Ophthalmic surgery management method and system based on eye movement and pupil analysis

Abstract

The application provides an ophthalmic surgery management method and system based on eyeball movement and pupil analysis, and relates to the technical field of medical informatization and operating room intelligent management. The method comprises the following steps: collecting physiological data of the eyes of a patient based on a sensing device; analyzing and processing the eyeball movement and / or pupil change characteristics in the data to generate real-time state information representing the physiological and psychological conditions of the patient; generating a collaborative instruction set containing at least two operation instructions pointing to different posts based on the real-time state information and a preset collaborative strategy; and distributing each operation instruction to the corresponding terminal device according to the correspondence between the post and the terminal, thereby realizing real-time and objective analysis of the intraoperative state of the patient, automatically and synchronously scheduling multiple posts for collaborative intervention, and improving the collaborative efficiency, response speed and safety of day surgery.

Description

Technical Field

[0001] This invention relates to the fields of medical information technology and intelligent operating room management, and in particular to a method and system for managing ophthalmic surgeries based on eye movement and pupil analysis. Background Technology

[0002] Day surgery for ophthalmology is becoming increasingly popular due to its advantages such as rapid recovery and low cost. However, the short operation time and the patient's conscious state place higher demands on real-time monitoring and team collaboration during the procedure. Currently, multi-positional collaboration in the operating room (such as doctors, anesthesiologists, and nurses) mainly relies on the personal experience and verbal communication of medical staff, lacking objective, real-time, and continuous monitoring methods for the patient's physiological and psychological state during surgery. Especially in ophthalmology surgery, subtle changes such as involuntary eye movements and abnormal pupillary responses to stimuli are often early indicators of physiological tension, pain, or drug reactions, but this information is difficult to continuously and accurately capture and interpret with the naked eye. This makes it difficult for the team to identify subtle changes in the patient's condition in a timely and accurate manner and coordinate targeted interventions across different positions, potentially affecting surgical safety and efficiency.

[0003] Therefore, there is an urgent need for an intelligent management solution that can analyze the patient's status in real time and automatically trigger multi-position collaborative operations to improve the response speed, collaborative efficiency and overall safety of day ophthalmic surgery. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, the first aspect of this invention proposes an ophthalmic surgery management method based on eye movement and pupil analysis, comprising: S1: Based on the surgical environment, the sensing device collects physiological data of the patient's eyes and generates eye physiological data; S2: Analyze and process the eye movement and / or pupillary change characteristics reflected in the ocular physiological data to generate real-time status information representing the patient's current physiological and psychological condition; wherein, S2 includes: S21a: Extract eye movement trajectory sequences from ocular physiological data; S22a: Perform pattern recognition processing on the eye movement trajectory sequence to obtain the patient's eye movement pattern characteristics; S23a: Based on eye movement pattern features, match and classify them with multiple reference patterns in the database to generate real-time status information; and / or S21b: Extract time-series data of pupil diameter from ocular physiological data; S22b: Calculate and analyze the rate of change and stability of pupil diameter time series data to generate pupil response characteristic parameters; S23b: Identify and classify the pupil response patterns to light or drug stimulation reflected in the pupil response characteristic parameters to generate real-time status information characterizing the patient's specific physiological state. S3: Based on real-time status information and combined with preset collaboration strategies, perform analysis and mapping processing to generate a collaboration instruction set containing at least two operation instructions that point to different positions. S4: Based on the collaborative instruction set and the preset correspondence between positions and terminals, perform instruction distribution processing and send each operation instruction to the terminal device of the corresponding position.

[0005] Compared with the prior art, the beneficial effects of the present invention are as follows: By constructing an automated closed-loop process from data acquisition and status analysis to collaborative command generation and distribution, the problems of reliance on subjective experience, response lag, and insufficient coordination mentioned in the background technology are effectively solved. First, sensing devices based on the surgical environment collect physiological data from the patient's eyes, generating ocular physiological data. This provides a data foundation for objective and continuous monitoring of the patient's status, overcoming the shortcomings of discontinuous and inaccurate manual observation. Next, by analyzing the eye movement and / or pupillary change characteristics reflected in the ocular physiological data, specific eye movement patterns can be identified from the eye movement trajectory sequence, and / or pupillary response characteristic parameters can be analyzed from the pupil diameter time series data. This step transforms the raw physiological signals into real-time status information that characterizes the patient's current physiological and psychological state, achieving automated and intelligent identification of key states such as patient tension, pain response, or drug effects.

[0006] Subsequently, based on this real-time status information, and combined with a preset collaborative strategy, analysis and mapping are performed to generate a collaborative instruction set containing at least two operational instructions pointing to different positions. This means that the system can automatically plan and generate a set of composite instructions that require simultaneous or sequential execution by multiple positions (such as nurses, anesthesiologists, and equipment operators) based on a single status judgment, achieving a leap from "status recognition" to "multi-threaded collaborative action planning." Finally, based on the collaborative instruction set and the preset correspondence between positions and terminals, instruction distribution is performed, sending each operational instruction to the terminal device of the corresponding position. This ensures that the parsed status information and decisions can reach the relevant responsible persons without delay, driving them to execute specific operations, forming a complete collaborative link of "collection-analysis-decision-execution."

[0007] The entire method, through the close connection and synergy of each step, achieves real-time and accurate perception of the patient's intraoperative status, and automatically and efficiently dispatches resources from multiple positions for collaborative intervention, significantly improving the intelligence level and team collaboration efficiency of day ophthalmic surgery management, thereby ensuring the safety and smooth progress of the surgery. Attached Figure Description

[0008] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0009] Figure 1 The diagram shown is a flowchart of an ophthalmic surgery management method based on eye movement and pupil analysis provided in an embodiment of the present invention. Figure 2 The diagram shown is a structural schematic of an ophthalmic surgical management system based on eye movement and pupil analysis provided in an embodiment of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0011] The specific embodiments of the present invention will be described below.

[0012] Example 1 like Figure 1 As shown, this invention proposes a method for managing ophthalmic surgery based on eye movement and pupil analysis, including: S1: Based on the surgical environment, the sensing device collects physiological data of the patient's eyes and generates eye physiological data; S2: Analyze and process the eye movement and / or pupillary change characteristics reflected in the ocular physiological data to generate real-time status information representing the patient's current physiological and psychological condition; wherein, S2 includes: S21a: Extract eye movement trajectory sequences from ocular physiological data; S22a: Perform pattern recognition processing on the eye movement trajectory sequence to obtain the patient's eye movement pattern characteristics; S23a: Based on eye movement pattern features, match and classify them with multiple reference patterns in the database to generate real-time status information; and / or S21b: Extract time-series data of pupil diameter from ocular physiological data; S22b: Calculate and analyze the rate of change and stability of pupil diameter time series data to generate pupil response characteristic parameters; S23b: Identify and classify the pupil response patterns to light or drug stimulation reflected in the pupil response characteristic parameters to generate real-time status information characterizing the patient's specific physiological state. S3: Based on real-time status information and combined with preset collaboration strategies, perform analysis and mapping processing to generate a collaboration instruction set containing at least two operation instructions that point to different positions. S4: Based on the collaborative instruction set and the preset correspondence between positions and terminals, perform instruction distribution processing and send each operation instruction to the terminal device of the corresponding position.

[0013] The ophthalmic surgery management method based on eye movement and pupil analysis provided by this invention begins with the objective and continuous capture of the patient's intraoperative ocular vital signs. The "surgical environment-based sensing device" mentioned in step S1 refers to a dedicated data acquisition device deployed in the operating room. Examples include a high-frame-rate eye tracker (a camera system capable of accurately recording changes in eye position hundreds of times per second) or an infrared pupillometer (a device that uses an infrared light source and sensors to non-contactly measure pupil diameter). These sensing devices are typically integrated into the optical path of the surgical microscope or fixed near the patient's head using a separate flexible support to ensure they remain stably oriented towards the patient's eyes throughout the surgery without interfering with the surgeon's field of vision. They work continuously, collecting and generating "ocular physiological data," which is represented in a computer as a structured time-series dataset. This data may include fields such as timestamps, the position data of the left / right eyeballs in a two-dimensional or three-dimensional coordinate system (X, Y, possibly including the Z axis), and measurements of the diameter of the left / right pupils (usually in millimeters or pixels). This step fundamentally replaces the subjective model that relies on intermittent visual observation by medical staff, providing a high-precision, quantifiable raw signal basis for subsequent analysis.

[0014] After obtaining the raw data, step S2 enters the crucial "analysis and processing" stage, the purpose of which is to extract features reflecting the patient's core state from the raw "ocular physiological data." Here, the analysis of "eye movement" features is typically achieved by processing "eye movement trajectory sequences." An eye movement trajectory sequence is an array of eye center point coordinates arranged in chronological order, output by the eye tracker. Performing "pattern recognition processing" on this sequence means applying specific algorithms to identify its movement patterns. For example, a Hidden Markov Model (HMM) can be used to identify different types of eye movement states (such as fixation, saccades, and smooth tracking), and to calculate the duration and transition probability of each state; or a Dynamic Time Warping (DTW) algorithm can be applied to compare the current trajectory with preset typical trajectory templates such as "tremor" and "saccade abnormalities." Through this processing, quantified "eye movement pattern features" are obtained, such as parameters like "saccade frequency," "fixation point stability index," and "tremor amplitude." Subsequently, these features are "matched and classified" with "multiple reference patterns in the database." The database pre-stores a set of feature patterns associated with specific state labels (such as "calm", "anxious", "pain response"). The matching process can involve calculating the Euclidean distance or cosine similarity between the current feature vector and each reference pattern vector, and then using a classifier such as K-Nearest Neighbors (KNN) or Support Vector Machine (SVM) to classify it into the most similar state category, thereby outputting structured "real-time state information".

[0015] Simultaneously, or as an alternative, analytical processing also targets "pupil change characteristics." This involves extracting "pupil diameter time series data" from ocular physiological data, i.e., the waveform of pupil diameter changes over time. "Calculation and analysis of the rate of change and stability" quantifies its dynamic characteristics using mathematical methods. For example, the "rate of change" can be obtained by calculating the absolute value or derivative of the diameter difference between adjacent time points to reflect the speed of the pupil's response to stimulation; "stability" can be assessed by calculating the variance or root mean square error of the diameter over a period of time to reflect whether the pupil baseline is stable. These calculations generate specific "pupil response characteristic parameters" such as "average dilation rate," "contraction delay time," and "baseline fluctuation standard deviation." Furthermore, these parameters are "identified and classified" to determine the pupil's "response pattern." For example, identifying a "sluggish light reflex" pattern might manifest as a significantly prolonged latency and insufficient contraction amplitude of the pupil when the surgical microscope light is enhanced; identifying a "drug-induced dilation" pattern might manifest as continuous and slow pupil dilation in the absence of light stimulation. By comparing with known pathological or pharmacological response model libraries, real-time status information such as "suspected oversedation" or "possible nerve stimulation" can be generated to characterize "specific physiological states".

[0016] Step S3 is the decision-making center, which transforms the state judgment into an executable collaborative action plan. The "pre-defined collaborative strategy" is a rule base pre-established by the medical team based on clinical experience, existing in the logical form of "if [state], then [execute the set of instructions]". For example, a strategy might be: "If the state is 'moderate anxiety with nystagmus,' then generate the instructions: 1. Notify the nurse to provide reassurance, 2. Notify the anesthesiologist to check the depth of sedation." "Analysis and mapping processing" executes this logical judgment: using the real-time state information output from step S2 as a condition, it iterates through and matches all rule conditions in the strategy base to find the most matching one or more strategies. After a successful match, the system instantiates a "collaborative instruction set" based on the "rule" part of the strategy. This instruction set must contain "at least two operation instructions pointing to different positions," for example, one instruction is "play the pre-defined reassurance voice," with the target position identified as "nurse"; the other instruction is "adjust the vital signs monitoring interval to high-frequency mode," with the target position identified as "anesthesiologist." This forces the system to engage in multi-threaded collaborative thinking.

[0017] Finally, step S4 is responsible for accurately delivering the instructions. The system maintains a mapping table of "positions and terminals," for example, recording the "target terminal device address" corresponding to the "anesthesiologist position," which might be the IP address of the anesthesia workstation computer and a certain software interface. During instruction distribution, the system parses the "target position identifier" of each operation instruction in the collaborative instruction set, queries this mapping table to obtain the specific network address or device identifier, and then sends the formatted instruction message (such as JSON or XML format) to that address via the hospital intranet or a dedicated wireless communication protocol. After receiving the instruction, the corresponding terminal device (such as a computer, tablet, smartwatch, or augmented reality glasses) notifies the staff through screen pop-ups, sound prompts, or vibrations.

[0018] In summary, the method defined in this invention achieves a significant technological leap through the tight integration of the four steps described above. It utilizes sophisticated sensing equipment to objectively and continuously collect high-dimensional physiological signals such as eye movements and pupil changes, laying the mathematical foundation for precise analysis. Next, through advanced signal processing and pattern recognition algorithms, the raw data stream is transformed into state information with clear clinical semantics, completing the leap from "seeing" to "understanding." Then, based on a pre-set clinical strategy knowledge base, single state judgments are automatically mapped into a set of complex instructions requiring multi-role collaboration, achieving intelligent transition from "cognition" to "decision-making." Finally, through precise networked instruction distribution, decisions are ensured to reach the relevant terminals instantly and accurately, driving synchronized action among team members and completing the closed loop from "decision-making" to "execution." This process fundamentally changes the traditional collaborative model that relies on manual observation, experience-based judgment, and verbal communication. It enables the entire team to act like a sophisticated instrument, automatically, synchronously, and accurately responding to subtle changes in the patient's condition based on the same objective data, thereby greatly improving the safety, efficiency, and response speed of day ophthalmic surgery.

[0019] In some implementations, S3 includes: S31a: Determine the status category of real-time status information; S32a: When the status category is determined to be the first category, the first instruction generation logic is triggered to generate a first set of collaborative instructions, which includes sending a voice reassurance instruction to the nurse's terminal and sending an instruction to increase the monitoring frequency to the anesthesiologist's terminal. S33a: When the state category is determined to be the second category, the second instruction generation logic is triggered to generate a second set of collaborative instructions, which includes sending a pause instruction to the surgical equipment control system and sending an optical positioning adjustment instruction to the doctor's operating terminal; wherein, the first set of collaborative instructions and the second set of collaborative instructions constitute a collaborative instruction set.

[0020] This embodiment further reveals a specific and efficient logical implementation method within step S3, namely, triggering different predefined instruction generation paths through state classification. In step S31a, "judging the state category of real-time state information" means that the system uses a classification function or decision rule to summarize real-time state information containing multi-dimensional parameters (such as "anxiety index: 0.7, pain probability: 0.2, pupil stability: low") into a discrete category label. This judgment process may be based on threshold comparison: for example, when the "anxiety index" is higher than a certain threshold and the "pain probability" is lower than another threshold, it is judged as "Category 1"; when the "eye movement disorder" increases sharply or the "pupil light reflex disappears", it is judged as "Category 2". Here, "Category 1" and "Category 2" are predefined sets of states with different clinical priorities and intervention logics. For example, Category 1 may represent "non-emergency states requiring attention" (such as mild anxiety, mild discomfort), and Category 2 may represent "emergency states requiring immediate intervention" (such as sudden body movement risk, severe pain response, pupillary crisis).

[0021] When the system determines the status category to be Category 1, it "triggers the first instruction generation logic." This is a pre-defined program submodule that internally encodes a standardized response scheme for Category 1 statuses. This logic "generates a first set of collaborative instructions, including sending a voice reassurance instruction to the nurse's terminal and an instruction to increase the monitoring frequency to the anesthesiologist's terminal." The "voice reassurance instruction" is a specific, executable command; its content might specify playing a numbered reassurance voice message (such as "Please relax, take a deep breath"), or trigger a prominent prompt on the nurse's terminal reminding the nurse to execute the standard reassurance script. The "increase monitoring frequency instruction" is equally specific; it might be a standardized command sent to the anesthesia monitoring control system, requesting it to adjust the interval between automatic recording and alarm checks of key parameters such as blood pressure and heart rate from the usual once per minute to once every thirty seconds. These two instructions are generated simultaneously, targeting different positions, forming a collaborative response package for "non-emergency situations requiring attention."

[0022] Accordingly, when the condition is classified as Category II, the system "triggers the second instruction generation logic." This is a distinct submodule designed for emergency situations. The generated "second collaborative instruction subset" is more interventionist, including, for example, "sending a pause instruction to the surgical equipment control system and an optical positioning adjustment instruction to the doctor's operating terminal." The "pause instruction" is a direct control signal whose protocol format must be compatible with the control interface of the specific surgical equipment (such as an phacoemulsification machine, vitrectomy machine, or laser therapy device). The instruction content is either "immediately pause the current output" or "enter standby safety mode." The "optical positioning adjustment instruction" is sent to the operating terminal of the surgical microscope or navigation system used by the doctor. Its content may include bringing up a specific positioning calibration interface or highlighting anatomical structure markers that require the doctor's reconfirmation. The collaborative goal of these two instructions is to halt potentially dangerous procedures in the shortest possible time and provide the doctor with technical support for reassessment and adjustment. Ultimately, depending on the real-time status, the system dynamically selects to generate either the first or second collaborative instruction subset, both of which are instances constituting the complete "collaborative instruction set."

[0023] This implementation method, by introducing a clear state classification and condition triggering mechanism, produces clear and efficient technical results. Firstly, through automated state category judgment, it assigns clear action priorities to complex real-time state information, enabling the system to quickly distinguish between situations requiring mild handling and crises requiring urgent intervention. Triggering independent instruction generation logic for different categories means that the system internally achieves modularization and specialization of response strategies; each type of logic can be independently optimized to best suit its corresponding clinical scenario. Pre-setting a subset of instructions for the first category, including reassurance and enhanced monitoring, embodies a preventative and supportive collaborative approach, aiming to smoothly transition non-emergency states and prevent their escalation. Pre-setting a subset of instructions for the second category, including equipment suspension and operational adjustments, embodies a decisive and direct safety intervention approach, aiming to control risks immediately and create a safe window for human decision-making. This hierarchical response mechanism ensures precise matching of system resources and response intensity, avoiding overreaction or underreaction. This allows the entire multi-position collaborative management process to possess both the delicacy to handle routine fluctuations and the decisiveness to deal with sudden risks, thereby comprehensively improving the stability and safety of the surgical procedure.

[0024] In some implementations, before S1, the following is also included: S0: Based on the ocular physiological data and corresponding manually labeled state data from historical surgical cases, the machine learning model is trained and validated to obtain the trained state analysis model; S2 can also be: inputting eye physiological data into the state analysis model, and generating real-time state information through the forward computation processing of the state analysis model.

[0025] This embodiment introduces a data-driven artificial intelligence method to enhance the accuracy and intelligence of state analysis, the core of which lies in constructing and applying a "state analysis model." This begins with an offline model building phase S0. This phase relies on "ocular physiological data from historical surgical cases and their corresponding manually labeled state data." Historical ocular physiological data is a raw dataset collected and archived from a large number of previous day ophthalmic surgeries using the aforementioned sensing devices. The "manually labeled state data" is crucial; it requires experienced clinical experts (such as surgeons and anesthesiologists) to qualitatively or semi-quantitatively label the patient's state reflected in the ocular physiological data for a specific time period after reviewing surgical videos, synchronized vital sign records, and anesthesia records. For example, a 30-second data segment might be labeled with tags such as "calm and cooperative," "moderate anxiety," "painful pre-movement aura," or "drug onset period." These labels constitute the "truth values" required for supervised learning.

[0026] Training and validating a machine learning model is a systematic engineering process. First, a suitable initial machine learning model architecture needs to be selected. Given that ocular physiological data is time-series data, recurrent neural networks (RNNs), especially their improved forms such as Long Short-Term Memory (LSTM) networks or gated recurrent units (GRUs), are often chosen because they effectively capture temporal dependencies. One-dimensional convolutional neural networks (1D-CNNs) can also be used to extract local temporal features, or a hybrid model combining CNNs and RNNs can be used. Next, the prepared labeled historical dataset is divided into non-overlapping training, validation, and test sets. The training process (i.e., iterative optimization of internal parameters) typically uses backpropagation and an optimizer (such as Adam). In each training round, the model reads a batch of data from the training set (i.e., a batch of ocular physiological data sequences), performs forward propagation to calculate the predicted state labels, then calculates the difference (loss) between the predicted labels and the actual manually labeled labels, calculates the gradient of the loss relative to the parameters of each layer of the model through backpropagation, and finally, the optimizer updates the model parameters based on the gradient to reduce the loss. This process is repeated. "Until the difference between its output and manually labeled state data meets the accuracy requirements" means that training stops when the model's performance metrics (such as classification accuracy) on an independent validation set reach a preset threshold and no longer significantly improve. At this point, the "trained state parsing model" is obtained. This model has internalized the mapping function between the raw eye physiological data sequence and complex state categories.

[0027] In the application phase, this embodiment provides another implementation option for step S2: "Inputting ocular physiological data into the state resolution model, and generating real-time state information through the forward computation processing of the state resolution model." This means that during real-time surgical operation, the system only needs to perform necessary preprocessing (such as normalization) on the latest segment of ocular physiological data obtained from the sensing device (e.g., a data window from the past 10 seconds) according to the same format as during training, and then directly feed it into this pre-trained state resolution model. The trained neurons in each layer of the model will automatically perform feature extraction, abstraction, and classification, and finally directly generate the "real-time state information" corresponding to the current moment at the output layer, such as the probability distribution of various states or the most likely category label. This is equivalent to encapsulating multiple steps such as feature engineering, pattern recognition, and classification decision-making into an end-to-end neural network and completing them all at once.

[0028] This implementation approach brings about a fundamental technological advancement. By utilizing a large amount of historical data with expert annotations for model training, it establishes the system's state analysis capabilities based on statistical regularities and collective clinical experience, greatly improving the objectivity and repeatability of judgments and reducing biases caused by individual experience differences. Integrating complex analysis processes into a single trained model makes real-time analysis extremely efficient and unified, requiring only a single forward computation, significantly reducing computational latency and meeting the requirements for real-time response during surgery. More importantly, deep neural networks possess the ability to automatically learn high-level abstract features, discovering and utilizing subtle patterns and state associations that are difficult for the human eye to discern but do exist in the data, thereby achieving a deeper and more accurate state insight than methods based on fixed rules or simple feature analysis. Furthermore, the model has the potential for continuous learning and evolution. With the accumulation of more labeled data, the model can be retrained and updated periodically, allowing the system's analysis capabilities to continuously adapt to new clinical discoveries and technological developments. Therefore, this state analysis path based on a machine learning model provides a powerful, evolvable, and highly intelligent core analysis engine for the method of this invention, serving as the scientific cornerstone for the entire system to achieve accurate collaborative decision-making.

[0029] In some implementations, S0 includes: S0a: Filter, denoise, and extract features from historical eye physiological data to generate a standardized historical feature dataset; S0b: The standardized historical feature dataset and the corresponding manually labeled state data are used as training sample pairs and input into the initial machine learning model; S0c: The internal parameters of the initial machine learning model are adjusted through an iterative optimization algorithm until the difference between its output and the manually labeled state data meets the accuracy requirements, thus generating a trained state analysis model.

[0030] This embodiment clearly depicts a standardized and reproducible construction path from raw data to a high-performance model. Step S0a, "filtering, denoising, and extracting features from historical ocular physiological data to generate a standardized historical feature dataset," is the primary step in ensuring data quality and model training effectiveness. Historical ocular physiological data inevitably contains noise during acquisition, such as electromagnetic interference from the operating room environment (manifested as high-frequency spikes) and motion artifacts caused by patient blinking or slight head movements (manifested as instantaneous changes in data values ​​or baseline drift). Therefore, "filtering and denoising" is a necessary preprocessing step. For example, a Butterworth low-pass filter can be used, with a cutoff frequency set (e.g., 10Hz) to filter out high-frequency noise above the physiological ocular movement frequency; for baseline drift, a high-pass filter or direct subtraction of the moving average baseline can be used for correction. After denoising, "feature extraction" is performed. This does not involve directly feeding all data points to the model, but rather purposefully calculating a series of clinically and engineeringally significant indicators. These features can be extracted from multiple dimensions: time-domain features, such as the mean, variance, and peak value of eye movement velocity; frequency-domain features, such as the energy proportion of the signal power spectrum obtained by Fast Fourier Transform (FFT) in a specific frequency band (e.g., the 0-2Hz tremor band); and specialized features based on the aforementioned analytical logic, such as the "sliding window variance of the pupil diameter change rate" and the "average distance between consecutive fixation points (saccade amplitude)". The calculated raw feature values ​​may vary greatly in scale and range, so "standardization" is required. Common methods include Z-score standardization (making the feature mean 0 and the standard deviation 1) or max-min normalization (scaling the feature to the [0, 1] interval). The final "standardized historical feature dataset" is a clean and well-organized numerical matrix, where each row represents a historical data sample and each column represents a standardized feature.

[0031] Step S0b, "using the standardized historical feature dataset and the corresponding manually labeled state data as training sample pairs and inputting them into the initial machine learning model," is a crucial step in constructing supervised learning samples. Here, the composition of the "training sample pair" is very clear: the input is the feature vector of a sample processed by S0a (i.e., a row in the standardized historical feature dataset), and the output is the "manually labeled state data" annotated by experts for the corresponding time segment. This state data is typically encoded as a one-hot vector or a category index. For example, for a classification problem, if there are three categories of states: "calm," "anxious," and "painful," then the "anxious" state might be encoded as [0, 1, 0]. Thousands upon thousands of such (feature vector, state label) pairs are combined to form the training dataset, which is then input into the initial machine learning model. The initial machine learning model is a model to be trained with a determined number of network layers, number of neurons per layer, and activation function type (such as ReLU, Softmax), but all weights and bias parameters are random initial values ​​or default values.

[0032] Step S0c, "adjusting the internal parameters of the initial machine learning model through iterative optimization algorithms until the difference between its output and the manually labeled state data meets the accuracy requirements," is the core process of model learning. "Iterative optimization algorithms" typically refer to a family of algorithms based on gradient descent. Taking stochastic gradient descent (SGD) or its variants (such as Adam) as an example: In each iteration (or processing a batch of data), the algorithm performs the following operations: 1) Forward propagation: Inputting the feature vectors of a batch of training samples into the model, and calculating the predicted state output through each layer. 2) Loss calculation: Using a loss function (such as the cross-entropy loss function), comparing the model's predicted output with the actual "manually labeled state data" of this batch of samples, calculating a scalar loss value that quantifies the degree of error in the current model's prediction. 3) Backward propagation: Using the chain rule, calculating the gradient of the loss function relative to each internal parameter (weight and bias) of the model from the output layer back to the input layer. The gradient indicates the direction and approximate magnitude by which each parameter should be adjusted to reduce the loss. 4) Parameter update: The optimizer updates all parameters in the model based on the calculated gradient and the preset learning rate. This cycle repeats, and the model "learns" on a large number of samples. Training continues "until the difference between its output and manually labeled state data meets the accuracy requirements," which is usually determined by monitoring the model's performance on a "validation set" that was not used in training. Training stops when the accuracy on the validation set no longer improves over several consecutive training cycles, or when the loss value drops below a predetermined threshold. At this point, the model's internal parameters have been sufficiently tuned to generalize the mapping relationship from features to states well, thus "generating a fully trained state analysis model."

[0033] This meticulous training process design ensures the quality and reliability of the final model. Thorough filtering and denoising of the raw data effectively eliminates interfering information, improves the signal-to-noise ratio, and provides clean input for the model to learn realistic physiological patterns, which is the foundation for the model's strong robustness. Systematic feature extraction and standardization not only transform the raw data into a more informative representation but also eliminate the negative impact of feature scale differences on model training, accelerate the convergence of the optimization process, and enable unified processing of data from different sources. Precisely pairing standardized features with expert annotations to form sample pairs provides the model with clear and unambiguous learning objectives, ensuring the consistency between the model's learned knowledge system and clinical cognition. The use of iterative optimization algorithms for automatic parameter adjustment is a data-driven, adaptive learning process that enables the model to automatically discover and fit highly nonlinear, multi-factor intertwined correlations between features and complex states from massive samples—corresponding patterns whose complexity far exceeds that of manual rules. By establishing explicit stopping criteria based on validation set performance, overfitting of the model on the training set is effectively prevented, ensuring that the trained state-based analytical model has good generalization ability and practical value for new data and cases. Therefore, this standardized set of data processing and model training implementation methods is a necessary technical guarantee for generating a truly effective, reliable, and clinically applicable intelligent analytical module.

[0034] In some implementations, S4 includes: S41: Analyze the content of each operation instruction in the collaborative instruction set and the target job identifier; S42: Based on the target job identifier, query the job-terminal mapping table to determine the corresponding target terminal device address; S43: Encapsulate and send the operation instructions based on the target terminal device address.

[0035] This embodiment refines step S4 of the instruction distribution process, depicting a clear path from logical instruction to physical delivery. Step S41, "parses the content of each operation instruction and target position identifier in the collaborative instruction set," is the primary step in data processing. The collaborative instruction set typically exists in the system as a structured data collection, such as a JSON array containing multiple entries. Each entry represents an independent operation instruction, and its data structure may include fields such as "instruction_id" (used for unique identification and tracking), "target_role" (target position identifier, such as "circulating nurse," "anesthesia specialist," "instrument nurse"), "action_type" (operation type, such as "notification," "control," "adjustment"), and "action_parameters" (specific parameters, such as the audio file number to be played, the target value for equipment adjustment). The parsing process involves the instruction distribution module reading this data set, traversing each entry, and extracting the aforementioned key fields using string parsing or object attribute access methods. This step breaks down the packaged instruction into independent, machine-executable task units, and clarifies the "destination" (target job identifier) ​​and "task" (content) of each unit, laying the foundation for accurate routing in the future.

[0036] Subsequently, step S42, "Based on the target job identifier, query the job-terminal mapping table to determine the corresponding target terminal device address," addresses the "addressing" problem of the instruction. The job-terminal mapping table is a core configuration database or file that dynamically maintains the association between organizational roles and physical / logical computing devices. This table typically contains at least two columns: one is the "job identifier," whose value corresponds to the `target_role` value in the operation instruction; the other is the "terminal device address," the format of which depends on the network communication protocol. For example, for IP-based network communication, the address might be a combination of IP address and port number; for message queues, the address might be a subscription topic (e.g., "OR / Team / Nurse / Command"); for the hospital's internal device bus, the address might be a device code. Once the system extracts the target job identifier (e.g., "Anesthesiologist") from the instruction, it uses this as the query key to retrieve information from the mapping table. The query process may call a database query interface (e.g., an SQL query) or read a hash table in memory, resulting in one or more specific, network-addressable target terminal device addresses bound to that job. This design decouples personnel roles from specific devices, greatly improving the system's flexibility—when a doctor uses different workstations, only the address corresponding to that doctor's role in the mapping table needs to be updated, without modifying any program code.

[0037] After obtaining the exact network address, proceed to step S43: "Encapsulate and send the operation instructions according to the target terminal device address." This step implements the "packaging" and "delivery" of the instructions. Encapsulation refers to reorganizing the parsed instruction content (action_type and action_parameters) into a data packet conforming to the protocol specifications according to the communication protocol and application layer data format supported by the target terminal. For example, if the target terminal supports a RESTful API, the encapsulation process may convert the instruction content into JSON format and use it as the payload (body) of an HTTP POST request; if the MQTT protocol is used, it is encapsulated into a message payload (Payload) under a specified topic. The encapsulation process usually also adds necessary protocol header information, such as message sequence number, timestamp, checksum, etc., to ensure the reliability and traceability of transmission. The sending process is completed by the system's network communication module, which calls the underlying network socket library and sends the encapsulated data packet to the network through transport layer protocols such as TCP or UDP according to the target terminal device address (IP and port). For wireless terminals (such as tablets and augmented reality glasses), the sending will be carried out through a Wi-Fi access point. The terminal device has a corresponding service program that listens to the network port or subscribes to message topics. After receiving the data packet, it decapsulates it and finally executes the operation required by the instruction, such as popping up a prompt box, issuing a voice message, or controlling external devices.

[0038] Through the clear division and connection of these three steps, this implementation method constructs a robust, flexible, and efficient instruction distribution system. The line-by-line parsing of the instruction set ensures that the system can handle complex collaborative tasks with fine granularity, and each instruction can be managed and tracked independently. Addressing is achieved by querying a dynamic job-terminal mapping table, making the deployment and maintenance of the entire system extremely flexible. Personnel changes, equipment replacements, or role adjustments will not affect the core logic; only a configuration table needs to be maintained, which significantly reduces system maintenance costs and improves adaptability. Standardized encapsulation and transmission based on target addresses and protocols ensure that the system can seamlessly integrate with various heterogeneous hardware terminals and software systems in the operating room, achieving technical compatibility and scalability. The entire process begins with semantic parsing of instructions, proceeds through logical addressing, and ultimately transforms into data flow on the physical network, forming a reliable data transmission chain. This ensures that the collaborative instruction set generated by the intelligent decision-making center can accurately, promptly, and error-free reach every designated job terminal, transforming abstract collaborative strategies into synchronous, concrete actions distributed throughout the operating room. This is a key technological bridge ensuring the smooth operation of the entire multi-job collaborative management closed loop.

[0039] In some implementations, the terminal device includes augmented reality glasses; and after S43, it also includes: S44: Based on the received operation instructions, the augmented reality glasses drive the display module of the augmented reality glasses to render and display the operation instructions as text or icons overlaid on the real surgical field of view.

[0040] This embodiment further defines a specific form of the terminal device and its unique command presentation method, namely, introducing augmented reality glasses as a terminal for information reception and display. Augmented reality glasses (AR glasses) are a type of head-mounted transparent display device. Their core feature lies in the optical superposition of images generated by micro-display elements with the user's real-world view through transparent waveguides or optical lenses located in front of the eyes. This allows the user to see digital information (text, icons, 3D models, etc.) superimposed on the real environment while observing it. When such a device is connected to the system as a terminal device, it is assigned a job identifier (e.g., "chief surgeon" or "first assistant") and, like other terminals, registers its network address (usually a Wi-Fi IP address and the service port running on the device) in the job-terminal mapping table.

[0041] After instruction distribution step S43 is completed, the operation instruction has been sent to the augmented reality glasses. Step S44 describes the processing and feedback process on the glasses side: "Based on the received operation instruction, the augmented reality glasses drive the display module of the augmented reality glasses to render and display the operation instruction content as text or icons overlaid on the real surgical field of view." Specifically, after receiving the instruction data packet through the network interface, the computing unit built into the augmented reality glasses (such as an ARM processor) first decapsulates and parses it to extract the substantive content of the instruction. This content may be designed to include display metadata, such as: "display_type": "text", "content": "Patient vital signs fluctuate, please pay attention", "anchor": "world_space", "position": "near_patient_head"; or "display_type": "icon", "icon_id": "warning_triangular", "anchor": "screen_space", "position": "top_right".

[0042] Next, the "display module driving the augmented reality glasses" is a process involving hardware and software collaboration. The display module of augmented reality glasses typically includes a graphics processing unit (GPU) and a spatial positioning system. The spatial positioning system (implementing SLAM, i.e., simultaneous localization and mapping, through a built-in camera, inertial measurement unit, and depth sensor) tracks the six degrees of freedom pose (position and orientation) of the glasses in the operating room space in real time and builds an understanding of the surrounding environment. The graphics rendering engine (the core software of the display module) determines how the virtual information is presented based on the display type and anchoring information in the received instructions. If anchored to "world_space," the rendering engine "locks" the virtual text or icon to a specific three-dimensional coordinate in the real world (for example, a prompt icon floating above the patient's forehead), and the information appears fixed in that real location regardless of how the user moves their head. If anchored to "screen_space," the information is fixed in a corner of the display screen (such as the upper right corner) and moves with the user's field of view. Finally, the rendering engine projects the generated virtual image onto the transparent display lens via an optical engine, precisely blending it with the real surgical scene seen through the lens ("real surgical view"), completing the "overlay" and "rendering display." The wearer then sees prompts that appear to float in the real environment.

[0043] This implementation method revolutionizes the human-computer interaction paradigm for collaborative instructions by introducing spatial computing and augmented reality technologies. Augmented reality glasses, as wearable, hands-free interactive terminals, allow medical staff to receive information without shifting their gaze or interrupting their workflow while focusing on delicate surgical procedures or monitoring equipment. This perfectly meets the high demands of continuity and attention in sterile surgical environments. Transforming operational instructions into intuitive text or icons and overlaying them onto the real field of vision achieves deep spatial binding and contextual integration between information and the task scenario. This directly links the semantics of the instructions to the physical environment, significantly reducing the cognitive processing time required from receiving information to understanding its context. This intuitive, spatial presentation significantly reduces the risk of misreading or missing traditional text notifications in high-pressure, complex surgical environments, improving the accuracy and efficiency of information delivery. It creates a highly personalized, immersive information space for medical staff in each role, enabling collaborative instructions from the central system to be seamlessly integrated into their individual workflows in a natural, non-invasive, and highly context-relevant manner. Ultimately, this implementation method not only completes the transmission of information, but also optimizes the process of information consumption and digestion, upgrading collaborative management from the traditional "interruption-response" model to an "augmentation-perception" model, providing an optimal, augmented reality-driven interactive interface for rapid and accurate team collaboration at critical moments.

[0044] In some implementations, S3 includes: S31b: Perform a strategy index query based on real-time status information and the current surgical stage identifier obtained from the surgical scheduling system; S32b: Based on the query results, retrieve the corresponding collaborative strategy template from the strategy library; S33b: Use parameters from real-time status information to populate variables in the cooperative strategy template and instantiate a cooperative instruction set.

[0045] This embodiment provides another specific implementation of step S3, which emphasizes that the selection of collaborative strategies is closely linked to the current specific surgical stage, thereby enabling the system's decision-making to have a deep context-aware capability. This implementation begins with step S31b: "Based on real-time status information and the current surgical stage identifier obtained from the surgical scheduling system, perform a strategy index query." Here, the "surgical scheduling system" is the core module of the hospital's operating room information management system. It is not only responsible for scheduling the operating rooms but also typically defines standardized surgical process stages. These stage identifiers, such as "ANES_INDUCTION" (anesthesia induction), "INCISION_CORNEAL" (corneal incision), "PHACO_EMULSIFY" (phacoemulsification), "IOL_INSERT" (intraocular lens implantation), and "WOUND_CLOSURE" (incision closure), represent key links in the surgical process with different technical characteristics and risk points. The system retrieves the "current surgical stage identifier" corresponding to the currently ongoing surgery from the surgical scheduling system in real time through the application programming interface (API). The system then combines real-time status information (e.g., parsed "patient anxiety level: high; pupillary light reflex: sluggish") with this stage identifier (e.g., "PHACO_EMULSIFY") to form a composite query condition. This condition is used to execute a "policy index query" within a pre-built "policy library." The policy library can be a relational database table or a knowledge base of a rules engine, and its index key is typically a combination of "status category" and "surgical stage."

[0046] Subsequently, step S32b: "Based on the query results, call the corresponding collaboration strategy template from the strategy library." When the index query successfully matches a record, the "collaboration strategy template" pointed to by that record is loaded into memory. The collaboration strategy template is not the final instruction, but rather an instruction generation framework or blueprint containing variable placeholders. It describes which roles typically need to participate in collaboration and what type of operation needs to be performed under this specific combination of state and surgical stage, but the parameters of the specific operation are open.

[0047] Finally, step S33b: "Use parameters from the real-time status information to populate the variables in the collaborative strategy template and instantiate a collaborative instruction set." Besides category labels, the real-time status information typically includes specific quantitative or refined parameters. For example, the "High_Anxiety" status might be accompanied by an "Anxiety Index" value of 85 (out of 100), or the "Slow Pupil Light Reflex" status might be accompanied by a "Lateness Prolongation Percentage" of 50%. The task of step S33b is to populate these specific parameter values ​​into the variable placeholders of the loaded collaborative strategy template, thereby instantiating the abstract template into concrete, executable operation instructions. Continuing the example above, the system might, based on a preset mapping relationship, map "Anxiety Index 85" to comfort_level = "High-Intensity Reassurance Script," set assessment_focus to "Sedation Depth and Respiratory Rate," and generate a specific alert_message = "Patient Anxiety, Pay Attention to Operational Stability" based on the stage and status. After population, each action in the template is transformed into a complete operation instruction, and all these instructions together constitute the final "collaborative instruction set" sent.

[0048] This implementation method, which deeply integrates surgical stages into the decision-making logic, significantly improves the accuracy and clinical relevance of collaborative management. By introducing surgical stage identifiers, the system gains the ability to understand the key context of "where the surgery is currently progressing." This allows for differentiated interpretations of the same physiological state (such as pupillary changes) based on the specific surgical stage in which it occurs (e.g., whether it occurs during incision creation or during lens implantation), leading to more accurate clinical judgment. Strategy indexing based on the composite key of "state + stage" enables the system to quickly locate a response tailored to the specific situation from a structured knowledge base. This provides a reliable basis for decision-making and avoids the mismatch issues that may arise from generic strategies. Calling predefined collaborative strategy templates essentially solidifies best clinical practices and team collaboration models for various typical situations into the system, ensuring the professionalism and standardization of response solutions. Populating template variables with specific parameters from real-time data injects personalized and precise adjustments into the standardized response framework, ensuring that the generated instruction set conforms to the standard process while reflecting subtle differences in the current situation. Ultimately, this method endows the entire collaborative instruction generation mechanism with a high degree of contextual intelligence. It can dynamically generate multi-position collaborative instructions that best meet the current actual needs based on "how the patient is doing at this moment" and "what key step the surgery is in". This achieves a leap from static response to dynamic contextual response, significantly improving the timeliness and appropriateness of intervention measures, as well as the overall effectiveness and safety of the entire surgical team's collaborative operations.

[0049] In some implementations, the current surgical stage identifier obtained from the surgical scheduling system is acquired in the following way: Based on the operational status sensor data of key equipment in the operating room, analyze the start-up and shutdown status of the equipment; Based on the stage-marked operation log recorded on the doctor's operating terminal, the current stage is analyzed; Based on the analysis of the integrated equipment start-up and shutdown status and the results of the current stage analysis, a unique identifier for the current surgical stage is determined through arbitration logic.

[0050] This embodiment further reveals the specific method for obtaining the "current surgical stage identifier from the surgical scheduling system," providing a robust method that does not solely rely on the scheduling system's preset timeline but instead uses multi-source data fusion for real-time, automatic judgment. This implementation indicates that the current surgical stage identifier can be determined by combining two independent analysis paths and using arbitration logic. The first path is "analysis of equipment start-up and shutdown status based on the operational status sensor data of key equipment in the operating room." Key equipment directly related to ophthalmic surgery in the operating room, such as phacoemulsification / vitrectomy machines, surgical microscopes, front-end laser treatment devices, and corneal topography instruments, typically provides digital signals reflecting their operational status through their internal controllers. This "operational status sensor data" can be obtained in real-time through the equipment's communication interface (such as Ethernet or USB) or by added sensors (such as detectors monitoring laser light emission or acoustic sensors listening to the operating frequency of the phacoemulsification handpiece). The data may include the equipment's power-on / off status, current operating mode (such as "infusion / aspiration mode," "phacoemulsification mode," or "laser firing"), and key parameter settings (such as ultrasonic energy percentage and negative pressure setpoint). Time-series analysis of this data allows for the deduction of the surgical progress stages. For example, continuous detection of the phacoemulsification handpiece in a high-energy (e.g., 80%) intermittently triggered state, coupled with continuous operation of the irrigation pump, strongly suggests that the surgery is in the "lens nucleus emulsification and ablation" stage. Switching the microscope's optical path to "posterior reflective illumination" mode may indicate entry into the "posterior capsule polishing" stage. This analysis based on physical signals provides an objective and continuous lateral portrait of the surgical progress.

[0051] The second approach is to "analyze the current stage based on the stage-marked operation log recorded on the doctor's operating terminal." Doctor's operating terminals, such as touchscreen consoles, foot switch controllers, or voice input systems, typically integrate functionality that allows the surgical team to actively mark surgical stages, in addition to controlling the equipment. For example, the screen might have a set of buttons labeled "Incision Completed," "Water Separation Completed," "Cortical Aspiration Completed," and "Implantation Completed." Each time the doctor or scrub nurse clicks the corresponding button, a "stage-marked operation log" with a precise timestamp is recorded in the system's database or log file. This approach provides a direct statement of the surgical team's subjective perception and intent, offering the most explicit stage indication information. Analyzing these logs—that is, finding the latest valid stage-marked record—directly yields the "current surgical stage identifier" as perceived by the team.

[0052] Because a single data source may fail (device sensor malfunction, doctor forgets to manually mark) or be contradictory (device shows emulsification, but doctor just marked "implantation complete"), the following step is required: "Based on the combined analysis of device start / stop status and current stage resolution, a unique current surgical stage identifier is determined through arbitration logic." Arbitration logic is a set of predefined conflict resolution and decision-making rules. For example, a simple arbitration logic might be: First, check if a valid manual stage mark exists within a short time window (e.g., 30 seconds); if so, prioritize that mark as the current stage identifier, as it represents the operator's direct intent. If no manual mark exists, rely entirely on the stage inferred by the device status analysis module. More complex logic can handle conflicts: when the manual mark and device status inference are significantly inconsistent (e.g., manually marked "incision closed," but the device status continues to show "ultrasound emulsification energy output"), the arbitration logic can trigger a "stage conflict" warning, sending it to the circulating nurse terminal for manual confirmation; simultaneously, before receiving confirmation, the system can adopt a more conservative or general stage identifier (e.g., "surgery in progress"), or temporarily retain the previous undisputed stage identifier. Through such arbitration, the system can combine the advantages of objective signals and subjective statements to output a more reliable and unique identifier of the current surgical stage.

[0053] This multi-source sensing and intelligent arbitration approach significantly enhances the accuracy, real-time performance, and anti-interference capabilities of the system's autonomous judgment of surgical stages. By monitoring and analyzing the underlying operational status signals of key medical equipment, the system obtains an objective, physical-world-based progress indicator independent of any information system interface. This provides tamper-proof physical evidence for stage judgment, enhancing the system's perception depth. Simultaneously, actively collecting and parsing the marked logs from the doctor's operating terminal respects and incorporates the highest-level cognitive intent of the surgical team, ensuring the system's judgment is synchronized with human decision-making processes and improving the naturalness of human-machine collaboration. Designing and implementing arbitration logic to reconcile these two potentially inconsistent information sources demonstrates the system's intelligent decision-making ability in the face of uncertainty and conflicting information. It can assess the real-time credibility of different information sources, improve accuracy by leveraging complementarity when information is complete, and handle errors through rules when information is contradictory or missing, thus ensuring the robustness of the output results. This method reduces the absolute dependence on a single information source (such as the theoretical timeline of the scheduling system or the doctor's continuous error-free manual operation), making the acquisition of the "current surgical stage identifier" a dynamic, adaptive, and highly available process. Ultimately, it provides high-quality, highly reliable context fuel for the refined context-aware collaborative decision-making engine, ensuring that the system can more realistically and timely understand "the exact steps happening on the operating table," thus laying a solid and credible factual foundation for generating the most appropriate collaborative instructions.

[0054] Example 2 like Figure 2 As shown, in a second aspect, the present invention provides an ophthalmic surgical management system based on eye movement and pupil analysis. The system is used to execute the method provided in any of the above embodiments, and includes: The data acquisition module is used to perform step S1: based on the sensing device of the surgical environment, it collects physiological data of the patient's eyes and generates eye physiological data; The state analysis module is used to execute step S2: analyze and process the eye movement and / or pupil change characteristics reflected in the ocular physiological data to generate real-time state information representing the patient's current physiological and psychological state; wherein, S2 includes: S21a: Extract eye movement trajectory sequences from ocular physiological data; S22a: Perform pattern recognition processing on the eye movement trajectory sequence to obtain the patient's eye movement pattern characteristics; S23a: Based on eye movement pattern features, match and classify them with multiple reference patterns in the database to generate real-time status information; and / or S21b: Extract time-series data of pupil diameter from ocular physiological data; S22b: Calculate and analyze the rate of change and stability of pupil diameter time series data to generate pupil response characteristic parameters; S23b: Identify and classify the pupil response patterns to light or drug stimulation reflected in the pupil response characteristic parameters to generate real-time status information characterizing the patient's specific physiological state. The collaborative instruction generation module is used to execute step S3: based on real-time status information and combined with preset collaborative strategies, perform analysis and mapping processing to generate a collaborative instruction set containing at least two operation instructions that point to different positions respectively; The instruction distribution module is used to execute step S4: based on the collaborative instruction set and the preset correspondence between positions and terminals, perform instruction distribution processing and send each operation instruction to the terminal device of the corresponding position.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for managing ophthalmic surgery based on eye movement and pupil analysis, characterized in that, include: S1: Based on the surgical environment, the sensing device collects physiological data of the patient's eyes and generates eye physiological data; S2: Analyze and process the eye movement and / or pupillary change characteristics reflected in the ocular physiological data to generate real-time status information representing the patient's current physiological and psychological condition; wherein, S2 includes: S21a: Extract eye movement trajectory sequences from ocular physiological data; S22a: Perform pattern recognition processing on the eye movement trajectory sequence to obtain the patient's eye movement pattern characteristics; S23a: Based on eye movement pattern features, match and classify them with multiple reference patterns in the database to generate real-time status information; and / or S21b: Extract time-series data of pupil diameter from ocular physiological data; S22b: Calculate and analyze the rate of change and stability of pupil diameter time series data to generate pupil response characteristic parameters; S23b: Identify and classify the pupil response patterns to light or drug stimulation reflected in the pupil response characteristic parameters to generate real-time status information characterizing the patient's specific physiological state. S3: Based on real-time status information and combined with preset collaboration strategies, perform analysis and mapping processing to generate a collaboration instruction set containing at least two operation instructions that point to different positions. S4: Based on the collaborative instruction set and the preset correspondence between positions and terminals, perform instruction distribution processing and send each operation instruction to the terminal device of the corresponding position.

2. The ophthalmic surgery management method based on eye movement and pupil analysis according to claim 1, characterized in that, S3 include: S31a: Determine the status category of real-time status information; S32a: When the status category is determined to be the first category, the first instruction generation logic is triggered to generate a first set of collaborative instructions, which includes sending a voice reassurance instruction to the nurse's terminal and sending an instruction to increase the monitoring frequency to the anesthesiologist's terminal. S33a: When the state category is determined to be the second category, the second instruction generation logic is triggered to generate a second set of collaborative instructions, which includes sending a pause instruction to the surgical equipment control system and sending an optical positioning adjustment instruction to the doctor's operating terminal; wherein, the first set of collaborative instructions and the second set of collaborative instructions constitute a collaborative instruction set.

3. The ophthalmic surgery management method based on eye movement and pupil analysis according to claim 1, characterized in that, Before S1, it also includes: S0: Based on the ocular physiological data and corresponding manually labeled state data from historical surgical cases, the machine learning model is trained and validated to obtain the trained state analysis model; S2 can also be: inputting eye physiological data into the state analysis model, and generating real-time state information through the forward computation processing of the state analysis model.

4. The ophthalmic surgery management method based on eye movement and pupil analysis according to claim 3, characterized in that, S0 includes: S0a: Filter, denoise, and extract features from historical eye physiological data to generate a standardized historical feature dataset; S0b: The standardized historical feature dataset and the corresponding manually labeled state data are used as training sample pairs and input into the initial machine learning model; S0c: The internal parameters of the initial machine learning model are adjusted through an iterative optimization algorithm until the difference between its output and the manually labeled state data meets the accuracy requirements, thus generating a trained state analysis model.

5. The ophthalmic surgery management method based on eye movement and pupil analysis according to claim 1, characterized in that, S4 include: S41: Analyze the content of each operation instruction in the collaborative instruction set and the target job identifier; S42: Based on the target job identifier, query the job-terminal mapping table to determine the corresponding target terminal device address; S43: Encapsulate and send the operation instructions based on the target terminal device address.

6. The ophthalmic surgery management method based on eye movement and pupil analysis according to claim 5, characterized in that, Terminal devices include augmented reality glasses; Following S43 are: S44: Based on the received operation instructions, the augmented reality glasses drive the display module of the augmented reality glasses to render and display the operation instructions as text or icons overlaid on the real surgical field of view.

7. The ophthalmic surgery management method based on eye movement and pupil analysis according to claim 1, characterized in that, S3 include: S31b: Perform a strategy index query based on real-time status information and the current surgical stage identifier obtained from the surgical scheduling system; S32b: Based on the query results, retrieve the corresponding collaborative strategy template from the strategy library; S33b: Use parameters from real-time status information to populate variables in the cooperative strategy template and instantiate a cooperative instruction set.

8. The ophthalmic surgery management method based on eye movement and pupil analysis according to claim 7, characterized in that, The current surgical stage identifier, obtained from the surgical scheduling system, is acquired through the following methods: Based on the operational status sensor data of key equipment in the operating room, analyze the start-up and shutdown status of the equipment; Based on the stage-marked operation log recorded on the doctor's operating terminal, the current stage is analyzed; Based on the analysis of the integrated equipment start-up and shutdown status and the results of the current stage analysis, a unique identifier for the current surgical stage is determined through arbitration logic.

9. An ophthalmic surgical management system based on eye movement and pupil analysis, characterized in that, The system is used to perform the method according to any one of claims 1 to 8, the system comprising: The data acquisition module is used to perform step S1: based on the sensing device of the surgical environment, it collects physiological data of the patient's eyes and generates eye physiological data; The state analysis module is used to execute step S2: analyze and process the eye movement and / or pupil change characteristics reflected in the ocular physiological data to generate real-time state information representing the patient's current physiological and psychological state; wherein, S2 includes: S21a: Extract eye movement trajectory sequences from ocular physiological data; S22a: Perform pattern recognition processing on the eye movement trajectory sequence to obtain the patient's eye movement pattern characteristics; S23a: Based on eye movement pattern features, match and classify them with multiple reference patterns in the database to generate real-time status information; and / or S21b: Extract time-series data of pupil diameter from ocular physiological data; S22b: Calculate and analyze the rate of change and stability of pupil diameter time series data to generate pupil response characteristic parameters; S23b: Identify and classify the pupil response patterns to light or drug stimulation reflected in the pupil response characteristic parameters to generate real-time status information characterizing the patient's specific physiological state. The collaborative instruction generation module is used to execute step S3: based on real-time status information and combined with preset collaborative strategies, perform analysis and mapping processing to generate a collaborative instruction set containing at least two operation instructions that point to different positions. The instruction distribution module is used to execute step S4: based on the collaborative instruction set and the preset correspondence between positions and terminals, perform instruction distribution processing and send each operation instruction to the terminal device of the corresponding position.

Images