obtaining a set of fluctuation segments from the vital signs data by screening out artifact fluctuations that are routinely filtered out includes: decomposing the waveform channels of each lead of the vital signs data to form a single-channel waveform; extracting artifact fluctuations before and after the monitoring alarm from the single-channel waveform to generate candidate fluctuation segments; comparing the first segment of the single-channel waveform with the first segment of the waveform to remove contact artifact segments from the candidate fluctuation segments to obtain suspected fluctuation segments; and merging segments with similar morphologies based on the suspected fluctuation segments to obtain a set of fluctuation segments.

[0030] The waveform channels of each lead in the vital signs data are decomposed to form single-channel waveforms. The ECG-related portion of the vital signs data is acquired by the monitor's ECG module using a standard 5-lead configuration. Electrodes are attached to the right upper limb, left upper limb, right lower limb, left lower limb, and the V1 position on the chest. The raw voltage waveforms of leads II and aVF are continuously retained: the voltage in lead II is the potential difference between the electrodes in the left lower limb and right upper limb, while the voltage in lead aVF is the potential difference between the electrodes in the left lower limb and the average potentials of the electrodes in the right and left upper limbs. The two calculation methods differ, resulting in different waveform characteristics for the cardiac electrical activity at the same moment. The heart rate value recorded in the vital signs data is the result of R-wave recognition by the ECG module on these two raw waveforms. The raw waveforms and the heart rate value in the vital signs data are stored in parallel without overwriting each other. The single-channel waveforms correspond to leads II and aVF, respectively. The voltage value of each sampling point is output by a 12-bit analog-to-digital converter, with a quantization range covering ±5 millivolts and a sampling rate of 250 Hz. Each waveform is saved as a continuous voltage sequence, not merged with the sequence of the other lead, generating a total of 500 sampling points per second. The filtering used in the R-wave recognition process is related to heart rate calculation. The single-channel waveform is the raw voltage before filtering, retaining the high-frequency and low-frequency components that were filtered out during heart rate calculation. The start and end times of the acquisition of the two single-channel waveforms are from the same source as the timestamp of the heart rate channel in the vital signs data, both originating from the same timing circuit inside the monitor. The single-channel waveform reading at any moment can be correlated with the vital signs data record at the same moment. The two serve as mutual references and verifications. It is precisely this correspondence that allows for accurate location of the center moment when extracting fluctuation segments.

[0031] Candidate fluctuation segments are generated by extracting artifact fluctuations before and after a monitoring alarm from a single-channel waveform. The single-channel waveform corresponds only to two ECG signals: lead II and lead aVF. The monitoring alarm is triggered when the reading of any channel in the vital signs data exceeds a preset physiological range. The physiological range threshold used to trigger the alarm is the same as the threshold for marking suspected abnormal readings in the vital signs data. Both heart rate and blood oxygen saturation trigger alarms, sharing the same set of judgment criteria. Alarms are also triggered when the readings of systolic blood pressure, diastolic blood pressure, and respiratory rate exceed their respective physiological ranges, but these three channels do not have corresponding single-channel waveforms. For these alarms, no candidate windows are generated; only the abnormality markings within the vital signs data themselves are recorded. Using the alarm trigger time as the center time, a 60-second voltage sequence (30 seconds before and 30 seconds after) is extracted from each of the two leads of the single-channel waveform as the candidate window for this alarm. If both heart rate and blood oxygen saturation channels exceed their respective thresholds at the same time, a separate set of candidate windows is generated for each of the two alarms. A continuous segment within a candidate window whose single-channel waveform voltage amplitude exceeds three times the normal range for that lead on that day is marked as a candidate fluctuation segment. The normal range for that day is the standard deviation of a continuous record for that lead that does not overlap with any candidate window on that day. A maximum of three candidate fluctuation segments can be marked within a single candidate window. If more than three are marked, only the three with the largest amplitudes are retained. If two alarms are less than 60 seconds apart, causing candidate windows to overlap, the overlapping portion is only truncated once and is not counted twice as candidate fluctuation segments corresponding to the two candidate windows. Candidate fluctuation segments corresponding to alarms triggered by the heart rate channel are usually concentrated before the alarm midpoint, while alarms triggered by the blood oxygen saturation channel are more likely to show significant amplitude exceeding the limit after the midpoint.

[0032] Suspected fluctuation segments were obtained by comparing the initial waveform of the single-channel waveform with the candidate fluctuation segments to eliminate contact artifact segments. The initial waveform was taken from a continuous recording within the first 30 minutes after admission of the lead in the single-channel waveform, and did not overlap with any candidate fluctuation segments. This served as a reference waveform when the electrode contact was good. If the amplitude of the reference waveform also dropped sharply within the first 30 minutes, the second 30-minute window after admission was used to re-extract the waveform until a continuous recording without a sharp drop was found. Contact artifacts in the single-channel waveform manifest in one of two ways: first, the amplitude drops sharply to near zero and lasts for more than 2 seconds, which is more common when the electrode is completely detached; second, there are jumps between adjacent sampling points that are far beyond the normal range of the reference waveform, which is more common when there is intermittent contact with a false electrode. The criterion for judging the jump is the standard deviation of the difference between all adjacent points in the reference waveform. A jump that exceeds 8 times the standard deviation is judged as a jump-type contact artifact. Genuine physiological fluctuations typically manifest as gradual or regular changes in the overall amplitude of a single-channel waveform. They do not exhibit sudden drops to zero or instantaneous jumps far exceeding the normal range at a single sampling point. This is the main difference between contact artifacts and genuine physiological signals at the single-channel waveform level. If any of the aforementioned contact artifacts appear in a candidate fluctuation segment, it is identified as a contact artifact segment and removed. The remaining segments are considered suspected fluctuation segments. When retained, the original start and end times and amplitude data are not changed; only the contact artifact marker is removed. Therefore, the number of suspected fluctuation segments always does not exceed the total number of candidate fluctuation segments marked in the corresponding candidate window. The difference between the two is the number of contact artifact segments removed in this round. The removed segments will not be included in any subsequent statistics as suspected fluctuation segments.

[0033] A set of wave segments is obtained by merging suspected wave segments with similar morphologies. Two suspected wave segments are considered similar in morphology if they are taken from the same lead, have a duration difference of no more than 2 seconds, and a correlation coefficient of amplitude envelope of no less than 0.8. Before comparison, the time axes of the two waveforms are scaled proportionally to a uniform 100 sampling points, and the amplitudes at corresponding positions are compared by aligning the first points. The correlation coefficient is calculated using the Pearson correlation coefficient formula. Suspected wave segments in lead II and lead aVF are not included in the morphology comparison. Multiple suspected wave segments with similar morphologies are merged into a single record in the wave segment set. The record takes the longest duration segment among all the segments involved in the merging as the representative waveform, and the start and end times of the remaining segments are listed separately. The record also indicates which lead the representative waveform was taken from. All suspected wave segments merged from the same record must originate from the same lead. This labeling allows the records corresponding to lead II and lead aVF to still be distinguishable within the wave segment set. The number of records for the two leads is usually not equal; the lead with more pronounced fluctuations corresponds to more records. If a suspected fluctuation segment does not meet the morphological similarity condition with any existing records in the fluctuation segment set, it becomes a new record in the fluctuation segment set, representing the complete waveform of that segment itself. A patient's fluctuation segment set during their entire hospitalization typically contains dozens to hundreds of records, the exact number depending on the alarm triggering frequency. Patients with severe fluctuations correspond to more records, which is directly related to the stability of the patient's condition. During stable periods, the growth rate of the fluctuation segment set slows significantly.

[0034] A wave rhythm sequence is generated based on the analysis of wave segment sets. Records in the wave segment set are grouped according to their respective leads, and within each group, they are arranged sequentially according to the start time of their representative waveforms. The difference in start time between two adjacent records within the same group is defined as an interval. The specific value of the interval is quite large, ranging from tens of seconds to several hours, depending on the actual time interval between two adjacent alarm segments in the wave segment set; there is no fixed normal interval. When there is only one record in the wave segment set, there are no adjacent records to compare. The first interval can only be calculated when the second record appears. This situation is more common in the early stages of monitoring and gradually disappears as monitoring progresses. The two leads independently go through this initial stage from nothing to something, without affecting or advancing each other. Let I_k be the k-th interval and I_{k-1} be the preceding interval. The rate of change of rhythm ρ_k = I_k / I_{k-1}: ρ_k close to 1 indicates a stable fluctuation rhythm, while ρ_k significantly deviating from 1 indicates a sudden increase in the frequency of fluctuations or a sudden long absence of fluctuations, often corresponding to a short-term fluctuation in the body's regulation. The first interval has no earlier intervals for comparison, and the rate of change of rhythm only has values ​​from the second interval onwards. The fluctuating rhythm sequence consists of two parts: the interval and the corresponding rate of change of rhythm. Each position corresponds to one fluctuation, and the position stores the interval length from this fluctuation to the previous fluctuation and the corresponding rate of change. Leads II and aVF generate their own fluctuating rhythm sequences. If the rate of change of rhythm in the same time period of the two fluctuating rhythm sequences is in the same direction, increases or decreases simultaneously, it suggests that the rhythm change during this period is more likely to come from a real physiological process, rather than local interference from a single lead.

[0035] An autonomic nervous system abnormality index is formed by combining the monitoring interruption load correction fluctuation rhythm sequence. The rate of change of the rhythm corresponding to each position in the fluctuation rhythm sequence is reduced according to the monitoring interruption load of the window corresponding to that position: let the monitoring interruption load be W, the maximum monitoring interruption load during the monitoring period of the patient be W_m, and the reduction coefficient β = 1 - W / W_m: when the monitoring interruption load is at the highest level of the entire process, β drops to 0, and the rate of change of the rhythm is completely reduced to 0 at that time; when the monitoring interruption load is 0, β is 1, and the rate of change of the rhythm remains unchanged. The reduced rate of change of rhythm is calculated separately for the two fluctuating rhythm sequences in lead II and lead aVF, denoted as β·ρ_II and β·ρ_F. When both values ​​corresponding to the same alarm exist, the arithmetic mean is taken, and the autonomic nervous abnormality component A = (β·ρ_II + β·ρ_F) / 2. When only one value exists, A is directly taken as the value of that value. If no value exists in either of the two paths, no component is generated, and A is the value of the autonomic nervous abnormality index at the corresponding position. The autonomic nervous system abnormality index strings together all the components at each location into a line. Each location corresponds to the moment when the original fluctuation occurs. The number of locations is consistent with the total number of fluctuation events after merging and counting. The index retains the same time granularity as the fluctuation rhythm sequence, so there is no need to re-divide the time window. The autonomic nervous system abnormality index increases naturally with the length of hospital stay and the alarm frequency, without setting an upper limit. It keeps in sync with the growth rhythm of the fluctuation rhythm sequence and there will be no disconnect between the two in terms of the number of locations. When the patient's condition changes drastically or the alarms increase or decrease suddenly, the two continue to extend synchronously according to their actual fluctuations. When the patient is discharged or monitoring is terminated, the autonomic nervous system abnormality index stops extending and there is no need for retroactive calculation.

[0036] Step S103: Based on the hematoma and edema range, the expansion rate of the lesion during the negative period is obtained by statistical analysis. Cushing's reaction is identified through vital sign data to determine signs of intracranial hypertension. The signs of intracranial hypertension are compared with the lesion expansion rate to form signs of deterioration risk.

[0037] In some embodiments, the space-occupying lesion expansion rate is obtained by statistically analyzing the negative period expansion amount based on the hematoma and edema range, including: forming a re-examination time sequence by reading the time of each re-examination based on the hematoma and edema range; identifying false negative re-examination segments by screening the re-examination time sequence for re-examination intervals where the imaging conclusion is no progression; measuring the difference in the hematoma and edema range between adjacent re-examinations based on the false negative re-examination segments to obtain the negative period expansion amount; and converting the negative period expansion amount into the space-occupying lesion expansion rate according to the false negative re-examination segments.

[0038] The time sequence of follow-up examinations is formed by retrieving the time of each examination based on the extent of hematoma and edema. Each record of the hematoma and edema extent carries the corresponding collection time of the follow-up examination. This time is recorded by the same monitoring system as the timestamp of vital signs data, and is read directly without re-measuring or recalculating. The length of the follow-up examination time sequence for the same patient is equal to the total number of head CT scans actually performed during their hospitalization, which is usually between 3 and 5 for most patients. The interval between two adjacent times is not fixed, ranging from less than 24 hours to more than a week. The follow-up examination time sequence thus preserves the true time distribution of each follow-up examination without assuming equal intervals. The follow-up examination time sequence is arranged in chronological order, with each position corresponding to one follow-up examination. Several adjacent intervals are formed between positions, with the total number of intervals being one less than the total number of follow-up examination times. These intervals are the basic units for subsequent assessment of whether there is hidden progression between follow-up examinations. If a record of the same hematoma and edema extent is clinically reviewed and a revised conclusion is issued due to image quality issues, the corresponding time in the follow-up time series will not be counted repeatedly; it will still be counted as one record based on the original acquisition time. This avoids the same follow-up examination being mistakenly counted as two adjacent intervals. If a patient only received one head CT follow-up examination during hospitalization, the follow-up time series will only have one location and will not generate any adjacent intervals. Subsequent interval-based judgments cannot be carried out in such patients; a comparable object will only be available after the second follow-up examination. If the same hematoma and edema extent involves multiple lesions, it will still be counted as the same location in the follow-up time series based on the same acquisition time. It will not be split into multiple records based on the number of lesions. In the case of multiple lesions, the total hematoma and edema volume of all lesions will be included in the record.

[0039] For example, determining false negative re-examination segments for re-examination intervals where the screening image conclusion of the re-examination time sequence is no progress includes: reading the image report conclusion of each re-examination from the re-examination time sequence to obtain a re-examination conclusion sequence; using the re-examination conclusion sequence to retrieve conclusions containing the description of no progress to generate a set of negative conclusions; screening re-examinations that contradict adjacent conclusions based on the set of negative conclusions to obtain inconsistent negative conclusions; and summarizing the inconsistent negative conclusions according to the corresponding re-examination intervals to determine false negative re-examination segments.

[0040] The follow-up conclusion sequence is obtained by retrieving the imaging report conclusions from each follow-up CT scan. After each head CT scan, the imaging report written by the radiologist is archived synchronously with the medical record system. Each record in the follow-up conclusion sequence corresponds to a moment in the follow-up time sequence. The original text of the impression description field in the report is extracted and retained verbatim without deletion or rewriting. Although the examination findings section in the report also includes a textual description of the lesion size, different physicians use significantly different words and levels of detail when describing the same lesion, making it unsuitable for direct use as a basis for automated judgment. Only the more formatted and conclusive content of the impression description is used. The follow-up conclusion sequence and the follow-up time sequence correspond one-to-one using the same set of position numbers. The follow-up conclusion sequence at position N records the corresponding moment in the follow-up time sequence at position N. The two share the same numbering and do not require a separate mapping relationship. When radiologists write reports, they rely on a visual comparison between the current and previous follow-up examinations. The follow-up conclusion sequence retains only the textual judgments derived from this comparison, excluding any inferences not previously expressed by the physician. Variations in wording by physicians are also preserved in the follow-up conclusion sequence without uniform rewriting or standardization. If two reports are issued by different physicians for a particular follow-up examination, only the first one archived is retained in the follow-up conclusion sequence. Both treatment records are kept separately, and the earlier one is not considered definitive, thus avoiding conflicting conclusions at the same time. Once the follow-up conclusion sequence is established, it serves as the sole source for subsequent text retrieval; other paragraphs of the original report are not referenced. Even if the original report is subsequently supplemented with clinical additions, the corresponding records in the follow-up conclusion sequence are not updated.

[0041] A negative conclusion set is generated by retrieving conclusion sequences containing descriptions of "no progression." A pre-defined keyword list for "no progression" includes common descriptions such as "no significant progression," "no new developments," "roughly the same as before," and "similar lesion size." During the search, each conclusion sequence is scanned, and records matching any term in the keyword list are included in the negative conclusion set. The keyword list is adjusted quarterly based on the hospital's latest reporting terminology. If a patient's hospitalization spans a keyword list adjustment period, the keyword list in effect at the time of retrieval is used, without retrospectively retrieving completed sections. Some reports may contain both descriptions of "no progression" and descriptions of new developments; in such cases, the patient is still classified as negative based on the matching of the "no progression" keyword. Specific discrepancies will be determined in subsequent steps based on quantitative data. In most patients' hospitalization follow-up conclusion sequences, negative descriptions appear in more than half of the entries. Therefore, the number of records in the negative conclusion set is usually significantly greater than the number of records classified as non-negative, which aligns with the clinical reality that most cases of cerebral hemorrhage are in a relatively stable phase. Each record in the negative conclusion set retains its original position number in the re-examination conclusion sequence, which is used in the next step to locate the corresponding hematoma and edema range record without needing to be searched again. The retrieval process only targets the text content and does not involve any numerical comparison. Numerical verification is left to the next step based on the negative conclusion set. The retrieval process itself does not judge the authenticity of the textual conclusions; it is only responsible for selecting records that meet the keyword conditions. The authenticity judgment is left to the subsequent steps to complete in combination with quantitative values. The responsibilities of the two stages are separated and do not overstep each other.

[0042] Discrepancies between negative and adjacent negative conclusions are identified through screening of the negative conclusion set. If a re-examination falls within the negative conclusion set, the corresponding hematoma and edema range records for that re-examination and the previous re-examination are retrieved. The difference between the hematoma volume and the edema volume is calculated. If any difference exceeds the measurement error range and is still considered an increase, the negative conclusion of that re-examination is determined to contradict the quantitative value, generating a discrepancy negative conclusion. The measurement error range is twice the standard deviation of the hematoma volume obtained from 10 repeated scans of the same phantom on the device, serving as the boundary between actual progression and measurement noise. Fluctuations with differences less than this boundary do not generate discrepancies negative conclusions. Re-examinations deemed not contradictory remain in the negative conclusion set as they were, maintaining their original state and not participating in subsequent comparisons. The number of discrepancies negative conclusions is usually significantly less than the total number of records in the negative conclusion set. Re-examinations where the textual conclusion matches the quantitative value account for the majority. Each discrepancy negative conclusion carries the start and end times of the corresponding re-examination interval for further aggregation and merging. If a negative result set includes a follow-up examination that happens to be the patient's first follow-up examination since hospitalization, and there is no previous follow-up examination for comparison, this follow-up examination will not be included in the judgment and will be skipped directly. The comparison will only begin when a second follow-up examination occurs. In the case where the edema volume increases while the hematoma volume remains unchanged within the hematoma and edema range, the judgment will also be included according to the rule that the difference exceeds the measurement error range. The judgment will be triggered as long as either the hematoma or edema dimension meets the condition, and both are not required to be met at the same time. This relaxed treatment is to avoid omitting cases where one dimension progresses significantly, but the other dimension of the hematoma and edema range is temporarily stable.

[0043] Discrepancies between negative and positive results are grouped into corresponding retest intervals to determine false negative retest segments. For each retest interval corresponding to a discrepancy between negative and positive results, the start time is taken from the time of the previous retest, and the end time is taken from the time the current retest occurred. When two adjacent retest intervals corresponding to discrepancies between negative and positive results are contiguous, they are merged into a continuous false negative retest segment. The start and end times of the merged segment are taken as the earliest start time and the latest end time among all the intervals involved in the merge. For example... Figure 2As shown, a hollow circle (2) indicates no progression in the follow-up report, while a solid circle (3) indicates other conclusions. If a follow-up report still uses a hollow circle but the measured volume significantly increases compared to the previous report, a double-layered hollow circle (5) indicates a non-negative conclusion. Only one non-negative conclusion is most common in false-negative follow-up reports. Reports spanning more than two follow-ups and consisting of multiple non-negative conclusions are relatively rare, usually occurring when the disease is progressively worsening and the quantitative difference in each follow-up just exceeds the measurement error range. The duration of a false-negative follow-up period is calculated by subtracting the start time from the end time. The duration varies for different patients and different periods, ranging from a few days to over a month. The duration itself serves as one of the reference points for subsequently determining the duration of latent disease progression. Each false negative retest segment carries the start and end times and the numbers of all inconsistent negative conclusions it involves, allowing subsequent verification to trace back to which specific retests led to the contradictory conclusions. Even if two false negative retest segments are temporally adjacent, they are not merged and remain independently. A patient may have zero, one, or multiple false negative retest segments during hospitalization. For the zero segment, the written conclusions and quantitative values ​​for all retests for that patient remain consistent, indicating the mechanism is functioning normally and no cases requiring attention have been identified.

[0044] The negative period expansion is calculated by measuring the difference in hematoma and edema extent between adjacent false-negative follow-up examinations. Each false-negative follow-up examination segment has a corresponding hematoma and edema extent record at both the start and end times. The hematoma volume at the end time is denoted as V_e, and the hematoma volume at the start time as V_s. The hematoma dimensional expansion ΔV = V_e - V_s. The edema volume is subtracted in the same way to obtain the edema dimensional expansion. The difference between the two hematoma and edema extent records constitutes the values ​​of these two dimensions of the negative period expansion. When a false-negative follow-up examination segment spans more than two follow-up examinations, the hematoma and edema extent records corresponding to the intermediate follow-up examinations are only used to confirm the continuity of the lesion's trend. The expansion is still calculated by subtracting the earliest and latest records of the entire segment, without accumulating from each follow-up examination: if the intermediate follow-up examination shows an increase relative to the earliest, or the latest follow-up examination shows a larger increase relative to the intermediate, the calculation method for the entire segment remains unchanged. The specific increase of the intermediate follow-up examination is not included in the result separately. A negative period expansion volume falling between a few milliliters and tens of milliliters is common in false-negative follow-up examinations. A volume significantly exceeding this range indicates substantial actual progression of the lesion during that period. The calculation of the negative period expansion volume only accepts input from false-negative follow-up examinations. Between two consecutive follow-up examinations not identified as false-negative follow-up examinations, even if there are changes in hematoma or edema volume, no negative period expansion volume record is generated. These changes, representing normal progression or absorption, are handled through routine clinical follow-up procedures. The two dimensions of the negative period expansion volume are retained independently and not combined into a single value. These values ​​are used for subsequent conversion into rates, preventing the forced merging of two distinct volume changes—hematoma and edema—into a meaningless intermediate number.

[0045] The expansion rate of the lesion during the negative period is calculated by converting the false negative re-examination segment into the negative period expansion amount. Let the hematoma dimension value of the negative period expansion be ΔV, the duration of the false negative re-examination segment be τ, and the hematoma dimension expansion rate v = ΔV / τ, with the unit being daily volume change. The edema dimension is converted in the same way; both dimensions are retained separately and not combined into a single value. For example... Figure 2 As shown in the figure, this figure presents the identification mechanism of false negative re-examination segments by comparing the re-examination sequence with the measured volume curve: the horizontal axis is the re-examination time sequence 1, and its scale is the number of re-examinations (in times), that is, the order of each re-examination in the sequence, arranged at equal intervals according to the order of re-examination, without scaling according to the actual time interval; the vertical axis is the measured volume (in milliliters). For ease of illustration, only the hematoma volume dimension is plotted. The trend of edema volume is similar and is not plotted separately. The measured volumes corresponding to all re-examinations are connected in sequence to form the measured value curve of hematoma and edema range 4; the difference in the hematoma and edema range between two adjacent re-examinations that are judged as false negative re-examination segments 6 is the negative period expansion amount 7, and the converted rate of change is the space expansion rate 8. Figure 2 The deviation between the textual conclusions and the measured values ​​is presented visually using curves, facilitating the localization of latent progression intervals. A single patient may experience multiple false-negative follow-up examinations during hospitalization. Each segment has its own independently calculated lesion expansion rate, without merging or averaging. Subsequent steps use each segment's time interval for evaluation, rather than taking a single average value for the entire hospitalization period. The lesion expansion rate is only effective within the corresponding false-negative follow-up examination time range. Gaps between segments are not interpolated or continue from the previous segment's value; these gaps are treated as missing lesion expansion rates in subsequent steps. Each negative period expansion amount strictly corresponds to a generated lesion expansion rate; the two are perfectly matched, preventing situations where there is negative period expansion amount but a missing lesion expansion rate, thus avoiding rate flattening.

[0046] The Cushing reaction is used to identify signs of intracranial hypertension (ICH) by analyzing vital signs data. The Cushing reaction is the clinical sign used to diagnose IH, clinically manifested as the simultaneous presence of elevated blood pressure, bradycardia, and irregular breathing. These three indicators correspond to systolic and diastolic blood pressure, and heart rate and respiratory rate related channels in vital signs data, respectively. All three must fall within the same 1-hour window to be considered a sign of IH. Elevated blood pressure is determined by pulse pressure, calculated as systolic blood pressure minus diastolic blood pressure. The starting value of the pulse pressure window is the difference between systolic and diastolic blood pressure at the first sampling point within that window. An increase of more than 20 mmHg in pulse pressure within a 1-hour window compared to the start of the window satisfies the blood pressure condition. Bradycardia is defined as a decrease in heart rate exceeding 15 beats / min within the window compared to the start of the window. Irregular breathing is defined as the percentage of consecutive respiratory rate readings within the window where the difference exceeds 8 beats / min, which accounts for more than one-third of all readings within the window. The systolic and diastolic blood pressure, heart rate, and respiratory rate readings used for the three criteria are directly taken from the corresponding vital signs data channels without any additional filtering or smoothing. If all three conditions are met within a given one-hour window, that window is recorded as a sign of intracranial hypertension. If any one of the three conditions is missing, it is not considered a sign of intracranial hypertension for that window. There are no compromise records for partial satisfaction; the judgment result is only either satisfied or not satisfied. Intracranial hypertension signs are judged independently window by window. Whether two consecutive windows are satisfied or not does not affect each other. A patient may repeatedly experience multiple intracranial hypertension sign windows during hospitalization, or may not experience them at all, depending on the actual changes in vital signs data.

[0047] The signs of intracranial hypertension are compared with the rate of lesion expansion to form a risk of deterioration. Each false negative follow-up segment corresponding to the lesion expansion rate has a clear start and end time. The comparison method is to check whether the start time of the window corresponding to each sign of intracranial hypertension falls within the start and end range of any false negative follow-up segment to which the lesion expansion rate belongs. If it falls within the range, it is determined that the segment and the window constitute a match. The risk of deterioration is composed of four items: the matched false negative follow-up segments and their corresponding lesion expansion rate values, the number of matched intracranial hypertension sign windows, and a list of the start times of each of the matched windows. The list of start times is not further filtered or sorted, allowing clinicians to determine which time point is most worthy of attention. A false negative follow-up segment may match zero or one or more windows of signs of intracranial hypertension within its start and end range. When zero windows are matched, the false negative follow-up segment is still retained independently and no record of signs of deterioration risk is generated. The two types of signals have completely different sources: one originates from the combined interpretation of imaging and vital signs, and the other originates from the rhythm analysis of artifact fluctuations. It is rare for them to appear simultaneously within the same time range, and when they do, they are more noteworthy than viewing either signal alone. Therefore, signs of deterioration risk only correspond to time periods that are supported by both the rate of lesion expansion and signs of intracranial hypertension. They are the result of mutual confirmation of the two types of clues, not the product of either clue being triggered alone. This requirement of dual confirmation is also the core design of this method that distinguishes it from single-signal warning. The appearance of either type of signal alone is insufficient to constitute a sign of deterioration risk; the two types of signals must truly overlap in time to be recorded.

[0048] Step S104: Obtain the synchronous overthreshold window by taking the intersection of the signs of worsening risk and the autonomic nervous system abnormality index; determine the risk level based on the statistical critical zone dwell time of the synchronous overthreshold window; and generate an early warning instruction by matching the treatment path according to the risk level.

[0049] Specifically, the intersection of the signs of worsening risk and the autonomic nervous system abnormality index is used to obtain the synchronous overthreshold window. Two indicators are considered synchronous if they both fall within the same 1-hour window and simultaneously meet their respective overthreshold conditions. When the autonomic nervous system abnormality component corresponding to a certain position in the autonomic nervous system abnormality index exceeds 1.5, the 1-hour window in which that component occurs is considered the overthreshold window for that indicator. The value 1.5 is taken from a distance above the upper quartile of the index's distribution during the monitoring period of most patients. Components exceeding this level are relatively rare in previous data. Multiple components may meet the conditions within a window; the presence of any one is considered an overthreshold for that window. The overthreshold window for signs of worsening risk is directly taken from the starting time of the recorded intracranial hypertension sign matching window. These windows are already at a 1-hour granularity and do not require additional conversion. The two types of signals have completely different sources: one originates from the combined interpretation of imaging and vital signs, and the other originates from the rhythm analysis of artifact fluctuations. It is uncommon for both to simultaneously exceed the threshold within the same hour, and when this occurs, it is more noteworthy than viewing either signal alone. The synchronous overthreshold window is obtained by taking the intersection of the two. Only when both the deterioration risk sign and the autonomic nervous system abnormality index are judged to be overthreshold within the same 1-hour window will this window be included in the synchronous overthreshold window. If either one is overthreshold alone, it will not be included. Window where the deterioration risk sign is satisfied but the autonomic nervous system abnormality index is not satisfied, or vice versa, will not be included in the synchronous overthreshold window. This method requires that the two types of heterogeneous signals truly overlap. It is precisely by relying on this strict threshold that it is implemented. The judgment criteria will not be relaxed just because a single signal seems strong enough.

[0050] In some embodiments, determining the risk level based on the critical band dwell time statistically determined by the synchronous over-threshold window includes: retrieving a threshold gap list by retrieving the difference between each indicator within the window and a preset threshold based on the synchronous over-threshold window; filtering out indicators with continuously increasing differences from the threshold gap list to generate near-over-limit indicators; determining the disorder level by statistically determining the fluctuation dispersion of each indicator based on the synchronous over-threshold window; and determining the risk level by statistically determining the critical band dwell time statistically determined by using the disorder level and the near-over-limit indicators.

[0051] A threshold gap list is obtained by retrieving the differences between each indicator within the synchronous overthreshold window and the preset threshold. The synchronous overthreshold window is accessed one by one, and the readings of each indicator within the window are compared with the thresholds. The indicators involved include five vital signs channels: heart rate, systolic blood pressure, diastolic blood pressure, respiratory rate, and blood oxygen saturation; and two composite indicators: the lesion expansion rate (taking the larger of two dimensions) and the autonomic nervous system abnormality index, for a total of seven items. Each indicator has a preset threshold: a heart rate of 60 bpm, a respiratory rate of 8 bpm, and blood oxygen saturation of 90% are considered low-risk; a systolic blood pressure of 160 mmHg, a diastolic blood pressure of 100 mmHg, a lesion expansion rate of 2 ml / day, and an autonomic nervous system abnormality index of 1.5 are considered high-risk. All thresholds are taken as the boundary values ​​of the indicator leaning towards the dangerous direction. These seven thresholds are uniformly applied within the synchronous overthreshold window and are not adjusted according to individual patient differences. Within the synchronous out-of-threshold window, each reading of each indicator is calculated using the rules described above. The difference is either the reading minus the threshold or the threshold minus the reading, depending on the hazard direction of the indicator. The differences of the seven indicators are arranged chronologically into seven sub-sequences, which are then combined to form a threshold gap list. The synchronous out-of-threshold window is therefore the only range for data collection in the threshold gap list; readings outside the window are not included in the statistics. For the same indicator in the threshold gap list, there may be multiple differences within the same window. For multiple readings of the same indicator within the corresponding window, the differences are retained individually and not aggregated or averaged. The occupancy expansion rate is updated only after each false negative retest period. In windows where the rate has not been updated, the most recent value is used and is not considered a missing reading. The update frequency of the occupancy expansion rate is already much lower than that of the vital signs channel; using the most recent value ensures that each indicator in the threshold gap list can be obtained in every window.

[0052] Indicators with continuously increasing differences are selected from the threshold gap list to generate near-limit indicators. If there are fewer than three records of the difference for a certain indicator in the threshold gap list, it is also impossible to determine whether it is continuously increasing. This situation is treated as not meeting the condition, and a conclusion is not forced based on fewer than three records. When judging the trend, only the differences recorded in the threshold gap list are considered. If the difference corresponding to three consecutive readings of a certain indicator in chronological order is larger than the previous one, it is determined that the difference of that indicator is continuously increasing. Three readings are used to avoid random fluctuations in a single reading being mistaken for a trend. For the three consecutive readings of the occupancy expansion rate, the values ​​of the three most recent false negative retest segments are used instead. Indicators that meet this condition are included in the near-limit indicators. Within a synchronous threshold window, multiple indicators may meet the condition at the same time, or none of them may meet it. Indicators with consistently flat or fluctuating differences in the threshold gap list are also not included in the near-limit indicators, as such fluctuations are more likely to be normal physiological fluctuations rather than signals of continuous deterioration. Regarding heart rate, if the difference between three consecutive readings within a certain window is 2, 5, and 9 beats / min, respectively, and increases sequentially, the heart rate is considered a near-exceeding indicator. If the third difference drops back to 4 beats / min, the condition of continuous increase is not met, and the heart rate is not included in the near-exceeding indicator. The near-exceeding indicator record also retains the specific values ​​of the three most recent differences for the corresponding indicator, allowing people to see the magnitude of the increase at a glance rather than just a binary judgment. Indicators with larger magnitudes often correspond to situations with faster deterioration, allowing clinicians to prioritize the one with the largest magnitude when multiple near-exceeding indicators appear simultaneously. The near-exceeding indicators themselves do not distinguish priorities; the accompanying values ​​are for further manual evaluation. The number of near-exceeding indicators within a simultaneous exceeding threshold window and the subsequent disorder grading jointly determine the severity of that window.

[0053] The disorder classification is determined based on the fluctuation dispersion of each indicator within the synchronous over-threshold window. The fluctuation dispersion within the synchronous over-threshold window is represented by the coefficient of variation (CV). The standard deviation of all readings for the indicator within the window is denoted as σ_w, and the mean is μ_w, so CV = σ_w / μ_w. For indicators with a negative or near-zero mean, CV is replaced by dividing σ_w by the preset threshold θ for that indicator, i.e., CV = σ_w / θ, avoiding direct comparison between the dimensionless standard deviation and the dimensionless 0.15 threshold. If the number of readings for an indicator within the synchronous over-threshold window is less than 3, the coefficient of variation for that indicator is not calculated, and it is not included in the disorder indicator count. For the remaining indicators within the synchronous over-threshold window that meet the threshold, a coefficient of variation is calculated for each. Indicators with a coefficient of variation exceeding 0.15 are counted as a disorder indicator, while indicators with low coefficients of variation are not included. When the number of disorder indicators is 0 or 1, the disorder is classified as mild; 2 to 3 indicators are classified as moderate; and 4 or more indicators are classified as severe. These three levels cover all possible disorder indicators. A severe disorder typically corresponds to significant fluctuations in multiple systems within a short period, while a mild disorder is more commonly seen in normal physiological fluctuations that already exist during periods of stable condition. Considering a synchronous overthreshold window, if the coefficients of variation for three of the seven indicators—heart rate, respiratory rate, and autonomic nervous system abnormality index—all exceed 0.15, and the disorder indicator count is 3, falling within the range of 2 to 3 indicators, the disorder classification for this synchronous overthreshold window is determined to be moderate. The coefficients of variation for the other four indicators do not exceed 0.15 and are not included in the current disorder indicator count. If any of these three indicators falls below 0.15, the disorder classification for this window will be downgraded to mild. Therefore, the disorder classification is determined independently for each window and does not carry the results of the previous window. A severe disorder in the previous window does not automatically result in a severe disorder classification in the next window.

[0054] Risk levels are determined by calculating the duration of stay in the critical zone using disorder grading and near-limit indicators. A critical zone is defined as a state interval where both conditions are met: a disorder grading of moderate or higher and at least one near-limit indicator. Window segments where only the disorder grading is met but no near-limit indicators are present or show a continuous increasing trend are not included in the critical zone. Simple indicator fluctuations may only be temporary disturbances caused by nursing procedures and do not necessarily represent true deterioration. Only fluctuations accompanied by a continuous trend approaching the danger boundary are more likely to reflect a worsening of the patient's condition. The duration of stay in the critical zone is the total duration of consecutive windows within the critical zone state. Each occurrence is counted separately. If either the disorder grading or near-limit indicator condition is no longer met, and the patient exits the critical zone, the accumulated duration is immediately calculated. The next duration is calculated from the moment the patient re-enters the critical zone; the two segments are not combined. A stay of no more than 2 hours is classified as Level 1 risk; more than 2 hours but no more than 6 hours is classified as Level 2; and more than 6 hours is classified as Level 3. Level 3 is the highest risk level given by this method. There are no intermediate levels between the three levels, nor are there any levels higher than Level 3. If a patient remains within the borderline zone for 4 hours and 30 minutes, starting at 9:00 AM and ending at 1:30 PM on the same day, falling within the 2-6 hour range, the corresponding risk level is classified as Level 2. If the patient is still within the borderline zone at the end of the monitoring, the stay duration is calculated based on the actual end time of the monitoring, without waiting for the borderline zone to naturally dissipate. Therefore, the risk level still has a clear value at the end of the monitoring, leaving no gaps.

[0055] Warning instructions are generated based on the matching of treatment pathways according to risk levels. A pre-defined correspondence exists between risk levels and treatment pathways: Level 1 corresponds to a scrolling notification on the nurse's station screen; Level 2 corresponds to a pop-up on the responsible nurse's terminal; and Level 3 corresponds to simultaneous pop-ups on both the nurse's and physician's terminals with continuous audio alerts. Once a risk level is determined, a treatment pathway is immediately matched according to the correspondence; there is no delay in generation or batch processing based on accumulated risk levels. Each risk level update triggers a new matching process. Warning instructions consist of three parts: the risk level, the corresponding treatment pathway name, and the specific instruction text. Level 3 instructions additionally include a threshold duration value for triggering the Level 3 determination, allowing physicians to understand how long the patient's condition has deviated from the expected course. The Level 3 pathway involves two examinations: an expedited repeat head CT scan and a bedside examination related to intracranial pressure (such as transcranial ultrasound to measure the diameter of the optic nerve sheath). These verify the accuracy of evidence regarding the lesion's expansion rate and signs of intracranial hypertension. The on-call physician chooses one based on the bedside verification results; the instruction text does not pre-specify which option to choose. Level 3 warning instructions require mandatory action, leaving no room for physician discretion. However, the physician remains responsible for the decision to choose between the two options after initiation. Level 1 and Level 2 warning instructions only provide suggestions and do not mandate immediate action. Only one warning instruction is valid for the same patient within the same critical zone at any given time. If the risk level escalates from Level 2 to Level 3 within the zone, the original Level 2 warning instruction is invalidated, and a new Level 3 instruction is generated. Two instructions are not retained concurrently. The invalidated Level 2 warning instruction is retained for later review and is not physically deleted. This allows for a complete reconstruction of the entire process and specific timing of the risk level escalation from Level 2 to Level 3, and both versions of the warning instruction before and after the upgrade can be fully traced without any breaks.

[0056] Step S105: Use the warning command to push the warning message to establish a terminal delivery record, summarize the silent non-response items and confirmation delay in the terminal delivery record to obtain the warning failure list, evaluate the delivery status through the warning failure list and output the warning review prompt.

[0057] Specifically, a terminal delivery record is established using early warning instructions to push early warning messages. Early warning instructions are pushed to different terminals according to the recorded treatment path: Level 1 paths are pushed to the central monitoring screen at the nurse station, displayed only in a scrolling manner, showing the bed number and risk level color-coded indicators (Level 1: green, Level 2: yellow, Level 3: red), with the color changing accordingly as the risk level changes, without requiring nurses to actively refresh the page; Level 2 paths are pushed to the responsible nurse's mobile terminal as a pop-up window with vibration alerts; Level 3 paths are pushed simultaneously to both the responsible nurse's and the on-duty physician's mobile terminals, with a pop-up window accompanied by a continuous sound alert until action is taken. When an early warning instruction is converted into an early warning message, the risk level, treatment path, specific instruction text, and the critical zone dwell time carried by Level 3 instructions are retained intact, plus a delivery identifier generated specifically for this push, used to uniquely distinguish this push from other pushes in the record. When a terminal receives and successfully renders an alert message, delivery is considered successful. The terminal delivery record includes the delivery identifier, target terminal type, and actual delivery time. If the terminal is offline or the push fails at the network layer, the same identifier is recorded, but the delivery time is left blank. A failure mark is added, and failed pushes are not automatically retried, waiting for the next rule-triggered re-push. If the same alert instruction is pushed again according to the re-push rule, a new delivery identifier is generated each time. Multiple pushes are recorded separately in the terminal delivery record, without overwriting each other, and are stored independently. The terminal delivery record thus completely preserves the entire delivery trajectory of an alert instruction from the first push to the final resolution. For two terminals in a three-level path, each receiving an alert message generates its own terminal delivery record. The delivery identifiers of the two records are different, and the confirmations from nurses and doctors are not interchangeable.

[0058] In some embodiments, summarizing silent and unresponsive entries and confirmation delays in the terminal delivery records to obtain a list of warning failures includes: reading the handling status of each warning from the terminal delivery records to establish a warning handling record; screening the warning handling records for silent and unresponsive entries and entries where the indicators did not improve after handling to obtain suspected failure entries; statistically analyzing the confirmation delays of the suspected failure entries by time period to generate a time period failure distribution; and classifying the time period failure distribution by failure type to obtain a list of warning failures.

[0059] The alert handling record is established by reading the handling status of each alert from the terminal delivery record. Each terminal delivery record is accessed sequentially, and the corresponding manual operation on the terminal is read. There are three types of operations: tapping the pop-up window to view details, pressing the mute button to stop the sound, and checking "handled" at the bottom of the pop-up window. The mute operation corresponds to a muted state, and checking "handled" corresponds to a handled state. Tapping to view does not change the status and remains recorded as "not responding." Since the nurse station screen has no interactive buttons, the corresponding terminal delivery record always remains in a "not responding" state. The alert handling record is established sequentially based on the delivery identifier in the terminal delivery record. Each record carries three items: handling status, status confirmation time, and terminal category. The status confirmation time is the actual time the operation occurred, not the push notification time. For records where no operation has yet occurred, the "not responding" status remains unchanged, and the fields are reserved for future updates; no speculative values ​​are pre-filled. If a terminal delivery record is left blank due to a failed push, the corresponding warning handling record will also not generate a handling status until the next repeated push actually delivers the record. Both types of records are linked using the same set of delivery identifiers; the absence of one terminal delivery record results in no corresponding warning handling record. In the three-tiered pathway, nurses and physicians operate their respective terminals independently, creating two separate warning handling records that do not substitute for each other. Confirmation by either party will not change the status of the other's record. The two warning handling records are also independently statistically analyzed in subsequent steps and are not merged due to belonging to the same warning instruction. From this step onward, all steps using warning handling records follow this separate approach, preventing the merging of different terminals into a single unified record and the loss of individual operational details.

[0060] For example, screening silent unresponsive entries and entries where indicators did not improve after handling in the early warning handling records to obtain suspected failed entries includes: establishing an entry handling chain by retrieving the confirmation delay of each entry and handling records based on the early warning handling records; identifying silent unresponsive entries by marking entries that were not handled after repeated early warnings from the entry handling chain; identifying entries where related indicators did not improve within a preset time period by screening the entry handling chain; and merging the silent unresponsive entries and the entries where indicators did not improve after handling as early warning entries to obtain suspected failed entries.

[0061] Based on the early warning handling records, the confirmation delay of each item is retrieved and the handling records are used to establish an item handling chain. Each record in the early warning handling records corresponds to an item according to the delivery identifier. An early warning may undergo several status changes from push to final closure. The item handling chain is the complete record that strings these changes together. Even if an early warning instruction triggers multiple pushes, it still corresponds to only one item handling chain. The confirmation delay is taken from the early warning handling records at the moment when the status of the item first changes from unresponsive to muted or handled, denoted as t_c, and subtracted from the corresponding delivery time t_d in the terminal delivery record. D = t_c - t_d, in minutes. Items whose status remains unresponsive are not calculated for D. The handling record part retrieves all operations after the status is determined in the early warning handling records, including whether there are supplementary remarks, whether there is secondary muting, and whether the handling personnel have been changed. It is written into the item handling chain in chronological order. The timestamp is from the same source as the vital signs data and hematoma and edema images, so no cross-system time conversion is required when comparing. A Level 2 alert was pushed out at 10:00 AM, muted at 10:06 AM, with a confirmation delay of 6 minutes. It was then muted again at 10:40 AM. This second muting time was recorded in the handling log sublist. If the same alert instruction generates multiple delivery markers due to repeated pushes, the confirmation delays for each marker and the handling record are calculated separately. The item handling chain is grouped according to the alert instruction, and within each group, sub-chains are further subdivided according to delivery markers. For Level 3 pathways, the delivery markers of the nurse and physician terminals correspond to their respective independent sub-chains, which are not merged.

[0062] Silenced, unresponsive entries are identified by marking repeated warnings in the entry handling chain and identifying those that have not been handled. The cumulative number of silences for each warning instruction in the entry handling chain is checked one by one. Nursing shifts are divided into day shifts, night shifts, and evening shifts. The night shift has the fewest staff, and a single nurse often monitors multiple patients simultaneously. Therefore, the proportion of silenced, unresponsive entries in the entry handling chain is usually higher during the night shift than during the day shift, which is directly related to staffing levels. The repeated warning rule is as follows: if a warning has not been handled within 5 minutes of confirmation, it is re-pushed according to the original handling path, up to a maximum of 3 times. If it is still unhandled after the 4th time, the push is stopped, and subsequent warnings are identified retrospectively based on the number of silences. If a warning instruction in an entry handling chain has accumulated 3 or more silenced states within a group, and none of the sub-chains within that group have ever shown a handled state, then that warning instruction is determined to be a silenced, unresponsive entry. This determination is based entirely on the order of status changes recorded in the entry handling chain and does not incorporate any subjective evaluation. The "Silent and Unresponsive" entry record additionally includes the time of the last instance of silence. This serves as a reference point for verifying whether the on-duty personnel ultimately took corrective action. The number of times silence occurred does not affect whether it is classified into this category; classification only considers whether the "already handled" status has appeared in the entry's handling chain. For Level 3 alerts, the nurse and physician sub-chains are counted independently. Only when both sub-chains in the entry's handling chain have accumulated three instances of silence and neither has reached the "already handled" status will this Level 3 alert be classified as a "Silent and Unresponsive" entry. If either party has even experienced a "already handled" status only once, the entire Level 3 alert will not be classified into this category. Therefore, "Silent and Unresponsive" entries are strictly limited to alerts that have genuinely not been responded to, rather than alerts that have merely been silenced.

[0063] Based on the item handling chain, items whose relevant indicators did not improve within a preset time period are identified as having no improvement after handling. Once an item handling chain shows a "handled" status, it checks whether the original indicator that triggered the warning falls below its threshold within a 2-hour window: warnings triggered by elevated blood pressure check pulse pressure, warnings triggered by abnormal heart rate check heart rate, warnings triggered by autonomic nervous system abnormality index check the index, and so on for other indicators, each compared to its respective trigger threshold. If the triggered indicator does not fall below its threshold within the 2-hour window, or if it briefly falls and then exceeds it again, it is considered that the indicator has not improved, and the handling in the item handling chain is recorded as invalid. This warning instruction is included in the "no improvement after handling" item. The record of "no improvement after handling" items also retains the name of the triggering indicator and the time of the operation marked as "handled," for direct retrieval in subsequent shift-based statistical confirmation of time delays, without needing to revisit the item handling chain. If monitoring ends less than 2 hours after the "handled" status appears, based on the available data, those clearly showing no improvement are recorded; those with insufficient data are not categorized and not presumed to be without improvement. Regarding the three-level warning triggered by elevated blood pressure, if the pulse pressure remains above the threshold for two hours after treatment, even if the treatment chain shows that the physician has completed the treatment, it is still classified as an item where the indicator has not improved after treatment. The treatment action itself does not equate to its effectiveness. The focus of the item "no improvement after treatment" is whether the indicator itself has actually decreased, not whether any treatment action was performed. The warning categorized as "no improvement after treatment" is different from the warning categorized as "silent and unresponsive": the former indicates that someone has indeed treated the patient, but the effect was unsatisfactory, while the latter indicates that no treatment has ever been performed.

[0064] Entries marked as "silent and unresponsive" and those marked as "no improvement after handling" are merged into a list of suspected failure entries. If an alert has been silent less than three times and has never been marked as "handled," it is not temporarily classified as either a "silent and unresponsive" entry or a "no improvement after handling" entry. These will be reviewed together at the end of the observation period. The merging method is to summarize each alert instruction individually, with each record carrying an additional failure type label, labeled as either "silent and unresponsive" or "no improvement after handling." The original exclusive fields for each type of entry ("silent and unresponsive" or "no improvement after handling") are retained along with the record. Both types of records are written together into this list of suspected failure entries, without needing to be stored in separate files. In principle, the same alert instruction will only fall into one of these two categories. Since alerts already classified as "silent and unresponsive" have never been marked as "handled," there is no question of whether the indicators improved after handling. The two classifications are inherently mutually exclusive in terms of time sequence; there is no situation where the same alert belongs to both categories. Each record in the suspected failure entries retains the original risk level information and initial push time of the corresponding warning instruction, and retains all original fields of the silent non-response entries or entries where the indicators did not improve after treatment. The level, time, and original fields are all taken from the source record as is, without recalculation or rewriting. A patient triggered a total of 12 warning instructions throughout their hospitalization. Among them, 3 fell into the silent non-response entries, 2 fell into the entries where the indicators did not improve after treatment, and the remaining 7 were neither silent more than 3 times nor did the indicators improve after treatment. They were not classified into any category of suspected failure entries. These 7 warnings do not constitute failure samples for this patient. Therefore, the suspected failure entries only cover the portion of warnings that were truly not effectively responded to or effectively treated throughout the hospitalization.

[0065] The confirmation delay of suspected failure items is statistically analyzed by time period to generate a time-based failure distribution. Each suspected failure item is then assigned to its corresponding shift. Nursing shifts are divided into three shifts: day shift (8:00 AM to 4:00 PM), night shift (4:00 PM to midnight), and night shift (12:00 AM to 8:00 AM). The staffing and work rhythm differ among the three shifts, with the day shift having the most staff and the night shift having the fewest. The confirmation delay of suspected failure items is calculated separately for each shift. The shift to which a suspected failure item is assigned is determined by the shift in which the corresponding warning instruction was initially sent. Even if a duplicate instruction is sent across to the next shift, the assignment remains based on the initial instruction and does not change as the handling process progresses. For items classified as silent and unresponsive, the confirmation delay is calculated by subtracting the initial push time from the last silent occurrence. For items classified as having no improvement after handling, the confirmation delay is calculated by subtracting the initial push time from the time the operation was marked as handled. Both types of delays have the same starting point but different ending points, and are uniformly converted to minutes. The two types of delays are not compared to each other, nor are they combined into a single overall failure delay index. The time-based failure distribution records the number of suspected failure items and the average confirmation delay for each shift. Each of the three shifts forms an independent statistical record for the time-based failure distribution, and they are not summed together. During a certain observation period, the number of suspected failure items is often the highest among the three shifts, consistent with the reality that night shifts are the most understaffed and have relatively slower response times. Day shifts typically have the fewest items, and the number and confirmation delay of the night shift are often between the two. The time-based failure distribution across the three shifts shows a decreasing trend with manpower allocation.

[0066] The failure distribution over time periods is categorized by failure type to obtain a warning failure list. The warning failure list is essentially a cross-tabulation. One dimension represents the three shifts: day shift, night shift, and evening shift. The other dimension represents two failure types: silent response and no improvement after handling. The cross-tabulation results in six cells. The failure distribution over time periods provides overall statistics for each shift. Based on this, each shift is further divided into two types, and the number of entries and average confirmation delay are calculated for each type. Originally, one failure distribution record per shift is split into two, each corresponding to one failure type and entering into the warning failure list. If only one failure type appears in the failure distribution for a particular shift, and the number of records for the other type in the corresponding cell is 0, the average is not calculated, but the cell is still generated and retained as a historical record of the operational quality for that shift during that period. This is for future review and comparison. The cell is not deleted simply because the number is zero, thus maintaining the fixed arrangement of the six cells in the warning failure list and ensuring that the number and position of cells remain unchanged each time the data is reviewed. Once the six cells of the early warning failure list are arranged, they are always presented in a fixed order, allowing managers to easily check the same positions each time without needing to adapt to a new layout. Within the same statistical period, different shifts, and even within the same shift, different failure types, can be directly compared side-by-side in terms of the number of entries and average delay. For example, in a certain week's statistics, the number of entries for failure types that showed no improvement after the night shift was 5, with an average delay of 22 minutes. These two figures, placed side-by-side in the early warning failure list, make the difference immediately apparent.

[0067] The system assesses delivery status through a warning failure list and outputs warning review prompts. These prompts are not targeted at individual patients, but rather reflect the operational quality of the warning system over a specific time period, from the execution of the warning instruction to the warning failure list stage. This differs significantly from steps S101 to S104, which focus on individual patient conditions, and is the only part of the entire methodology that shifts to the system level. If a nurse manages too many patients, even if every warning is seen, the response may be generally slow; such problems only become apparent across patients. The assessment rules are as follows: if any cell in the warning failure list has 3 or more entries, a corresponding warning review prompt is triggered; cells with an average confirmation delay exceeding 15 minutes, even if they have fewer than 3 entries, also trigger a prompt. Meeting either rule is sufficient; if both rules are met, only one prompt is generated, preventing duplicate triggering due to both rules being met. The warning review prompts include four items: the shift that triggered the prompt, the type of failure, the number of corresponding items, and the average confirmation delay. These are left to departmental managers to handle at the scheduling or training level; this method does not replace managers making specific rectification decisions. For example, a weekly warning failure list generated two prompts: For the night shift, there were 5 items with no improvement after handling, and an average confirmation delay of 22 minutes. Even though both the number of items exceeding 3 and the average delay exceeding 15 rules were met, a separate prompt was still generated, stating: night shift, failure type: no improvement after handling, number of items: 5, average confirmation delay: 22 minutes; For the day shift, there were 2 items with no response to silent operation, and an average confirmation delay of 18 minutes. Only the average delay exceeding 15 triggered a separate prompt, stating: day shift, failure type: no response to silent operation, number of items: 2, average confirmation delay: 18 minutes. Both prompts were sent to managers for handling.

[0068] In order to implement the above-described method embodiments, a dynamic early warning method for cerebral hemorrhage is proposed to achieve the corresponding functions and technical effects. Figure 3 This paper shows a structural block diagram of a dynamic early warning device 300 for cerebral hemorrhage provided in an embodiment of this application, including: Data acquisition unit 301 is used to acquire vital sign data from bedside monitoring and hematoma and edema images from head CT scans, determine the hematoma and edema area expansion boundary from the hematoma and edema images to delineate the hematoma and edema range, and construct a dynamic monitoring dataset by marking the interruption period of the vital sign data according to a preset time window. Anomaly identification unit 302 is used to statistically analyze the distribution of interruption periods in the dynamic monitoring dataset to obtain the monitoring interruption load, screen out the pseudo-fluctuations that are routinely filtered out from the vital sign data to obtain a set of fluctuation segments, parse the rhythm based on the set of fluctuation segments to generate a fluctuation rhythm sequence, and combine the monitoring interruption load to correct the fluctuation rhythm sequence to form an autonomic nervous system abnormality index. The sign comparison unit 303 is used to obtain the lesion expansion rate by statistically analyzing the expansion amount during the negative period of the hematoma and edema range, identify the Cushing's reaction through the vital sign data to determine the signs of intracranial hypertension, and compare the signs of intracranial hypertension with the lesion expansion rate to form signs of deterioration risk. Risk grading unit 304 is used to obtain a synchronous over-threshold window by taking the intersection of the worsening risk signs and the autonomic nervous system abnormality index, determine the risk level based on the statistical critical zone dwell time of the synchronous over-threshold window, and generate an early warning instruction by matching the treatment path according to the risk level. The early warning verification unit 305 is used to push early warning messages using the early warning command to establish a terminal delivery record, summarize the silent non-response items and confirmation delays in the terminal delivery record to obtain an early warning failure list, evaluate the delivery status through the early warning failure list, and output an early warning review prompt.

[0069] The aforementioned dynamic early warning device 300 for cerebral hemorrhage can implement one of the dynamic early warning methods for cerebral hemorrhage described in the above method embodiments. The options in the above method embodiments are also applicable to this embodiment and will not be detailed here. The remaining contents of this application embodiment can be referred to the contents of the above method embodiments, and will not be repeated in this embodiment.

[0070] The purpose of the above embodiments is to reproduce and derive the technical solution of the present invention by way of example, and to fully describe the technical solution, purpose and effect of the present invention. The purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosure of the present invention, and not to limit the scope of protection of the present invention.

Claims

1. A method for dynamic early warning of cerebral hemorrhage, characterized in that, include: Acquire bedside vital sign data and hematoma and edema images from head CT scans; determine the hematoma and edema area expansion boundary from the hematoma and edema images to delineate the hematoma and edema range; and construct a dynamic monitoring dataset by marking the interruption period of the vital sign data according to a preset time window. The monitoring interruption load is obtained by statistically analyzing the distribution of interruption periods in the dynamic monitoring dataset. The set of fluctuation segments is obtained by screening the abnormal fluctuations that are routinely filtered out from the vital sign data. The rhythm is analyzed based on the set of fluctuation segments to generate a fluctuation rhythm sequence. The fluctuation rhythm sequence is corrected by combining the monitoring interruption load to form an autonomic nervous system abnormality index. Based on the hematoma and edema range, the expansion rate of the space-occupying lesion is obtained by statistically examining the expansion amount during the negative period. The Cushing's reaction is identified through the vital signs data to determine the signs of intracranial hypertension. The signs of intracranial hypertension are compared with the space-occupying lesion expansion rate to form signs of deterioration risk. The intersection of the aforementioned signs of deterioration risk and the autonomic nervous system abnormality index is used to obtain a synchronous over-threshold window. The risk level is determined by statistically analyzing the critical zone dwell time based on the synchronous over-threshold window. An early warning instruction is generated by matching the treatment path according to the risk level. The system uses the warning command to push warning messages and establishes a terminal delivery record. It summarizes the silent and unresponsive items and the confirmation delay in the terminal delivery record to obtain a warning failure list. The system evaluates the delivery status through the warning failure list and outputs a warning review prompt.

2. The method according to claim 1, characterized in that, A dynamic monitoring dataset is constructed based on the interruption periods of the vital sign data marked by a preset time window, including: The timestamps of each collection moment are read from the vital signs data to form a vital signs time sequence stream; A list of interruption periods is obtained for the interruption periods in the time-series identification of the vital signs that exceed a preset threshold. The interruption period list is categorized and labeled according to the direction of changes in vital signs before and after the interruption to obtain an interruption marker file; The interruption marker file is aligned with the vital signs time series according to a preset time window and then merged to construct a dynamic monitoring dataset.

3. The method according to claim 1, characterized in that, The monitoring interruption load is obtained by statistically analyzing the distribution of interruption periods in the dynamic monitoring dataset, including: The interruption period sequence is obtained by retrieving the interruption period from the dynamic monitoring dataset. An interruption frequency curve is generated based on the number of interruptions per unit time according to the interruption time sequence. Based on the interruption frequency curve, periodic interruptions are filtered out to form an unplanned interruption distribution; The monitoring interruption load is obtained by accumulating the frequency of the unplanned interruptions distributed across each preset time window.

4. The method according to claim 1, characterized in that, The set of fluctuation segments obtained from the routinely filtered artifact fluctuations of the vital sign data includes: The vital signs data are decomposed into waveform channels of each lead to form a single-channel waveform; Candidate fluctuation segments are generated by extracting the artifact fluctuations before and after the monitoring alarm from the single-channel waveform. By comparing the first segment of the single-channel waveform with the previous segment of the waveform, contact artifact segments in the candidate wave segments are eliminated to obtain suspected wave segments; A set of wave segments is obtained by merging segments with similar morphologies based on the suspected wave segments.

5. The method according to claim 1, characterized in that, The lesion expansion rate is obtained by statistically analyzing the expansion amount during the negative period of the hematoma and edema range, including: Based on the hematoma and edema range, the time of each follow-up examination is recorded to form a sequence of follow-up examination times; For the re-examination time sequence screening imaging conclusion of no progress, a false negative re-examination segment is determined; The negative period extension is obtained by measuring the difference in the hematoma and edema range between adjacent re-examinations based on the false negative re-examination segment. The negative period expansion amount is converted into the placebo expansion rate based on the false negative retest segment.

6. The method according to claim 1, characterized in that, The risk level is determined based on the statistical critical band dwell time of the synchronous over-threshold window, including: Based on the synchronous over-threshold window, the difference between each indicator in the window and the preset threshold is retrieved to obtain a threshold gap list. From the threshold gap list, indicators with continuously increasing differences are filtered out to generate near-exceeding indicators; Based on the synchronous over-threshold window, the fluctuation dispersion of each indicator is statistically analyzed to determine the disorder level; The risk level is determined by using the disorder classification and the duration of the critical zone near the out-of-limit index.

7. The method according to claim 1, characterized in that, By summarizing the silent and unresponsive entries and the acknowledgment delays in the terminal delivery records, a list of early warning failures is obtained, including: The processing status of each warning is read from the terminal delivery record to establish a warning processing record; By screening the warning response records for silent, unresponsive entries and entries where the indicators did not improve after the response, suspected failure entries were identified. The failure distribution by time period is generated by statistically analyzing the confirmation delay of the suspected failure items in different time periods; The failure distribution over the specified time period is categorized by failure type to obtain a list of early warning failures.

8. The method according to claim 5, characterized in that, For the follow-up time sequence screening imaging conclusion of no progression, a false negative follow-up segment is determined, including: The image report conclusions of each re-examination are read from the re-examination time sequence to obtain the re-examination conclusion sequence; The negative conclusion set is generated by retrieving conclusions containing statements of no progress using the aforementioned review conclusion sequence; Based on the screening of the set of negative conclusions, a re-examination of contradictory conclusions yielded inconsistent negative conclusions; The non-negative conclusions are summarized according to the corresponding re-examination intervals to determine the false negative re-examination segment.

9. The method according to claim 7, characterized in that, By screening the warning response records for silent, unresponsive entries and entries where post-response indicators showed no improvement, suspected failure entries were identified, including: Based on the aforementioned early warning handling records, the confirmation delay of each item and the handling records are retrieved to establish an item handling chain; Entries that have not been processed after repeated warnings are identified from the entry processing chain markers; Entries that remain silent and unresponsive are identified. Based on the item-based treatment chain screening, if the relevant indicators do not improve within a preset time period, items for which no improvement is expected after treatment will be identified. The silent, unresponsive entries and the entries where the indicators did not improve after the action were merged into warning entries to obtain suspected failure entries.

10. A dynamic early warning device for cerebral hemorrhage, characterized in that, include: The data acquisition unit is used to acquire vital sign data from bedside monitoring and hematoma and edema images from head CT scans. It determines the hematoma and edema area expansion boundary from the hematoma and edema images to define the range of hematoma and edema. It constructs a dynamic monitoring dataset by marking the interruption period of the vital sign data according to a preset time window. An anomaly identification unit is used to statistically analyze the distribution of interruption periods in the dynamic monitoring dataset to obtain the monitoring interruption load, screen out the pseudo-variables that are routinely filtered out from the vital signs data to obtain a set of fluctuation segments, parse the rhythm based on the set of fluctuation segments to generate a fluctuation rhythm sequence, and combine the monitoring interruption load to correct the fluctuation rhythm sequence to form an autonomic nervous system abnormality index. The sign comparison unit is used to obtain the lesion expansion rate by statistically analyzing the expansion amount during the negative period of the hematoma and edema range, identify the Cushing's reaction through the vital sign data to determine the signs of intracranial hypertension, and compare the signs of intracranial hypertension with the lesion expansion rate to form signs of deterioration risk. The risk grading unit is used to obtain a synchronous over-threshold window by taking the intersection of the worsening risk signs and the autonomic nervous system abnormality index, determine the risk level based on the statistical critical zone dwell time of the synchronous over-threshold window, and generate an early warning instruction by matching the treatment path according to the risk level. The early warning verification unit is used to push early warning messages using the early warning command to establish a terminal delivery record, summarize the silent non-response items and confirmation delays in the terminal delivery record to obtain an early warning failure list, evaluate the delivery status through the early warning failure list, and output an early warning review prompt.

Images