Methods, devices, electronic equipment, and storage media for determining the probability of an event occurring.
By aligning and extracting the patient's baseline static features and multi-channel time-series data along the time axis, and combining sequence coding networks and task branches, the problem of low accuracy in postoperative adverse event assessment in existing technologies has been solved, achieving efficient and accurate assessment and dynamic trend prediction of multiple types of events.
Patent Information
- Application Number
- CN202611388958.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-08
- Publication Date
- 2026-11-10
AI Technical Summary
In existing technologies, the accuracy of assessing the probability of postoperative adverse events based on patients' static characteristic data or intraoperative and postoperative data is low.
By acquiring patients' baseline static feature data and multi-channel time-series data, time axis alignment and statistical feature extraction within a preset time window are performed. The data are then input into a sequence coding network, and multiple task branches are used to evaluate the probability of occurrence of various predefined events.
It enables parallel assessment of multiple types of postoperative adverse events, improving the accuracy of event probability assessment and the ability to predict dynamic trends.
Smart Images

Figure CN123072767A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of artificial intelligence technology and data processing technology, and in particular to a method, apparatus, electronic device and storage medium for determining the probability of an event occurring. Background Technology
[0002] With the development of artificial intelligence technology, the assessment of adverse events following surgery has become an important research direction in the field of clinical decision support. Adverse events refer to unexpected conditions that occur during the recovery process after a surgical procedure. Accurately assessing the probability of these events is of significant technical value for optimizing perioperative management strategies.
[0003] Among existing assessment methods, one type assesses the probability of each adverse event based solely on the patient's static characteristics (such as age, gender, weight, and past medical history); another type assesses the probability of each adverse event based on intraoperative data and postoperative recovery data after the patient's surgical procedure. However, both of these methods rely solely on the patient's static characteristics or intraoperative data and postoperative recovery data to assess the probability of each adverse event, resulting in low accuracy in assessing the probability of each adverse event.
[0004] Therefore, how to accurately assess the probability of occurrence of each adverse event is a technical problem that urgently needs to be solved in this field. Summary of the Invention
[0005] This application provides a method, apparatus, electronic device, and storage medium for determining the probability of an event, in order to solve the technical problem that the prior art has low accuracy in assessing the probability of adverse events.
[0006] In a first aspect, embodiments of this application provide a method for determining the probability of an event occurring, including: Baseline static feature data of the target object is acquired, as well as multi-channel time-series data of the target object, including ventilation parameter sequences, cumulative drug dose sequences, and vital sign parameter sequences. The ventilation parameter sequence, the cumulative drug dose sequence, and the vital signs parameter sequence are aligned along the time axis, and statistical features of each sequence are extracted at a preset time window granularity to obtain the process time sequence feature sequence. The baseline static feature data and the process time-series feature sequence are input into a preset sequence coding network to obtain the fused features output by the sequence coding network; The fusion features are input into multiple preset task branches to obtain the occurrence probability value of the corresponding category of predefined events output by each task branch; wherein each task branch corresponds to one category of the predefined events.
[0007] Secondly, embodiments of this application provide an apparatus for determining the probability of an event occurring, comprising: The acquisition module is used to acquire baseline static feature data of the target object and acquire multi-channel time-series data of the target object, wherein the multi-channel time-series data includes ventilation parameter sequences, drug cumulative dose sequences and vital sign parameter sequences; The alignment module is used to align the ventilation parameter sequence, the cumulative drug dose sequence, and the vital sign parameter sequence along the time axis, and extract the statistical features of each sequence at a preset time window granularity to obtain the process time sequence feature sequence. The first processing module is used to input the baseline static feature data and the process time-series feature sequence into a preset sequence coding network to obtain the fusion features output by the sequence coding network; The second processing module is used to input the fusion features into a plurality of preset task branches respectively, and obtain the occurrence probability value of the corresponding category of predefined events output by each task branch; wherein, each task branch corresponds to a category of the predefined events.
[0008] Thirdly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for determining the probability of occurrence of the aforementioned events.
[0009] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method for determining the probability of occurrence of the aforementioned events.
[0010] Based on this, one technical solution provided by this application includes: acquiring baseline static feature data of a target object, and acquiring multi-channel time-series data of the target object, wherein the multi-channel time-series data includes ventilation parameter sequences, drug cumulative dose sequences, and vital sign parameter sequences; aligning the ventilation parameter sequences, drug cumulative dose sequences, and vital sign parameter sequences along the time axis, and extracting statistical features of each sequence at a preset time window granularity to obtain a process time-series feature sequence; inputting the baseline static feature data and the process time-series feature sequence into a preset sequence coding network to obtain fusion features output by the sequence coding network; inputting the fusion features into multiple preset task branches to obtain the occurrence probability values of predefined events of corresponding categories output by each task branch; wherein each task branch corresponds to one type of predefined event. This scheme acquires baseline static feature data and multiple time-series data (including ventilation parameter sequences, cumulative drug dose sequences, and vital sign parameter sequences) of the target object. After time axis alignment and statistical feature extraction of the above-mentioned multiple time-series data, it is input into a sequence coding network along with the baseline static feature data for fusion encoding. Then, through multiple task branches, it outputs the occurrence probability values of multiple predefined events. This scheme can make full use of the multidimensional dynamic time-series data generated by the target object during dynamic support, realize the parallel evaluation of the occurrence probability of multiple predefined events, and effectively improve the accuracy of event occurrence probability evaluation. Attached Figure Description
[0011] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0012] Figure 1 This is a schematic diagram of a system for determining the probability of an event occurring in one embodiment of this application; Figure 2 This is a flowchart illustrating a method for determining the probability of an event occurring in one embodiment of this application; Figure 3 This is another flowchart illustrating a method for determining the probability of an event occurring in one embodiment of this application; Figure 4 This is a schematic diagram of an event occurrence probability determination device according to an embodiment of this application; Figure 5 This is a schematic diagram of an electronic device according to an embodiment of this application. Detailed Implementation
[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0014] This application provides a method for determining the probability of an event occurring, which can be applied to, for example... Figure 1 In a system for determining the probability of an event occurring, as shown. Figure 1 As shown, the system for determining the probability of an event includes a data source layer, a feature engineering layer, a prediction model layer, a counterfactual reasoning engine layer, and an output layer. These layers work together to implement the steps for determining the probability of an event. Please refer to the following for further details. Figures 2 to 3 The steps of the method for determining the probability of an event to be observed provided in the embodiments of this application are described in detail.
[0015] Firstly, such as Figure 2 and Figure 3 As shown, this application provides a method for determining the probability of an event occurring, which is then applied to... Figure 1 Taking the determination of the probability of an event occurring in a system as an example, the following steps are included: S101: Acquire baseline static characteristic data of the target object, and acquire multi-channel time-series data of the target object, including ventilation parameter sequences, cumulative drug dose sequences, and vital sign parameter sequences.
[0016] As an example, the target group refers to an individual receiving ventilation support and medication management, such as a patient undergoing surgery. Baseline static characteristic data refers to individual characteristics of the target group that are determined before the start of ventilation support and medication management (such as surgery) and remain substantially unchanged over time during the process. These characteristics include, but are not limited to, age, sex, weight, height, body surface area, past medical history, physiological status score, type of management, and expected duration of management. These baseline static characteristic data are inherent attributes of the target group and do not change during management; they can be directly obtained through hospital information systems or anesthesia information systems.
[0017] The first time interval refers to the time range from the start of the ventilation support and medication administration process to the current time. For example, if the current time is 30 minutes after the start of the administration process, then the first time interval is the time range from the start of administration (minute 0) to the current time (minute 30).
[0018] Multi-channel time-series data refers to serialized data from multiple different data channels recorded in chronological order within the first time interval. Multi-channel time-series data includes, but is not limited to, ventilation parameter sequences, cumulative drug dose sequences, and vital sign parameter sequences.
[0019] A ventilation parameter sequence refers to a sequence of operating parameter values from a ventilation device, recorded sequentially according to the acquisition time within the first time interval. The ventilation parameter sequence includes time-series data of at least one of the following parameters within the first time interval: tidal volume, peak airway pressure, plateau airway pressure, positive end-expiratory pressure, respiratory rate, inspiratory-to-expiratory ratio, concentration of inhaled anesthetic gas, concentration of exhaled anesthetic gas, minimum effective alveolar concentration, and fresh gas flow rate.
[0020] A drug cumulative dose sequence refers to the curve data showing the cumulative dosage of various drugs administered to a target subject within a first time interval as a function of the data collection time. Each data point in the drug cumulative dose sequence corresponds to the total cumulative dosage of various drugs administered from the start of the administration process to that data collection time. These drugs include, but are not limited to, induction drugs, maintenance drugs, analgesics, muscle relaxants, and vasoactive drugs. The drug cumulative dose sequence has a monotonically non-decreasing characteristic, meaning that the cumulative dosage either increases or remains constant over time and does not decrease. Taking remifentanil as an example, the data points in the drug cumulative dose sequence are as follows: cumulative dose of 0 μg at minute 0, 50 μg at minute 5, 110 μg at minute 10, and 180 μg at minute 15, with the sequence increasing over time.
[0021] A vital signs parameter sequence refers to a numerical sequence of physiological monitoring parameters of a target subject recorded in chronological order within a first time interval. The vital signs parameter sequence includes time-series data of at least one of the following parameters within the first time interval: systolic blood pressure, diastolic blood pressure, mean arterial pressure, heart rate, blood oxygen saturation, end-tidal carbon dioxide partial pressure, bispectral index (BTI), and body temperature.
[0022] In one possible implementation, the data source layer can acquire ventilation parameter sequences through the standard data interface of ventilation equipment (such as a ventilator or anesthesia machine); acquire cumulative drug dose sequences through a drug infusion recording interface (such as the information interface of a drug infusion pump); and acquire vital sign parameter sequences through the data interface of monitoring equipment (such as a multi-parameter monitor). Optionally, the standard data interface, the drug infusion recording interface, and the data interface of the monitoring equipment all conform to the HL7 or FHIR communication protocol to achieve standardized transmission and sharing of data between different medical devices.
[0023] In step S101, the data source layer provides a complete input data foundation for subsequent steps by acquiring baseline static feature data and multiple time-series data within the first time interval. The baseline static feature data provides the inherent attribute information of the target object, while the multiple time-series data provides the dynamic change information of the target object during the management process. The two complement each other, enabling subsequent processing to utilize both the inherent attributes and real-time dynamic change information of the target object simultaneously.
[0024] S102: Align the ventilation parameter sequence, drug cumulative dose sequence, and vital sign parameter sequence with time axis, and extract the statistical features of each sequence at a preset time window granularity to obtain the process time series feature sequence.
[0025] As an example, timeline alignment refers to mapping each data point in the ventilation parameter sequence, drug cumulative dose sequence, and vital sign parameter sequence to a unified timeline according to their respective recording timestamps. This ensures that three data streams from different data sources, with different sampling frequencies and data recording times, correspond one-to-one under the same time reference. Since the ventilation parameter sequence originates from the ventilation equipment, the drug cumulative dose sequence from the drug infusion recording equipment, and the vital sign parameter sequence from the monitoring equipment, their respective data sampling start times and sampling frequencies may differ (e.g., the ventilation equipment records data every 2 seconds, the monitoring equipment every 1 second, and the drug infusion equipment every 10 seconds). Timeline alignment eliminates the time misalignment between the three data streams, allowing ventilation parameters, drug cumulative doses, and vital sign data at the same moment to correspond to each other.
[0026] A preset time window refers to a pre-defined time length for extracting segments of time-series data. For example, the preset time window length can be set to 10 minutes. In another example, the preset time window length can be any value within the range of 5 to 15 minutes, such as 5 minutes, 8 minutes, 12 minutes, or 15 minutes. Extracting statistical features at the preset time window granularity means: on a unified time axis, using the preset time window as the basic unit, sequentially extracting time segments within each preset time window from the aligned three-channel time-series data, and extracting statistical quantities from the data within each preset time window that reflect the distribution characteristics and variation patterns of the data within that preset time window.
[0027] A process time-series feature sequence refers to a data sequence formed by arranging the statistical features extracted from each preset time window in chronological order. Each element in the process time-series feature sequence corresponds to the statistical feature expression of data within a preset time window, and the length of the process time-series feature sequence is equal to the number of time windows extracted. Taking a preset time window length of 10 minutes and a sliding step of 5 minutes as an example, when the evaluation is triggered 30 minutes after the start of the surgical procedure, a unified time axis covers the time range from the start of the surgical procedure to the 30th minute. With a window length of 10 minutes and a sliding step of 5 minutes, six preset time windows can be extracted, corresponding to 0-10 minutes, 5-15 minutes, 10-20 minutes, 15-25 minutes, 20-30 minutes, and 25-35 minutes respectively (the part of the last window that exceeds the current time is filled with valid data or truncated). Statistical features are extracted from each preset time window to obtain six process time-series features, which are then arranged in chronological order to form the process time-series feature sequence.
[0028] In step S102, the feature engineering layer performs time axis alignment and window statistical feature extraction on the three time series data, transforming the original high-frequency, heterogeneous, and asynchronous time series data into a synchronous, fixed-dimensional feature sequence containing statistical information, providing input data with a unified format and condensed information for the processing of the sequence coding network.
[0029] S103: Input the baseline static feature data and the process time series feature sequence into the preset sequence coding network to obtain the fused features output by the sequence coding network.
[0030] As an example, a sequence coding network is a pre-trained neural network model located in the prediction model layer, used to encode serialized input data to extract high-level features. A sequence coding network can receive baseline static feature data and process temporal feature sequences as input, and through its internal nonlinear transformation mechanism, encodes input data of different modalities (including static and temporal features) into fixed-dimensional feature vectors. Baseline static feature data reflects the inherent attribute information of the target object, while process temporal feature sequences reflect the dynamic changes of the target object during management. These two types of data differ significantly in data type and feature distribution. The sequence coding network can map these two different modalities of data into a unified feature space for joint representation.
[0031] Fusion features refer to the comprehensive feature representation output by a sequence coding network after jointly encoding baseline static feature data and process temporal feature sequences. Fusion features integrate the static individual information of the target object with the dynamic process information in the management process. It is a compact vectorized representation of the overall state of the target object from the start of the management process to the current moment. This fusion feature includes cumulative information and temporal dependencies in the management process.
[0032] In step S103, the prediction model layer fuses and encodes the baseline static feature data and the process time-series feature sequence through a sequence coding network, realizing the joint expression of static and dynamic information at the feature level, enabling the task branch to evaluate the probability of occurrence of various predefined events based on unified information.
[0033] S104: Input the fused features into multiple preset task branches respectively, and obtain the occurrence probability value of the corresponding category of predefined event output by each task branch; wherein, each task branch corresponds to a category of predefined event.
[0034] As an example, a task branch refers to a pre-defined network structure located within the prediction model layer. Each task branch corresponds to a predefined class of events and is used to independently map the fused features to output the probability value of the predefined event for that class. Multiple task branches are independent of each other, each with its own network parameters, and do not affect each other. By setting multiple task branches, multiple classes of predefined events can be processed simultaneously.
[0035] Predefined events refer to one or more pre-determined categories of events that require risk assessment after the management process has concluded. Each category of predefined events corresponds to an independent assessment task. Predefined events include, but are not limited to, at least one of the following: nausea and vomiting, delayed awakening, chills, acute pain, respiratory depression, delirium, pulmonary complications, and hemodynamic events. Different categories of predefined events have different clinical characteristics and assessment criteria. By setting independent task branches for each category of predefined events, independent assessments can be conducted based on the characteristics of each category.
[0036] The probability value is a numerical value between 0 and 1, output after mapping the fused features through task branches. It represents the likelihood of a predefined event of the corresponding category occurring after the management process ends. The closer the value is to 1, the higher the probability of the predefined event occurring; the closer the value is to 0, the lower the probability of the predefined event occurring.
[0037] In step S104, the prediction model layer maps the same fused feature through multiple independent task branches, enabling it to output the probability values of multiple predefined events at once, thus achieving multi-task parallel evaluation. Each task branch processes the same fused feature independently, thereby enabling evaluation for each predefined event, improving evaluation efficiency and the accuracy of event probability evaluation.
[0038] In summary, the method for determining the probability of the aforementioned events can be implemented through an event probability determination system. Specifically: the data source layer acquires baseline static feature data of the target object and multiple time-series data of the target object, including ventilation parameter sequences, drug cumulative dose sequences, and vital sign parameter sequences; the feature engineering layer aligns the ventilation parameter sequences, drug cumulative dose sequences, and vital sign parameter sequences along the time axis and extracts statistical features of each sequence at a preset time window granularity to obtain the process time-series feature sequence; the prediction model layer inputs the baseline static feature data and the process time-series feature sequence into a preset sequence coding network to obtain the fusion features output by the sequence coding network; the fusion features are then input into multiple preset task branches to obtain the probability values of predefined events of the corresponding categories output by each task branch; wherein each task branch corresponds to a predefined event category. This scheme acquires baseline static feature data and multiple time-series data (including ventilation parameter sequences, cumulative drug dose sequences, and vital sign parameter sequences) of the target object through the data source layer. After time axis alignment and statistical feature extraction of the above-mentioned multiple time-series data through the feature engineering layer, the prediction model layer inputs the extracted process time-series feature sequences and baseline static feature data into the sequence coding network for fusion encoding. Then, through multiple task branches of the prediction model layer, it outputs the occurrence probability values of multiple predefined events. This scheme can make full use of the multi-dimensional dynamic time-series data generated by the target object during the management process, realize the parallel evaluation of the occurrence probability of multiple predefined events, and effectively improve the accuracy of event occurrence probability evaluation.
[0039] It should be noted that the method for determining the probability of an event occurrence also includes repeatedly executing steps S101 to S104 at preset time intervals. Specifically, steps S101 to S104 can be repeatedly executed at preset time intervals to obtain the probability values of predefined events of various categories at each preset time interval, forming a probability dynamic trajectory of the target object within the first time interval. Here, the preset time interval refers to the pre-set time length between two adjacent assessment trigger times. For example, if the preset time interval is 5 minutes, an assessment is triggered at the 5th, 10th, 15th, 20th, etc., after the management process begins. In another example, the preset time interval can also be 10 minutes. The probability dynamic trajectory refers to a sequence of probability values obtained from multiple rolling assessments arranged in chronological order. The probability dynamic trajectory can reflect the trend of the probability values of various predefined events changing as the management process progresses.
[0040] As can be seen, by repeatedly performing assessments at preset time intervals and forming a dynamic probability trajectory, operators can intuitively observe the trends in the occurrence probability values of various predefined events as the management process progresses and medication accumulates, thus providing a dynamic reference for decision-making adjustments during the management process. Taking nausea and vomiting as an example, the probability value of occurrence is 0.3 at the 5th minute, 0.45 at the 10th minute, and 0.55 at the 15th minute. These values, in chronological order, constitute the dynamic probability trajectory of nausea and vomiting.
[0041] In one embodiment, the ventilation parameter sequence, cumulative drug dose sequence, and vital sign parameter sequence are time-axis aligned, and statistical features of each sequence are extracted at a preset time window granularity to obtain a process time-series feature sequence, including the following steps: S1021: Time-axis alignment of ventilation parameter sequences, cumulative drug dose sequences, and vital sign parameter sequences is performed to obtain aligned three-channel time-series data.
[0042] As an example, the feature engineering layer obtains the first set of timestamps for each data point in the ventilation parameter sequence, the second set of timestamps for each data point in the drug cumulative dose sequence, and the third set of timestamps for each data point in the vital signs parameter sequence; it establishes a unified timeline that covers the union of the first, second, and third timestamp sets and is arranged in chronological order; for each time point on the unified timeline, if the original data of a certain sequence exists at that time point, the original data is used directly; if the original data does not exist, the data value of that time point is estimated from the data of the adjacent time points of the sequence through linear interpolation, thereby obtaining the aligned three-way time series data.
[0043] S1022: Using a preset time window as the granularity, extract the statistical features within each preset time window from the aligned three-channel time series data to obtain the process time series features under each preset time window.
[0044] As an example, the feature engineering layer slides along a unified time axis with a preset time window size and a preset sliding step size to determine the start and end times of each preset time window. For each preset time window, the feature engineering layer extracts the corresponding ventilation parameter data segment, drug cumulative dose data segment, and vital sign parameter data segment from the aligned three-channel time series data.
[0045] For example, with a preset time window length of 10 minutes and a sliding step size of 5 minutes, when the feature engineering layer triggers the evaluation at the 30th minute after the start of the management process, it extracts the ventilation parameter data segment, drug cumulative dose data segment, and vital sign parameter data segment corresponding to 25-35 minutes from the aligned three-channel time series data.
[0046] The feature engineering layer extracts statistics from the ventilation parameter data segments captured within each preset time window, serving as the ventilation window features corresponding to each preset time window. These statistics include, but are not limited to: mean, median, standard deviation, maximum, minimum, trend slope, and fluctuation amplitude within the window.
[0047] Taking tidal volume as an example, within a preset time window, the ventilation parameter data segment contains multiple tidal volume data points. The feature engineering layer calculates the in-window mean, in-window standard deviation, and in-window maximum of the tidal volume within the preset time window, and uses the calculated statistics as components of the ventilation window characteristics corresponding to the preset time window. For example, within a preset time window of 0-10 minutes, the tidal volume data points are 500mL, 510mL, 495mL, 505mL, and 490mL, respectively. The feature engineering layer calculates the in-window mean as 500mL, the in-window standard deviation as 8.37mL, and the in-window maximum as 510mL.
[0048] The feature engineering layer extracts dose features from the cumulative drug dose data segments captured within each preset time window, serving as the drug window features corresponding to each preset time window. The dose features of the drug window features include, but are not limited to, the dosing rate per unit time within the preset time window, the area under the cumulative dose curve within the preset time window, and the normalized ratio of the cumulative dose at the end of the preset time window to the target subject's weight or body surface area.
[0049] Taking remifentanil as an example, within a preset time window, the cumulative drug dose data segment includes the cumulative dose value at the start and end times. The feature engineering layer calculates the dosing rate per unit time and the area under the cumulative dose curve within this preset time window, and uses the calculated dose characteristics as a component of the drug window characteristics corresponding to this preset time window. For example, within a preset time window of 10-20 minutes, the initial cumulative dose is 110 μg, the final cumulative dose is 180 μg, the preset time window length is 10 minutes, and the feature engineering layer calculates the dosing rate per unit time to be 7 μg / min.
[0050] The feature engineering layer extracts statistics from the vital sign parameter data segments captured within each preset time window, serving as the vital sign window features corresponding to each preset time window. These statistics include, but are not limited to, the mean, median, standard deviation, maximum, minimum, trend slope, fluctuation amplitude, and the percentage of time exceeding a preset threshold within the window.
[0051] Taking blood oxygen saturation as an example, with a preset threshold of 95%, within a preset time window, a vital sign parameter data segment contains multiple blood oxygen saturation data points. The feature engineering layer counts the number of data points with blood oxygen saturation below 95%, multiplies this number by the sampling interval to obtain the duration of low blood oxygen. This duration of low blood oxygen is then divided by the preset time window length to obtain the percentage of low blood oxygen duration. The calculated mean within the window and the percentage of low blood oxygen duration are used as components of the vital sign window features corresponding to that preset time window. For example, within a preset time window of 0-10 minutes, the total duration corresponding to data points with blood oxygen saturation below 95% is 2 minutes, the preset time window length is 10 minutes, and the feature engineering layer calculates the percentage of low blood oxygen duration to be 20%.
[0052] The feature engineering layer fuses the ventilation window features, drug window features, and vital sign window features corresponding to the same preset time window to obtain the process time series features under each preset time window. Fusion means concatenating the components of the ventilation window features, drug window features, and vital sign window features to form a single vector containing all feature components, which serves as the process time series feature under that preset time window.
[0053] Since the ventilation window feature is derived from ventilation parameter data segments, the drug window feature is derived from drug cumulative dose data segments, and the vital signs window feature is derived from vital sign parameter data segments, and all three data segments are extracted from three aligned time-series data within the same preset time window, the process time-series feature obtained by splicing the three can simultaneously reflect the ventilation status, drug cumulative status, and vital sign status of the target object within the preset time window. The three are interconnected and together constitute a complete state description under the preset time window.
[0054] Taking a preset time window as an example, the ventilation window features extracted by the feature engineering layer include four components (intra-window mean of tidal volume, intra-window mean of respiratory rate, intra-window standard deviation of tidal volume, and intra-window standard deviation of respiratory rate); the drug window features include two components (dose administration rate per unit time and area under the cumulative dose curve); and the vital signs window features include two components (intra-window mean of heart rate and intra-window mean of mean arterial pressure). The feature engineering layer concatenates these eight components into a single vector containing eight components, which serves as the process temporal feature within that preset time window.
[0055] In this embodiment, the feature engineering layer extracts features from the three time-series data separately and then fuses them, so that the process time-series features can comprehensively represent the three dimensions of information of the target object within the preset time window: ventilation status, drug accumulation status, and vital signs status, effectively avoiding the loss of information in a single dimension.
[0056] S1023: Generate a process timing feature sequence based on the process timing features under each preset time window.
[0057] As an example, the feature engineering layer arranges the process timing features of the first preset time window as the first element of the process timing feature sequence, arranges the process timing features of the second preset time window as the second element of the process timing feature sequence, and so on, until the process timing features of all preset time windows are arranged in chronological order to obtain the process timing feature sequence.
[0058] Taking a preset time window length of 10 minutes and a sliding step of 5 minutes as an example, when the evaluation is triggered 30 minutes after the start of the management process, the feature engineering layer has obtained the process timing features corresponding to the 6 preset time windows in step S1022, which correspond to the 6 preset time windows of 0-10 minutes, 5-15 minutes, 10-20 minutes, 15-25 minutes, 20-30 minutes and 25-35 minutes respectively. The feature engineering layer arranges these 6 process time-series features in chronological order (i.e., from the earliest time window to the latest time window). The process time-series features corresponding to 0-10 minutes are arranged as the first element of the sequence, the process time-series features corresponding to 5-15 minutes are arranged as the second element, the process time-series features corresponding to 10-20 minutes are arranged as the third element, the process time-series features corresponding to 15-25 minutes are arranged as the fourth element, the process time-series features corresponding to 20-30 minutes are arranged as the fifth element, and the process time-series features corresponding to 25-35 minutes are arranged as the sixth element, thus forming a process time-series feature sequence containing 6 process time-series features.
[0059] In the above embodiments, the feature engineering layer solves the problem of temporal asynchrony of multi-source heterogeneous data from different devices by performing time axis alignment and window feature extraction on the three time series data. Time axis alignment ensures that the three data have the same set of time points on the same time base; windowing converts long-term series data into fixed-length feature sequences, providing a data format that the sequence coding network can effectively process, and ensuring the real-time performance and computational efficiency of feature extraction.
[0060] In one embodiment, the process involves inputting baseline static feature data and process time-series feature sequences into a preset sequence coding network to obtain the fused features output by the sequence coding network, including the following steps: S1031: Map the baseline static feature data to a static feature vector.
[0061] As an example, the prediction model layer performs numerical encoding and normalization on each feature item in the baseline static feature data. For discrete features, the prediction model layer performs one-hot encoding to convert them into numerical form; for continuous features, the prediction model layer performs normalization to give them uniform dimensions. The prediction model layer combines all the encoded and normalized feature values into a fixed-dimensional vector, which serves as the static feature vector.
[0062] For example, the baseline static feature data includes three features: gender, age, and weight. Gender is a discrete feature, with "male" encoded as 0 and "female" as 1; age is a continuous feature, assuming an age range of 0-100 years, with an actual age of 65 years, and a normalized value of 0.65; weight is a continuous feature, assuming a weight range of 0-100 kg, with an actual weight of 70 kg, and a normalized value of 0.70. The prediction model layer combines these three feature values to form a static feature vector, which contains three components: [0, 0.65, 0.70].
[0063] S1032: Concatenate the static feature vector with each process time-series feature in the process time-series feature sequence to obtain multiple fused input vectors.
[0064] As an example, the prediction model layer acquires each process time-series feature in the process time-series feature sequence, and each process time-series feature is distinguished according to its corresponding preset time window. For each process time-series feature in the process time-series feature sequence, the prediction model layer sequentially concatenates each component of the process time-series feature with each component of the static feature vector to form a fused input vector. The prediction model layer performs the above concatenation operation on each process time-series feature in the process time-series feature sequence to obtain multiple fused input vectors, each fused input vector corresponding to a preset time window.
[0065] For example, the process time-series feature sequence contains process time-series features corresponding to 6 preset time windows, each process time-series feature having 20 components. The static feature vector has 3 components. The prediction model layer sequentially concatenates the 20 components of the process time-series features of the first preset time window with the 3 components of the static feature vector to form a fused input vector containing 23 components; the prediction model layer sequentially concatenates the 20 components of the process time-series features of the second preset time window with the 3 components of the static feature vector to form another fused input vector containing 23 components; the prediction model layer sequentially performs the concatenation operation on each of the 6 preset time windows, resulting in a total of 6 fused input vectors, each corresponding to one of the 6 preset time windows.
[0066] S1033: Input each fused input vector into the sequence coding network in chronological order so that the sequence coding network can perform state transfer and update of the fused input vector at each time step, and output the hidden state vector at the last time step as the fused feature.
[0067] As an example, the prediction model layer uses the fusion input vector of the first preset time window as the input of the sequence coding network at the first time step, the fusion input vector of the second preset time window as the input of the sequence coding network at the second time step, and so on, until all the fusion input vectors are input into the sequence coding network in sequence.
[0068] Sequence coding networks include gated recurrent unit networks or Transformer networks.
[0069] When the sequence coding network is a gated recurrent unit (GRU) network, the GRU network, through its internal update and reset gate mechanisms, receives the fused input vector at the current time step and the hidden state from the previous time step at each time step, and outputs the hidden state at the current time step. At the first time step, the GRU network receives the first fused input vector and the initial hidden state (which is a zero vector), and calculates the first hidden state through the GRU. At the second time step, the GRU network receives the second fused input vector and the first hidden state, and calculates the second hidden state through the GRU. The GRU network repeats this process until the last time step, outputting the hidden state vector at the last time step as the fused feature. Because the update and reset gates of the GRU network can control the forgetting and retention of information, the fused feature can effectively preserve long-distance temporal dependencies.
[0070] When the sequence encoding network is a Transformer network, the Transformer network models the relationships between the fused input vectors at each time step through a self-attention mechanism, outputs the encoded vectors at each time step, and takes the encoded vector at the last time step as the fused feature. The self-attention mechanism can capture the global dependencies between time steps, enabling the fused feature to integrate information from each time step.
[0071] Taking six fused input vectors as an example, the prediction model layer inputs the fused input vector from the first time step into the sequence coding network, which outputs the first hidden state. The prediction model layer then inputs the fused input vector from the second time step into the sequence coding network, which combines the first hidden state to output the second hidden state. This process continues until the prediction model layer inputs the fused input vector from the sixth time step into the sequence coding network, which combines the fifth hidden state to output the sixth hidden state, serving as the fused feature. This fused feature integrates static individual information (gender, age, weight) from the static feature vector with dynamic process information from the process time-series features across all six preset time windows, providing a compact vectorized representation of the overall state of the target object from the start of the management process to the current time step.
[0072] In this embodiment, the prediction model layer concatenates the static feature vector with each process time-series feature and inputs it into the sequence coding network. This allows the sequence coding network to simultaneously acquire static information and dynamic information of the current window at each time step, and gradually accumulate historical time-series information through a state transfer mechanism. The final output fusion feature integrates static information and full-process time-series information.
[0073] In one embodiment, the fused features are input into multiple preset task branches, and the occurrence probability value of a predefined event of the corresponding category output by each task branch is obtained, including: S1041: Input the fused features into each task branch respectively, so that each task branch performs linear transformation and sigmoid function activation on the fused features respectively, and outputs the occurrence probability value of the predefined event of the corresponding category respectively.
[0074] Each task branch is a single-layer fully connected network with an independent weight matrix and bias vector. The linear transformation refers to the operation of adding the bias vector after performing matrix multiplication on the fused features and the weight matrix. Through the linear transformation, the fused features can be mapped from the original feature space to the output space of the task branch.
[0075] As an example, the prediction model layer inputs the fused features into each task branch, and each task branch performs a linear transformation on the fused features, that is, multiplies them with the weight matrix of its respective task branch and adds a bias vector to obtain a linear output result. The linear output result is then activated by the sigmoid function to output the occurrence probability value of the predefined event of the corresponding category.
[0076] The mathematical expression for multi-task prediction is Equation (1): In equation (1), For the first Class predefined events at the current moment The probability value of occurrence, Indicates the current moment. For the category index of predefined events, For the sigmoid function, For the first The weight matrix of the task branches corresponding to the predefined events of the class. For the first The bias vector of the task branch corresponding to the predefined event of the class. To fuse features, the mathematical expression for the sigmoid function is: .
[0077] For example, taking predefined events including nausea and vomiting and delayed awakening as an example, the prediction model layer will fuse features. Input the task branch corresponding to nausea and vomiting and the task branch corresponding to wake-up delay, respectively. The task branch corresponding to nausea and vomiting is determined by its weight matrix. and bias vector right Perform a linear transformation to obtain After activation by the sigmoid function, the probability value of nausea and vomiting is obtained. The task branch corresponding to the wake-up delay is determined by its weight matrix. and bias vector right Perform a linear transformation to obtain After activation by the sigmoid function, the probability value of the occurrence of wake-up delay is obtained. Assume that the probability of nausea and vomiting is calculated. The probability value for delayed awakening is 0.65. If the value is 0.32, it indicates that the probability of nausea and vomiting occurring after the management process is completed is 65%, and the probability of delayed awakening occurring after the management process is 32%.
[0078] In this embodiment, the prediction model layer performs linear transformation and sigmoid function activation on the same fused feature through multiple task branches. Each task branch independently outputs the probability value of a predefined event, thereby outputting the probability values of multiple predefined events at once, achieving multi-task parallel evaluation. Each task branch independently processes the same fused feature, enabling separate evaluation for each predefined event, improving evaluation efficiency and the accuracy of event probability assessment.
[0079] In one embodiment, after inputting the fused features into multiple preset task branches and obtaining the occurrence probability values of predefined events of the corresponding categories output by each task branch, the method further includes the following steps: S201: When the probability value of at least one type of predefined event exceeds the first preset threshold, attribution analysis is performed on the contribution of each feature dimension of the input side of the sequence coding network to the current output to obtain the contribution index of each feature dimension; wherein, the feature dimension includes each component in the static feature vector formed after the baseline static feature data is mapped, and each process time series feature component in each time window of the process time series feature sequence. S202: Select feature dimensions whose contribution index exceeds the preset contribution threshold from each feature dimension and use them as candidate feature dimensions; S203: Based on the candidate feature dimensions, determine the feature dimensions that belong to the adjustable variables and use them as adjustable feature dimensions; S203: Generate risk mitigation measures information for the target object based on adjustable feature dimensions.
[0080] As an example, the first preset threshold refers to a pre-set probability threshold used to determine whether the probability value of the occurrence of a certain type of predefined event reaches the risk level that requires triggering attribution analysis. The value of the first preset threshold ranges from 0.3 to 0.6. For example, the first preset threshold can be set to 0.4 or 0.5, etc., and is not limited here.
[0081] The input feature dimensions fall into two categories: one is the components of the static feature vector formed after mapping the baseline static feature data, such as "age component," "gender component," and "weight component"; the other is the process time-series feature components within each time window of the process time-series feature sequence, such as "mean tidal volume component within the 3rd time window" and "dosing rate component per unit time in the 5th time window." Attribution analysis calculates the influence of each component in these two types of feature dimensions on the current prediction result (i.e., the probability value of occurrence), outputting a contribution index for each feature dimension. The contribution index is a numerical measure used to quantify the degree of contribution of each feature dimension. The larger the contribution, the more significant the influence of that feature dimension on the current prediction result.
[0082] The preset contribution threshold refers to a pre-defined critical value used to filter contribution indicators. After obtaining the contribution indicators for each feature dimension in step S201, feature dimensions whose contribution indicators exceed the preset contribution threshold are retained as candidate feature dimensions. Candidate feature dimensions are a set of feature dimensions with significant contributions, representing the feature factors that contribute the most to the current high-risk prediction result. The contribution indicator can be attention weight, permutation importance, or SHAP value, etc., and is not limited here. When the contribution indicator is attention weight, the preset contribution threshold can be set to the average of the attention weights of each feature dimension; when the contribution indicator is SHAP value, the preset contribution threshold can be set to the average of the absolute values of all feature SHAP values.
[0083] Attention weights refer to the weight values assigned to each input position by the self-attention mechanism in a Transformer network, representing the importance of the information at that input position in the encoding process. Permutation importance is a metric that assesses the importance of a feature dimension by randomly shuffling its values and observing the change in the model output. SHAP (Shapley Additive exPlanations) values are calculated based on Shapley values in cooperative game theory, determining the marginal contribution of each feature to the model output. SHAP values simulate all possible feature combinations, calculating the average marginal contribution of each feature to the prediction result, thus quantifying the influence of each feature dimension on the probability of occurrence.
[0084] Adjustable variables refer to parameters that can be actively adjusted by operators based on clinical judgment during the management process, excluding static characteristics that cannot be changed during management (such as age, gender, management type, etc.). Adjustable variables include, but are not limited to, at least one of the following: tidal volume, positive end-expiratory pressure, respiratory rate, inspiratory-to-expiratory ratio, infusion rate of opioids, infusion rate or target concentration of sedatives, supplemental dose of muscle relaxants, infusion rate of vasoactive drugs, fluid infusion rate, and target temperature management value. Based on the candidate feature dimensions, feature dimensions that belong to adjustable variables are further screened as adjustable feature dimensions. Adjustable feature dimensions represent feature factors that contribute significantly to the current predictive results and can be actively intervened by operators during the management process.
[0085] The specific steps for generating risk mitigation measures information for the target object based on the adjustable feature dimension are as follows: First, determine the current value of the adjustable variable in the process time-series feature sequence based on the adjustable variable corresponding to the adjustable feature dimension; then, obtain the preset adjustment rule library associated with the adjustable variable, which records the correspondence between different value ranges of the adjustable variable and mitigation measure texts; finally, match the current value with each value range in the preset adjustment rule library, and obtain the mitigation measure text corresponding to the successfully matched value range as the risk mitigation measure information.
[0086] For example, assuming the probability of nausea and vomiting is 0.68, and the first preset threshold is set to 0.5, the attribution analysis process is triggered because 0.68 is greater than 0.5. This involves calculating the various feature dimensions of the input side of the sequence coding network. These input feature dimensions include components such as "age," "gender," and "weight" from the baseline static feature data, as well as process time-series feature components within each time window of the process time-series, such as the "mean tidal volume within the third time window" and the "dosing rate per unit time within the fifth time window." By calculating the contribution of each of these components to the probability of nausea and vomiting (0.68), the contribution index corresponding to each feature dimension is output. The contribution index of the "mean tidal volume within the third time window" ranks highest among all feature dimensions, indicating that the mean tidal volume within the window contributes most significantly to the current high-risk prediction of nausea and vomiting.
[0087] The preset contribution threshold can be set as the average of the contribution indices of all feature dimensions. Attribution analysis yielded the following contribution indices for each feature dimension: 0.05, 0.03, 0.12, 0.08, 0.15, 0.07, 0.22, 0.09, 0.11, and 0.06, with an average of 0.098. Among these, the contribution index of "the mean component of tidal volume within the third time window" is 0.22, exceeding the preset contribution threshold of 0.098, and therefore it was selected as a candidate feature dimension. Other feature dimensions with contribution indices exceeding 0.098 were also selected as candidate feature dimensions. Among the candidate feature dimensions, the adjustable variable corresponding to the mean component of tidal volume within the third time window is "tidal volume," a ventilation parameter that operators can actively adjust during management; therefore, it was determined as an adjustable feature dimension.
[0088] Subsequently, based on the adjustable variable "tidal volume" corresponding to the adjustable feature dimension, the current value of tidal volume in the process time-series feature sequence is determined. Assume the current tidal volume is 5.2 mL / kg, and this low tidal volume state has persisted for more than 30 minutes. The preset adjustment rule base records: when the tidal volume is below 6 mL / kg for more than 30 minutes, the corresponding mitigation measure text is "It is recommended to adjust the tidal volume to 6-8 mL / kg"; when the tidal volume is within the 6-8 mL / kg range, the corresponding mitigation measure text is "The tidal volume is currently within a safe range; continue monitoring." The current value of 5.2 mL / kg is matched with the value ranges in the preset adjustment rule base. The current value falls within the range of "below 6 mL / kg for more than 30 minutes," therefore, the mitigation measure text corresponding to the successfully matched value range, "It is recommended to adjust the tidal volume to 6-8 mL / kg," is obtained as the risk mitigation measure information.
[0089] In summary, the counterfactual reasoning engine layer triggers attribution analysis when the probability of at least one predefined event exceeds a first preset threshold. This quantitatively evaluates the contribution of each feature dimension from the sequence coding network input to the current output, accurately identifying the feature dimensions that contribute most significantly to the current high-risk prediction result. This effectively avoids indiscriminate feature processing and improves the targeting of subsequent analysis. By filtering through a preset contribution threshold, feature dimensions with contribution indicators exceeding the threshold are retained as candidate feature dimensions, while feature dimensions with lower contributions are filtered out. This effectively reduces the computational complexity of subsequent processing while ensuring the statistical significance of the contributions of candidate feature dimensions. By further filtering the candidate feature dimensions to identify adjustable variables, the scope of the screening is narrowed from all feature dimensions to specific parameters that operators can actively intervene in during the management process. Static features that cannot be changed are excluded, ensuring that the final risk mitigation measures information is actionable. Finally, by matching the current value of the adjustable variable corresponding to the adjustable feature dimension with the value range in the preset adjustment rule base, the mitigation measure text corresponding to the successfully matched value range is obtained as the risk mitigation measure information. This directly links the prediction results with actionable suggestions, realizing the leap from "risk warning" to "intervention suggestion" and providing clear intervention guidance for operators.
[0090] In one embodiment, such as Figure 3 As shown, after inputting the fused features into multiple preset task branches and obtaining the occurrence probability values of predefined events of the corresponding categories output by each task branch, the process also includes the following steps: S301: When the probability value of at least one type of predefined event exceeds the second preset threshold, identify the adjustable variable corresponding to the current moment and the current value of each adjustable variable; S302: For each adjustable variable, generate multiple candidate adjustment values within the limited range of values corresponding to the adjustable variable, and obtain a set of candidate adjustment values corresponding to each adjustable variable; S303: Select one candidate adjustment value from the candidate adjustment value set corresponding to each adjustable variable, combine the selected candidate adjustment values of each adjustable variable into a group of candidate adjustment schemes, and iterate through all combination methods to generate multiple groups of candidate adjustment schemes. S304: For each set of candidate adjustment schemes, starting from the current time, generate simulation process time series data within a preset future time period according to the candidate adjustment scheme; S305: Input the baseline static feature data, the multi-channel time series data collected before the current time and the simulation process time series data into the sequence coding network and multiple task branches in sequence to obtain the inference probability values of various predefined events under each group of candidate adjustment schemes; S306: Determine the recommended adjustment scheme based on the inferred probability values corresponding to each group of candidate adjustment schemes.
[0091] As an example, the second preset threshold refers to a pre-set critical value for the probability of occurrence used to trigger the counterfactual analysis process. The second preset threshold is greater than the first preset threshold. The value of the second preset threshold ranges from 0.5 to 0.8. For example, when the first preset threshold is 0.5, the second preset threshold can be set to 0.6. That is, the counterfactual analysis process is only initiated when the probability of occurrence of a certain type of predefined event exceeds a higher risk level (the second preset threshold). If the probability of occurrence only exceeds the first preset threshold but not the second preset threshold, only the attribution analysis and risk mitigation measure generation process (i.e., S201 to S204) is executed, and the counterfactual analysis is not performed.
[0092] In step S301, the method for identifying the adjustable variables corresponding to the current time and the current value of each adjustable variable can be as follows: obtain a preset adjustable variable feature library, which records the feature identifiers corresponding to each adjustable variable; match the feature identifiers of each feature dimension of the input side of the sequence coding network with each feature identifier in the adjustable variable feature library; identify the parameters corresponding to the successfully matched feature dimensions as adjustable variables, and extract the current value of each adjustable variable from the multi-channel time series data at the current time.
[0093] Specifically, the adjustable variable feature library pre-stores the feature identifiers for "tidal volume" as "intra-window mean of tidal volume", "remifentanil infusion rate" as "remifentanil infusion rate", and "positive end-expiratory pressure" as "intra-window mean of positive end-expiratory pressure". The input side of the sequence coding network contains multiple feature dimensions, including "intra-window mean of tidal volume", "remifentanil infusion rate", "age component", and "gender component". The feature identifiers of each feature dimension on the input side are compared one by one with the feature identifiers in the adjustable variable feature library. If the "intra-window mean of tidal volume" matches the feature identifier corresponding to "tidal volume" in the adjustable variable feature library, then "tidal volume" is identified as an adjustable variable. If the "remifentanil infusion rate" matches the feature identifier corresponding to "remifentanil infusion rate" in the adjustable variable feature library, then "remifentanil infusion rate" is identified as an adjustable variable. The "age component" and "gender component" have no matching items in the adjustable variable feature library and are not identified as adjustable variables. After a successful match, the current value of each identified adjustable variable is extracted from the multi-channel time series data at the current time. For example, the current value of tidal volume is extracted as 5.2 mL / kg from the ventilation parameter sequence at the current time, and the current value of remifentanil infusion rate is extracted as 0.3 μg / kg / min from the drug infusion record at the current time.
[0094] In step S302, the defined value range refers to the pre-set safe adjustment range for each adjustable variable, determined by the drug's instructions, clinical guidelines, or clinical expert knowledge. For example, the defined value range for remifentanil infusion rate is 0.05-0.5 μg / kg / min; the defined value range for tidal volume is 4-12 mL / kg. For each adjustable variable, based on the current value of the adjustable variable, at least one downward adjustment candidate value and at least one upward adjustment candidate value are generated within the defined value range corresponding to that adjustable variable, resulting in a set of candidate adjustment values for each adjustable variable. A downward adjustment candidate value is an adjustment value less than the current value, and an upward adjustment candidate value is an adjustment value greater than the current value.
[0095] Specifically, taking the remifentanil infusion rate as an example, assuming the current value is 0.3 μg / kg / min and the range is limited to 0.05-0.5 μg / kg / min, then the following candidate values can be generated: 0.15 μg / kg / min and 0.20 μg / kg / min for downward adjustment, and 0.40 μg / kg / min and 0.45 μg / kg / min for upward adjustment. A total of 4 candidate adjustment values are obtained, which constitute the set of candidate adjustment values corresponding to the remifentanil infusion rate.
[0096] In step S303, a candidate adjustment scheme refers to a complete adjustment plan formed by combining the candidate adjustment values of each adjustable variable. Specifically, a candidate adjustment value is selected from the set of candidate adjustment values corresponding to each adjustable variable, and the selected candidate adjustment values of each adjustable variable are combined into a candidate adjustment scheme. All combinations are iterated to generate multiple candidate adjustment schemes. When the number of adjustable variables is N and the number of candidate adjustment values in the set of candidate adjustment values corresponding to each adjustable variable is M, M to the power of N candidate adjustment schemes are generated, where N and M are both positive integers. For example, when there are two adjustable variables (tidal volume and remifentanil infusion rate), and each adjustable variable has three candidate adjustment values, M to the power of N candidate adjustment schemes are generated. =9 candidate adjustment schemes.
[0097] When M raised to the power of N exceeds a preset combination threshold, a strategy needs to be adopted to reduce the number of candidate adjustment schemes to avoid combination explosion. The preset combination threshold refers to a pre-defined maximum number of candidate adjustment schemes used to limit combination explosion. Specifically, at least one of the following strategies can be employed: The variable-wise search strategy involves first fixing other variables, searching for the optimal value of a single variable, and then iteratively optimizing each variable sequentially. For example, suppose there are three adjustable variables (remifentanil infusion rate, tidal volume, and positive end-expiratory pressure), each with four candidate adjustment values, and a preset combination threshold of 32. Theoretically, the number of candidate adjustment schemes that can be generated is... =64 groups, exceeding the preset combination threshold of 32. If a variable-wise search strategy is adopted, firstly, tidal volume and positive end-expiratory pressure are fixed at their current values, and only the remifentanil infusion rate is adjusted to determine the optimal value (that is, the candidate value with the lowest inferred probability value) among the four candidate values; then, the remifentanil infusion rate is fixed at this optimal value, and the positive end-expiratory pressure is fixed at its current value, and the tidal volume is adjusted to determine the optimal value among the four candidate values; finally, the remifentanil infusion rate and tidal volume are fixed at the aforementioned optimal values, and the positive end-expiratory pressure is adjusted to determine the optimal value among the four candidate values. Through variable-wise search, three candidate adjustment schemes are finally obtained (one optimal value is retained for each variable), instead of 64 groups.
[0098] Greedy search strategy: In each round, the variable with the largest decrease in risk is selected for adjustment, and the adjustment schemes are gradually stacked. Taking the above three adjustable variables as an example, firstly, the inferred probability value corresponding to each scheme is calculated from all 64 candidate adjustment schemes. The variable with the largest decrease in inferred probability value is selected for single-step adjustment, and the adjustment value of this variable is fixed. Then, based on this, the inferred probability values corresponding to each candidate value of the remaining variables are calculated again, and the next variable with the largest decrease in inferred probability value is selected for adjustment. The above process is repeated until all variables are adjusted, and finally, one recommended adjustment scheme is obtained.
[0099] The pruning strategy based on contraindications rules eliminates unreasonable combinations based on prior clinical knowledge. For example, a combination of significantly decreasing the infusion rate of remifentanil and simultaneously decreasing the infusion rate of sedative drugs may violate the minimum clinical requirements for analgesia and sedation, and is therefore excluded; a combination of significantly increasing tidal volume and simultaneously significantly increasing positive end-expiratory pressure may lead to excessive airway pressure, and is also excluded. Through these pruning rules, combinations that do not conform to prior clinical knowledge are eliminated from 64 candidate protocols, retaining reasonable combinations for subsequent extrapolation and calculation.
[0100] The three strategies described above can be used individually or in combination. For example, a tabu-based pruning strategy can be used first to eliminate obviously unreasonable combinations, and then a variable-wise search strategy or a greedy search strategy can be used to further reduce the number of candidate solutions for the remaining combinations, thereby effectively controlling computational complexity while ensuring search coverage.
[0101] In step S304, the preset future duration refers to the length of the future time window pre-set in the counterfactual simulation. For example, the preset future duration can be set to 30 minutes. The simulation process time series data refers to the multiple time series data generated from the current moment to the preset future duration under the candidate adjustment scheme. The specific method for generating the simulation process time series data within the preset future duration according to the candidate adjustment scheme, starting from the current moment, can be as follows: For drug-related adjustable variables, based on the end data of the current drug cumulative dose sequence, the dosage for a preset future duration is accumulated according to the candidate adjustment values of the drug-related adjustable variable in the candidate adjustment scheme, generating a simulated drug cumulative dose subsequence. Specifically, the cumulative dose values of various drugs at the current time are obtained as the starting cumulative dose, and the cumulative dose is gradually increased according to the infusion rate set in the candidate adjustment scheme within the preset future duration, generating a simulated drug cumulative dose curve from the current time to the end of the preset future duration.
[0102] For example, taking remifentanil as an example, the current cumulative dose is 200 μg, the preset future duration is 30 minutes, and the remifentanil infusion rate in the candidate adjustment protocol is 0.2 μg / kg / min (the target subject's weight is 70 kg). Then the dosage in the next 30 minutes is 0.2 × 70 × 30 = 420 μg. The simulated drug cumulative dose subsequence starts from 200 μg and linearly accumulates to 620 μg after 30 minutes.
[0103] For adjustable ventilation variables, a subsequence of simulated ventilation parameters for a preset future duration is generated based on the candidate adjustment values of the adjustable ventilation variables in the candidate adjustment schemes. Specifically, the ventilation parameter values set in the candidate adjustment schemes (such as tidal volume of 6.0 mL / kg) are used as the target parameter values for the preset future duration, and time-series data of ventilation parameters from the current moment to the end of the preset future duration are generated in chronological order.
[0104] For example, if the tidal volume is adjusted from 5.2 mL / kg to 6.0 mL / kg in the candidate adjustment scheme, the simulated ventilation parameter subsequence will remain at 6.0 mL / kg for the preset future duration of 30 minutes. If the transition process of the adjustment is considered, a gradual sequence from 5.2 mL / kg to 6.0 mL / kg can be set at the beginning of the preset future duration.
[0105] For adjustable variables related to vital signs, the response of the target subject's physiological parameters to dose adjustments is estimated based on pharmacokinetic or pharmacodynamic models, generating a simulated vital sign parameter subsequence over a preset future time period. A pharmacokinetic model is a mathematical model used to describe the absorption, distribution, metabolism, and excretion of drugs in the body; a pharmacodynamic model is a mathematical model used to describe the relationship between drug concentration and effect. Specifically, by inputting drug dose adjustment information from candidate adjustment schemes, the pharmacokinetic model estimates the change in drug concentration over time, and then the pharmacodynamic model estimates the impact of this concentration change on vital sign parameters such as blood pressure and heart rate of the target subject, generating time-series data of simulated vital sign parameters over a preset future time period.
[0106] For example, in the candidate adjustment regimen, the norepinephrine infusion rate is increased from 0.1 μg / kg / min to 0.15 μg / kg / min. The change curve of norepinephrine blood concentration within 30 minutes is estimated by the pharmacokinetic model, and the effect of this concentration change on mean arterial pressure is estimated by the pharmacodynamic model, generating simulated mean arterial pressure time series data within 30 minutes.
[0107] The three types of simulated subsequences (simulated drug cumulative dose subsequence, simulated ventilation parameter subsequence, and simulated vital sign parameter subsequence) together constitute the complete simulation process time series data. This simulation process time series data is continuous in time with the real historical time series data, reflecting the simulated state of the target object within a preset future time period after the implementation of the candidate adjustment scheme, providing a complete input data foundation for subsequent calculation and extrapolation of probability values. After step S304 is completed, the simulation process time series data corresponding to each group of candidate adjustment schemes are output.
[0108] In step S305, the inferred probability value refers to the inferred prediction value of the probability of occurrence of various predefined events under the candidate adjustment scheme. The specific method for obtaining the inferred probability values of various predefined events under each candidate adjustment scheme by sequentially inputting baseline static feature data, multi-channel time-series data collected before the current time, and simulation process time-series data into the sequence coding network and multiple task branches is as follows: taking the current time as the dividing point, the ventilation parameter sequence, drug cumulative dose sequence, and vital sign parameter sequence collected before the current time are used as real historical time-series data; the real historical time-series data and simulation process time-series data are spliced together in chronological order to form complete inferred time-series data; the inferred process time-series feature sequence is extracted from the inferred time-series data, and the extraction method is the same as that in step S102, which is to perform time axis alignment and extract statistical features with a preset time window as the granularity; the baseline static feature data and the inferred process time-series feature sequence are input into the sequence coding network to obtain the inferred fusion features output by the sequence coding network; the inferred fusion features are input into multiple task branches respectively to obtain the inferred probability values of the corresponding categories of predefined events output by each task branch.
[0109] The mathematical expression for counterfactual reasoning is Equation (2): (2) In equation (2), In the first Group candidate adjustment scheme The inferred probability value of a predefined event. Indicates the current moment. For the category index of predefined events, For the first Group candidate adjustment plan, For the sigmoid function, For the first The weight matrix of the task branches corresponding to the predefined events of the class. Let be the nonlinear mapping function represented by the sequence coding network. These are the network parameters of the sequence coding network. For the baseline static feature vector, From the start time to the current time Real historical time-series data, Indicates the first Under the group candidate adjustment scheme, from the current moment To the future Time series data of the simulation process during the period, This refers to the complete inferred time series data formed by splicing real historical time series data and simulated process time series data in chronological order.
[0110] In step S306, based on the projected probability values corresponding to each group of candidate adjustment schemes, a group can be selected as the recommended adjustment scheme from multiple candidate adjustment schemes by comprehensively comparing the magnitudes of the projected probability values of each group.
[0111] In summary, the counterfactual reasoning engine layer triggers the counterfactual inference process when the probability value of at least one type of predefined event exceeds a second preset threshold, initiating counterfactual inference only when the risk level reaches a high level. This avoids the waste of computational resources caused by performing counterfactual inference on all prediction results, thus improving the system's operational efficiency. By generating candidate adjustment values within the limited range of each adjustable variable and combining them into candidate adjustment schemes, it ensures that all candidate schemes are within a safe range. By generating simulated process time series data within a preset future duration based on the candidate adjustment schemes, starting from the current moment, it achieves quantitative simulation of future state changes after the implementation of candidate adjustment schemes. By concatenating real historical time series data with simulated process time series data and inputting the result into a sequence coding network and multiple task branches to obtain inference probability values, the calculation of inference probability values is based on the currently accumulated real state, while also considering the future changes brought about by candidate adjustment schemes. By determining the recommended adjustment scheme based on the inference probability values corresponding to each group of candidate adjustment schemes, it achieves the goal of finding the optimal adjustment direction within the range of adjustable variable values, providing operators with quantitative decision support.
[0112] In one embodiment, determining the recommended adjustment scheme based on the inferred probability values corresponding to each group of candidate adjustment schemes includes the following steps: S3061: Obtain the set of constraints for the target object. The set of constraints includes at least one of the following: safe dose limits for various drugs, safe ranges for ventilation parameters, acceptable ranges for physiological parameters, hemodynamic stability boundary conditions, and incompatibilities. S3062: Eliminate candidate adjustment schemes that do not meet the constraint conditions from multiple candidate adjustment schemes; S3063: Obtain the preset weights of various predefined events; S3064: For each group of candidate adjustment schemes that satisfy the set of constraints, calculate the weighted sum of the inferred probability values of various predefined events and the corresponding preset weights to obtain the comprehensive evaluation value of each group of candidate adjustment schemes; S3065: Select the candidate adjustment scheme with the smallest comprehensive evaluation value as the recommended adjustment scheme.
[0113] As an example, the constraint set is a pre-defined set of safety constraints that candidate adjustment schemes must satisfy. The safe dose limits for various drugs refer to the upper limits that various drugs should not exceed in terms of both unit time and cumulative dosage; for example, the infusion rate of remifentanil should not exceed 0.5 μg / kg / min. The safe range for ventilation parameters refers to the range within which ventilation parameters should be maintained; for example, tidal volume should be maintained between 4-12 mL / kg, and peak airway pressure should be below 40 cmH2O. The acceptable range for physiological parameters refers to the range within which the physiological parameters of the target subject should be maintained; for example, the bispectral index of EEG should be maintained between 40-60. The hemodynamic stability boundary condition refers to the requirement that the hemodynamic parameters of the target subject should remain stable; for example, the expected mean arterial pressure should not be lower than 70% of the baseline. The incompatibility condition refers to the requirement that known drug incompatibilities should not be violated.
[0114] Preset weights refer to the weight coefficients that are pre-set for various predefined events. These preset weights are pre-set according to the severity of each predefined event; the higher the severity, the greater the corresponding preset weight.
[0115] Specifically, the comprehensive evaluation value refers to the comprehensive score calculated by weighting the inferred probability values of various predefined events with their corresponding preset weights. The mathematical expression for the optimal solution search is Equation (3): In equation (3), This is the optimal candidate adjustment scheme (i.e., the recommended adjustment scheme). For the first Group candidate adjustment plan, For the first Predefined weights for class-predefined events In the first Group candidate adjustment scheme The inferred probability value of a predefined event. This is the set of constraints.
[0116] For example, taking the existence of two predefined events—nausea and vomiting, and delayed awakening—as an example, the preset weight for nausea and vomiting is set to 0.4, and the preset weight for delayed awakening is set to 0.6. For candidate scheme A, the inferred probability of nausea and vomiting is 0.3, and the probability of delayed awakening is 0.4, with a comprehensive evaluation value of 0.4×0.3+0.6×0.4=0.36; for candidate scheme B, the inferred probability of nausea and vomiting is 0.5, and the probability of delayed awakening is 0.2, with a comprehensive evaluation value of 0.4×0.5+0.6×0.2=0.32. If both candidate schemes satisfy the constraint set, then candidate scheme B, with the smaller comprehensive evaluation value, is selected as the recommended adjustment scheme.
[0117] In one embodiment, the following steps are also included: S307: For the recommended adjustment scheme, calculate the difference between the occurrence probability value and the inferred probability value of each type of predefined event to obtain the probability decrease of each type of predefined event; S308: Based on the probability decrease rate and the occurrence probability values of various predefined events, estimate the duration required for the occurrence probability value to decrease to below the preset safety threshold; S309: Generate recommended adjustment plans, the projected probability values corresponding to the recommended adjustment plans, the probability decrease rate and duration, and side effect risk warning information for the recommended adjustment plans.
[0118] As an example, the probability decrease refers to the numerical amount by which the probability of a certain type of predefined event decreases after the implementation of the recommended adjustment plan compared to the current probability. Specifically, the probability decrease is calculated as the difference between the current probability and the projected probability under the recommended adjustment plan. The larger the probability decrease, the more significant the effect of the recommended adjustment plan in reducing the risk of that type of event.
[0119] A preset safety threshold is a pre-set critical value used to determine whether the probability of an occurrence has fallen below a safe level. For example, the preset safety threshold can be set to 0.5.
[0120] Duration refers to the length of time required for the probability of occurrence to decrease to below a preset safety threshold from the current moment. The duration is estimated by linearly estimating the time required for the probability of occurrence to decrease to below the preset safety threshold based on the difference between the current probability of occurrence and the preset safety threshold, and by extrapolating the decreasing trend of the probability value relative to the current probability value.
[0121] Side effect risk warning information refers to the warning content regarding the potential negative effects that may occur after adjusting the adjustable variables involved in the recommended adjustment plan.
[0122] For example, taking a recommended adjustment protocol as an example, the currently predicted probability of nausea and vomiting is 0.68, and the projected probability of nausea and vomiting under the recommended adjustment protocol is 0.38. Therefore, the probability of nausea and vomiting decreases by 0.68 - 0.38 = 0.30, or approximately 44%. Based on the probability of occurrence (0.68), the projected probability (0.38), and the preset safety threshold (0.5), the estimated duration required for the probability of occurrence to decrease to below 0.5 is approximately 20 minutes. The recommended adjustment protocol is generated: "Reduce the remifentanil infusion rate from 0.3 μg / kg / min to 0.2 μg / kg / min and adjust the tidal volume from 5.2 mL / kg to 6.0 mL / kg," along with the corresponding projected probability (0.38), probability decrease (0.30), and duration (20 minutes). Simultaneously, side effect risk warnings are generated: "Insufficient analgesia may occur; monitoring of patient pain scores is recommended. Increased airway pressure may occur; monitoring of peak airway pressure is recommended."
[0123] In this embodiment, by calculating the probability decrease rate, the actual effect of the recommended adjustment plan on reducing the risk of various predefined events can be quantified; by estimating the duration, the time required for the risk to decrease to a safe level can be predicted; by generating recommended adjustment plans, projected probability values, probability decrease rates, duration, and side effect risk warning information, complete decision support information including quantified effects and risk warnings is provided to operators, enabling them to comprehensively evaluate the expected effects and potential risks of the recommended adjustment plan, thereby making more reasonable decisions.
[0124] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.
[0125] Secondly, embodiments of this application provide an apparatus for determining the probability of an event occurring, which corresponds one-to-one with the methods for determining the probability of an event occurring in the above embodiments. For example... Figure 4 As shown, the device for determining the probability of this event includes an acquisition module 401, an alignment module 402, a first processing module 403, and a second processing module 404. Detailed descriptions of each functional module are as follows: The acquisition module 401 is used to acquire baseline static feature data of the target object and acquire multi-channel time series data of the target object, wherein the multi-channel time series data includes ventilation parameter sequence, drug cumulative dose sequence and vital sign parameter sequence; Alignment module 402 is used to align the ventilation parameter sequence, the cumulative drug dose sequence and the vital sign parameter sequence along the time axis, and extract the statistical features of each sequence at a preset time window granularity to obtain the process time sequence feature sequence. The first processing module 403 is used to input the baseline static feature data and the process time-series feature sequence into a preset sequence coding network to obtain the fusion features output by the sequence coding network; The second processing module 404 is used to input the fusion features into a plurality of preset task branches respectively, and obtain the occurrence probability value of the corresponding category of predefined events output by each task branch; wherein, each task branch corresponds to a category of the predefined events.
[0126] In one embodiment, the alignment module 402 is further configured to: The ventilation parameter sequence, the cumulative drug dose sequence, and the vital signs parameter sequence are time-axis aligned to obtain three time-series data streams after alignment. Using the preset time window as the granularity, statistical features within each preset time window are extracted from the aligned three-channel time series data to obtain the process time series features under each preset time window. Based on the process timing characteristics under each preset time window, the process timing characteristic sequence is generated.
[0127] In one embodiment, the first processing module 403 is further configured to: Map the baseline static feature data to a static feature vector; The static feature vector is concatenated with each process time-series feature in the process time-series feature sequence to obtain multiple fused input vectors; The fusion input vectors are input into the sequence coding network in chronological order, so that the sequence coding network performs state transfer and update on the fusion input vectors at each time step, and outputs the hidden state vector at the last time step as the fusion feature. The sequence coding network includes a gated recurrent unit network or a Transformer network.
[0128] In one embodiment, the second processing module 404 is further configured to: The fused features are input into each of the task branches, so that each task branch performs linear transformation and sigmoid function activation on the fused features, and outputs the occurrence probability value of the predefined event of the corresponding category. Each of the task branches is a single-layer fully connected network with an independent weight matrix and bias vector.
[0129] In one embodiment, the device further includes a counterfactual reasoning module (not shown in the figure); The counterfactual reasoning module is used for: When the probability value of occurrence of at least one of the predefined events exceeds a first preset threshold, attribution analysis is performed on the contribution of each feature dimension of the input side of the sequence coding network to the current output to obtain the contribution index of each feature dimension; wherein, the feature dimension includes each component in the static feature vector formed by mapping the baseline static feature data, and each process time series feature component in each time window of the process time series feature sequence. From each of the aforementioned feature dimensions, feature dimensions whose contribution index exceeds a preset contribution threshold are selected as candidate feature dimensions; Based on the candidate feature dimensions, the feature dimensions belonging to the adjustable variables are determined and used as the adjustable feature dimensions. Based on the adjustable feature dimensions, risk mitigation measures information for the target object is generated.
[0130] In one embodiment, the counterfactual reasoning module is further configured to: When the probability value of occurrence of at least one of the predefined events exceeds a second preset threshold, identify the adjustable variable corresponding to the current time and the current value of each adjustable variable; For each adjustable variable, within the limited range of values corresponding to the adjustable variable, multiple candidate adjustment values are generated to obtain a set of candidate adjustment values corresponding to each adjustable variable; Select one candidate adjustment value from the set of candidate adjustment values corresponding to each of the adjustable variables, combine the selected candidate adjustment values of each adjustable variable into a set of candidate adjustment schemes, and iterate through all combinations to generate multiple sets of candidate adjustment schemes. For each group of candidate adjustment schemes, starting from the current time, simulated process time series data within a preset future time period is generated according to the candidate adjustment scheme; The baseline static feature data, the multi-channel time series data collected before the current time, and the simulation process time series data are sequentially input into the sequence coding network and the multiple task branches to obtain the inferred probability values of various predefined events under each group of candidate adjustment schemes; Based on the inferred probability values corresponding to the candidate adjustment schemes in each group, the recommended adjustment scheme is determined.
[0131] In one embodiment, the counterfactual reasoning module is further configured to: Obtain the set of constraints for the target object, wherein the set of constraints includes at least one of the following: safe dose limits for various drugs, safe ranges for ventilation parameters, acceptable ranges for physiological parameters, hemodynamic stability boundary conditions, and incompatibilities. Eliminate candidate adjustment schemes that do not meet the set of constraints from the multiple sets of candidate adjustment schemes; Obtain the preset weights of each of the predefined events; For each group of candidate adjustment schemes that satisfy the set of constraints, calculate the weighted sum of the inferred probability value of each type of predefined event and the corresponding preset weight to obtain the comprehensive evaluation value of each group of candidate adjustment schemes; The candidate adjustment scheme with the smallest comprehensive evaluation value is selected as the recommended adjustment scheme.
[0132] Specific limitations regarding the device for determining the probability of an event can be found in the limitations of the method for determining the probability of an event described above, and will not be repeated here. Each module in the aforementioned device for determining the probability of an event can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the electronic device, or stored in the memory of the electronic device in software form, so that the processor can call and execute the operations corresponding to each module.
[0133] Thirdly, embodiments of this application provide an electronic device whose internal structure diagram can be as follows: Figure 5 As shown, the electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the steps of a method for determining the probability of an event occurring.
[0134] In one embodiment, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the following steps: Baseline static feature data of the target object is acquired, as well as multi-channel time-series data of the target object, including ventilation parameter sequences, cumulative drug dose sequences, and vital sign parameter sequences. The ventilation parameter sequence, the cumulative drug dose sequence, and the vital signs parameter sequence are aligned along the time axis, and statistical features of each sequence are extracted at a preset time window granularity to obtain the process time sequence feature sequence. The baseline static feature data and the process time-series feature sequence are input into a preset sequence coding network to obtain the fused features output by the sequence coding network; The fusion features are input into multiple preset task branches to obtain the occurrence probability value of the corresponding category of predefined events output by each task branch; wherein each task branch corresponds to one category of the predefined events.
[0135] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the following steps: Baseline static feature data of the target object is acquired, as well as multi-channel time-series data of the target object, including ventilation parameter sequences, cumulative drug dose sequences, and vital sign parameter sequences. The ventilation parameter sequence, the cumulative drug dose sequence, and the vital signs parameter sequence are aligned along the time axis, and statistical features of each sequence are extracted at a preset time window granularity to obtain the process time sequence feature sequence. The baseline static feature data and the process time-series feature sequence are input into a preset sequence coding network to obtain the fused features output by the sequence coding network; The fusion features are input into multiple preset task branches to obtain the occurrence probability value of the corresponding category of predefined events output by each task branch; wherein each task branch corresponds to one category of the predefined events.
[0136] It should be noted that the steps described above that can be implemented by computer-readable storage media or electronic devices can be referred to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.
[0137] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. A method for determining the probability of an event occurring, characterized in that, include: Baseline static feature data of the target object is acquired, as well as multi-channel time-series data of the target object, including ventilation parameter sequences, cumulative drug dose sequences, and vital sign parameter sequences. The ventilation parameter sequence, the cumulative drug dose sequence, and the vital signs parameter sequence are aligned along the time axis, and statistical features of each sequence are extracted at a preset time window granularity to obtain the process time sequence feature sequence. The baseline static feature data and the process time-series feature sequence are input into a preset sequence coding network to obtain the fused features output by the sequence coding network; The fusion features are input into multiple preset task branches to obtain the occurrence probability value of the corresponding category of predefined events output by each task branch; wherein each task branch corresponds to one category of the predefined events.
2. The method for determining the probability of an event occurring according to claim 1, characterized in that, The ventilation parameter sequence, the cumulative drug dose sequence, and the vital signs parameter sequence are time-axis aligned, and statistical features of each sequence are extracted at a preset time window granularity to obtain a process time-series feature sequence, including: The ventilation parameter sequence, the cumulative drug dose sequence, and the vital signs parameter sequence are time-axis aligned to obtain three time-series data streams after alignment. Using the preset time window as the granularity, statistical features within each preset time window are extracted from the aligned three-channel time series data to obtain the process time series features under each preset time window. Based on the process timing characteristics under each preset time window, the process timing characteristic sequence is generated.
3. The method for determining the probability of an event occurring according to claim 1, characterized in that, The step of inputting the baseline static feature data and the process time-series feature sequence into a preset sequence coding network to obtain the fused features output by the sequence coding network includes: Map the baseline static feature data to a static feature vector; The static feature vector is concatenated with each process time-series feature in the process time-series feature sequence to obtain multiple fused input vectors; The fusion input vectors are input into the sequence coding network in chronological order, so that the sequence coding network performs state transfer and update on the fusion input vectors at each time step, and outputs the hidden state vector at the last time step as the fusion feature. The sequence coding network includes a gated recurrent unit network or a Transformer network.
4. The method for determining the probability of an event occurring according to claim 1, characterized in that, The step of inputting the fused features into multiple preset task branches and obtaining the occurrence probability value of a predefined event of the corresponding category output by each task branch includes: The fused features are input into each of the task branches, so that each task branch performs linear transformation and sigmoid function activation on the fused features, and outputs the occurrence probability value of the predefined event of the corresponding category. Each of the task branches is a single-layer fully connected network with an independent weight matrix and bias vector.
5. The method for determining the probability of an event occurring according to claim 1, characterized in that, After inputting the fused features into multiple preset task branches and obtaining the occurrence probability value of a predefined event of the corresponding category output by each task branch, the method further includes: When the probability value of occurrence of at least one of the predefined events exceeds a first preset threshold, attribution analysis is performed on the contribution of each feature dimension of the input side of the sequence coding network to the current output to obtain the contribution index of each feature dimension; wherein, the feature dimension includes each component in the static feature vector formed by mapping the baseline static feature data, and each process time series feature component in each time window of the process time series feature sequence. From each of the aforementioned feature dimensions, feature dimensions whose contribution index exceeds a preset contribution threshold are selected as candidate feature dimensions; Based on the candidate feature dimensions, the feature dimensions belonging to the adjustable variables are determined and used as the adjustable feature dimensions. Based on the adjustable feature dimensions, risk mitigation measures information for the target object is generated.
6. The method for determining the probability of an event occurrence according to any one of claims 1 to 5, characterized in that, After inputting the fused features into multiple preset task branches and obtaining the occurrence probability value of a predefined event of the corresponding category output by each task branch, the method further includes: When the probability value of occurrence of at least one of the predefined events exceeds a second preset threshold, identify the adjustable variable corresponding to the current time and the current value of each adjustable variable; For each adjustable variable, within the limited range of values corresponding to the adjustable variable, multiple candidate adjustment values are generated to obtain a set of candidate adjustment values corresponding to each adjustable variable; Select one candidate adjustment value from the set of candidate adjustment values corresponding to each of the adjustable variables, combine the selected candidate adjustment values of each adjustable variable into a set of candidate adjustment schemes, and iterate through all combinations to generate multiple sets of candidate adjustment schemes. For each group of candidate adjustment schemes, starting from the current time, simulated process time series data within a preset future time period is generated according to the candidate adjustment scheme; The baseline static feature data, the multi-channel time series data collected before the current time, and the simulation process time series data are sequentially input into the sequence coding network and the multiple task branches to obtain the inferred probability values of various predefined events under each group of candidate adjustment schemes; Based on the inferred probability values corresponding to the candidate adjustment schemes in each group, the recommended adjustment scheme is determined.
7. The method for determining the probability of an event occurring according to claim 6, characterized in that, The step of determining the recommended adjustment scheme based on the inferred probability value corresponding to each group of candidate adjustment schemes includes: Obtain the set of constraints for the target object, wherein the set of constraints includes at least one of the following: safe dose limits for various drugs, safe ranges for ventilation parameters, acceptable ranges for physiological parameters, hemodynamic stability boundary conditions, and incompatibilities. Eliminate candidate adjustment schemes that do not meet the set of constraints from the multiple sets of candidate adjustment schemes; Obtain the preset weights of each of the predefined events; For each group of candidate adjustment schemes that satisfy the set of constraints, calculate the weighted sum of the inferred probability value of each type of predefined event and the corresponding preset weight to obtain the comprehensive evaluation value of each group of candidate adjustment schemes; The candidate adjustment scheme with the smallest comprehensive evaluation value is selected as the recommended adjustment scheme.
8. A device for determining the probability of an event occurring, characterized in that, include: The acquisition module is used to acquire baseline static feature data of the target object and acquire multi-channel time-series data of the target object, wherein the multi-channel time-series data includes ventilation parameter sequences, drug cumulative dose sequences and vital sign parameter sequences; The alignment module is used to align the ventilation parameter sequence, the cumulative drug dose sequence, and the vital sign parameter sequence along the time axis, and extract the statistical features of each sequence at a preset time window granularity to obtain the process time sequence feature sequence. The first processing module is used to input the baseline static feature data and the process time-series feature sequence into a preset sequence coding network to obtain the fusion features output by the sequence coding network; The second processing module is used to input the fusion features into a plurality of preset task branches respectively, and obtain the occurrence probability value of the corresponding category of predefined events output by each task branch; wherein, each task branch corresponds to a category of the predefined events.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for determining the probability of an event as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method for determining the probability of an event as described in any one of claims 1 to 7.
