A method and system for early warning of postoperative deterioration of a neurosurgical critically ill patient

Abstract

The application belongs to the technical field of medical information technology and clinical intensive care, and discloses a neurosurgery critical patient postoperative condition deterioration early warning method and system, which comprises a patient information input module, a vital sign acquisition module, a neural function evaluation module, a drainage state monitoring module, a test index access module, a data preprocessing module, a condition deterioration feature extraction module, a risk assessment module, an early warning output module and an early warning treatment record module. Through deep integration of intracranial multi-modal monitoring equipment and the neural function evaluation module, the intracranial pressure fluctuation characteristics and the consciousness level decline rate specific to postoperative neurosurgery are included in the feature extraction system. Meanwhile, the dynamic baseline adaptive unit can automatically adjust the integrated threshold according to different postoperative stages, the weight of the specialist precursor symptoms of intracranial hypertension and consciousness disorders is improved, early specific deterioration early signals of the nervous system can be captured from the numerical false normal, and the false negative rate and the false positive rate are greatly reduced.

Description

Technical Field

[0001] This invention belongs to the fields of medical information technology and clinical intensive care technology, specifically a method and system for early warning of postoperative deterioration of condition in critically ill neurosurgical patients. Background Technology

[0002] Postoperative management of critically ill neurosurgical patients is one of the core challenges in clinical intensive care. Following craniotomy and intracranial procedures, patients face a range of serious complications, including intracranial rebleeding, worsening cerebral edema, intracranial infection, hydrocephalus, seizures, and delayed cerebral ischemia. These complications often occur suddenly and insidiously; missing the optimal intervention window can lead to irreversible neurological damage or even death.

[0003] Currently, clinical monitoring and early warning of postoperative deterioration in neurosurgical patients mainly rely on the following methods: First, there is the issue of simplistic early warning scoring systems. While general early warning tools such as the Modified Early Warning Score (MEWS) are widely used in clinical practice, their parameter settings are primarily geared towards general internal medicine patients, lacking specificity for neurosurgical characteristics. Postoperative deterioration in neurosurgery often presents with neurological-specific manifestations such as increased intracranial pressure, decreased level of consciousness, and pupillary changes as the initial signals. However, general scoring systems do not give sufficient weight to these specialized indicators, resulting in low early warning sensitivity.

[0004] Secondly, there are post-hoc diagnostic imaging examinations. CT and MRI are the gold standard for diagnosing intracranial complications, but these examinations are usually only performed after the patient has obvious clinical symptoms. They are post-hoc confirmations rather than pre-hoc warnings, and they are time-consuming, making it difficult to meet the timeliness requirements of emergency intervention. Summary of the Invention

[0005] To address the problems mentioned in the background section, this invention provides a method and system for early warning of postoperative deterioration in critically ill neurosurgical patients.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for early warning of postoperative deterioration in critically ill neurosurgical patients, comprising: The patient information entry module is used to enter the patient's basic information, surgical information, past medical history and postoperative medical orders data after surgery, to establish a baseline for the patient's individualized health record, and to provide patient identity association and individualized reference benchmarks for subsequent modules; The vital signs acquisition module is used to collect continuous vital signs data (heart rate, blood pressure, respiratory rate, blood oxygen saturation, body temperature) and intracranial specialized monitoring data (intracranial pressure, brain temperature, brain tissue oxygen partial pressure) of patients in real time through bedside monitors and intracranial multimodal monitoring equipment, and realizes automatic docking and real-time data retrieval with the monitoring equipment through the data interface. The neurological function assessment module is used to input neurological function assessment data such as Glasgow Coma Scale score, pupillary light reflex, limb muscle strength and language function through nurse mobile terminals or bedside workstations, and supports time-series recording and trend visualization of multiple assessment results. The drainage status monitoring module is used to monitor the drainage volume, drainage fluid properties, and drainage tube patency in real time through weighing sensors, color sensors, and pressure sensors, and automatically triggers local warnings when abnormalities occur. The test indicator access module is used to automatically access the patient's multiple postoperative test results data through the hospital information system interface, including blood routine, coagulation function, inflammatory indicators and cerebrospinal fluid biochemical indicators, to realize the automatic synchronization and updating of test data. The data preprocessing module is used to perform unified formatting, outlier removal, missing value imputation and time alignment on the multi-source heterogeneous data obtained by the above modules, and generate a standardized time series dataset for subsequent analysis. The disease deterioration feature extraction module is used to extract multidimensional disease deterioration features from the standardized time series data, including vital sign trend features, intracranial pressure fluctuation features, neurological function score change features, drainage status abnormality features, and test index deviation features. The risk assessment module is used to build a pre-trained disease deterioration risk prediction model, receive multi-dimensional features output by the disease deterioration feature extraction module, and output the current patient's disease deterioration risk level and the probability of deterioration within a preset time window through model inference. The early warning output module is used to generate graded early warning information according to the risk level and deterioration probability output by the risk assessment module and according to the preset graded early warning rules, and push it to medical staff through multiple channels such as the nurse station screen, mobile terminal, SMS and sound and light alarm. The early warning and response record module is used to record the response time, response measures and results of medical staff to early warning information, and to track and record the patient's condition outcome data after the early warning is triggered. The response results and condition outcome are fed back to the risk assessment module for incremental model learning and continuous optimization.

[0007] Preferably, the vital signs acquisition module also integrates a dynamic baseline adaptive unit, which is used to automatically adjust the normal reference range of each vital sign parameter and intracranial pressure parameter according to different time stages after the patient's surgery, so as to realize individualized dynamic threshold warning and avoid false positive or false negative warnings caused by using fixed thresholds.

[0008] Preferably, the neurological function assessment module also integrates a GCS score trend analysis subunit, which is used to automatically calculate the total score change trend and the score change trend of each sub-item of the GCS score in multiple consecutive tests. When the GCS score drops more than a preset value within a preset time window or when a single item score shows abnormal fluctuations, an early warning signal for neurological function deterioration is automatically generated.

[0009] A method for an early warning system for postoperative deterioration of condition in critically ill neurosurgical patients as described above includes the following steps: S1. Obtain basic information, surgical information, past medical history and postoperative medical orders of postoperative patients through the patient information entry module, and establish a baseline for individualized health records of patients. S2. The patient's continuous vital signs data, including heart rate, blood pressure, respiratory rate, blood oxygen saturation and body temperature, are collected in real time through the vital signs acquisition module, and intracranial pressure, brain temperature and brain tissue oxygen partial pressure data are collected through the intracranial multimodal monitoring device. S3. Obtain neurological function assessment data such as Glasgow Coma Scale score, pupillary light reflex, limb muscle strength and language function of the patient through the neurological function assessment module. S4. Obtain data on postoperative drainage volume, drainage fluid characteristics, and drainage tube patency status through the drainage status monitoring module; S5. Access the patient's postoperative test results data through the test indicator access module, including complete blood count, coagulation function, inflammatory markers and cerebrospinal fluid biochemical indicators. S6. The data preprocessing module performs unified formatting, outlier removal, missing value imputation and time alignment on the multi-source heterogeneous data obtained in steps S1 to S5 to generate a standardized time series dataset. S7. Extract multidimensional deterioration features from the standardized time-series data through the disease deterioration feature extraction module, including vital sign trend features, intracranial pressure fluctuation features, neurological function score change features, drainage status abnormality features, and test index deviation features. S8. Input the multidimensional features extracted in step S7 into the pre-trained disease deterioration risk prediction model through the risk assessment module, and output the current patient's disease deterioration risk level and the probability of deterioration within the future preset time window. S9. Based on the risk level and probability of deterioration output in step S8, the early warning output module generates graded early warning information and pushes it to the terminal of medical staff. S10. The early warning and handling record module records the medical staff's handling of the early warning information and the patient's subsequent condition outcome data. The handling results are fed back to the condition deterioration risk prediction model for incremental learning and model optimization.

[0010] Preferably, the intracranial multimodal monitoring device described in S2 includes a ventricular intracranial pressure monitor, a brain tissue oxygen partial pressure monitor, and an intracranial temperature monitoring probe. The intracranial pressure, brain temperature, and brain tissue oxygen partial pressure data are continuously collected and uploaded in real time at a sampling frequency of not less than 1 Hz.

[0011] Preferably, the continuous vital sign data acquisition described in S2 further includes dynamic trend analysis of the time-series vital sign data, specifically including: S201. Using the sliding time window method, calculate the mean, standard deviation, slope of change, and extreme values ​​of each vital sign parameter within a preset time window. S202. Set corresponding baseline reference intervals according to different time stages after surgery (0-6 hours, 6-24 hours, 24-72 hours and more than 72 hours after surgery); S203. Calculate the deviation of the current vital signs parameters from the individualized baseline, and use the deviation as an early warning feature of disease deterioration.

[0012] Preferably, the neurological function assessment data described in S3 is obtained in the following manner: S301. Manually enter the Glasgow Coma Scale and its sub-items (eye opening response, verbal response, motor response) through the nurse's mobile terminal. S302. Automatically collect pupil diameter and light reflection velocity data using a pupil measuring instrument; S303. Record the assessment results of limb muscle strength, language function and swallowing function through a standardized neurological function assessment scale. S304. Organize the results of multiple neurological function assessments into a time series, calculate the trend and rate of change of GCS scores, and use them as an important predictor of neurological function deterioration.

[0013] Preferably, the drainage status monitoring described in S4 specifically includes: S401. Monitor the cumulative flow rate and flow rate per unit time of the drainage fluid in real time using a weighing sensor or a volume sensor. S402. Obtain information on the color and characteristics (bloody, purulent, clear, etc.) of the drainage fluid through a color sensor or by input by medical staff; S403. Monitor the negative pressure state inside the drainage tube using a pressure sensor to determine if there is a risk of blockage or dislodgement in the drainage tube; S404. When the drainage volume exceeds the preset threshold within a unit time, the properties of the drainage fluid change abnormally, or the pressure of the drainage tube is abnormal, an abnormal drainage status warning signal is generated.

[0014] Preferably, the data preprocessing described in S6 specifically includes: S601. Perform unified encoding and structured conversion on data from different sources and in different formats to generate standardized time-series data records; S602. Employ an outlier detection algorithm based on statistical methods to identify and mark abnormal data points caused by sensor drift, human input errors, etc. S603. Use linear interpolation or hot card imputation methods based on similar patient groups to reasonably imput missing data; S604. Using the patient's admission time as a benchmark, align and resample all data along a unified time axis to generate a standardized time-series dataset with equal time intervals.

[0015] Preferably, the extraction of disease deterioration features in S7 specifically includes: S701. Vital Sign Trend Characteristics: Extract the changing trend, fluctuation amplitude, percentage of time exceeding the threshold, and joint change pattern of each vital sign parameter within a preset time window. S702. Characteristics of intracranial pressure fluctuations: Extract the mean, peak, trough, pressure wave morphology (A wave, B wave, C wave), and calculated value of cerebral perfusion pressure of intracranial pressure. S703. Characteristics of changes in neurological function scores: Extract the total score change trend, the score change trend of each sub-item, the rate of score decline, and the characteristics of abnormal pupillary changes in the GCS score. S704. Abnormal characteristics of drainage status: abnormal increase or sudden decrease in the amount of drainage, sudden change in the color and properties of the drainage fluid, and abnormal pressure in the drainage tube. S705. Detection index deviation characteristics: Extract the dynamic trends of inflammatory indicators such as white blood cell count, neutrophil ratio, C-reactive protein, and procalcitonin, as well as the deviation characteristics of coagulation indicators such as prothrombin time and activated partial thromboplastin time.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Through the deep integration of intracranial multimodal monitoring equipment and neurological function assessment module, the intracranial pressure fluctuation characteristics and the rate of decline in consciousness level unique to neurosurgery are incorporated into the feature extraction system. At the same time, the dynamic baseline adaptive unit can automatically adjust the individualized threshold according to different postoperative stages, which increases the weight of intracranial hypertension and prodromal symptoms of consciousness disorders. It can capture early electrical signals of specific deterioration of the nervous system from falsely normal values, and significantly reduce the false alarm rate and false alarm rate. 2. The system achieves continuous data stream acquisition at the second level through the vital signs acquisition module and the drainage status monitoring module. Combined with the disease deterioration feature extraction module, it performs real-time calculations on the slope of changes in vital signs and the precursor features of changes in drainage fluid characteristics. Finally, the pre-trained risk prediction model directly outputs the probability of deterioration within a preset time window in the future. This significantly advances the warning threshold. The system can issue graded alarms in advance when the patient has not yet shown obvious clinical symptoms and has not yet reached the indications for imaging examinations, so that medical staff can provide timely treatment to the patient. Attached Figure Description

[0017] Figure 1 This is the main flowchart of the core early warning process of this invention; Figure 2 This is a detailed flowchart of the data preprocessing and feature extraction process of the present invention; Figure 3 This is a flowchart illustrating the internal reasoning and interpretability of the risk assessment model of this invention. Figure 4 This is a flowchart illustrating the continuous optimization process of the model in this invention. Figure 5 This is a flowchart of the closed-loop feedback process of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figures 1 to 5 As shown, this invention provides a method and system for early warning of postoperative deterioration in critically ill neurosurgical patients. The system includes: The patient information entry module is used to enter the patient's basic information after surgery (including age, gender, BMI index and baseline blood pressure level), surgical information (including surgical type, approach, intraoperative blood loss, operation time and intraoperative adverse events), past medical history (including chronic disease history such as hypertension, diabetes, coronary heart disease and anticoagulant drug use history) and postoperative medical order data (including the use plan of sedative drugs, antihypertensive drugs and antiepileptic drugs), and to establish a patient's individualized health record baseline. This baseline not only includes static demographic and clinical characteristics, but also includes physiological reserve index calculated based on preoperative health status, providing patient identity association and individualized reference benchmarks for subsequent modules; The vital signs acquisition module is used to collect continuous vital signs data (heart rate, blood pressure, respiratory rate, blood oxygen saturation, body temperature) and intracranial specialized monitoring data (intracranial pressure, brain temperature, brain tissue oxygen partial pressure) of patients in real time through bedside monitors and intracranial multimodal monitoring devices. It also achieves automatic docking and real-time data retrieval with monitoring devices through a data interface (supporting HL7 and DICOM protocols). The data retrieval adopts a disconnection reconnection and data caching mechanism to ensure data integrity under network fluctuation conditions. The neurological function assessment module is used to input neurological function assessment data such as Glasgow Coma Scale (GCS), pupillary light reflex (including diameter, symmetry and reflex speed), limb muscle strength (using a 0-5 grade muscle strength grading method) and language function (including spontaneous language, comprehension and naming ability) through nurse mobile terminals or bedside workstations. It also supports time-series recording and trend visualization of multiple assessment results. The assessment interface uses structured templates and intelligent prompts to reduce human input errors. The drainage status monitoring module is used to monitor the drainage volume, drainage fluid characteristics (clear, light red, bloody, purulent, turbid, etc.) and drainage tube patency status (patency, partial patency, blockage, dislodged) of the postoperative drainage tube in real time through a weighing sensor (accuracy 0.5g), a color sensor (supporting RGB three channels and color difference analysis), and a pressure sensor (range -20cmH2O to +20cmH2O, accuracy 0.5cmH2O). It automatically triggers local early warning when abnormalities occur. The sensor data adopts a redundancy verification mechanism, and automatically switches to the backup sensor channel when a single sensor fails. The laboratory test result access module is used to automatically access multiple postoperative test results data of patients through the hospital information system interface (following the HL7 FHIR standard), including complete blood count (white blood cell count, neutrophil percentage, hemoglobin and platelet count), coagulation function (prothrombin time, activated partial thromboplastin time, international normalized ratio and fibrinogen), inflammatory markers (C-reactive protein, procalcitonin and interleukin-6) and cerebrospinal fluid biochemical indicators (quantitative protein, quantitative glucose and chloride), to achieve automatic synchronization and updating of test data. The update frequency is dynamically triggered according to the execution time of the test order. The data preprocessing module is used to perform unified formatting (converting data from different sources into a unified data object model), outlier removal (dual screening based on a combination of medical reasonable range thresholds and statistical outlier detection), missing value imputation (using mean imputation, linear imputation, or multiple imputation methods according to the missing rate), and time alignment processing (resampling each data stream to an equally spaced time axis based on the absolute time after the patient's surgery), generating a standardized time-series dataset for subsequent analysis. The disease deterioration feature extraction module is used to extract multidimensional disease deterioration features from standardized time-series data, including vital sign trend features (including coefficient of variation, area exceeding threshold, and joint change index), intracranial pressure fluctuation features (including pressure wave morphology classification and cerebral perfusion pressure derived features), neurological function score change features (including GCS slope and individual fluctuation amplitude), drainage status abnormality features (including cumulative volume deviation and trait mutation index), and test index deviation features (including dynamic change rate and multiples exceeding normal range). The risk assessment module is used to build a pre-trained disease deterioration risk prediction model. This model is built based on a temporal neural network and attention mechanism. It receives multi-dimensional features output by the disease deterioration feature extraction module and outputs the current patient's disease deterioration risk level (low risk, medium risk, high risk, and very high risk) and the probability of deterioration within a future preset time window (default is set to three prediction time windows: the next 4 hours, 8 hours, and 24 hours). The early warning output module is used to generate graded early warning information according to the risk level and probability of deterioration output by the risk assessment module, and according to the preset graded early warning rules. It is then pushed to medical staff through multiple channels, including the nurse station screen (displaying a global overview and highlighting beds), mobile terminals (pushing structured early warning cards), SMS (sending key summaries), and sound and light alarms (distinguishing different risk levels by color and frequency). The early warning and response recording module is used to record the response time of medical staff to early warning information (the time difference between the issuance of the early warning and the first confirmation operation), response measures (including but not limited to medication adjustment, imaging re-examination, nursing upgrade or surgical intervention), and response results (effective, partially effective or ineffective measures). It also tracks and records the patient's condition outcome data after the early warning is triggered (including the type of deterioration event, the time of occurrence and the final prognosis). The response results and condition outcome are fed back to the risk assessment module for incremental model learning and continuous optimization.

[0020] The vital signs acquisition module also integrates a dynamic baseline adaptive unit, which automatically adjusts the normal reference ranges of various vital signs parameters and intracranial pressure parameters according to different time stages after the operation (0-6 hours, 6-24 hours, 24-72 hours and more than 72 hours after the operation). This adaptive unit uses the moving percentile method based on quantile regression, combined with the patient's preoperative baseline values, to dynamically calculate the individualized warning threshold (upper warning limit and lower warning limit) under the current time window, so as to realize individualized dynamic threshold warning and avoid false positive or false negative warnings caused by using fixed thresholds. At the same time, the unit also has a trend-sensitive acceleration function. When a parameter shifts in an abnormal direction in three consecutive sampling points, the corresponding threshold range is dynamically tightened.

[0021] The neurological function assessment module also integrates a GCS score trend analysis subunit, which is used to automatically calculate the total score change trend of GCS scores over multiple consecutive times (using linear regression to calculate the slope) and the change trend of each sub-item score (eye opening, language, motor). When the GCS score drops more than a preset value (e.g., 2 points) within a preset time window (e.g., within 4 hours) or when a single item score (especially the motor response item) shows abnormal fluctuations (e.g., from compliant movement to abnormal flexion or hyperextension), an early warning signal for neurological function deterioration is automatically generated. This subunit also introduces the concept of GCS change acceleration, and dynamically increases the warning priority when the rate of score decline shows an accelerating trend.

[0022] The drainage status monitoring module also integrates a multi-dimensional abnormality fusion judgment subunit, which is used to simultaneously fuse monitoring data from three dimensions: abnormal drainage volume (drainage volume per unit time exceeds 1.5 times the upper limit of the normal reference value at the same postoperative stage or is lower than 0.5 times the lower limit), abnormal drainage fluid properties (the color difference value of the color sensor deviates from the Euclidean distance of the corresponding normal property color gamut in the baseline library by more than the set threshold), and abnormal drainage tube pressure (pressure value exceeds the range of -15cmH2O to +10cmH2O or the pressure waveform shows a continuous negative pressure plateau). Only when the abnormal judgment results of at least two dimensions are consistent (i.e., the three-out-of-two logic is satisfied) will a drainage-related warning be triggered to reduce invalid warnings caused by false alarms from a single sensor. At the same time, this subunit also records the combination patterns of three-dimensional abnormalities to distinguish between bleeding events, infectious events, or mechanical obstruction.

[0023] The risk assessment module also integrates a model interpretability output unit, which outputs the top N features (N defaults to 5) that contribute the most to the current prediction result, along with their contribution weights (calculated based on SHAP or integral gradient method), while outputting the risk level and the probability of deterioration. This presents the causes of the risk in a visual way (using a waterfall plot or force-force diagram) to assist clinical decision-making and make the early warning basis traceable and understandable.

[0024] The early warning and response record module also integrates a model incremental learning unit, which periodically (e.g., every 7 days or when the cumulative number of newly labeled samples reaches 200) uses the complete data sequence before and after the early warning (including all monitoring data from 24 hours before the early warning to 48 hours after the early warning), medical staff's response records, and patient condition outcome data as new labeled samples to retrain the disease deterioration risk prediction model offline or perform online incremental learning (using elastic weight consolidation and experience playback mechanisms to prevent catastrophic amnesia), so as to achieve continuous optimization of the early warning model as clinical data accumulates.

[0025] The method includes the following steps: S1. Obtain the patient's basic information, surgical information, past medical history and postoperative medical orders through the patient information entry module, and establish a patient's individualized health record baseline. The baseline construction process also includes a comprehensive score of the patient's preoperative physiological reserve, and uses this score as the individualized denominator factor for all subsequent feature normalization processing. S2. The patient's continuous vital signs data, including heart rate, blood pressure, respiratory rate, blood oxygen saturation and body temperature, are collected in real time through the vital signs acquisition module, and intracranial pressure, brain temperature and brain tissue oxygen partial pressure data are collected through the intracranial multimodal monitoring device. S3. Obtain neurological function assessment data such as Glasgow Coma Scale score, pupillary light reflex, limb muscle strength and language function of the patient through the neurological function assessment module. S4. Obtain data on postoperative drainage volume, drainage fluid characteristics, and drainage tube patency status through the drainage status monitoring module; S5. Access the patient's postoperative test results data through the test indicator access module, including complete blood count, coagulation function, inflammatory markers and cerebrospinal fluid biochemical indicators. S6. The data preprocessing module performs unified formatting, outlier removal, missing value imputation and time alignment on the multi-source heterogeneous data obtained in steps S1 to S5 to generate a standardized time series dataset. S7. Extract multidimensional features of disease deterioration from standardized time-series data through the disease deterioration feature extraction module, including vital sign trend features, intracranial pressure fluctuation features, neurological function score change features, drainage status abnormality features, and test index deviation features. S8. Input the multidimensional features extracted in step S7 into the pre-trained disease deterioration risk prediction model through the risk assessment module, and output the current patient's disease deterioration risk level and the probability of deterioration within the future preset time window. S9. Based on the risk level and probability of deterioration output in step S8, the early warning output module generates graded early warning information and pushes it to the terminal of medical staff. S10. The early warning and handling record module records the medical staff's handling of the early warning information and the patient's subsequent condition outcome data. The handling results are fed back to the condition deterioration risk prediction model for incremental learning and model optimization.

[0026] The intracranial multimodal monitoring equipment in S2 includes a ventricular intracranial pressure monitor (with drainage function), a brain tissue oxygen partial pressure monitor (using fluorescence quenching method), and an intracranial temperature monitoring probe (thermometer type). Intracranial pressure, brain temperature, and brain tissue oxygen partial pressure data are continuously collected and uploaded in real time at a sampling frequency of not less than 1Hz. All intracranial monitoring data are accompanied by device status indicators and signal quality indices, which are used for automatic marking and removal of low-quality data in the subsequent preprocessing stage.

[0027] Continuous vital sign data acquisition in S2 also includes dynamic trend analysis of time-series vital sign data, specifically including: S201. Using a sliding time window method (the window length can be configured to 15 minutes, 30 minutes, 60 minutes and 120 minutes, and the step size is fixed at 1 minute), the mean, standard deviation, slope of change (using the least squares linear fitting method) and extreme values ​​(maximum and minimum values ​​and the time of occurrence) of each vital sign parameter within the preset time window are calculated. At the same time, the baroreflex sensitivity index is calculated for the coupling relationship between heart rate and blood pressure. S202. Set corresponding baseline reference intervals according to different time stages after the operation of the patient (0-6 hours, 6-24 hours, 24-72 hours and more than 72 hours after the operation). The baseline reference interval not only includes the population reference range, but also embeds the patient's individualized preoperative baseline value as the central anchor point. The Bayesian framework is used to weight and fuse the population prior and the individual prior to generate a hybrid baseline. S203. Calculate the deviation of the current vital signs parameters from the individualized baseline (using the standard score Z-score and relative percentage deviation), and use the deviation as an early warning feature of disease deterioration. At the same time, record the duration of deviation and the area of ​​deviation (time-deviation integral) as quantitative indicators in feature engineering and input them into the subsequent model.

[0028] Neurological function assessment data in S3 were obtained through the following methods: S301. The Glasgow Coma Scale and its sub-items (eye opening response 1-4 points, verbal response 1-5 points, motor response 1-6 points) are manually entered through the nurse's mobile terminal. The entry interface uses a graphical selector and is accompanied by standardized operation video guidance to reduce variability among raters. S302. The pupil diameter (accurate to 0.1 mm) and light reflection velocity (contraction rate and dilation rate) data are automatically collected by the pupil measuring instrument (automatic infrared pupil measuring instrument), and the pupil asymmetry index is automatically calculated. When the asymmetry index exceeds 0.3 mm, it is automatically marked as pupil abnormality. S303. The results of assessments of limb muscle strength (using bilateral comparative recording to distinguish the differences between the left and right sides), language function (assessed from three dimensions: fluency, comprehension and retelling ability) and swallowing function (using the Kubota test grading) were recorded using standardized neurological function assessment scales. S304. Organize the results of multiple neurological function assessments into a time series, calculate the trend of GCS score changes (the slope is calculated after noise reduction using the exponential smoothing method) and the rate of change (i.e. the amount of change in score per unit time), and at the same time, perform a time-series correlation analysis between abnormal pupillary changes (such as unilateral dilation or loss of light reflex) and GCS decline as an important predictor of neurological function deterioration.

[0029] The drainage status monitoring in S4 specifically includes: S401. The cumulative flow rate and flow rate per unit time (such as the flow rate per hour) of the drainage fluid are monitored in real time by a weighing sensor or a capacity sensor (using capacitive liquid level detection), and the cumulative amount is integrated over time to generate a cumulative drainage flow rate curve. S402: Automatically acquires the reflectance curve of drainage fluid in the visible light band (430nm-660nm) through a color sensor (based on reflectance spectrum analysis), converts it into color space coordinates, calculates the matching degree with the built-in normal characteristic color gamut library, and automatically classifies the color and characteristics of drainage fluid (bloody, purulent, clear, turbid, bile-colored, etc.). At the same time, it supports the review and input of medical staff as a reference standard for the sensor's judgment results. S403. Monitor the negative pressure state in the drainage tube using a pressure sensor (sampling frequency of 10Hz), extract pressure waveform characteristics (including average negative pressure value, pressure fluctuation amplitude and waveform slope), and calculate the drainage tube patency index by combining the pressure difference between the proximal and distal ends of the drainage tube to determine whether there is a risk of blockage (pressure continuously rising to a positive value) or tube dislodgement (pressure suddenly dropping to zero or close to atmospheric pressure). S404. When the drainage volume exceeds the preset threshold within a unit of time (e.g., drainage volume > 50 ml per hour or an increase of > 100% compared to the previous period), the characteristics of the drainage fluid change abnormally (the color sensor judgment result jumps to the bloody or purulent category), or the drainage tube pressure is abnormal (pressure continuously > +5 cmH2O or < -18 cmH2O), an abnormal drainage status warning signal is generated. Each abnormal signal is accompanied by a severity rating (mild, moderate, severe).

[0030] Data preprocessing in S6 specifically includes: S601. Perform unified encoding and structured conversion on data from different sources and in different formats (convert JSON, XML and HL7 messages into DataFrame format) to generate standardized time-series data records. Each record includes a timestamp, patient identifier, parameter type, parameter value and data source marker. S602. An outlier detection algorithm based on statistical methods (combining interquartile range method and local anomaly factor method) is used to identify and mark abnormal data points caused by sensor drift, human input errors, etc. For data points marked as abnormal, the original value is retained and a quality flag is added instead of deleting them directly, so that they can be selected as needed in the feature extraction stage. S603. Use linear interpolation (for transient missing data, the missing window is less than 5 sampling points) or hot card imputation method based on similar patient groups (for longer-term missing data, retrieve data with similar demographic characteristics and disease stages from historical databases for matching and imputation) to reasonably imput the missing data, and generate imputation confidence labels after imputation. S604. Using the patient's admission or the end of surgery as a unified time reference zero point, align and resample all data according to a unified time axis (using equal interval 1-minute resampling, and a hybrid filling strategy combining forward filling and linear interpolation) to generate a standardized time-series dataset with equal time intervals, ensuring that the time matching accuracy between each data stream is within 1 second.

[0031] The specific features extracted in S7 for disease deterioration include: S701. Vital Sign Trend Characteristics: Extract the trend of each vital sign parameter within a preset time window (slope of linear trend and curvature of quadratic trend), fluctuation amplitude (maximum and minimum difference and amplitude variation coefficient), percentage of time exceeding threshold (the percentage of the cumulative time of each parameter exceeding the individualized dynamic threshold to the total window duration), and joint change patterns (such as the phase plane trajectory area of ​​heart rate-blood pressure, the cross-correlation hysteresis coefficient between blood oxygen saturation and respiratory rate). S702. Intracranial pressure fluctuation characteristics: Extract the mean, peak, trough values ​​of intracranial pressure, pressure wave morphology classification (identify the frequency and duration of A wave (plateau wave), B wave (rhythmic wave), and C wave (irregular wave) occurrences and the calculated value of cerebral perfusion pressure (calculate cerebral perfusion pressure = mean arterial pressure - intracranial pressure, and extract the cumulative duration of cerebral perfusion pressure below 60 mmHg). S703. Characteristics of changes in neurological function scores: Extract the overall score change trend (time series slope), the score change trend of each sub-item (the trend of eye opening, language, and motor are calculated separately), the score decline rate (the change between two adjacent assessments is calculated by the difference method and divided by the time interval), and the characteristics of abnormal pupillary changes (duration of pupil size asymmetry and duration of loss of light reflex). S704. Abnormal characteristics of drainage status: Extract abnormal increases or decreases in drainage volume (the rate of change of drainage volume per unit time exceeds twice the standard deviation of the normal fluctuation range), sudden changes in the color characteristics of drainage fluid (the Euclidean distance abrupt change index of the color space coordinates), and abnormal drainage tube pressure (the cumulative duration of negative pressure value deviating from the baseline by more than 10 cmH2O). The abnormal indicators of the three dimensions are cross-combined and coded. S705. Deviation characteristics of test indicators: Extract the dynamic trend of inflammatory indicators such as white blood cell count, neutrophil ratio, C-reactive protein, and procalcitonin (calculate the percentage change rate between two adjacent test results), as well as the deviation characteristics of coagulation indicators such as prothrombin time and activated partial thromboplastin time (the ratio of the measured value to the upper limit of the normal reference range is used as the deviation factor), and extract the correlation characteristics between multiple test indicators (such as the ratio of the synchronous rise slope of white blood cells and C-reactive protein).

[0032] The construction and training process of the disease exacerbation risk prediction model in S8 specifically includes: S801. Collect complete clinical data of patients after historical neurosurgery, including vital sign time series data, neurological function assessment records, drainage monitoring data, test results, and postoperative condition deterioration event labels (preset target events include six categories: intracranial rebleeding, worsening cerebral edema, intracranial infection, hydrocephalus, epileptic seizures, and delayed cerebral ischemia, each event has clear diagnostic criteria based on imaging or clinical manifestations), construct a training dataset, and perform time alignment on positive and negative samples. Use the warning time as the anchor point to extract a fixed length of feature sequence forward and a fixed length backward to determine whether the target event has occurred. S802. Perform feature engineering on the training dataset, extract the multidimensional disease deterioration features from step S7, and normalize the features (using Z-score standardization) and perform dimensionality reduction (using principal component analysis or recursive elimination based on feature importance to retain principal components with a cumulative contribution rate of 95%). At the same time, handle the class imbalance problem (using synthetic minority class oversampling technology or Focal Loss weighted strategy). S803. Construct a hybrid deep learning model based on temporal neural networks and attention mechanisms. The specific architecture is as follows: The bottom layer uses a bidirectional long short-term memory network (Bi-LSTM, with a hidden layer dimension of 128) or a gated recurrent unit (GRU, with a hidden layer dimension of 64) to capture the temporal dependencies of multi-source monitoring data. The upper layer introduces a multi-head self-attention mechanism (with 4 attention heads) to automatically identify key time windows (such as 4-6 hours before the warning) and key feature subsets that have a high contribution to the prediction of disease deterioration. The top layer outputs the probability of each category through a fully connected layer and a Softmax activation function. S804. Using postoperative deterioration events as labels (using multi-label classification strategies or one-to-one / one-to-many OVR strategies to handle multiple types of events), the hybrid deep learning model is trained using the cross-entropy loss function (combined with class weight factors) as the loss function. The optimizer is AdamW (the learning rate adopts a cosine annealing decay strategy), and the generalization ability of the model is evaluated using the time series cross-validation method (using forward chain validation, that is, dividing the training set and validation set in time order to avoid future data leakage). S805. Deploy the trained model as a real-time inference service (using ONNX or TensorRT acceleration framework, with end-to-end inference latency less than 100ms). Receive real-time multidimensional feature data using a sliding time window (window length is the monitoring data of the past 4 hours, with a step size of 1 minute). Output the current patient's disease deterioration risk level (four levels: low risk, medium risk, high risk, and very high risk, divided according to the percentile threshold of the Softmax output probability) and the probability of deterioration within a future preset time window (4 / 8 / 24 hours).

[0033] The disease deterioration risk prediction model also integrates a model interpretability module. This module outputs the top N features (N defaults to 5, but can be configured) that contribute the most to the current risk prediction result, along with their contribution weights (represented by standard SHAP values, with positive values ​​indicating promotion of deterioration and negative values ​​indicating inhibition of deterioration), using the SHAP (SHapley Additive exPlanations) value calculation method while outputting the risk level. These features are then visualized on the warning interface in the form of radar charts or bar graphs to help medical staff understand the basis for the warning and develop targeted intervention plans. At the same time, this module can generate readable text explanations, such as the main driving factors for the warning being: increased frequency of intracranial pressure A waves (weight +0.32), decreased GCS motor score (weight +0.28), and prolonged duration of cerebral perfusion pressure below the threshold (weight +0.21).

[0034] The tiered early warning system in S9 specifically includes: S901. When the risk level is low (probability of deterioration < 20%), the system only records data and provides routine monitoring prompts without triggering active alarms. However, it continuously tracks the characteristic trend in the background. If the predicted probability of the sliding window shows a monotonically increasing trend for three consecutive times, it will automatically upgrade to the attention level and display a yellow background prompt on the monitoring interface. S902. When the risk level is medium risk (probability of deterioration is 20%-50%), the system sends a yellow warning notification to the responsible nurse's mobile terminal, prompting to strengthen rounds and close observation. The warning notification includes a summary of the current abnormal characteristics and a list of items that are recommended to be monitored. At the same time, the warning information is displayed on the nurse station screen with a yellow flashing border, and the reminder is repeated every 30 minutes until it is confirmed to be viewed. S903. When the risk level is high (probability of deterioration is 50%-80%), the system will send an orange alert to the responsible nurse and the on-duty doctor at the same time, and push specific abnormal indicators and suggested intervention measures (such as suggesting immediate re-examination of head CT or suggesting adjustment of dehydration drug dosage). The alert information is accompanied by a one-click quick operation entry (such as direct call, opening emergency examination request form, etc.), and is displayed on the nurse station screen in orange with continuous flashing, and is repeated every 10 minutes until the treatment is confirmed. S904. When the risk level is extremely high (80% probability of deterioration), the system immediately triggers a red emergency warning and notifies the emergency team (including the responsible nurse, on-duty doctor, neurosurgery second-line and anesthesiology standby) through multiple channels such as audible and visual alarms (using an independent audible and visual alarm, with a red flashing light and a continuous buzzer), SMS and telephone. It also automatically pops up a summary interface of the patient's key monitoring data (integrating current real-time vital signs, the latest test results, historical imaging comparisons and risk contribution characteristic waterfall charts), and automatically records the warning trigger time and starts the emergency timer.

[0035] The early warning response records and closed-loop feedback in S10 specifically include: S1001. Record the response time of medical staff after receiving the warning information (the time difference from the warning push to the first click to view details or confirm receipt), treatment measures (recorded through a structured form, including but not limited to adjusting the type and dosage of medication, emergency CT / MRI examination, lumbar puncture and manometry, surgical intervention, cardiopulmonary resuscitation support, etc., with free text supplementary descriptions supported), and treatment results (reported by medical staff after a preset time, with evaluation indicators including symptom improvement, stable monitoring parameters, or further deterioration). S1002. Track and record patient condition data after warnings are triggered, including whether the condition has worsened (based on whether the six preset categories of events occur within 24 hours of triggering), the type and time of the worsening event (accurate to the minute), and the patient's final prognosis (including the endpoint of improvement and discharge, recovery, residual neurological dysfunction, vegetative state or death, etc.). All outcome data must be reviewed and confirmed by the attending physician before they can be entered into the database as valid annotation samples. S1003. Take the complete data sequence (including original waveforms, preprocessed data, feature vectors, and intermediate calculation results), treatment records (structured treatment codes and effect scores), and disease outcome data (event labels and outcome labels) from 24 hours before the warning is triggered to 48 hours after the warning is triggered as new labeled samples. According to the set update strategy (such as offline batch update once a week, or triggering online incremental update when the cumulative number of new high-quality samples reaches a preset number), the data is periodically fed back into the disease deterioration risk prediction model for incremental training (using the EWC elastic weight consolidation algorithm to prevent forgetting previous knowledge) and model iterative optimization (comparing the AUC-ROC and F1-score of the new and old models on the validation set through A / B testing, and deploying the best one). This enables the warning system to continuously improve itself, so that the predictive performance of the model gradually improves with the accumulation of clinical data.

[0036] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0037] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A pre-operative deterioration early warning system for critically ill neurosurgical patients, characterized in that, include: The patient information entry module is used to enter the patient's basic information, surgical information, past medical history and postoperative medical orders data after surgery, to establish a baseline for the patient's individualized health record, and to provide patient identity association and individualized reference benchmarks for subsequent modules; The vital signs acquisition module is used to collect continuous vital signs data and intracranial specialist monitoring data of patients in real time through bedside monitors and intracranial multimodal monitoring equipment, and realizes automatic docking and real-time data retrieval with the monitoring equipment through the data interface; The neurological function assessment module is used to input neurological function assessment data such as Glasgow Coma Scale score, pupillary light reflex, limb muscle strength and language function through nurse mobile terminals or bedside workstations, and supports time-series recording and trend visualization of multiple assessment results. The drainage status monitoring module is used to monitor the drainage volume, drainage fluid properties, and drainage tube patency in real time through weighing sensors, color sensors, and pressure sensors, and automatically triggers local warnings when abnormalities occur. The test indicator access module is used to automatically access the patient's multiple postoperative test results data through the hospital information system interface, including blood routine, coagulation function, inflammatory indicators and cerebrospinal fluid biochemical indicators, to realize the automatic synchronization and updating of test data. The data preprocessing module is used to perform unified formatting, outlier removal, missing value imputation and time alignment on the multi-source heterogeneous data obtained by the above modules, and generate a standardized time series dataset for subsequent analysis. The disease deterioration feature extraction module is used to extract multidimensional disease deterioration features from the standardized time-series data, including vital sign trend features, intracranial pressure fluctuation features, neurological function score change features, drainage status abnormality features, and test index deviation features. The risk assessment module is used to build a pre-trained disease deterioration risk prediction model, receive multi-dimensional features output by the disease deterioration feature extraction module, and output the current patient's disease deterioration risk level and the probability of deterioration within a preset time window through model inference. The early warning output module is used to generate graded early warning information according to the risk level and deterioration probability output by the risk assessment module and according to the preset graded early warning rules, and push it to medical staff through multiple channels such as the nurse station screen, mobile terminal, SMS and sound and light alarm. The early warning and response record module is used to record the response time, response measures and results of medical staff to early warning information, and to track and record the patient's condition outcome data after the early warning is triggered. The response results and condition outcome are fed back to the risk assessment module for incremental model learning and continuous optimization.

2. The early warning system for postoperative deterioration of critically ill neurosurgical patients according to claim 1, characterized in that: The vital signs acquisition module also integrates a dynamic baseline adaptive unit, which is used to automatically adjust the normal reference range of various vital signs parameters and intracranial pressure parameters according to different time stages after the operation of the patient, so as to realize individualized dynamic threshold warning and avoid false positive or false negative warnings caused by using fixed thresholds.

3. The early warning system for postoperative deterioration of critically ill neurosurgical patients according to claim 2, characterized in that: The neurological function assessment module also integrates a GCS score trend analysis subunit, which is used to automatically calculate the total score change trend and the score change trend of each sub-item of GCS scores in multiple consecutive tests. When the GCS score drops more than the preset value within a preset time window or when a single item score shows abnormal fluctuations, an early warning signal for deterioration of neurological function is automatically generated.

4. A method applicable to the early warning system for postoperative deterioration of condition in critically ill neurosurgical patients as described in any one of claims 1-3, characterized in that, Includes the following steps: S1. Obtain basic information, surgical information, medical history and postoperative medical orders of postoperative patients through the patient information entry module, and establish a baseline for individualized health records of patients. S2. The patient's continuous vital signs data, including heart rate, blood pressure, respiratory rate, blood oxygen saturation and body temperature, are collected in real time through the vital signs acquisition module, and intracranial pressure, brain temperature and brain tissue oxygen partial pressure data are collected through the intracranial multimodal monitoring device. S3. Obtain neurological function assessment data such as Glasgow Coma Scale score, pupillary light reflex, limb muscle strength and language function of the patient through the neurological function assessment module. S4. Obtain data on postoperative drainage volume, drainage fluid characteristics, and drainage tube patency status through the drainage status monitoring module; S5. Access the patient's postoperative test results data through the test indicator access module, including complete blood count, coagulation function, inflammatory markers and cerebrospinal fluid biochemical indicators. S6. The data preprocessing module performs unified formatting, outlier removal, missing value imputation and time alignment on the multi-source heterogeneous data obtained in steps S1 to S5 to generate a standardized time series dataset. S7. Extract multidimensional deterioration features from the standardized time-series data through the disease deterioration feature extraction module, including vital sign trend features, intracranial pressure fluctuation features, neurological function score change features, drainage status abnormality features, and test index deviation features. S8. Input the multidimensional features extracted in step S7 into the pre-trained disease deterioration risk prediction model through the risk assessment module, and output the current patient's disease deterioration risk level and the probability of deterioration within the future preset time window. S9. Based on the risk level and probability of deterioration output in step S8, the early warning output module generates graded early warning information and pushes it to the terminal of medical staff. S10. The early warning and handling record module records the medical staff's handling of the early warning information and the patient's subsequent condition outcome data. The handling results are fed back to the condition deterioration risk prediction model for incremental learning and model optimization.

5. The method of the early warning system for postoperative deterioration of critically ill neurosurgical patients according to claim 4, characterized in that: The intracranial multimodal monitoring device described in S2 includes a ventricular intracranial pressure monitor, a brain tissue oxygen partial pressure monitor, and an intracranial temperature monitoring probe. The intracranial pressure, brain temperature, and brain tissue oxygen partial pressure data are continuously collected and uploaded in real time at a sampling frequency of not less than 1 Hz.

6. The method of the early warning system for postoperative deterioration of critically ill neurosurgical patients according to claim 5, characterized in that: The continuous vital sign data acquisition described in S2 also includes dynamic trend analysis of the time-series vital sign data, specifically including: S201. Using the sliding time window method, calculate the mean, standard deviation, slope of change, and extreme values ​​of each vital sign parameter within a preset time window. S202. Set corresponding baseline reference intervals according to different time stages after the patient's surgery; S203. Calculate the deviation of the current vital signs parameters from the individualized baseline, and use the deviation as an early warning feature of disease deterioration.

7. The method of the early warning system for postoperative deterioration of critically ill neurosurgical patients according to claim 6, characterized in that: The neurological function assessment data described in S3 were obtained through the following methods: S301. Manually input Glasgow Coma Scale scores and their sub-items via a nurse's mobile terminal; S302. Automatically collect pupil diameter and light reflection velocity data using a pupil measuring instrument; S303. Record the assessment results of limb muscle strength, language function and swallowing function through a standardized neurological function assessment scale. S304. Organize the results of multiple neurological function assessments into a time series, calculate the trend and rate of change of GCS scores, and use them as an important predictor of neurological function deterioration.

8. The method of the early warning system for postoperative deterioration of critically ill neurosurgical patients according to claim 7, characterized in that: The drainage status monitoring described in S4 specifically includes: S401. Monitor the cumulative flow rate and flow rate per unit time of the drainage fluid in real time using a weighing sensor or a volume sensor. S402. Obtain the color and properties of the drainage fluid through a color sensor or by inputting information by medical staff. S403. Monitor the negative pressure state inside the drainage tube using a pressure sensor to determine if there is a risk of blockage or dislodgement in the drainage tube; S404. When the drainage volume exceeds the preset threshold within a unit time, the properties of the drainage fluid change abnormally, or the pressure of the drainage tube is abnormal, an abnormal drainage status warning signal is generated.

9. The method of the early warning system for postoperative deterioration of critically ill neurosurgical patients according to claim 8, characterized in that: The data preprocessing described in S6 specifically includes: S601. Perform unified encoding and structured conversion on data from different sources and in different formats to generate standardized time-series data records; S602. Employ an outlier detection algorithm based on statistical methods to identify and mark abnormal data points caused by sensor drift, human input errors, etc. S603. Use linear interpolation or hot card imputation methods based on similar patient groups to reasonably imput missing data; S604. Using the patient's admission time as a benchmark, align and resample all data along a unified time axis to generate a standardized time-series dataset with equal time intervals.

10. The method of the early warning system for postoperative deterioration of critically ill neurosurgical patients according to claim 9, characterized in that: The extraction of disease deterioration features described in S7 specifically includes: S701. Vital Sign Trend Characteristics: Extract the changing trend, fluctuation amplitude, percentage of time exceeding the threshold, and joint change pattern of each vital sign parameter within a preset time window. S702. Characteristics of intracranial pressure fluctuations: Extract the mean, peak, trough, pressure wave morphology, and calculated value of cerebral perfusion pressure of intracranial pressure. S703. Characteristics of changes in neurological function scores: Extract the total score change trend, the score change trend of each sub-item, the rate of score decline, and the characteristics of abnormal pupillary changes in the GCS score. S704. Abnormal characteristics of drainage status: abnormal increase or sudden decrease in the amount of drainage, sudden change in the color and properties of the drainage fluid, and abnormal pressure in the drainage tube. S705. Detection index deviation characteristics: Extract the dynamic trends of inflammatory indicators such as white blood cell count, neutrophil ratio, C-reactive protein, and procalcitonin, as well as the deviation characteristics of coagulation indicators such as prothrombin time and activated partial thromboplastin time.

Images