Anesthesia response characteristic analysis system based on visual identification

Abstract

The invention relates to the technical field of medical data analysis, in particular to an anesthesia response characteristic analysis system based on visual identification. The system comprises an anesthesia multi-source acquisition module, an anesthesia characterization analysis module, an anesthesia reaction analysis module, an anesthesia characteristic analysis module, an anesthesia state evaluation module, a multi-dimensional visual acquisition device and an associated medical data interface in an anesthesia scene, the method comprises the following steps: collecting human visual feature data and synchronous medical monitoring data in an anesthesia process, carrying out noise reduction, time sequence alignment and format standardization preprocessing, and carrying out extraction and characterization processing at the same time to generate anesthesia state visual characterization data; correlation analysis is carried out in combination with the visual representation data of the anesthesia state, and meanwhile, an anesthesia reaction characteristic dynamic model is constructed and is iteratively optimized to generate anesthesia reaction characteristic analysis data; and outputting an anesthesia state evaluation report and clinical intervention suggestions according to the anesthesia reaction characteristic analysis data. According to the method, intervention opportunity lag caused by lack of specific guidance can be avoided.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis technology, and in particular to a visual recognition-based anesthesia response characteristic analysis system. Background Technology

[0002] In recent years, with the continuous improvement of clinical medicine's requirements for precision and personalization in anesthesia, the need for real-time monitoring, characteristic analysis, and intervention guidance of patient responses during anesthesia has become increasingly urgent, placing higher demands on the timeliness, accuracy, and data fusion analysis capabilities of anesthesia response identification. Traditional anesthesia response monitoring relies heavily on single-dimensional detection of physiological indicators (such as heart rate, blood pressure, and blood oxygen saturation), lacking comprehensive capture of visual features such as facial expressions and limb movements, making it difficult to fully reflect changes in the patient's anesthesia status. At the same time, existing methods are mostly based on fixed clinical threshold judgments, failing to form dynamic and personalized anesthesia response characteristic analysis models, which can easily lead to delayed intervention timing or judgment bias. Summary of the Invention

[0003] Therefore, it is necessary for the present invention to provide a visual recognition-based anesthesia response characteristic analysis system to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a visual recognition-based anesthesia response characteristic analysis system includes the following modules: The anesthesia multi-source acquisition module is used to acquire multi-dimensional visual acquisition devices and related medical data interfaces in the anesthesia scenario. Based on the above devices and interfaces, it acquires human visual feature data and synchronous medical monitoring data during the anesthesia process, and performs noise reduction, time sequence alignment and format standardization preprocessing to generate integrated multi-source fusion data of the anesthesia scenario. The anesthesia characterization analysis module is used to extract and characterize visual features from multi-source fusion data of anesthesia scenarios to generate visual characterization data of anesthesia status. The anesthetic response analysis module is used to perform correlation analysis by combining visual representation data of anesthetic state with preset clinical anesthetic response evaluation indicators to generate preliminary anesthetic response analysis results including probability distribution of response types. The anesthetic characteristics analysis module is used to construct a dynamic model of anesthetic response characteristics based on the preliminary analysis results of anesthetic response, and to iteratively optimize and generate anesthetic response characteristic analysis data. The anesthesia status assessment module is used to analyze data based on the characteristics of anesthesia response and output an anesthesia status assessment report and clinical intervention recommendations.

[0005] Furthermore, the anesthesia multi-source acquisition module includes the following functions: Acquire medical data interfaces for facial expression capture devices, eye dynamic capture devices, limb posture monitoring devices, anesthesia monitors, and vital sign monitors in anesthesia scenarios. The aforementioned visual acquisition devices continuously collect visual information corresponding to the characteristics of facial muscle movement, eyelid opening and closing, eye movement trajectory, limb movement range and posture changes during anesthesia. Simultaneously, medical monitoring data corresponding to heart rate, blood pressure, respiratory rate and anesthetic drug concentration are collected through the medical data interface, which together constitute the original anesthesia-related data. Image denoising, blur correction, and feature point enhancement are performed on the visual information in the original anesthesia-related data; outlier removal and missing value completion are performed on the medical monitoring data. The processed visual information and medical monitoring data are aligned sequentially according to timestamps, and the data format and sampling frequency are unified to generate integrated multi-source fusion data of the anesthesia scene.

[0006] Furthermore, the anesthesia characterization analysis module includes the following functions: Visual feature data and medical monitoring-related data were separated from multi-source fusion data in anesthesia scenarios; Hierarchical feature extraction is performed on visual feature data. First, basic visual features corresponding to facial contours and limb contours are extracted. Then, deep dynamic features corresponding to facial micro-expressions, eye tremor frequency, and subtle limb twitches are mined to generate a multi-dimensional visual feature set. Feature filtering and dimensionality reduction are performed on the multi-dimensional visual feature set to remove redundant and interfering features and retain the core visual features that are highly correlated with the anesthetic response. The core visual features are correlated with the corresponding medical monitoring data to construct a mapping relationship between visual features and anesthesia physiological state, and generate visual representation data of anesthesia state. The validity of the visual representation data of the anesthesia state was verified, invalid representation information was removed, and the visual representation data of the anesthesia state was ensured to reflect the probability distribution of response types during the anesthesia process.

[0007] Furthermore, the anesthesia response analysis module includes the following functions: Clinical anesthesia response evaluation indicators were extracted from clinical anesthesia diagnosis and treatment guidelines, covering key dimensions corresponding to sedation depth level, analgesic effect assessment, signs of adverse reactions, and degree of neuromuscular blockade. The core visual features in the visual representation data of anesthesia status are matched one by one with each indicator in the clinical anesthesia response evaluation index to establish the correspondence between visual features and evaluation indicators. By analyzing Pearson correlation coefficients and calculating mutual information, the correlation strength between core visual features and various evaluation indicators is quantified, and feature-indicator combinations with correlation strengths higher than preset thresholds are selected. Multidimensional correlation analysis is performed based on the selected feature-indicator combinations to explore the dynamic pattern between changes in visual features and fluctuations in anesthesia response indicators, calculate the probability of occurrence of different response types, and generate preliminary analysis results of anesthesia response including the probability distribution of response types.

[0008] Furthermore, the anesthesia characteristic analysis module includes the following functions: We collected a large amount of clinical anesthesia case data, covering anesthesia response data, synchronous visual feature data and related probability statistics corresponding to different anesthesia methods, different patient physical conditions and different surgical types, and constructed an anesthesia response characteristic sample library. Based on the preliminary analysis results of the anesthetic response, the input feature dimensions of the model are determined. The core visual features, related medical monitoring data and feature association probabilities are used as input variables, and the actual values ​​and occurrence probabilities of clinical anesthetic response evaluation indicators are used as output variables to construct an initial dynamic model of anesthetic response characteristics. The gradient descent algorithm combined with cross-validation is used to train the initial dynamic model of anesthesia response characteristics, and the model parameters are continuously adjusted to minimize the prediction error. The preliminary analysis results of anesthesia response are input into the trained dynamic model of anesthesia response characteristics, and the deviation is corrected through dynamic iterative calculation of the dynamic model of anesthesia response characteristics to generate anesthesia response characteristic analysis data. The anesthetic response characteristics analysis data output by the model were clinically validated, and the dynamic model of anesthetic response characteristics was continuously optimized by combining the evaluation feedback of anesthesiologists.

[0009] Furthermore, the evaluation feedback continuously optimizes the dynamic model of anesthetic response characteristics, including: Visual feature sequence data from multi-source fusion data of anesthesia scenarios are acquired and divided into time windows. The visual feature change trend, mutation point information and feature occurrence probability are extracted within each time window. By combining synchronously collected medical monitoring data, the temporal correlation between visual feature mutation points and vital sign fluctuations and anesthetic drug infusion nodes is analyzed, the probability of occurrence of related events is calculated, and feature-physiology-drug correlation data containing temporal correlation probabilities is generated. An anesthesia response time-series change map is constructed based on associated data. The map includes visual feature trajectories, physiological index curves, drug infusion curves, correlation markers between the three and related probability information. Dynamic patterns in the anesthesia response time-series change map are then mined. The mined dynamic patterns and corresponding probability information are integrated into the dynamic model of anesthesia response characteristics to continuously optimize the anesthesia response characteristic analysis data.

[0010] Furthermore, the generation of feature-physiological-drug association data containing temporal correlation probabilities includes: The visual feature sequence data in the multi-source fusion data of the anesthesia scene is preprocessed to remove abnormal sequence segments caused by equipment shaking and light changes. The missing sequence data is filled in by interpolation to ensure the continuity, integrity and accuracy of probability calculation of the sequence data. The time windows are divided according to a preset time interval. The length of each time window is dynamically adjusted according to different stages of the anesthesia process. The length of the window during the anesthesia induction and recovery periods is shortened, while the length of the window during the maintenance period is appropriately extended. Statistical analysis is performed on the visual feature data within each time window to calculate the feature mean, variance, rate of change, peak value and probability of feature occurrence, extract the feature change trend, and identify visual feature mutation points within the time window through anomaly detection algorithm, and record the time location, feature change magnitude and mutation probability of visual feature mutation points. The visual feature change trends, mutation point information, and related probability parameters within each time window are mapped one-to-one with the heart rate, blood pressure, blood oxygen saturation, and anesthetic drug concentration collected at the same time, and a correlation index is established in the time dimension. By calculating the time difference between visual feature mutation points, physiological indicator fluctuation nodes, and drug infusion nodes, the causal relationship among the three is analyzed, generating feature-physiology-drug association data that includes temporal correlation strength, response delay time, and probability of occurrence of associated events.

[0011] Furthermore, the dynamic patterns include the time difference between changes in visual features and fluctuations in physiological indicators, the response period of visual features after drug adjustment, and the probability of occurrence of various patterns.

[0012] Furthermore, the process of mining the dynamic patterns includes: With the time axis as the horizontal axis, a visual feature trajectory layer, a physiological index curve layer, and a drug infusion curve layer are constructed respectively. The visual feature trajectory layer uses the numerical changes and occurrence probabilities of core visual features as the vertical axis, the physiological index curve layer uses the values ​​of various vital signs parameters and their normal probability ranges as the vertical axis, and the drug infusion curve layer uses the infusion rate, cumulative dose, and dose adjustment probability of anesthetic drugs as the vertical axis. In each layer, key nodes are marked. The visual feature trajectory layer marks abrupt change points, feature peak points and corresponding probabilities. The physiological index curve layer marks the normal range threshold line, abnormal fluctuation points and the probability of abnormal occurrence. The drug infusion curve layer marks drug type switching points, dose adjustment points and adjustment probabilities. By linking and marking the key nodes of time synchronization in the three layers through the correlation index, the correspondence and correlation probability between changes in visual features, fluctuations in physiological indicators, and drug adjustments are clarified, forming a complete time sequence map of anesthetic response changes. The anesthesia response time-series change map was analyzed to identify recurring feature-physiological-drug change patterns. The lead time of visual feature changes relative to physiological index fluctuations under different patterns, the response period for visual features to reach a stable state after drug adjustment, and the probability of occurrence of various patterns were calculated. The mined features, including the lead time range, the mean response period, and the probability of pattern occurrence, were integrated as supplementary features into the training process of the dynamic model of anesthesia response characteristics to continuously optimize the anesthesia response characteristic analysis data.

[0013] Furthermore, the process of continuously optimizing the anesthetic response characteristic analysis data during the training of the dynamic model of anesthetic response characteristics includes: The parameterized supplementary features, together with the original core visual features and associated medical monitoring data, constitute the model input feature set, expanding the input dimension of the dynamic model of anesthesia response characteristics. The network structure of the dynamic model of anesthesia response characteristics was adjusted by adding a temporal feature processing layer and a probability calculation unit to learn and fuse the temporal regularity parameters and probability information corresponding to the supplementary features. The optimized dynamic model of anesthesia response characteristics was retrained using newly added supplementary features. By comparing the prediction error, accuracy, recall, and probability prediction precision of the dynamic model of anesthesia response characteristics before and after training, the impact of the integration of temporal patterns and probability information in the supplementary features on the model performance was verified. Through dynamic iterative calculation of the dynamic model of anesthesia response characteristics, anesthesia response characteristic analysis data with optimized probability assessment was generated.

[0014] The beneficial effects of this invention are: The anesthesia response characteristic analysis system based on visual recognition proposed in this invention consists of an anesthesia multi-source acquisition module, an anesthesia characterization analysis module, an anesthesia response analysis module, an anesthesia characteristic analysis module, and an anesthesia state assessment module. Compared with existing technologies, the beneficial effects of this application lie in the comprehensive acquisition of human visual feature data (such as facial expressions, limb movements, and eye status) and synchronous medical monitoring data (such as heart rate, blood pressure, and blood oxygen saturation) by integrating multi-dimensional visual acquisition devices (such as high-definition cameras and motion capture devices) with associated medical data interfaces. This breaks through the limitations of traditional monitoring that relies solely on a single physiological indicator, achieving multi-source coverage of "visual features + physiological data." Noise reduction processing filters out equipment noise and environmental interference, time alignment ensures the time synchronization of data from different sources, and format standardization eliminates data format differences. The generated integrated multi-source fusion data possesses integrity, consistency, and reliability, avoiding analytical biases caused by the disorder of the original data.

[0015] Secondly, targeted feature extraction (such as facial muscle contraction amplitude, limb movement frequency, and pupil change trends) is performed on the visual information in the multi-source fusion data of the anesthesia scenario. Through representation processing, the abstract visual information is transformed into a quantifiable and analyzable data form. The generated visual representation data of the anesthesia state can capture subtle fluctuations in the patient's state during anesthesia (such as facial micro-movements and premonitory limb agitation when anesthesia is too light). Compared to traditional monitoring methods that rely solely on physiological indicators, visual representation data supplements the non-invasive state assessment dimension, reflecting potential changes in the patient's anesthesia state in advance, even before significant changes in physiological indicators appear. This provides rich feature evidence for subsequent correlation analysis and dynamic modeling, improving the timeliness of anesthesia response identification. By correlating visual representation data of the anesthesia state with pre-defined clinical anesthesia response evaluation indicators (such as anesthesia depth grading and adverse reaction judgment criteria), this approach moves beyond simple threshold comparisons of single indicators. Instead, it generates preliminary analysis results including probability distributions of response types (e.g., "60% probability of too shallow anesthesia, 30% probability of normal anesthesia, and 10% probability of too deep anesthesia") through comprehensive analysis of multi-dimensional features. This probabilistic output mode, compared to the traditional binary "yes / no" judgment, better reflects the ambiguity and individual differences in anesthesia responses, avoiding judgment biases caused by fixed thresholds. Then, based on the preliminary analysis results of anesthesia responses, and combined with individual patient characteristics (such as age, weight, and anesthesia history) and real-time monitoring data, a dynamic model of anesthesia response characteristics is constructed. This model can capture the dynamic evolution trend of the patient's anesthesia state in real time, rather than being limited to static threshold judgments. By continuously incorporating new monitoring data and analysis results for iterative optimization, the model can constantly adapt to the individual response patterns of patients, gradually improving analytical accuracy. The generated anesthetic response characteristic analysis data is dynamic and personalized, avoiding the drawbacks of traditional fixed models that apply a "one-size-fits-all" approach to different patients. This allows anesthetic response analysis to shift from "static judgment" to "dynamic adaptation," significantly improving the relevance and adaptability of the analysis. Finally, based on the anesthetic response characteristic analysis data, the output anesthetic status assessment report includes key information such as the patient's current depth of anesthesia, response type, and potential risks. Compared to traditional reports that only list physiological indicators, this report is more clinically interpretable and helps medical staff quickly grasp the full picture of the patient's anesthetic status. Simultaneously, the clinical intervention recommendations generated by combining the dynamic model analysis results (such as adjusting anesthetic drug dosage and optimizing ventilation parameters) are clearly practical and targeted, avoiding the delayed intervention or blind operation caused by the lack of specific guidance in traditional monitoring. The combination of the assessment report and intervention recommendations realizes the transformation from "data monitoring" to "clinical guidance," allowing monitoring data to directly serve anesthesia diagnosis and treatment decisions, helping medical staff adjust treatment plans in a timely manner. Attached Figure Description

[0016] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the modules of the visual recognition-based anesthesia response characteristic analysis system of the present invention; Figure 2 for Figure 1 A functional flowchart of the multi-source data acquisition module for traditional Chinese medicine anesthesia. Detailed Implementation

[0017] The technical system of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0018] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor systems and / or microcontroller systems.

[0019] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0020] To achieve the above objectives, please refer to Figures 1 to 2 This invention provides a visual recognition-based system for analyzing anesthesia response characteristics. In the embodiments of this invention, please refer to... Figure 1 The diagram shown is a schematic of the modules of the visual recognition-based anesthesia response characteristic analysis system of the present invention. In this example, the visual recognition-based anesthesia response characteristic analysis system includes the following modules: The anesthesia multi-source acquisition module is used to acquire multi-dimensional visual acquisition devices and related medical data interfaces in the anesthesia scenario. Based on the above devices and interfaces, it acquires human visual feature data and synchronous medical monitoring data during the anesthesia process, and performs noise reduction, time sequence alignment and format standardization preprocessing to generate integrated multi-source fusion data of the anesthesia scenario. In this embodiment of the invention, multi-dimensional visual acquisition devices and associated medical data interfaces for anesthesia scenarios are acquired through interface adaptation tools: the facial expression acquisition device uses a USB 3.0 interface (transmission rate 5Gbps), the eye dynamic capture device uses an HDMI interface (supports 1080P@60fps), and the limb posture monitoring device uses an RS485 interface (baud rate 9600bps); the anesthesia monitor uses an Ethernet RJ45 interface, and the vital signs monitor uses an HL7 interface. The visual acquisition devices are installed in fixed positions: the facial acquisition device is 1.5m from the head of the bed, the eye capture device is 1.2m from the face, and the limb monitoring device is fixed at the joints of the limbs, simultaneously acquiring visual feature data such as facial muscle movement, eyelid opening and closing, eye trajectory, and limb posture (sampling frequency 30 frames / second). The medical data interface simultaneously acquires heart rate (40-180 beats / minute), blood pressure (60 / 40-200 / 120mmHg), respiratory rate (10-30 breaths / minute), and anesthetic drug concentration (0-10μg / mL) (sampling frequency 5Hz). Image processing tools were used to denoise the visual data using Gaussian filtering (5×5 kernels) and perform Laplacian operator blur correction. Medical data was used to remove outliers using the 3σ criterion and fill in missing values ​​using the mean of 5 adjacent points. The two types of data were time-series aligned using a timestamp synchronization tool (accurate to milliseconds), uniformly converted to JSON format, and a sampling frequency of 5Hz was set to generate integrated multi-source fusion data of the anesthesia scene, containing 3600 records over 2 hours. Each record contains 15 visual features and 4 medical monitoring data, with a data integrity of 99.5%.

[0021] The anesthesia characterization analysis module is used to extract and characterize visual features from multi-source fusion data of anesthesia scenarios to generate visual characterization data of anesthesia status. In this embodiment of the invention, visual features in multi-source fusion data of anesthesia scenarios are processed in layers using feature extraction tools: First, basic features such as facial contours (width 12-18cm, length 18-25cm) and limb contours (upper limbs 50-70cm, lower limbs 80-100cm) are extracted using the Canny edge detection algorithm (threshold 50-150) to generate 12 basic feature parameters; then, a micro-expression recognition algorithm (sampling interval 100ms) is used to capture twitching of the corners of the mouth (amplitude 0-10mm) and eye tremors (frequency 0-5 times / second), an eye tracking algorithm (50Hz sampling) is used to calculate the nystagmus frequency (0-4 times / second), and a limb motion analysis algorithm (accuracy 0.1mm) is used to identify subtle limb twitches (amplitude 0-5mm) to generate 18 deep dynamic features. A feature representation tool was used to integrate basic and deep features by timestamp, forming a 30-dimensional multi-dimensional visual feature set. Redundant features with variance <0.3 (such as facial contour length) were removed by analysis of variance (significance level 0.05), retaining 22 effective features. Based on the clinical anesthesia response feature mapping rules, the feature values ​​were converted into representation values. For example, eyelid opening angle >40° corresponds to "light anesthesia-related representation", 20-40° corresponds to "moderate anesthesia-related representation", and <20° corresponds to "deep anesthesia-related representation", generating visual representation data of anesthesia status, totaling 3600 records. Each record contains 22 representation parameters and corresponding anesthesia status association labels.

[0022] The anesthetic response analysis module is used to perform correlation analysis by combining visual representation data of anesthetic state with preset clinical anesthetic response evaluation indicators to generate preliminary anesthetic response analysis results including probability distribution of response types. In this embodiment of the invention, four core evaluation indicators are extracted from the clinical anesthesia response evaluation guidelines: sedation depth level (3 levels: superficial, moderate, deep), analgesic effect assessment (4 levels: none, mild, moderate, severe), signs of adverse reactions (3 categories: respiratory depression, blood pressure fluctuation, muscle rigidity), and neuromuscular blockade degree (4 levels: mild, moderate, deep, complete). The quantitative judgment criteria for each indicator are clearly defined (e.g., respiratory depression corresponds to a respiratory rate <10 breaths / minute). A correlation analysis tool is used to match 22 core visual features from the visual representation data of the anesthesia state with the evaluation indicators one by one. For example, a frowning degree >70% corresponds to severe analgesia, and a nystagmus frequency <1 time / second corresponds to deep neuromuscular blockade, establishing 28 feature-indicator matching pairs. The correlation strength is quantified through Pearson correlation coefficient analysis (threshold ±0.6) and mutual information calculation (threshold 0.5 bits), selecting 21 strongly correlated combinations (e.g., the correlation coefficient between eyelid opening / closing angle and sedation depth is -0.82). Multidimensional association analysis was performed based on matching pairs to count the frequency of evaluation indicators corresponding to different combinations of visual features. For example, when the eyelid opening angle was 20-40° and the nystagmus frequency was 1-3 times / second, moderate sedation occurred 1200 times. The total number of monitoring times was 3600, and the probability of occurrence was calculated to be 33.3%. Preliminary analysis results of anesthetic reaction containing probability distributions of 8 reaction types were generated, such as "probability of moderate sedation 33.3%, probability of moderate analgesia effect 28.5%".

[0023] The anesthetic characteristics analysis module is used to construct a dynamic model of anesthetic response characteristics based on the preliminary analysis results of anesthetic response, and to iteratively optimize and generate anesthetic response characteristic analysis data. In this embodiment of the invention, data from 970 clinical anesthesia cases (covering 3 anesthesia methods, 4 patient constitutions, and 5 surgical types) were collected, including anesthesia reaction data, synchronous visual feature data, and probability statistics, to construct an anesthesia reaction characteristic sample library. Based on the preliminary analysis results of anesthesia reactions, the model input feature dimensions were determined to be 25 dimensions: 21 core visual features, 3 related medical monitoring data (anesthetic drug concentration, heart rate, and respiratory rate), and 1 feature-related probability; the output variables were determined to be 8 dimensions: 4 actual values ​​of evaluation indicators and 4 probability of occurrence of indicators. An initial dynamic model (3-layer neural network: 25 neurons in the input layer, 40 neurons in the hidden layer, and 8 neurons in the output layer) was built using a model building tool, with ReLU (hidden layer) and Sigmoid (output layer) activation functions and an initial learning rate of 0.01. The gradient descent algorithm combined with 5-fold cross-validation was used to train the model, iterating 1000 times, reducing the loss value from 0.8 to 0.12, and achieving a validation set accuracy of 86%. The preliminary analysis results were input into the model, and the bias was corrected through 100 dynamic iterations. For example, the initial predicted probability of moderate sedation was 30%, which was corrected to 32.8% (0.5% error compared to the actual clinical level). A new temporal feature processing layer (LSTM structure, 20 memory units) and probability calculation unit were added, incorporating the temporal correlation between visual feature mutation points and drug infusion and physiological fluctuations (e.g., the eyelid opening angle decreases 2 seconds after the drug concentration increases). After retraining, the model accuracy improved to 94%, generating anesthetic response characteristic analysis data, including predicted values ​​of evaluation indicators, probability of occurrence, and error range (all <1.5%).

[0024] The anesthesia status assessment module is used to analyze data based on the characteristics of anesthesia response and output an anesthesia status assessment report and clinical intervention recommendations.

[0025] In this embodiment of the invention, an anesthesia response characteristic analysis data is structured and organized using a report generation tool to output an anesthesia status assessment report. The report includes an anesthesia process timeline (14:00:00-16:00:00), core visual feature change curves at each time point (e.g., eyelid opening angle decreases from 35° to 18°), physiological index fluctuation trends (e.g., heart rate decreases from 75 beats / minute to 62 beats / minute), anesthetic drug concentration change trajectory, and clearly marks the anesthesia response type and probability of occurrence at each stage (e.g., 35% probability of moderate sedation and 1.2% probability of respiratory depression between 14:30:00 and 14:40:00). Based on the dynamic patterns in the analyzed data, clinical intervention recommendations were generated: For the pattern that "when the rate of increase in drug concentration is >0.5 μg / mL / second, a rapid decrease in eyelid opening angle occurs in 80% of cases," it is recommended to control the drug infusion rate to within 0.3 μg / mL / second; for the pattern that "when the degree of frowning is >50% and the heart rate is >80 beats / minute, the probability of moderate pain is 60%," it is recommended to check analgesia-related parameters; for the pattern that "when signs of respiratory depression appear, the correlation probability of chest wall movement amplitude <3cm is 85%," it is recommended to strengthen the monitoring of respiratory-related visual features. The report is output in PDF format and includes three parts: data charts, pattern explanations, and intervention recommendations. Each recommendation is supported by specific analytical data (such as probability values ​​and correlation strength) to ensure the recommendations are targeted and actionable.

[0026] Furthermore, the anesthesia multi-source acquisition module includes the following functions: S101: Obtain medical data interfaces corresponding to facial expression capture devices, eye dynamic capture devices, limb posture monitoring devices, anesthesia monitors, and vital sign monitors in anesthesia scenarios. In this embodiment of the invention, data interfaces of various devices in anesthesia scenarios are obtained through a medical device standard communication interface adapter: the facial expression acquisition device uses a USB 3.0 data interface (transmission rate 5Gbps), the eye dynamic capture device uses an HDMI high-definition interface (supporting 1080P@60fps video transmission), the limb posture monitoring device uses an RS485 serial interface (baud rate 9600bps, 8 data bits, 1 stop bit); the anesthesia monitor is equipped with an Ethernet RJ45 interface (supporting TCP / IP protocol), and the vital signs monitor provides an HL7 data interface (following medical data transmission specifications). During the interface adaptation process, different types of interface signals are uniformly converted into a compatible digital signal format through an interface converter. The facial expression acquisition device interface is directly connected to the computer's PCIe slot, the eye motion capture device interface is connected to the data acquisition terminal through an HDMI to USB adapter, the limb posture monitoring device interface is connected to the local area network via an RS485 to Ethernet module, and the anesthesia monitor and vital signs monitor interfaces are connected to the same local area network through network cables. This ensures that all device interfaces can transmit data stably, with an interface connection success rate of 100% and a data transmission latency of ≤10ms.

[0027] S102: The above-mentioned visual acquisition device continuously collects visual information corresponding to the facial muscle movement characteristics, eyelid opening and closing status, eye movement trajectory, limb movement range and posture changes during the anesthesia process, and simultaneously collects medical monitoring data corresponding to heart rate, blood pressure, respiratory rate and anesthetic drug concentration through the medical data interface, which together constitute the original anesthesia-related data. In this embodiment of the invention, a facial expression acquisition device (installed 1.5m directly in front of the head of the anesthesia bed, with the lens vertically aimed at the face) continuously acquires facial muscle movement characteristics during anesthesia, with a sampling frequency of 30 frames / second, capturing dynamic visual information such as frowning, masseter muscle contraction, and twitching of the corners of the mouth; an eye dynamic capture device (installed beside the acquisition device, at a horizontal distance of 1.2m from the face) simultaneously acquires the eyelid opening and closing state (recording the opening and closing angle 0-100°) and eye movement trajectory (establishing a two-dimensional coordinate system with the pupil center as the origin to record the movement trajectory); and a limb posture monitoring device (fixed to the upper limb)... The following data were collected at the lower limb joints: limb range of motion (joint flexion angle 0-180°) and posture changes (standing, lying flat, rolling over, etc.), with a sampling frequency of 10Hz. Anesthetic drug concentration data (unit μg / mL, accuracy 0.01μg / mL) was collected via the Ethernet interface of the anesthesia monitor. Heart rate (unit beats / minute, accuracy 1 beat / minute), blood pressure (systolic / diastolic, unit mmHg, accuracy 1 mmHg), and respiratory rate (unit breaths / minute, accuracy 1 breath / minute) were simultaneously collected via the HL7 interface of the vital signs monitor, with a medical monitoring data sampling frequency of 5Hz. All devices were simultaneously started and collected data continuously for 2 hours, generating raw anesthesia-related data including 108,000 frames of facial expression images, 43,200 frames of eye dynamic video, 7,200 sets of limb posture data, and 3,600 sets of medical monitoring data. The data was stored on a 1TB solid-state drive and named according to "device type - acquisition time".

[0028] S103: Perform image denoising, blur correction and feature point enhancement on the visual information in the original anesthesia-related data, and perform outlier removal and missing value completion on the medical monitoring data; In this embodiment of the invention, the visual information in the original anesthesia-related data is preprocessed using a dedicated image processing tool: facial expression images are filtered by Gaussian filtering (5×5 kernel size, standard deviation 1.2) to remove high-frequency noise, and Laplacian operator is used for blur correction (enhancing image edge details). Histogram equalization is used to enhance key feature points such as the eyes and corners of the mouth to improve feature recognition. Eye dynamic videos are filtered by median filtering (3×3 window size) to remove intra-frame noise, and image registration technology is used to correct blur problems caused by shooting angle deviations. Limb posture images are filtered by bilateral filtering to remove background noise, and morphological dilation algorithm is used to enhance limb contour feature points. Statistical analysis methods were used to process the medical monitoring data: outliers in heart rate (e.g., data exceeding 190 beats / minute, which falls outside the 40-180 beats / minute range), blood pressure (e.g., systolic blood pressure >200 mmHg or <80 mmHg), respiratory rate (e.g., data <10 breaths / minute or >30 breaths / minute), and anesthetic drug concentration (e.g., data >10 μg / mL) were removed using the 3σ criterion. Missing values ​​were imputed using the mean of the five adjacent data points (e.g., if the heart rate data for time 14:25:30 was missing, it was imputed using the mean of the five heart rate data points from 14:25:28 to 14:25:32, which was 68 beats / minute), ensuring the continuity and completeness of the medical monitoring data. The accuracy of the preprocessed data was ≥99.5%.

[0029] S104: Align the processed visual information with the medical monitoring data according to the timestamp, unify the data format and sampling frequency, and generate integrated multi-source fusion data of the anesthesia scene.

[0030] In this embodiment of the invention, a timestamp synchronization tool is used to extract the timestamps (accurate to milliseconds) of the processed visual information and medical monitoring data. The visual information timestamp is extracted from the image EXIF ​​information or video frame header information (e.g., facial expression image timestamp 14:00:00.001), and the medical monitoring data timestamp is extracted from the data frame start field (e.g., heart rate data timestamp 14:00:00.002). A unified timeline is established with a 10ms time interval, and the visual information and medical monitoring data are matched according to their timestamps. For the visual information (30 frames / second), the sampling frequency is adjusted to 5Hz (consistent with the medical monitoring data) using interpolation. For the medical monitoring data, a data resampling method is used to maintain the 5Hz sampling frequency. The unified data format is JSON. Visual information is stored using the field "timestamp-device type-feature parameter-value" (e.g., "14:00:00.000-facial expression-frowning degree-60%)", while medical monitoring data is stored using the field "timestamp-monitoring indicator-value-unit" (e.g., "14:00:00.000-heart rate-72 beats / minute"). All time-series aligned data are integrated using a data fusion tool to generate multi-source fused data containing timestamps, visual feature data, vital sign data, and anesthetic drug concentration data. The total data volume reaches 8GB, and each fused data record contains 12 fields, ensuring data integrity and temporal consistency, providing a unified data foundation for subsequent analysis of anesthetic response characteristics.

[0031] Furthermore, the anesthesia characterization analysis module includes the following functions: Visual feature data and medical monitoring-related data were separated from multi-source fusion data in anesthesia scenarios; In this embodiment of the invention, a data classification tool is used to identify and separate fields in multi-source fusion data (JSON format) of anesthesia scenarios. Based on the data field identifiers, visual feature data and medical monitoring-related data are distinguished: data with the "Device Type" field as "Facial Expression Acquisition Device," "Eye Dynamics Capture Device," or "Limb Posture Monitoring Device" are filtered, and the "Feature Parameters" and "Values" fields are extracted to form visual feature data, including facial muscle movements (e.g., frowning degree 60%, masseter muscle contraction strength 45%), eyelid opening and closing (e.g., opening and closing angle 30°), and eye trajectory. Information such as X-axis offset of 5mm and Y-axis offset of 3mm, and limb posture (such as joint bending angle of 90°) were extracted, resulting in 72,000 visual feature records. Data with "monitoring indicators" fields of "heart rate," "blood pressure," "respiratory rate," and "anesthetic drug concentration" were filtered, and the "value" and "unit" fields were extracted to form medical monitoring associated data, including information such as heart rate of 72 beats / minute, systolic blood pressure of 120 mmHg, diastolic blood pressure of 80 mmHg, respiratory rate of 18 breaths / minute, and anesthetic drug concentration of 2.5 μg / mL, resulting in 3,600 medical monitoring records. During the separation process, a timestamp matching tool was used to ensure that each visual feature data point corresponded to medical monitoring data with the same timestamp, establishing a "timestamp-visual feature-medical monitoring" association index, which was stored in a data classification library. The separation accuracy reached 100%, providing a classification data foundation for subsequent feature processing.

[0032] Furthermore, hierarchical feature extraction is performed on the visual feature data. First, the basic visual features corresponding to facial contours and limb contours are extracted, and then the deep dynamic features corresponding to facial micro-expressions, eye tremor frequency, and subtle limb twitches are mined to generate a multi-dimensional visual feature set. In this embodiment of the invention, visual feature data is processed using a hierarchical feature extraction tool: The first layer extracts basic visual features. Facial contours (e.g., face width 15cm, length 20cm) are extracted from facial images using an edge detection algorithm (e.g., Canny operator, lower threshold 50, upper threshold 150). Limb contours (e.g., upper limb length 60cm, lower limb bending arc 45°) are extracted from limb images, generating a basic feature set containing 20 basic contour parameters. The second layer mines deep dynamic features. Facial micro-expressions (e.g., 5mm amplitude of slight twitching at the corner of the mouth, 2 times / second frequency of eye muscle tremors) are captured using a micro-expression recognition algorithm (sampling interval 100ms). Eye tremor frequency (e.g., 3 times / second) is calculated using an eye tracking algorithm (sampling frequency 50Hz). Fine limb twitches (e.g., 2mm amplitude of lower limb muscle twitching, 3-second interval) are identified using a limb motion analysis algorithm (displacement accuracy 0.1mm). A deep dynamic feature set containing 15 dynamic parameters is generated. The basic feature set and the deep dynamic feature set are integrated by timestamp to form a multi-dimensional visual feature set. Each record contains 35 feature parameters, such as "14:00:00.000-facial contour width 15cm-frowning degree 60%-eye tremor frequency 3 times / second-lower limb twitching amplitude 2mm". A total of 72,000 multi-dimensional visual feature records are generated and stored in the feature database.

[0033] Furthermore, feature selection and dimensionality reduction are performed on the multi-dimensional visual feature set to remove redundant and interfering features and retain the core visual features that are highly correlated with the anesthetic response. In this embodiment of the invention, a feature selection tool is used to process the multi-dimensional visual feature set (35 parameters): First, variance analysis (significance level 0.05) is used to calculate the variance of each feature, eliminating redundant features with variance < 0.5 (e.g., facial contour length variance 0.3, which changes very little during anesthesia and is therefore considered redundant), retaining 28 features with variance > 0.5; Second, correlation analysis (Pearson correlation coefficient threshold ± 0.3) is used to calculate the correlation coefficient between each feature and the concentration of anesthetic drugs, eliminating interfering features with absolute correlation coefficients < 0.3 (e.g., limb contour width correlation coefficient 0.2, which has a weak correlation with anesthetic response and is therefore considered redundant). To minimize interference, 22 features with an absolute correlation coefficient ≥ 0.3 were retained. The third step involved principal component analysis (with principal components contributing ≥ 85%) to reduce dimensionality, compressing the 22 features into 10 core visual features: frowning intensity (18%), eyelid opening and closing angle (15%), nystagmus frequency (14%), mouth corner twitching amplitude (12%), upper limb twitching frequency (10%), lower limb flexion angle change rate (9%), masseter muscle contraction strength (8%), eye corner twitching amplitude (7%), head micro-swaying frequency (6%), and hand micro-movement amplitude (6%). After dimensionality reduction, the core visual feature set was compressed to 32,000 entries, reducing feature dimensionality by 68%, while retaining key information highly correlated with anesthesia response.

[0034] Furthermore, the core visual features are correlated with the corresponding medical monitoring data to construct a mapping relationship between visual features and anesthesia physiological state, and to generate visual representation data of anesthesia state. In this embodiment of the invention, a correlation characterization tool is used to map and analyze core visual features with medical monitoring data: using anesthetic drug concentration as the baseline variable, the correlation degree between each core visual feature and medical monitoring indicator is calculated. For example, when the anesthetic drug concentration increases from 2.0 μg / mL to 3.0 μg / mL, the eyelid opening angle decreases from 50° to 20°, with a correlation degree of -0.8 (negative correlation); when the heart rate decreases from 75 beats / minute to 65 beats / minute, the nystagmus frequency decreases from 4 times / second to 1 time / minute, with a correlation degree of 0.9 (positive correlation). A mapping relationship table of "core visual features - medical monitoring indicators - correlation degree" is established, such as "eyelid opening angle - anesthetic drug concentration - correlation degree -0.8" and "nystagmus frequency - heart rate - correlation degree 0.9". Based on the mapping relationship, the core visual feature values ​​are converted into anesthesia physiological state representation values. For example, eyelid opening angle > 40° corresponds to "light anesthesia", 20°-40° corresponds to "moderate anesthesia", and < 20° corresponds to "deep anesthesia"; nystagmus frequency > 3 times / second corresponds to "high heart rate", 1-3 times / second corresponds to "normal heart rate", and < 1 time / second corresponds to "low heart rate". All representation results are integrated to generate visual representation data of anesthesia state, totaling 32,000 records. Each record includes core visual feature value, corresponding physiological state representation, and associated medical monitoring value, such as "14:00:00.000-eyelid opening angle 30°-moderate anesthesia state-anesthesia drug concentration 2.5μg / mL".

[0035] Furthermore, the validity of the visual representation data of the anesthesia state is verified by eliminating invalid representation information to ensure that the visual representation data of the anesthesia state can reflect the probability distribution of response types during the anesthesia process.

[0036] In this embodiment of the invention, the visual representation data of anesthesia status is validated using an effectiveness verification tool: The first step involves data consistency verification, comparing the matching degree between the visual representation results and medical monitoring data at the same time stamp. For example, when the visual representation is "moderate anesthesia status," the corresponding anesthetic drug concentration is checked to see if it is within the range of 2.0-3.0 μg / mL. If a record represents "moderate anesthesia status" but the drug concentration is 1.5 μg / mL (below 2.0 μg / mL), it is determined to be an invalid representation. A total of 120 invalid records were detected and removed. The second step involves probability distribution analysis, statistically analyzing the proportion of visual representation data for each anesthesia status (light, moderate, and deep), and calculating the probability distribution related to the reaction type. For example, the proportion of light anesthesia status representation data is 25%, moderate anesthesia 50%, and deep anesthesia 25%, ensuring that the probability distribution conforms to the physiological state change pattern during anesthesia. The third step involves repeated verification (selecting 10% of the samples for repeated representation) to confirm the consistency of the representation results, with a repeated verification accuracy rate ≥99%. Ultimately, 31,880 valid visual representation data of anesthesia status were retained, with an invalid data removal rate of 0.37%, ensuring that the data can accurately reflect the probability distribution of response types during anesthesia and provide reliable representation data for subsequent analysis of anesthesia response characteristics.

[0037] Furthermore, the anesthesia response analysis module includes the following functions: Clinical anesthesia response evaluation indicators were extracted from clinical anesthesia diagnosis and treatment guidelines, covering key dimensions corresponding to sedation depth level, analgesic effect assessment, signs of adverse reactions, and neuromuscular blockade degree. In this embodiment of the invention, key evaluation indicators are extracted from the clinical anesthesia response evaluation guidelines using an indicator extraction tool, clarifying four core dimensions and specific indicator parameters: The sedation depth level dimension is divided into three levels: light sedation (responding to auditory stimulation), moderate sedation (responding to tactile stimulation), and deep sedation (responding only to painful stimuli), each level corresponding to a specific response threshold; the analgesic effect assessment dimension includes no pain response (no limb movement upon stimulation), mild pain response (slight limb contraction upon stimulation), and moderate pain response (significant limb withdrawal upon stimulation). The evaluation criteria include four aspects: severe pain response (violent struggling of the limbs upon stimulation); adverse reaction signs, covering three categories and judgment thresholds: respiratory depression (respiratory rate <10 breaths / minute), blood pressure fluctuation (systolic blood pressure change >20 mmHg), and muscle rigidity (limb muscle stiffness >70%); and neuromuscular blockade degree, divided into four levels: complete blockade (no limb movement), deep blockade (only slight finger movement), moderate blockade (small limb movement), and mild blockade (large limb movement), with corresponding movement amplitude judgment criteria. The extracted evaluation indicators were organized in the format of "dimensional-indicator name-judgment criterion-threshold range" to form a clinical anesthesia response evaluation index table, containing 14 specific indicators, each with clearly defined quantitative judgment criteria, providing a standard reference for subsequent visual feature matching.

[0038] Furthermore, the core visual features in the visual representation data of the anesthesia state are matched one by one with each indicator in the clinical anesthesia response evaluation index to establish the correspondence between visual features and evaluation indicators. In this embodiment of the invention, an index matching tool is used to match 10 core visual features in the visual representation data of anesthesia status with 14 indicators in the clinical anesthesia response evaluation index table one by one: For the sedation depth level dimension, the eyelid opening angle (e.g., >40° corresponds to light sedation, 20°-40° corresponds to moderate sedation, <20° corresponds to deep sedation) and nystagmus frequency (e.g., >3 times / second corresponds to light sedation, 1-3 times / second corresponds to moderate sedation, <1 time / second corresponds to deep sedation) are matched with the sedation depth level; For the analgesia effect evaluation dimension, the degree of frowning (e.g., >70% corresponds to severe pain, 50%-70% corresponds to moderate pain, 30%-50% corresponds to mild pain, <30% corresponds to no pain) and limb twitching amplitude (e.g., >5mm corresponds to severe pain, 3-5mm corresponds to moderate pain, 1-3mm corresponds to mild pain, <1mm corresponds to no pain) are matched with the sedation depth level. Pain efficacy indicators were matched; in the dimension of adverse reaction signs, the amplitude of chest movement associated with respiratory rate (converted through limb posture characteristics, e.g., <3cm corresponds to respiratory depression), the degree of facial vasodilation associated with blood pressure (judged through facial features, e.g., dilation area >20% corresponds to blood pressure fluctuation), and the limb stiffness associated with muscle rigidity (judged through the rate of change of limb flexion angle, e.g., <5% / minute corresponds to muscle rigidity) were matched with adverse reaction signs; in the dimension of neuromuscular blockade, the frequency of upper limb twitching (e.g., >2 times / second corresponds to mild blockade, 1-2 times / second corresponds to moderate blockade, 0.5-1 times / second corresponds to deep blockade, and <0.5 times / second corresponds to complete blockade) and the amplitude of fine hand movements (e.g., >10mm corresponds to mild blockade, 5-10mm corresponds to moderate blockade, 1-5mm corresponds to deep blockade, and <1mm corresponds to complete blockade) were matched with the degree of neuromuscular blockade. A correspondence table of "core visual features - clinical evaluation indicators - matching criteria" was established, resulting in 22 feature-indicator matching pairs with a matching coverage rate of 100%.

[0039] Furthermore, the correlation strength between core visual features and various evaluation indicators is quantified through Pearson correlation coefficient analysis and mutual information calculation, and feature-indicator combinations with correlation strength higher than a preset threshold are selected. In this embodiment of the invention, 22 feature-index matching pairs are quantitatively analyzed using a correlation strength analysis tool: First, linear correlation strength is calculated using Pearson correlation coefficient analysis, with a threshold of 0.6 for the absolute value of the correlation coefficient. For example, the correlation coefficient between eyelid opening angle and sedation depth level is -0.82 (negative correlation, absolute value > 0.6), the correlation coefficient between frowning degree and analgesic effect assessment is 0.78 (positive correlation, absolute value > 0.6), and the correlation coefficient between upper limb twitching frequency and neuromuscular blockade degree is -0.75 (negative correlation, absolute value > 0.6). Matching pairs meeting these correlation strength criteria are retained. Second, nonlinear correlation strength is analyzed using mutual information calculation, with a threshold of 0.5 bits for the mutual information value. For example, the mutual information value between facial vasodilation degree and blood pressure fluctuation is 0.62 bits (> 0.5 bits), and the mutual information value between chest movement amplitude and respiratory inhibition is 0.58 bits (> 0.5 bits). Matching pairs meeting these nonlinear correlation criteria are retained. The intersection of the two analysis results was used to select 18 feature-indicator combinations with a correlation strength higher than the preset threshold. Four combinations that did not meet the threshold were removed (such as the frequency of head micro-swing and the degree of neuromuscular blockade, with a correlation coefficient of 0.45 and a mutual information value of 0.38). The average correlation strength of the selected combinations reached 0.72, ensuring a strong correlation between features and indicators and laying the foundation for subsequent multi-dimensional analysis.

[0040] Furthermore, based on the selected feature-indicator combination, a multi-dimensional correlation analysis is performed to explore the dynamic pattern between changes in visual features and fluctuations in anesthesia response indicators, calculate the probability of occurrence of different reaction types, and generate preliminary analysis results of anesthesia response including the probability distribution of reaction types.

[0041] In this embodiment of the invention, a multi-dimensional correlation analysis tool was used to dynamically mine the patterns of 18 feature-index combinations: taking time series as the axis, the correspondence between changes in core visual features and fluctuations in evaluation indicators was tracked. For example, when the eyelid opening angle decreased from 35° to 15° (within 10 minutes), the sedation depth level changed from moderate sedation to deep sedation, and the concentration of anesthetic drugs increased from 2.5 μg / mL to 3.5 μg / mL, forming a dynamic pattern of "feature change - index fluctuation - drug concentration correlation"; when the degree of frowning increased from 20% to 60% (within 5 minutes), the analgesic effect assessment changed from no pain to moderate pain, and the heart rate increased from 65 beats / minute to 85 beats / minute, summarizing the correspondence pattern of "pain response visual features - heart rate fluctuation". Based on dynamic patterns, probability calculation tools were used to statistically analyze the frequency of different reaction types. For example, within a 2-hour monitoring period, moderate sedation occurred 120 times, moderate pain 30 times, respiratory depression 5 times, and moderate neuromuscular blockade 90 times. Combined with a total of 360 monitoring sessions, the probability of each reaction type was calculated (moderate sedation 33.3%, moderate pain 8.3%, respiratory depression 1.4%, moderate blockade 25%). Integrating dynamic patterns and probability data, preliminary analysis results of anesthetic reactions were generated, including descriptions of 12 dynamic patterns, probability distribution tables for 8 reaction types, and corresponding characteristic change maps. For example, "When the eyelid opening angle is 20°-40°, the probability of moderate sedation is 33.3%, corresponding to an anesthetic drug concentration of 2.0-3.0 μg / mL." This ensures that the analysis results are quantifiable and traceable, providing a basis for further in-depth analysis of anesthetic reaction characteristics.

[0042] Furthermore, the anesthesia characteristic analysis module includes the following functions: We collected a large amount of clinical anesthesia case data, covering anesthesia response data, synchronous visual feature data and related probability statistics corresponding to different anesthesia methods, different patient physical conditions and different surgical types, and constructed an anesthesia response characteristic sample library. In this embodiment of the invention, a large number of case data were collected from anesthesia scenarios in three medical institutions using clinical data collection tools. These cases covered three anesthesia methods (general anesthesia, local anesthesia, and spinal anesthesia), four patient body types (weight 40-60kg, 60-80kg, 80-100kg, and >100kg), and five surgical types (abdominal surgery, orthopedic surgery, ophthalmic surgery, cardiothoracic surgery, and obstetrics and gynecology surgery), totaling 1000 complete anesthesia cases. Each case included anesthesia response data (real-time recording of sedation depth, analgesic effect, adverse reactions, and neuromuscular blockade degree, with a sampling interval of 1 minute), synchronous visual feature data (temporal data of 10 core visual features, with a sampling frequency of 5Hz), and relevant probability statistics (the probability of occurrence of each reaction type at different time points). For example, a case of abdominal surgery under general anesthesia (patient weight 70kg) included 60 sedation depth records (10 for light sedation, 35 for moderate sedation, and 15 for deep sedation) within 2 hours, 3000 visual feature data (eyelid opening angle 20-40° accounting for 60%), and 24 probability statistics (moderate sedation probability 58.3%). All case data were categorized and organized according to "anesthesia method - patient constitution - surgery type - data category." Data cleaning tools were used to remove cases with a missing rate >5% (a total of 30 cases were removed), ultimately constructing an anesthesia response characteristic sample library containing 970 valid cases, with a total data volume of 120GB. Each data entry included a timestamp, case number, and data source identifier, providing sufficient sample support for model construction.

[0043] Furthermore, based on the preliminary analysis results of the anesthetic response, the input feature dimensions of the model were determined. The core visual features, related medical monitoring data and feature association probabilities were used as input variables, and the actual values ​​and occurrence probabilities of clinical anesthetic response evaluation indicators were used as output variables to construct an initial dynamic model of anesthetic response characteristics. In this embodiment of the invention, based on the preliminary analysis results of the anesthetic response (18 sets of feature-indicator combinations, probability distributions of 8 response types), the model input feature dimensions are determined to be 15 dimensions: 10 core visual features (eyelid opening and closing angle, nystagmus frequency, etc.), 3 related medical monitoring data (anesthetic drug concentration, heart rate, respiratory rate), and 2 feature association probabilities (the association probability between visual features and sedation depth, and the association probability with analgesic effect). The input variable values ​​have a clear range (e.g., eyelid opening and closing angle 0-100°, anesthetic drug concentration 0-10μg / mL, association probability 0-100%). The model output variables are set to 8 dimensions: 4 actual values ​​of clinical anesthetic response evaluation indicators (sedation depth level 1-3, analgesic effect level 1-4, adverse reaction classification 0-1, neuromuscular blockade degree level 1-4), and 4 indicator occurrence probabilities (probability of each sedation depth level, probability of each analgesic effect level, etc.). The quantitative standards of the output variables are consistent with the clinical evaluation indicators (e.g., sedation depth level 3 corresponds to deep sedation). A dynamic model of initial anesthesia response characteristics was built using a model building tool. The model structure was a 3-layer neural network (15 neurons in the input layer, 32 neurons in the hidden layer, and 8 neurons in the output layer). The activation functions used were the Sigmoid function (output layer) and the ReLU function (hidden layer). The initial learning rate was set to 0.01, and the number of iterations was set to 1000. The initial values ​​of the model parameters (weights and biases) were generated through random initialization to ensure that the initial model could receive input features and output corresponding evaluation index results, laying the foundation for subsequent training.

[0044] Furthermore, the gradient descent algorithm combined with cross-validation is used to train the initial dynamic model of anesthesia response characteristics, and the model parameters are continuously adjusted to minimize the prediction error. The preliminary analysis results of anesthesia response are input into the trained dynamic model of anesthesia response characteristics, and the deviation is corrected through dynamic iteration calculation of the dynamic model of anesthesia response characteristics to generate anesthesia response characteristic analysis data. In this embodiment of the invention, a dynamic model of initial anesthesia response characteristics is trained using a model training tool: First, the sample database is divided into a training set (679 cases) and a validation set (291 cases) of 970 cases in a 7:3 ratio. The training set is used for parameter adjustment, and the validation set is used for error assessment. Second, the gradient descent algorithm is used to optimize the model parameters, with mean squared error (MSE) as the loss function. The loss value is calculated in each iteration, and the weights and biases are adjusted along the gradient direction. The learning rate is dynamically adjusted according to the number of iterations (reduced by 50% every 200 iterations, down to a minimum of 0.001). Third, a 5-fold cross-validation method is used, dividing the training set into 5 subsets. Four subsets are used for training and one subset for validation in each iteration, and this process is repeated 5 times. The average loss value is then used as the model performance indicator to ensure that the model does not overfit. When the model has been trained for 500 iterations, the loss value has decreased from the initial 0.8 to 0.15, and the validation set accuracy reaches 85%. The preliminary analysis results of the anesthetic response (such as an eyelid opening angle of 30°, anesthetic drug concentration of 2.5 μg / mL, and a moderate sedation probability of 33.3%) are input into the trained model. The model corrects the deviation through dynamic iterative calculation (updating the input feature weights in each iteration, and outputting stable results after 50 iterations). For example, the initial prediction of a moderate sedation probability of 30% is corrected to 32.8% after iteration, reducing the error from the actual clinical probability of 33.3% to 0.5%. Finally, the anesthetic response characteristic analysis data containing 8 output indicators is generated, with each data point accompanied by a prediction error value (all <1%) to ensure the accuracy of the analysis data.

[0045] Furthermore, the anesthetic response characteristic analysis data output by the model are clinically validated, and the dynamic model of anesthetic response characteristics is continuously optimized by combining the evaluation feedback of anesthesiologists.

[0046] In this embodiment of the invention, the anesthetic response characteristics analysis data output by the model are validated using clinical validation tools: 50 new anesthesia cases (covering different anesthesia methods, physical conditions, and surgical types) are selected, and the real-time input features of the cases (visual features, medical monitoring data) are input into the model to obtain the model's output analysis data (such as sedation depth level and probability of each response), while the anesthesiologist's real-time assessment results are recorded (judged according to clinical evaluation index standards). For example, in a case of orthopedic surgery under local anesthesia (patient weight 85kg), the model output "moderate sedation level, probability 35%", while the anesthesiologist's assessment result was "moderate sedation, actual probability 36%", with an error of 0.8%; in another case of cardiothoracic surgery under general anesthesia, the model output "respiratory depression probability 1.2%", while the physician's assessment of the actual probability was 1.3%, with an error of 0.1%. The validation results of the 50 cases show that the model's prediction accuracy reached 88%, with errors all <2%. Anesthesiologist feedback was collected (e.g., "The model's predictive sensitivity for deep sedation needs improvement," "The prediction bias for pain response probability is slightly large"). Based on this feedback, model parameters were adjusted: the weight of deep sedation-related features (e.g., nystagmus frequency) was increased (from 0.2 to 0.3), and the weight of the loss function for analgesic effect prediction was optimized (from 0.15 to 0.2). After adjustment, 50 cases were re-validated, improving the model accuracy to 92% and reducing the pain response probability prediction error from 1.5% to 0.8%, achieving continuous model optimization. Subsequently, the above validation and optimization process was repeated for every 100 new cases collected to ensure that model performance continuously improves with the accumulation of clinical data and always meets the needs of anesthetic response characteristic analysis.

[0047] Furthermore, the evaluation feedback continuously optimizes the dynamic model of anesthetic response characteristics, including: Visual feature sequence data from multi-source fusion data of anesthesia scenarios are acquired and divided into time windows. The visual feature change trend, mutation point information and feature occurrence probability are extracted within each time window. In this embodiment of the invention, visual feature sequence data is separated from multi-source fusion data of anesthesia scenarios using a data extraction tool. This data contains time-series records of 10 core visual features (sampling frequency 5Hz, 36,000 data points in 2 hours), such as eyelid opening and closing angles (35° at 14:00:00.000, 34° at 14:00:00.200, etc.) and nystagmus frequency (2 times / second at 14:00:00.000, 2 times / second at 14:00:00.200, etc.). A time windowing tool is used to divide the 36,000 data points into 720 time windows, each with a fixed duration of 10 seconds (each window containing 50 data points). For the data within each window, a trend analysis tool was used to extract the visual feature change trend: the linear regression slope (e.g., if the eyelid opening angle decreases from 35° to 30° within the 14:00:00-14:00:10 window, the slope is -0.5° / second, indicating a downward trend) and standard deviation (e.g., if the nystagmus frequency has a standard deviation of 0.2 times / second within the window, indicating a stable trend) were calculated); abrupt change point detection tools (with a threshold set at 20% of the feature mean) were used to identify abrupt change points. For example, within the 14:02:30-14:02:40 window, the frowning degree suddenly increased from 20% to 60% (exceeding the 20% threshold of the mean of 30%), which was identified as an abrupt change point, and the abrupt change time was recorded as 14:02:35, and the difference before and after the abrupt change was 40%; a probability statistics tool was used to calculate the probability of feature occurrence within each window. For example, if the proportion of records with an eyelid opening angle of 20-30° within the window is 80%, then the probability of occurrence of that angle range is 80%. Ultimately, each time window generates a feature time series report containing "change trend - mutation point information - feature occurrence probability". A total of 720 reports are generated for 720 windows, such as "Window 1 (14:00:00-14:00:10) - eyelid opening and closing angle decreasing trend (-0.5° / second) - no mutation point - 20-30° probability 80%", which provides windowed feature data for subsequent time series correlation analysis.

[0048] Furthermore, by combining the medical monitoring data collected simultaneously, the temporal correlation between visual feature mutation points and vital sign fluctuations and anesthetic drug infusion nodes is analyzed, the probability of occurrence of related events is calculated, and feature-physiology-drug correlation data containing temporal correlation probabilities is generated. In this embodiment of the invention, synchronous medical monitoring data (sampling frequency 5Hz, including heart rate, blood pressure, respiratory rate, and anesthetic drug concentration) is extracted from multi-source fusion data. For example, at 14:02:35, the heart rate is 75 beats / minute and the anesthetic drug concentration increases from 2.0 μg / mL to 3.0 μg / mL (drug infusion node). A time-series correlation analysis tool is used to match visual feature mutation points with medical monitoring data: taking the 14:02:35 frowning degree mutation point (from 20% to 60%) as the benchmark, medical monitoring data for 5 seconds before and after is extracted. It is found that the anesthetic drug concentration begins to rise 1 second before the mutation (14:02:34), and at the same time as the mutation (14:02:35), the heart rate rises to 85 beats / minute (a fluctuation of 13.3% from the average of 75 beats / minute in the previous 5 seconds). It is determined that there is a time-series correlation among the three. The frequency of such associated events was statistically analyzed using probability calculation tools: Within 720 time windows, 120 cases occurred simultaneously with visual feature mutations, drug infusion points, and vital sign fluctuations, with a calculated probability of 120 / 720 ≈ 16.7%. The association analysis results from all windows were integrated to generate feature-physiological-drug association data containing "mutation point time - associated medical data - association probability," such as "14:02:35 - frowning degree mutation - drug concentration rises to 3.0 μg / mL - heart rate 85 beats / minute - association probability 16.7%." A total of 120 association data points were generated, each with a timestamp and window number to ensure traceable temporal correlation.

[0049] Furthermore, an anesthesia response time-series change map is constructed based on the associated data. The map includes visual feature trajectories, physiological index curves, drug infusion curves, correlation markers between the three and related probability information. Dynamic patterns in the anesthesia response time-series change map are then mined, and the mined dynamic patterns and corresponding probability information are integrated into the dynamic model of anesthesia response characteristics to continuously optimize the anesthesia response characteristic analysis data.

[0050] In this embodiment of the invention, an anesthesia response time-series change map is constructed based on associated data using a map construction tool: the horizontal axis of the map is the time axis (divided into 10-second windows, with a total of 720 scales), and the vertical axis is divided into three levels—visual feature layer (plotting the change trajectory of 10 core visual features, such as the eyelid opening and closing angle curve decreasing from 35° to 20°), physiological indicator layer (plotting curves of heart rate, blood pressure, etc., such as the heart rate increasing from 75 beats / minute to 85 beats / minute), and drug infusion layer (plotting the anesthetic drug concentration curve, marking infusion nodes such as the concentration increase at 14:02:34); association markers are added at the time scales corresponding to associated events (such as marking "frowning mutation - drug infusion - heart rate fluctuation" association with a red dot at 14:02:35, with an association probability of 16.7%). Dynamic patterns in the anesthetic response profile were analyzed using pattern mining tools. It was found that when the rate of increase in anesthetic drug concentration was >0.5 μg / mL / second, 80% of the windows showed a decrease in eyelid opening angle (slope <-0.3° / second) and heart rate fluctuations >10 beats / minute. This led to the conclusion that the dynamic pattern was "rapid increase in drug concentration → decrease in visual features + heart rate fluctuations." The probability of this pattern occurring in all associated events was 80%. The 15 discovered dynamic patterns (including probability information) were integrated into the dynamic model of anesthetic response characteristics using a model update tool. The correlation weights between drug concentration and visual features / heart rate were adjusted (from 0.25 to 0.35), and a probability calculation module corresponding to the patterns was added (e.g., when the rate of increase in drug concentration is >0.5 μg / mL / second, the probability prediction of the corresponding features and heart rate is automatically triggered). The updated model reduced the prediction error for moderate sedation probability from 0.5% to 0.3%, and improved the accuracy of the anesthetic response characteristic analysis data by 3 percentage points, achieving continuous model optimization.

[0051] Furthermore, the generation of feature-physiological-drug association data containing temporal correlation probabilities includes: The visual feature sequence data in the multi-source fusion data of the anesthesia scene is preprocessed to remove abnormal sequence segments caused by equipment shaking and light changes. The missing sequence data is filled in by interpolation to ensure the continuity, integrity and accuracy of probability calculation of the sequence data. In this embodiment of the invention, a sequence preprocessing tool is used to process the visual feature sequence data (10 core features, sampling frequency 5Hz, 36,000 records in 2 hours) in the multi-source fusion data of the anesthesia scene: The first step is to remove abnormal sequence segments. The deviation of feature values ​​is calculated by sliding window detection (window size 5 data points). The deviation threshold is set to 3 times the standard deviation. For example, in the eyelid opening and closing angle sequence, a certain segment (14:15:20-14:15:22) suddenly increases from 30° to 80° (deviation 4.2 times the standard deviation), which is determined to be an anomaly caused by equipment shaking. The 10 data points in this 2-second period are directly removed. For blurred sequences caused by changes in light (such as irregular fluctuations in facial feature values ​​from 14:30:00-14:30:05), feature consistency verification (adjacent data change amplitude > 20% is considered abnormal) is used to identify and remove the 25 data points in this 5-second period. A total of 180 abnormal data points are removed. The second step is to complete the missing sequence data. For missing segments formed after anomaly removal (e.g., missing 14:15:20-14:15:22), linear interpolation is used to calculate the missing value. The missing value is then filled uniformly at time intervals, using the mean of the five data points before and after the missing segment as a benchmark. For example, if the missing data point before the segment is 30° (14:15:19) and the missing data point after is 28° (14:15:23), then the missing values ​​are filled as 29.5° (14:15:20), 29° (14:15:21), and 28.5° (14:15:22). For single-point missing data (e.g., missing eyelid opening angle data at 14:20:10), the mean of two adjacent points is used for completion (14:20:09 is 29°, 14:20:11 is 27°, so 28° is filled). After preprocessing, the sequence data achieves 100% continuity and 99.5% completeness, providing an accurate data foundation for subsequent probability calculations.

[0052] Furthermore, time windows are divided according to preset time intervals, and the length of each time window is dynamically adjusted according to different stages of the anesthesia process. The length of the window during the anesthesia induction and recovery periods is shortened, while the length of the window during the maintenance period is appropriately extended. In this embodiment of the invention, a dynamic windowing tool is used to divide time windows according to a preset time interval. First, the duration of the three stages of the anesthesia process is determined (30 minutes for induction, 90 minutes for maintenance, and 30 minutes for recovery). Then, the window length is adjusted according to the characteristics of each stage: During the anesthesia induction period (14:00:00-14:30:00), visual characteristics change frequently, so the window length is shortened to 5 seconds, with each window containing 25 data points (5Hz × 5 seconds), for a total of 360 windows; During the anesthesia maintenance period (14:30:00-16:00:00), visual characteristics are relatively stable, so the window length is extended to 15 seconds, with each window containing 75 data points, for a total of 360 windows; During the anesthesia recovery period (16:00:00-16:30:00), visual characteristics change frequently again, so the window length is restored to 5 seconds, for a total of 360 windows. When dividing the windows, a timestamp calibration tool was used to ensure seamless connection between windows of each stage. For example, the last window of the induction period (14:29:55-14:30:00) and the first window of the maintenance period (14:30:00-14:30:15) were precisely connected at 14:30:00, with no time overlap or gaps. At the same time, each window was labeled with a stage identifier (e.g., "Induction Period - Window 1 - 14:00:00-14:00:05" "Maintenance Period - Window 1 - 14:30:00-14:30:15"), generating a total of 1080 dynamic time windows. This ensured that the window division matched the characteristics of the anesthesia stages and improved the targeting of subsequent feature analysis.

[0053] Furthermore, statistical analysis is performed on the visual feature data within each time window to calculate the feature mean, variance, rate of change, peak value, and probability of feature occurrence, extract the feature change trend, and identify visual feature mutation points within the time window through anomaly detection algorithms, recording the time location, feature change amplitude, and mutation probability of visual feature mutation points. In this embodiment of the invention, statistical analysis tools are used to calculate the visual feature data within each time window: taking the induction period window (14:05:00-14:05:05, 25 eyelid opening and closing angle data) as an example, the calculated feature mean is 28° (summed from all data and divided by 25), and the variance is 2.5°. 2The average of the squared differences between each data point and the mean, the rate of change is -0.4° / second (the difference between the first and last data points within the window divided by the time length of 5 seconds), and the peak value is 32° (the maximum value within the window). Statistically, 92% of the records within this window show an eyelid opening angle of 25-30°, indicating a 92% probability of this characteristic occurring. The trend was extracted using a trend fitting tool (linear regression), and the slope of the fitted line for this window's data was -0.4° / second, classifying it as a "slow downward trend." Simultaneously, an anomaly detection algorithm (Isolation Forest algorithm, anomaly score threshold of 0.8) was used to identify mutation points: In the maintenance window (14:40:00-14:40:15, 75 frowning degree data), at a certain moment (14:40:10), the frowning degree suddenly increased from 15% to 55%, with an anomaly score of 0.85 (>0.8), which was determined to be a mutation point. The time position was recorded as 14:40:10, and the feature change range was 40% (55%-15%). It was found that a similar mutation point occurred once in this window. Combined with the total number of maintenance windows of 360, the probability of mutation occurrence was calculated to be 1 / 360≈0.28%. Each window generates a feature analysis report containing "statistical parameters - trend of change - mutation point information - mutation probability". A total of 1080 reports are generated from 1080 windows, such as "maintenance period - window 20 - 14:40:00 - 14:40:15 - average frowning degree 20% - downward trend - mutation point 14:40:10 - mutation probability 0.28%".

[0054] Furthermore, the visual feature change trends, mutation point information, and related probability parameters within each time window are correlated one-to-one with the heart rate, blood pressure, blood oxygen saturation, and anesthetic drug concentration collected at the same time, and a correlation index in the time dimension is established. In this embodiment of the invention, a correlation indexing tool is used to establish a correspondence between the feature analysis data of each time window and the medical monitoring data of the same period: First, the medical monitoring data within each window time interval (sampling frequency 5Hz, 25 data points for the induction period window and 75 data points for the maintenance period window) are extracted. For example, the medical monitoring data corresponding to the induction period window (14:05:00-14:05:05) are heart rate 72-75 beats / minute, blood pressure 120 / 80-122 / 82 mmHg, blood oxygen saturation 98%-99%, and anesthetic drug concentration 2.2-2.3 μg / mL. Then, using the unique identifier of the window (e.g., "induction period - window 30 - 14:05:00-14:05:05") as the index key, the visual feature change trend of the window (slow decrease in eyelid opening angle), mutation point information (no mutation point), and related probability parameters (25-30° probability 92%) are bound to the above medical monitoring data and stored in the correlation database. When creating the index, precise matching using timestamps ensures that the time intervals of visual feature data and medical monitoring data are completely consistent, without any misalignment. For example, the index record for "Maintenance Period - Window 20 - 14:40:00 - 14:40:15" includes "Inflection point of frown intensity 14:40:10 - Heart rate 78 beats / minute (at 14:40:10) - Anesthetic drug concentration 2.8 μg / mL (at 14:40:10)," generating a total of 1080 associated index records, achieving precise correlation between visual features and medical monitoring data in the time dimension.

[0055] Furthermore, by calculating the time difference between visual feature mutation points, physiological indicator fluctuation nodes, and drug infusion nodes, the causal relationship among the three is analyzed, generating feature-physiology-drug association data that includes temporal correlation strength, response delay time, and probability of occurrence of associated events.

[0056] In this embodiment of the invention, the time difference between visual feature mutation points and physiological index fluctuation points and drug infusion points is calculated using causal analysis tools: taking the frowning degree mutation point (14:40:10) during the maintenance window (14:40:00-14:40:15) as an example, the physiological index fluctuation points during the same period are extracted as follows: the time when the heart rate rises from 75 beats / minute to 78 beats / minute is 14:40:10 (time difference 0 seconds), and the time when the blood pressure rises from 125 / 85 mmHg to 130 / 88 mmHg is 14:40:11 (time difference 1 second); the drug infusion point is extracted as follows: the starting time when the anesthetic drug concentration rises from 2.5 μg / mL to 2.8 μg / mL is 14:40:08 (time difference 2 seconds). The correlation strength was calculated using a temporal correlation strength algorithm (based on mutual information values, ranging from 0 to 1): the mutual information value between the mutation point and the drug infusion node was 0.85 (strong correlation), the mutual information value with the heart rate fluctuation node was 0.9 (very strong correlation), and the mutual information value with the blood pressure fluctuation node was 0.7 (relatively strong correlation). Response delay times were recorded—the mutation point occurred 2 seconds after drug infusion, synchronized with heart rate fluctuation (0-second delay), and delayed by 1 second with blood pressure fluctuation. The frequency of this type of correlation event across all windows was statistically analyzed: 80 events similar to "drug infusion → visual mutation → physiological fluctuation" occurred across 1080 windows, resulting in a correlation event probability of 80 / 1080 ≈ 7.4%. By integrating all analysis results, feature-physiology-drug association data is generated, such as "14:40:10 - sudden change in frowning degree - drug infusion delay of 2 seconds (intensity 0.85) - heart rate fluctuation delay of 0 seconds (intensity 0.9) - blood pressure fluctuation delay of 1 second (intensity 0.7) - occurrence probability of 7.4%", a total of 80 association data are generated. Each data is accompanied by time difference, association strength and probability information to clarify the causal relationship among the three.

[0057] Furthermore, the dynamic patterns include the time difference between changes in visual features and fluctuations in physiological indicators, the response period of visual features after drug adjustment, and the probability of occurrence of various patterns.

[0058] In this embodiment of the invention, dynamic patterns are extracted from the early feature-physiology-drug correlation data to identify three core patterns and quantitative parameters: The first type is the time difference between changes in visual features and fluctuations in physiological indicators. For example, the time difference between the decrease in eyelid opening angle (from 30° to 25°) and the decrease in heart rate (from 75 beats / minute to 70 beats / minute) is 1.5 seconds. This type of pattern occurred 210 times in 1080 windows, with a probability of 210 / 1080≈19.4%. The second type is the response period of visual features after drug adjustment, such as the concentration of anesthetic drugs... After the concentration of the drug increased from 2.0 μg / mL to 2.5 μg / mL (drug adjustment point 14:20:00), the response period for the frowning intensity to decrease from 40% to 20% was 8 seconds (14:20:00-14:20:08). This type of pattern occurred 180 times, with a probability of 180 / 1080≈16.7%. The third type is a composite pattern (containing features of the first two types), such as the combination of visual feature changes leading physiological index fluctuations after drug adjustment, with a response period of 6 seconds and a lead time difference of 1 second, occurring 85 times, with a probability of 85 / 1080≈7.9%. All patterns were recorded with specific values ​​(time difference accurate to 0.1 seconds, response period accurate to 1 second) and corresponding probabilities, forming a dynamic pattern table containing 32 independent patterns, providing a pattern basis for subsequent atlas construction.

[0059] Furthermore, the process of mining the dynamic patterns includes: With the time axis as the horizontal axis, a visual feature trajectory layer, a physiological index curve layer, and a drug infusion curve layer are constructed respectively. The visual feature trajectory layer uses the numerical changes and occurrence probabilities of core visual features as the vertical axis, the physiological index curve layer uses the values ​​of various vital signs parameters and their normal probability ranges as the vertical axis, and the drug infusion curve layer uses the infusion rate, cumulative dose, and dose adjustment probability of anesthetic drugs as the vertical axis. In this embodiment of the invention, a layer-building tool is used to construct three parallel layers with the time axis (14:00:00-16:30:00, 10-second intervals, 900 scales in total) as the horizontal axis: the visual feature trajectory layer is divided into two parts on the vertical axis, with the core visual feature values ​​on the left (eyelid opening angle 0-100°, frowning degree 0-100%, etc.) and the feature occurrence probability on the right (0-100%), drawing eyelid opening angle curves (35° for 14:00:00, 28° for 14:05:00, etc.) and corresponding probability curves (92% probability for 14:00:00-14:00:10, 88% probability for 14:05:00-14:05:10, etc.); the physiological index curve layer has vital sign parameter values ​​on the left side of the vertical axis (heart rate 40-180 beats / minute, blood pressure 60 / 40-200 / 100 mmHg). The graph shows the following parameters: blood pressure 20 mmHg, blood oxygen saturation 90%-100%, normal probability range (0-100%) on the right, heart rate curve (75 bpm at 14:00:00, 72 bpm at 14:05:00, etc.) and normal range threshold line (heart rate 60-100 bpm, marked as normal probability 95%). The vertical axis of the drug infusion curve layer shows the anesthetic drug parameters (infusion rate 0-10 mL / h, cumulative dose 0-500 mL) on the left and the dose adjustment probability (0-100%) on the right. The graph shows the infusion rate curve (5 mL / h at 14:00:00, 8 mL / h at 14:20:00, etc.) and cumulative dose curve (0 mL at 14:00:00, 150 mL at 14:30:00, etc.) and marked the adjustment probability (adjustment node probability 16.7% at 14:20:00). The horizontal time scales of the three layers are perfectly aligned, and the vertical value range is set according to the parameter characteristics (to ensure that the curve does not overflow). The layer thickness is 5cm and the spacing is 2cm, forming a basic atlas framework.

[0060] Furthermore, key nodes are marked in each layer. The visual feature trajectory layer marks abrupt change points, feature peak points and corresponding probabilities. The physiological index curve layer marks the normal range threshold line, abnormal fluctuation points and the probability of abnormal occurrence. The drug infusion curve layer marks drug type switching points, dosage adjustment points and adjustment probabilities. In this embodiment of the invention, key nodes are marked on each layer using a node marking tool: In the visual feature trajectory layer, abrupt change points are marked with red triangles (e.g., the abrupt change point of frowning intensity at 14:40:10, with a side note "amplitude 40%, probability 0.28%)", and feature peak points are marked with blue circles (e.g., the peak value of eyelid opening angle at 14:15:30, 38°, with a side note "probability 90%)", for a total of 120 abrupt change points and 180 peak points; In the physiological index curve layer, the normal range threshold lines (heart rate 60-100 beats / minute, blood pressure 9) are marked with black dashed lines. (0 / 60-140 / 90 mmHg) Abnormal fluctuation points are marked with orange squares (e.g., 14:35:20 heart rate 105 bpm, with a note "abnormal probability 8%)", totaling 95 abnormal fluctuation points. In the drug infusion curve layer, drug type switching points are marked with green diamonds (e.g., 14:00:00 propofol initiation, with a note "switching probability 100%)", and dose adjustment points are marked with purple pentagrams (e.g., 14:20:00 concentration increased to 2.5 μg / mL, with a note "adjustment probability 16.7%)", totaling 3 switching points and 45 adjustment points. All nodes are labeled with specific values ​​(e.g., mutation amplitude, abnormal value deviation) and corresponding probabilities. The node size is set according to the probability ratio (the higher the probability, the larger the node) to ensure that key nodes are clearly identifiable.

[0061] Furthermore, by using the correlation index, the key nodes of time synchronization in the three layers are connected and marked to clarify the correspondence and correlation probability between changes in visual features, fluctuations in physiological indicators, and drug adjustments, thus forming a complete time sequence change map of anesthesia response. In this embodiment of the invention, by using a correlation connection tool and a previously established correlation index (unique identifier for the window), the key nodes for time synchronization in the three layers are connected with colored lines: the 14:20:00 drug dosage adjustment point (drug layer) is connected with the corresponding visual feature response node (14:20:00-14:20:08 decrease in frowning degree, visual layer) and the physiological indicator fluctuation node (14:20:08 decrease in heart rate, physiological layer) with a solid red line, marked "response cycle 8 seconds, correlation probability 16.7%"; the 14:40:10 visual feature mutation point (visual layer) is connected with the corresponding heart rate fluctuation node (14:40:10, physiological layer) and the previous drug adjustment node (14:40:08, drug layer) with a dashed blue line, marked "lead time difference 1 second, correlation probability 7.9%". A total of 320 correlation lines were drawn, each corresponding to a dynamic pattern. The pattern number, correlation strength (mutual information value 0.7-0.9), and probability of occurrence were labeled to form a complete time-series map of anesthesia response changes. The colors of the lines in the map are distinguished according to the correlation strength (red > 0.85, blue 0.75-0.85, black < 0.75) to ensure that the correlation relationships are presented intuitively without any time misalignment or missing nodes.

[0062] Furthermore, the anesthesia response time-series change map was analyzed to identify recurring feature-physiological-drug change patterns. The lead time of visual feature changes relative to physiological index fluctuations under different patterns, the response period for visual features to reach a stable state after drug adjustment, and the probability of occurrence of various patterns were calculated. The mined features, including the lead time range, the mean response period, and the probability of pattern occurrence, were integrated as supplementary features into the training process of the dynamic model of anesthesia response characteristics to continuously optimize the anesthesia response characteristic analysis data.

[0063] In this embodiment of the invention, a pattern recognition tool is used to analyze the anesthesia response time sequence change map to identify recurring feature-physiological-drug change patterns: Pattern 1 is "increased drug concentration → decreased frowning degree (response cycle 7 seconds) → decreased heart rate (lead time difference 1.2 seconds)", which appears 68 times in the map, with a probability of occurrence of 68 / 1080≈6.3%, a lead time range of 0.8-1.5 seconds, and a mean response cycle of 7.2 seconds; Pattern 2 is "stable drug concentration → slow decrease in eyelid opening and closing angle (no lead) → stable blood pressure", which appears 92 times, with a probability of occurrence of 8.5%, a response cycle of 12 seconds, and a lead time difference of 0 seconds; Pattern 3 is "decreased drug concentration → increased nystagmus frequency (response cycle 5 seconds) → increased heart rate (lead time difference 0.5 seconds)", which appears 45 times, with a probability of occurrence of 4.2%, a lead time range of 0.3-0.7 seconds, and a mean response cycle of 5.1 seconds. Eight repetitive patterns were identified, and the lead time range (accurate to 0.1 seconds), mean response period (accurate to 0.1 seconds), and probability of occurrence for each pattern were calculated. These parameters were incorporated as supplementary features (12 dimensions, such as mean lead time of 1.2 seconds and mean response period of 7.2 seconds for pattern 1) into the dynamic model of anesthesia response characteristics. The number of neurons in the model's input layer was increased from 15 to 27. After retraining, the model's prediction accuracy improved from 92% to 95%, and the error in the anesthesia response characteristic analysis data was reduced to within 0.3%, achieving continuous model optimization.

[0064] Furthermore, the process of continuously optimizing the anesthetic response characteristic analysis data during the training of the dynamic model of anesthetic response characteristics includes: The parameterized supplementary features, together with the original core visual features and associated medical monitoring data, constitute the model input feature set, expanding the input dimension of the dynamic model of anesthesia response characteristics. In this embodiment of the invention, a feature integration tool is used to merge the parameterized supplementary features with the original input features to construct an expanded model input feature set: the supplementary features have 12 dimensions, including the mean lead time of 8 types of repetitive patterns (e.g., 1.2 seconds for pattern 1 and 0.5 seconds for pattern 3), the mean response period of 3 types of patterns (e.g., 7.2 seconds for pattern 1 and 12 seconds for pattern 2), and the mean occurrence probability of 1 type of overall pattern (6.3%). Each supplementary feature is standardized from 0 to 1 (e.g., a normalized value of 0.6 for a mean lead time of 1.2 seconds and a normalized value of 0.48 for a mean response period of 7.2 seconds). The original input features include 10 core visual features (eyelid opening and closing angle, degree of frowning, etc., which have been standardized) and 3 related medical monitoring data (anesthetic drug concentration, heart rate, and respiratory rate, which have been standardized), for a total of 13 dimensions. The 12 supplementary feature dimensions and the 13 original feature dimensions were arranged in the order of "supplementary features - core visual features - medical monitoring data" to form a 25-dimensional model input feature set. Each feature dimension was labeled with its data source (e.g., "supplementary feature - mean of advance time in mode 1" and "core visual feature - eyelid opening and closing angle") and value range (0-1). The feature set was checked for completeness using a feature validation tool to ensure that no dimensions were missing or duplicated. The final generated input feature set can be directly used in the model input layer, and the data volume is consistent with the original feature set (970 cases, each containing 3600 feature records), providing a foundation for expanding the model dimensions.

[0065] Furthermore, the network structure of the dynamic model of anesthetic response characteristics was adjusted by adding a temporal feature processing layer and a probability calculation unit to learn and fuse the temporal regularity parameters and probability information corresponding to the supplementary features; In this embodiment of the invention, the original dynamic model of anesthesia response characteristics (3-layer neural network: 13 neurons in the input layer, 32 neurons in the hidden layer, and 8 neurons in the output layer) is optimized using a network structure adjustment tool: First, the number of neurons in the input layer is increased from 13 to 25, with 12 new neurons corresponding to 12 supplementary feature dimensions (such as "Mode 1 lead time mean neuron" and "Mode 1 response period mean neuron"). The input layer activation function remains ReLU, and the weights are initialized using a Xavier uniform distribution. Second, a temporal feature processing layer is added between the input layer and the hidden layer. This layer contains 40 neurons and uses long and short sequence features. The model employs a Last Memory (LSTM) structure (20 memory units, forgetting gate threshold 0.5) to learn temporal patterns in supplementary features (e.g., the temporal correlation between lead time and response cycle). The temporal feature processing layer and input layer are fully connected, with a weight learning rate of 0.008. In the third step, a probability calculation unit with 8 neurons (corresponding one-to-one with output layer neurons) is added between the hidden and output layers. This unit uses the Softmax activation function to fuse probabilistic information from supplementary features (e.g., pattern occurrence probability), converting the probability values ​​into weight coefficients recognizable by the output layer (e.g., a 6.3% pattern occurrence probability corresponds to a weight coefficient of 0.063). The adjusted model structure consists of 4 layers: 25 neurons in the input layer → 40 neurons in the temporal feature processing layer → 32 neurons in the hidden layer → 8 neurons in the probability calculation unit → 8 neurons in the output layer. All layers are fully connected, with an overall learning rate of 0.005 and 1500 iterations. A structure validation tool ensures parameter matching across layers (e.g., the input layer output dimension matches the input dimension of the temporal feature processing layer), eliminating dimensional incompatibility issues.

[0066] Furthermore, the optimized dynamic model of anesthesia response characteristics was retrained using the newly added supplementary features. By comparing the prediction error, accuracy, recall, and probability prediction accuracy of the dynamic model of anesthesia response characteristics before and after training, the impact of the integration of temporal patterns and probability information in the supplementary features on the model performance was verified. Through dynamic iterative calculation of the dynamic model of anesthesia response characteristics, anesthesia response characteristic analysis data with optimized probability assessment was generated.

[0067] In this embodiment of the invention, a model retraining tool is used to input the 25-dimensional input feature set into the optimized model for retraining: The first step involves dividing the model into a training set (679 cases) and a validation set (291 cases). The training set is used for parameter tuning, and the validation set is used for performance evaluation. The training batch size is set to 32, and the loss function is a weighted combination of cross-entropy loss (for classification output) and mean squared error (for probability output) (weight ratio 1:1). The second step uses the Adaptive Moment Estimation (Adam) optimizer for training, iterating 1500 times. Model performance metrics (prediction error, accuracy, and recall) are recorded every 100 iterations. The training process employed an early stopping strategy (training stopped if the validation set loss did not decrease after 50 consecutive iterations), ultimately stopping at 1200 iterations, at which point the model converged. The third step compared the model performance before and after training: before training, the model had a prediction error of 0.15, accuracy of 85%, recall of 82%, and probability prediction accuracy of 83%; after training, the model's prediction error decreased to 0.08, accuracy improved to 94%, recall improved to 91%, and probability prediction accuracy improved to 92%, with the most significant improvement in probability prediction accuracy (9 percentage points), indicating that the probability information in the supplementary features effectively optimized the model's probability output. A dynamic iterative calculation tool was used to input the preliminary analysis results of the anesthetic response (such as "eyelid opening and closing angle 0.3, mean advance time of Mode 1 0.6" from the 25-dimensional features) into the trained model. The model corrected the bias through 100 dynamic iterations (each iteration updating the weights of the memory unit and probability calculation unit of the temporal feature processing layer). For example, the initial predicted probability of moderate sedation was 30%, which was corrected to 32.9% after iteration (with an error of only 0.4% compared to the actual clinical probability of 33.3%). Finally, anesthetic response characteristic analysis data containing 8 output indicators (sedation depth level, probability of each response) was generated, with each data point accompanied by a probability assessment optimization label (such as "probability optimization - Mode 1 association"). Fifty new cases were selected for validation using a clinical validation tool. The model's prediction accuracy reached 93%, with errors all <1.5%, confirming the effect of supplementary feature integration on improving model performance and achieving continuous model optimization.

[0068] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0069] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A visual recognition-based anesthesia response characteristic analysis system, characterized in that, Includes the following modules: The anesthesia multi-source acquisition module is used to acquire multi-dimensional visual acquisition devices and related medical data interfaces in the anesthesia scenario. Based on the above devices and interfaces, it acquires human visual feature data and synchronous medical monitoring data during the anesthesia process, and performs noise reduction, time sequence alignment and format standardization preprocessing to generate integrated multi-source fusion data of the anesthesia scenario. The anesthesia characterization analysis module is used to extract and characterize visual features from multi-source fusion data of anesthesia scenarios to generate visual characterization data of anesthesia status. The anesthetic response analysis module is used to perform correlation analysis by combining visual representation data of anesthetic state with preset clinical anesthetic response evaluation indicators to generate preliminary anesthetic response analysis results including probability distribution of response types. The anesthetic characteristics analysis module is used to construct a dynamic model of anesthetic response characteristics based on the preliminary analysis results of anesthetic response, and to iteratively optimize and generate anesthetic response characteristic analysis data. The anesthesia status assessment module is used to analyze data based on the characteristics of anesthesia response and output an anesthesia status assessment report and clinical intervention recommendations.

2. The anesthesia response characteristic analysis system based on visual recognition according to claim 1, characterized in that, The anesthesia multi-source acquisition module includes the following functions: Acquire medical data interfaces for facial expression capture devices, eye dynamic capture devices, limb posture monitoring devices, anesthesia monitors, and vital sign monitors in anesthesia scenarios. The aforementioned visual acquisition devices continuously collect visual information corresponding to the characteristics of facial muscle movement, eyelid opening and closing, eye movement trajectory, limb movement range and posture changes during anesthesia. Simultaneously, medical monitoring data corresponding to heart rate, blood pressure, respiratory rate and anesthetic drug concentration are collected through the medical data interface, which together constitute the original anesthesia-related data. Image denoising, blur correction, and feature point enhancement are performed on the visual information in the original anesthesia-related data; outlier removal and missing value completion are performed on the medical monitoring data. The processed visual information and medical monitoring data are aligned sequentially according to timestamps, and the data format and sampling frequency are unified to generate integrated multi-source fusion data of the anesthesia scene.

3. The anesthesia response characteristic analysis system based on visual recognition according to claim 1, characterized in that, The anesthesia characterization analysis module includes the following functions: Visual feature data and medical monitoring-related data were separated from multi-source fusion data in anesthesia scenarios; Hierarchical feature extraction is performed on visual feature data. First, basic visual features corresponding to facial contours and limb contours are extracted. Then, deep dynamic features corresponding to facial micro-expressions, eye tremor frequency, and subtle limb twitches are mined to generate a multi-dimensional visual feature set. Feature filtering and dimensionality reduction are performed on the multi-dimensional visual feature set to remove redundant and interfering features and retain the core visual features that are highly correlated with the anesthetic response. The core visual features are correlated with the corresponding medical monitoring data to construct a mapping relationship between visual features and anesthesia physiological state, and generate visual representation data of anesthesia state. The validity of the visual representation data of the anesthesia state was verified, invalid representation information was removed, and the visual representation data of the anesthesia state was ensured to reflect the probability distribution of response types during the anesthesia process.

4. The anesthesia response characteristic analysis system based on visual recognition according to claim 1, characterized in that, The anesthesia response analysis module includes the following functions: Clinical anesthesia response evaluation indicators were extracted from clinical anesthesia diagnosis and treatment guidelines, covering key dimensions corresponding to sedation depth level, analgesic effect assessment, signs of adverse reactions, and neuromuscular blockade degree. The core visual features in the visual representation data of anesthesia status are matched one by one with each indicator in the clinical anesthesia response evaluation index to establish the correspondence between visual features and evaluation indicators. By analyzing Pearson correlation coefficients and calculating mutual information, the correlation strength between core visual features and various evaluation indicators is quantified, and feature-indicator combinations with correlation strengths higher than preset thresholds are selected. Multidimensional correlation analysis is performed based on the selected feature-indicator combinations to explore the dynamic pattern between changes in visual features and fluctuations in anesthesia response indicators, calculate the probability of occurrence of different response types, and generate preliminary analysis results of anesthesia response including the probability distribution of response types.

5. The anesthesia response characteristic analysis system based on visual recognition according to claim 1, characterized in that, The anesthesia characteristic analysis module includes the following functions: We collected a large amount of clinical anesthesia case data, covering anesthesia response data, synchronous visual feature data and related probability statistics corresponding to different anesthesia methods, different patient physical conditions and different surgical types, and constructed an anesthesia response characteristic sample library. Based on the preliminary analysis results of the anesthetic response, the input feature dimensions of the model are determined. The core visual features, related medical monitoring data and feature association probabilities are used as input variables, and the actual values ​​and occurrence probabilities of clinical anesthetic response evaluation indicators are used as output variables to construct an initial dynamic model of anesthetic response characteristics. The gradient descent algorithm combined with cross-validation is used to train the initial dynamic model of anesthesia response characteristics, and the model parameters are continuously adjusted to minimize the prediction error. The preliminary analysis results of anesthesia response are input into the trained dynamic model of anesthesia response characteristics, and the deviation is corrected through dynamic iterative calculation of the dynamic model of anesthesia response characteristics to generate anesthesia response characteristic analysis data. The anesthetic response characteristics analysis data output by the model were clinically validated, and the dynamic model of anesthetic response characteristics was continuously optimized by combining the evaluation feedback of anesthesiologists.

6. The visual recognition-based anesthesia response characteristic analysis system according to claim 5, characterized in that, The evaluation feedback continuously optimizes the dynamic model of anesthetic response characteristics, including: Visual feature sequence data from multi-source fusion data of anesthesia scenarios are acquired and divided into time windows. The visual feature change trend, mutation point information and feature occurrence probability are extracted within each time window. By combining synchronously collected medical monitoring data, the temporal correlation between visual feature mutation points and vital sign fluctuations and anesthetic drug infusion nodes is analyzed, the probability of occurrence of related events is calculated, and feature-physiology-drug correlation data containing temporal correlation probabilities is generated. An anesthesia response time-series change map is constructed based on associated data. The map includes visual feature trajectories, physiological index curves, drug infusion curves, correlation markers between the three and related probability information. Dynamic patterns in the anesthesia response time-series change map are then mined. The mined dynamic patterns and corresponding probability information are integrated into the dynamic model of anesthesia response characteristics to continuously optimize the anesthesia response characteristic analysis data.

7. The visual recognition-based anesthesia response characteristic analysis system according to claim 6, characterized in that, The generation of feature-physiological-drug association data containing temporal correlation probabilities includes: The visual feature sequence data in the multi-source fusion data of the anesthesia scene is preprocessed to remove abnormal sequence segments caused by equipment shaking and light changes. The missing sequence data is filled in by interpolation to ensure the continuity, integrity and accuracy of probability calculation of the sequence data. The time windows are divided according to a preset time interval. The length of each time window is dynamically adjusted according to different stages of the anesthesia process. The length of the window during the anesthesia induction and recovery periods is shortened, while the length of the window during the maintenance period is appropriately extended. Statistical analysis is performed on the visual feature data within each time window to calculate the feature mean, variance, rate of change, peak value and probability of feature occurrence, extract the feature change trend, and identify visual feature mutation points within the time window through anomaly detection algorithm, and record the time location, feature change magnitude and mutation probability of visual feature mutation points. The visual feature change trends, mutation point information, and related probability parameters within each time window are mapped one-to-one with the heart rate, blood pressure, blood oxygen saturation, and anesthetic drug concentration collected at the same time, and a correlation index is established in the time dimension. By calculating the time difference between visual feature mutation points, physiological indicator fluctuation nodes, and drug infusion nodes, the causal relationship among the three is analyzed, generating feature-physiology-drug association data that includes temporal correlation strength, response delay time, and probability of occurrence of associated events.

8. The anesthetic response characteristic analysis system based on visual recognition according to claim 6, characterized in that, The dynamic patterns include the time lag between changes in visual features and fluctuations in physiological indicators, the response cycle of visual features after drug adjustment, and the probability of occurrence of various patterns.

9. The visual recognition-based anesthesia response characteristic analysis system according to claim 8, characterized in that, The process of mining the dynamic patterns includes: With the time axis as the horizontal axis, a visual feature trajectory layer, a physiological index curve layer, and a drug infusion curve layer are constructed respectively. The visual feature trajectory layer uses the numerical changes and occurrence probabilities of core visual features as the vertical axis, the physiological index curve layer uses the values ​​of various vital signs parameters and their normal probability ranges as the vertical axis, and the drug infusion curve layer uses the infusion rate, cumulative dose, and dose adjustment probability of anesthetic drugs as the vertical axis. In each layer, key nodes are marked. The visual feature trajectory layer marks abrupt change points, feature peak points and corresponding probabilities. The physiological index curve layer marks the normal range threshold line, abnormal fluctuation points and the probability of abnormal occurrence. The drug infusion curve layer marks drug type switching points, dose adjustment points and adjustment probabilities. By linking and marking the key nodes of time synchronization in the three layers through the correlation index, the correspondence and correlation probability between changes in visual features, fluctuations in physiological indicators, and drug adjustments are clarified, forming a complete time sequence map of anesthetic response changes. The anesthesia response time-series change map was analyzed to identify recurring feature-physiological-drug change patterns. The lead time of visual feature changes relative to physiological index fluctuations under different patterns, the response period for visual features to reach a stable state after drug adjustment, and the probability of occurrence of various patterns were calculated. The mined features, including the lead time range, the mean response period, and the probability of pattern occurrence, were integrated as supplementary features into the training process of the dynamic model of anesthesia response characteristics to continuously optimize the anesthesia response characteristic analysis data.

10. The anesthetic response characteristic analysis system based on visual recognition according to claim 9, characterized in that, The training process of the dynamic model incorporating anesthesia response characteristics continuously optimizes the anesthesia response characteristic analysis data, including: The parameterized supplementary features, together with the original core visual features and associated medical monitoring data, constitute the model input feature set, expanding the input dimension of the dynamic model of anesthesia response characteristics. The network structure of the dynamic model of anesthesia response characteristics was adjusted by adding a temporal feature processing layer and a probability calculation unit to learn and fuse the temporal regularity parameters and probability information corresponding to the supplementary features. The optimized dynamic model of anesthesia response characteristics was retrained using newly added supplementary features. By comparing the prediction error, accuracy, recall, and probability prediction precision of the dynamic model of anesthesia response characteristics before and after training, the impact of the integration of temporal patterns and probability information in the supplementary features on the model performance was verified. Through dynamic iterative calculation of the dynamic model of anesthesia response characteristics, anesthesia response characteristic analysis data with optimized probability assessment was generated.

Images