Ophthalmology patient intelligent follow-up visit and nursing method for postoperative recovery

Abstract

The invention discloses an ophthalmology patient intelligent follow-up visit and nursing method for postoperative recovery, and relates to the technical field of intelligent medical health management, and the method comprises the steps: collecting postoperative multi-modal monitoring data, and carrying out the time-space synchronization and feature extraction to generate a standardized feature set; and inputting the feature set into a preset complication risk assessment model, and outputting a recovery process score, a complication probability and a risk level. And according to an evaluation result, a matched nursing intervention scheme library is retrieved from the nursing knowledge graph, a personalized nursing plan is generated in combination with personal data of the patient, and nursing guidance and follow-up visit reminding are dynamically optimized and output based on real-time monitoring data. According to the method, continuous and accurate evaluation of the recovery state and dynamic personalized generation of the nursing scheme are realized, and the accuracy and adaptability of postoperative management are improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent medical and health management technology, specifically an intelligent follow-up and nursing method for ophthalmology patients undergoing postoperative recovery. Background Technology

[0002] Postoperative monitoring and follow-up of ophthalmology patients during recovery primarily rely on regular outpatient visits. Current technology cannot achieve continuous data collection and comprehensive analysis of the recovery process, resulting in time-bound blind spots in assessment. Information obtained from outpatient visits is discrete and cannot accurately reflect the dynamic changes in vital signs during home recovery. Conventional remote monitoring systems often trigger alarms based on thresholds of single indicators, lacking the fusion and collaborative analysis of multi-dimensional information such as ocular signs, physiological parameters, and subjective symptoms. This leads to biased assessment conclusions and an inability to form a holistic quantitative judgment of the recovery process.

[0003] In terms of nursing interventions, existing protocols are typically static, general guidelines, or rely on manual adjustments based on the experience of healthcare professionals. This model struggles to match patients' constantly changing risk profiles and individualized needs. Existing medical knowledge bases are mostly item-based, lacking structured modeling and dynamic reasoning capabilities to understand the complex relationships between symptoms, nursing interventions, contraindications, and other elements. This results in delays in the generation and adjustment of nursing protocols, hindering accurate and automatic matching and adaptation to continuous risk assessment results.

[0004] There is a need for a technological approach that can continuously and multidimensionally assess postoperative recovery status intelligently, and dynamically generate and adjust personalized nursing plans accordingly. This approach needs to overcome the shortcomings of existing technologies, such as one-sided assessments and static interventions, and achieve closed-loop automated management from data collection and risk analysis to nursing decision-making, thereby improving the accuracy and efficiency of postoperative management. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes an intelligent follow-up and nursing method for ophthalmic patients undergoing postoperative recovery, comprising: Collect multimodal continuous monitoring data of ophthalmic patients during postoperative recovery; The multimodal continuous monitoring data is subjected to spatiotemporal synchronization and feature extraction processing to generate a standardized postoperative recovery feature set, which includes quantitative indicators of ocular signs, physiological state trend indicators, and subjective symptom feature vectors. The standardized set of postoperative recovery features is input into a preset complication risk assessment model, which outputs a recovery progress score, a probability value of potential complications, and a risk level label. Based on the recovery progress score, the probability value of potential complications, and the risk level marker, a matching nursing intervention program library is retrieved from the nursing knowledge graph; Based on the patient's personal historical data and real-time feedback, the plans in the nursing intervention plan library are personalized and adjusted to generate a personalized nursing plan. The personalized care plan is dynamically optimized by combining the latest monitoring data stream, generating executable nursing instructions and follow-up reminders.

[0006] Furthermore, the spatiotemporal synchronization and feature extraction processing of the multimodal continuous monitoring data to generate a standardized set of postoperative recovery features includes: The multimodal continuous monitoring data includes eye image sequences taken by the patient, time series of physiological parameters recorded by wearable devices, and self-symptom assessment reports entered by the patient through an interactive interface. The eye image sequence is subjected to image quality assessment and screening, and images that are blurry or have unacceptable angles are removed, while qualified eye images are retained; The qualified ocular images are segmented into lesion areas, analyzed for hyperemia, and feature extracted from healing areas. The quantitative indicators of ocular signs are calculated, including the proportion of inflammatory area, neovascularization density, and corneal edema index. The time series of physiological parameters recorded by the wearable device is aligned with the shooting time of each qualified eye image, and the statistical features of physiological parameters within the corresponding time window are extracted to form the physiological state trend index, which includes intraocular pressure fluctuation curve, heart rate variability rate and sleep quality index. Natural language processing is performed on the self-symptom assessment report to extract keywords and quantify symptom intensity, constructing the subjective symptom feature vector, which includes numerical representations of pain level, photophobia degree, and blurred vision score.

[0007] Further, the standardized set of postoperative recovery features is input into a preset complication risk assessment model, which outputs a recovery progress score, a probability value of potential complications, and a risk level label, including: The quantitative indicators of eye signs are input into the visual feature analysis sub-network of the model. The visual feature analysis sub-network is based on a convolutional neural network structure and is used to assess the degree of abnormality of the signs and output the abnormality score of the signs. The physiological state trend indicators are input into the time series analysis submodule of the model. The time series analysis submodule analyzes the changing trend and stability of the indicators and outputs the physiological risk coefficient. The subjective symptom feature vector is input into the subjective feedback fusion unit of the model. The subjective feedback fusion unit combines the patient's historical baseline data to assess the degree to which the symptoms deviate from the normal recovery trajectory and outputs the symptom deviation degree. The recovery process score is calculated by integrating the abnormal vital signs score, the physiological risk coefficient, and the symptom deviation score through an ensemble learning algorithm. Based on the abnormal physical signs score, physiological risk coefficient, and symptom deviation, the probability of infection risk, the probability of abnormal intraocular pressure, and the probability of delayed healing are obtained by querying a preset probability mapping table, and the highest value is taken as the probability value of potential complications. Based on the recovery progress score and the numerical range of the potential complication probability value, the corresponding risk level marker is matched from a preset rule base.

[0008] Furthermore, the step of retrieving a matching nursing intervention program library from the nursing knowledge graph based on the recovery progress score, potential complication probability value, and risk level marker includes: Using the risk level marker as the primary search key, the corresponding nursing strategy branch is located in the nursing knowledge graph; Using the recovery process score as a secondary screening criterion, a set of nursing plans applicable to the recovery stage is selected under the identified nursing strategy branch. Using the specific risk type indicated by the probability value of the potential complication as the guide for optimizing the plan, candidate plans containing targeted interventions are further screened from the set of nursing plans. The selected candidate solutions are analyzed in a structured manner to extract the core steps, applicable conditions, required resources and expected goals of the solutions, forming a preliminary nursing intervention solution library.

[0009] Furthermore, based on the patient's personal historical data and real-time feedback, the nursing intervention plan library is personalized and adjusted to generate a personalized nursing plan, including: Retrieve the patient's personal history data, which includes past medical history, drug allergy history, surgical details, and past recovery patterns; Each plan in the nursing intervention plan library is compared with the patient's personal historical data for compatibility, and plans that are contraindicated or have high risk of incompatibility are eliminated. Obtain real-time feedback from patients regarding their preferences for initial treatment options. This real-time preference feedback includes preferences for medication timing, convenience of follow-up visits, and acceptance of nursing procedures. Based on the real-time preference feedback, the execution details of the remaining plans are adjusted, including adjusting the medication reminder time, adjusting the priority of follow-up examination items, and adjusting the form of health education materials; By integrating the compatibility comparison results with the adjusted execution details, a personalized care plan is customized for the patient, including a specific execution schedule, key points of operation, and precautions.

[0010] Furthermore, obtaining real-time feedback on the patient's preferences for the initial treatment options includes: The core content and options of the selected nursing intervention plan library are presented to the patient through an interactive interface. Design a feedback collection method that combines structured and unstructured approaches. The structured approach includes questionnaire ratings with pre-defined options, while the unstructured approach includes open-ended voice or text input. It receives and analyzes patient feedback in real time, performs statistical analysis on structured scores, and performs sentiment analysis and key intent extraction on unstructured input. By combining the results of structured scoring with the analysis of unstructured inputs, the patient's acceptance of different treatment plans, tolerance to different nursing procedures, and preference for different follow-up frequencies are quantified to form the aforementioned real-time preference feedback.

[0011] Further, the subjective symptom feature vector is input into the model's subjective feedback fusion unit, which combines the patient's historical baseline data to assess the degree to which the symptoms deviate from the normal recovery trajectory and outputs the symptom deviation degree, including: Standardized change curves of various subjective symptoms during the typical postoperative recovery period corresponding to the type of ophthalmic surgery were extracted from the patient's historical baseline data. The scores of each item in the subjective symptom feature vector at the current time point are compared with the standard values ​​at the corresponding time points on the standard change curve, and the instantaneous deviation value of each symptom is calculated. Based on the weight of each symptom in the complication warning, the weighted sum of each instantaneous deviation value is calculated to obtain the comprehensive instantaneous deviation score; Analyze the changing trend of the subjective symptom feature vector over a period of time, calculate the angle or distance between the trend and the standard recovery trend, and obtain the trend deviation score; The comprehensive instantaneous deviation score and the trend deviation score are combined according to a preset ratio to obtain the final symptom deviation degree.

[0012] Furthermore, the personalized care plan is dynamically optimized by incorporating the latest monitoring data stream to generate executable nursing instructions and follow-up reminders, including: During the execution cycle of the personalized care plan, the latest monitoring data stream is continuously received and processed; The latest processed data is input into the complication risk assessment model to obtain an updated recovery progress score, potential complication probability value, and risk level label. Compare the updated data with the data used when the original plan was developed. If the risk level marker is upgraded or the probability value of potential complications exceeds the new threshold, the emergency revision process of the care plan is triggered. The emergency revision process includes: retrieving high-risk intervention plans from the nursing knowledge graph based on the new risk indications, and quickly integrating them into the original plan to form a revised nursing plan; If the risk does not change significantly, the intensity and frequency of subsequent care activities in the original plan will be fine-tuned based on the updated recovery progress score. The finalized nursing plan is translated into specific daily or weekly nursing instructions and follow-up reminders, and distributed through designated communication channels.

[0013] Furthermore, the selected candidate solutions undergo structured analysis to extract the core steps, applicable conditions, required resources, and expected goals of each solution, forming a preliminary nursing intervention solution library, including: Natural language processing is performed on the text description of each candidate solution to identify and label the key entities and relationships describing nursing actions, medication information, examination items, frequency and cycle; Based on the annotation results, the unstructured text solution is converted into a structured solution template. The solution template contains fixed fields, which are used to fill in the core steps, applicable conditions, required resources and expected goals. From the converted structured solution template, the list of key operation points of the core steps, the logical judgment statements of the applicable conditions, the list of required resources, and the quantitative indicators of the expected goals are automatically extracted. All candidate solutions are stored in a relational database with structured information and indexed according to dimensions such as solution identifier, risk level, and recovery stage, forming a nursing intervention solution library that can be queried and invoked programmatically.

[0014] Furthermore, the process of translating the finalized nursing plan into specific daily or weekly nursing guidance instructions and follow-up reminders, and distributing them through designated communication channels, includes: The finalized care plan is analyzed and broken down into atomic tasks with clearly defined execution time, content, and target. For each atomic task, generate corresponding natural language descriptions or graphical guidance content to form the nursing guidance instructions; For tasks requiring active patient participation, such as follow-up examinations and medication administration, a reminder message containing specific time, location, and precautions is generated, forming the aforementioned follow-up reminder; Based on the patient's preset channel preferences, different instructions and reminders are distributed to corresponding terminals, including smartphone application push notifications, SMS messages, telephone voice notifications, or emails; Establish a confirmation and feedback mechanism for instructions and reminders, track patients' receipt and confirmation status, and provide repeated reminders or escalation notifications for unconfirmed tasks.

[0015] Compared with the prior art, the beneficial effects of the present invention are: A standardized set of postoperative recovery features is input into a pre-defined complication risk assessment model. This model uses an ensemble algorithm to jointly analyze multi-dimensional features, simultaneously outputting a recovery progress score, a potential complication probability value, and a risk level label. This represents a shift from single-point alarms to a comprehensive situational awareness. The recovery progress score provides a continuous quantitative view of the longitudinal recovery trend, enabling healthcare professionals to accurately grasp the overall trajectory. The potential complication probability value provides a quantitative prediction for specific risks, improving the sensitivity and specificity of early warnings. The risk level label serves for rapid clinical triage decisions. This multi-dimensional parallel output provides clinical judgment with a much richer and more hierarchical basis for decision-making than a single indicator, enhancing the comprehensiveness and precision of the assessment.

[0016] Based on the multidimensional assessment results output by the model, a matching nursing intervention protocol library is retrieved from the nursing knowledge graph. The knowledge graph structures the semantic relationships between concepts such as signs, symptoms, complications, nursing procedures, and contraindications. The system uses real-time risk markers and probability values ​​to reason and retrieve information through relational paths within the knowledge graph network, dynamically associating and combining the set of intervention measures most relevant to the current patient's condition. Based on the patient's personal historical data and real-time feedback, the protocol library is personalized and adjusted. This represents an evolution of nursing protocols from static, general templates to dynamically personalized generation. The protocol matching process is data-driven, based on semantic associations rather than simple rules, ensuring the adaptability of intervention recommendations to complex clinical situations and enabling nursing guidance to keep pace with the dynamic changes in the patient's recovery status. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of the intelligent follow-up and nursing method for ophthalmic patients oriented towards postoperative recovery as described in this invention. Figure 2 A flowchart for generating a standardized set of postoperative recovery features; Figure 3 A flowchart for searching the nursing intervention protocol database; Figure 4 Comparison chart of elements for personalized adaptation of postoperative ophthalmic care plans; Figure 5 This is a chart showing the trend of ocular signs recovery after ophthalmic surgery. Detailed Implementation

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

[0019] Please see Figure 1 Multimodal continuous monitoring data is collected during the postoperative recovery period of ophthalmology patients. This data includes, but is not limited to, ocular images, physiological parameters, and patient subjective reports. The collected multimodal continuous monitoring data undergoes spatiotemporal synchronization and feature extraction processing to generate a standardized postoperative recovery feature set. This set encompasses quantitative indicators of ocular signs, physiological state trend indicators, and subjective symptom feature vectors. The standardized postoperative recovery feature set is input into a pre-built complication risk assessment model, which outputs a quantitative recovery progress score, a potential complication probability value, and a comprehensive risk level label. Based on the output recovery progress score, potential complication probability value, and risk level label, the system retrieves and matches corresponding nursing intervention plans from a pre-constructed nursing knowledge graph. Based on the patient's personal historical data and real-time feedback on the plans, the plans in the retrieved nursing intervention plan library are personalized and adjusted to generate a personalized nursing plan for the patient. Combined with the latest monitoring data stream, the ongoing personalized nursing plan is dynamically optimized, and executable nursing guidance instructions and follow-up reminders are generated and sent to the patient or nursing staff.

[0020] See Figure 2In one embodiment of the present invention, the multimodal continuous monitoring data includes a sequence of eye images taken autonomously by the patient using a smart device, a time series of physiological parameters recorded by a smart wearable device, and a self-symptom assessment report periodically filled out by the patient through a mobile application interface. The eye image sequence undergoes automated image quality assessment and screening, removing blurry images due to shakiness, out-of-focus issues, or shooting angles that do not meet preset requirements, retaining qualified eye images for subsequent analysis. Automated processing is performed on the qualified eye images, including lesion area segmentation, hyperemia analysis, and healing area feature extraction. Quantitative indicators of eye signs are calculated, specifically including the proportion of inflamed area, neovascularization density, and corneal edema index. The time series of physiological parameters recorded by the wearable device is aligned with the precise timestamp of each qualified eye image, and statistical features of physiological parameters within a specific time window centered on that time point are extracted to form physiological state trend indicators, such as intraocular pressure fluctuation curves, heart rate variability, and sleep quality index. Natural language processing was used to extract keywords describing symptoms such as pain, photophobia, and blurred vision from the self-symptom assessment report and quantify their intensity to construct a numerical subjective symptom feature vector. This vector contains numerical representations of pain level, photophobia degree, and blurred vision score.

[0021] In practice, the implementation involves processing multimodal continuous monitoring data to generate a standardized set of postoperative recovery features. The multimodal continuous monitoring data comprises three components: the first is a sequence of eye images taken by the patient independently using a smartphone application according to preset specifications; the second is a time series of physiological parameters continuously recorded by wearable devices such as smart bracelets or dedicated intraocular pressure monitors; and the third is a self-assessment report of symptoms periodically filled out by the patient through the application's interactive interface or inputted via voice. These data are correlated and aggregated using timestamps and patient identifiers.

[0022] In some embodiments, when processing eye image sequences, the system performs image quality assessment and filtering operations on each frame. The assessment is based on a sharpness score and composition compliance judgment. The sharpness score is calculated by analyzing the high-frequency component energy after performing a Fast Fourier Transform on the image. The composition compliance judgment is achieved by detecting whether there is a complete outline of both eyes in the image and whether the iris region is located in the center region of the image. The formula for calculating the sharpness score S_clarity is:

[0023] in: It is the result of the two-dimensional discrete Fourier transform of the image grayscale matrix; arrive and arrive A preset high-frequency component region is defined. Images with a sharpness score below the threshold or with non-compliant composition are marked as blurry or have an angle that does not meet the requirements and are removed from the subsequent process. Only eye images with a sharpness score that meets the standard and with compliant composition are retained as qualified eye images.

[0024] In practice, qualified ocular images are segmented into lesion regions, analyzed for hyperemia, and feature extracted from healing regions. Lesion region segmentation uses a semantic segmentation model based on the U-Net architecture to identify inflammatory or surgical incision regions in the image. Hyperemia analysis is achieved by extracting the red component intensity of the sclera region and comparing it with a healthy baseline. Feature extraction from healing regions focuses on the smoothness of the incision edge and the morphology of neovascularization. Based on these analysis results, quantitative indicators of ocular signs are calculated. Specifically, the quantitative indicators of ocular signs are quantified as the proportion of the area of ​​the inflammatory region to the entire visible sclera and cornea region as the proportion of the inflammatory region area, the total length of neovascularization per unit area as the neovascularization density, and the uniformity parameter of the grayscale texture of the corneal region as the corneal edema index.

[0025] It is understandable that when processing the time series of physiological parameters recorded by wearable devices, the system uses the precise shooting time of each qualified eye image as a reference point, extracts data segments within a specific time window before and after that time point from the physiological parameter time series, such as continuous intraocular pressure recordings from two hours before to one hour after the shooting time point, performs feature extraction operations on this data segment, and extracts features including the mean, standard deviation, maximum and minimum values ​​of intraocular pressure within the window to form an intraocular pressure fluctuation curve, analyzes the time domain and frequency domain features of the heartbeat interval sequence in the same window to calculate the heart rate variability, and integrates the sleep stage data of the previous night to generate a sleep quality index. These extracted statistical features together constitute a physiological state trend indicator.

[0026] In some embodiments, processing self-symptom assessment reports involves a natural language processing workflow. The system performs word segmentation, part-of-speech tagging, and entity recognition on the text reports input by the patient, identifying keywords such as "pain," "photophobia," and "blurred vision" and their modifying adverbs. For reports input by voice, speech recognition is performed first to convert the speech into text before processing. The system pre-sets a symptom intensity mapping dictionary, mapping the identified keywords and adverbs to numerical values. For example, "mild pain" is mapped to the numerical value 1, and "severe pain" is mapped to the numerical value 5. In this way, the text or voice report is converted into a structured numerical vector, namely the subjective symptom feature vector. The subjective symptom feature vector contains numerical values ​​for pain level, photophobia degree, and blurred vision score.

[0027] Optionally, multimodal data must undergo strict spatiotemporal synchronization before feature extraction. The system maintains a unified time server, and all data acquisition terminals must acquire and record this unified timestamp when collecting data. When the patient's smartphone application captures an eye image, it simultaneously records the latest physiological parameter snapshot value transmitted from the wearable device via Bluetooth at that moment. The timestamp of the submission time is also recorded when the self-symptom assessment report is submitted. The data processing backend aligns all data according to the unified timestamp to ensure that the quantitative indicators of eye signs, physiological state trend indicators, and subjective symptom feature vectors are comparable in the time dimension.

[0028] It is understandable that generating a standardized set of postoperative recovery features is an integration process. The system encapsulates the calculated quantitative indicators of ocular signs, physiological trend indicators, and constructed subjective symptom feature vectors according to a predefined field order and data structure, and normalizes the numerical indicators to eliminate the influence of different dimensions. Finally, it outputs a structured data object containing all standardized feature vectors. This structured data object is the standardized set of postoperative recovery features input to the subsequent model.

[0029] In one embodiment of the present invention, standard change curves of each subjective symptom within the typical postoperative recovery period corresponding to the ophthalmic surgery type are extracted from the patient's historical baseline data. The scores of each item in the subjective symptom feature vector at the current time point are compared with the standard reference values ​​at the corresponding time points on the standard change curves to calculate the immediate deviation value for each symptom. Based on the preset weight of each symptom in complication warning, the immediate deviation values ​​are weighted and summed to calculate a comprehensive immediate deviation score. The trend of change of the subjective symptom feature vector over a continuous period is analyzed, and the angle or distance between this trend and the standard recovery trend is calculated to obtain a trend deviation score. The calculated comprehensive immediate deviation score and the trend deviation score are fused according to a preset ratio to finally obtain the symptom deviation degree. This symptom deviation degree is then input into the subjective feedback fusion unit of the complication risk assessment model for subsequent comprehensive risk assessment.

[0030] In practice, the implementation method involves a specific process for calculating symptom deviation, which is a quantitative indicator used by the complication risk assessment model to evaluate the degree to which a patient's subjective symptoms deviate from the normal recovery trajectory. The system extracts standard change curves of each subjective symptom within the typical postoperative recovery period corresponding to the type of ophthalmic surgery from the stored patient historical baseline database. These standard change curves are generated based on statistical modeling of historical recovery data from a large number of patients undergoing similar surgeries, and are smooth curves with time on the horizontal axis and symptom intensity on the vertical axis.

[0031] In some embodiments, the scores of each item in the subjective symptom feature vector at the current time point are compared with the standard values ​​at the corresponding time points on the standard change curve. For each symptom in the subjective symptom feature vector, such as pain level, the system reads its current score value and simultaneously queries the standard value of the pain level corresponding to the same postoperative days on the standard change curve. The absolute value of the difference between the current score value and the standard value is calculated as the instantaneous deviation value for each symptom. Based on the preset weight of each symptom in complication warning, the instantaneous deviation values ​​are weighted and summed to calculate the comprehensive instantaneous deviation score. The formula is:

[0032] in; This represents the total number of symptom items contained in the subjective symptom feature vector; Indicates the first The preset weights of the symptoms and satisfy the following conditions ; Indicates the first The current score value of the symptom; This indicates the first value obtained from the standard variation curve. The standard values ​​for each symptom at the corresponding time point; That is, the first The immediate deviation value of each symptom.

[0033] Understandably, analyzing the changing trend of subjective symptom feature vectors over a period of time requires collecting subjective symptom feature vectors from multiple consecutive time points. The system extracts the score data of the same symptom from the seven most recent time points to form a short-term score sequence for that symptom. For each symptom, the system compares the short-term score sequence with the standard value sequence for the corresponding time period on the standard change curve. The trend similarity is quantified by calculating the cosine of the angle between the two sequence vectors in multidimensional space or the Euclidean distance. The trend deviation score is defined as 1 minus the trend similarity; therefore, the larger the trend deviation score, the greater the difference between the patient's actual symptom change trend and the standard recovery trend.

[0034] Optionally, the trend deviation score can also be calculated using regression analysis. The system performs linear fitting on the patient's short-term score sequence and the standard value sequence, respectively, to obtain the slopes of the two fitted lines. The trend deviation score can be obtained by calculating the absolute value of the difference between the slopes of the two fitted lines and then dividing it by the absolute value of the slope of the standard fitted line. This method is particularly suitable for assessing whether the recovery speed is accelerating or delayed.

[0035] In practice, the comprehensive immediate deviation score and the trend deviation score are fused according to a preset ratio to obtain the final symptom deviation degree. The preset ratio is determined during the model training phase through grid search and cross-validation; for example, the comprehensive immediate deviation score accounts for 60%, and the trend deviation score accounts for 40%. The fusion process is a linear weighted summation process, where the symptom deviation degree equals the comprehensive immediate deviation score multiplied by the weight coefficient α plus the trend deviation score multiplied by the weight coefficient (1-α). The final symptom deviation degree is a scalar value normalized to between 0 and 1. The higher the symptom deviation degree value, the more severe the deviation of the patient's subjective symptoms from the normal recovery trajectory.

[0036] In some embodiments, the subjective feedback fusion unit receives the calculated symptom deviation and an individualized correction factor extracted from the patient's historical baseline data. The individualized correction factor considers the impact of the patient's age and underlying diseases on pain sensitivity. The subjective feedback fusion unit multiplies the symptom deviation by the individualized correction factor to fine-tune the symptom deviation. The fine-tuned symptom deviation is then output to the subsequent fusion module of the complication risk assessment model, where it participates in the comprehensive recovery progress score calculation along with the physical sign abnormality score and physiological risk coefficient.

[0037] See Figure 3 In one embodiment of the present invention, risk level markers are used as the primary search key to locate the corresponding nursing strategy branch node in the network structure of the nursing knowledge graph. Recovery progress scores are used as a secondary filtering condition to filter a set of nursing plans suitable for that recovery stage under the located nursing strategy branch node. The specific risk type with the highest probability among potential complication probabilities is used as the plan optimization guide to further filter candidate plans containing targeted interventions from the above set of nursing plans. Natural language processing is performed on the text description of each selected candidate plan to identify and annotate key entities and relationships describing nursing actions, medication information, examination items, frequency, and cycle. Based on the annotation results, the unstructured text plan is converted into a structured plan template, which contains fixed fields for filling in core steps, applicable conditions, required resources, and expected goals. From the converted structured plan template, the list of key operational points for the core steps, logical judgment statements for applicable conditions, a list of required resources, and quantitative indicators of expected goals are automatically extracted. All candidate protocols are stored in a relational database with structured information and indexed according to multiple dimensions such as protocol identifier, risk level, and recovery stage, forming a nursing intervention protocol library that can be queried and invoked programmatically.

[0038] In practice, the implementation involves retrieving and structurally parsing a matching nursing intervention plan library from the nursing knowledge graph based on recovery progress scores, potential complication probability values, and risk level markers. Risk level markers may be "low," "medium," or "high." The recovery progress score is a value ranging from 0 to 100. Potential complication probability values ​​include probabilities of multiple specific risk types, such as infection risk, intraocular pressure abnormality risk, and delayed healing risk. The system uses the risk level marker as the primary search key to locate the corresponding nursing strategy branch in the nursing knowledge graph. The nursing knowledge graph is stored in a graph structure, where nodes represent nursing concepts or strategies, and edges represent relationships between concepts. For example, when the risk level marker is "medium," the system locates the node labeled "medium-risk nursing strategy" and its associated sub-branches in the graph.

[0039] In some embodiments, the recovery progress score is used as a secondary filtering criterion to filter a set of nursing plans applicable to the recovery stage under the located nursing strategy branch. The recovery progress score is mapped to discrete recovery stages; for example, a score of 0-30 corresponds to the "early recovery stage," a score of 31-70 corresponds to the "intermediate recovery stage," and a score of 71-100 corresponds to the "late recovery stage." The system traverses all nursing plan nodes under the located branch. Each nursing plan node is associated with an attribute label for the applicable recovery stage. Nursing plan nodes whose attribute labels match the current patient's recovery stage are filtered out to form a preliminary set of nursing plans. In a specific example scenario, the patient's recovery progress score is 65, which is mapped to the "intermediate recovery stage." Under the "intermediate risk nursing strategy" branch, the system filters out all nursing plans applicable to the "intermediate recovery stage," such as the "intermediate corneal repair nursing plan" and the "intermediate inflammation monitoring plan."

[0040] Understandably, the system uses the specific risk type indicated by the probability value of potential complications as a guide for optimizing the nursing plan. It further filters candidate plans from the set of nursing plans, comparing the probabilities of infection risk, abnormal intraocular pressure risk, and delayed healing risk, selecting the risk type with the highest value as the dominant risk. Each nursing plan node in the set of plans also includes intervention tags for different risk types. The system selects plans whose intervention tag includes the dominant risk as candidate plans. Data comparison shows that when the patient's infection risk probability is 0.8, the abnormal intraocular pressure risk probability is 0.3, and the delayed healing risk probability is 0.5, the dominant risk is "infection risk." The system filters plans from the initial nursing plan set that contain the "anti-infection intervention" tag; for example, the "intermediate corneal repair nursing plan" may contain this tag, while the "intermediate inflammation monitoring plan" is excluded if it does not.

[0041] In practice, the selected candidate solutions undergo structured parsing to extract the core steps, applicable conditions, required resources, and expected goals. The system applies natural language processing (NLP) to the text description of each candidate solution. For example, the text description might be: "Use antibiotic eye drops three times daily for one week; avoid rubbing your eyes; check your intraocular pressure weekly." The NLP process includes word segmentation, named entity recognition, and relation extraction. It identifies and labels key entities and relations describing nursing actions, medication information, examination items, frequency, and cycle. For instance, it identifies "antibiotic eye drops" as a medication entity, "three times daily" as a frequency entity, and "avoid rubbing your eyes" as a nursing action entity.

[0042] In some embodiments, unstructured text schemes are converted into structured scheme templates based on annotation results. The scheme templates contain fixed fields, including a "Core Steps List," "Applicable Condition Logic," "Required Resources List," and "Expected Target Indicators." The conversion process is achieved through fill-in rules; for example, all identified nursing action entities are sequentially filled into the "Core Steps List," and entities and relationships related to patient status descriptions are combined into logical statements and filled into the "Applicable Condition Logic." The formula for calculating the completeness of text information extraction during structured parsing is:

[0043] in: This indicates the completeness score of the extracted data. This indicates the number of entities successfully extracted from the solution text and populated into the template; This indicates the total number of entities to be extracted from the text using a predefined dictionary, and the extraction completeness score is used to evaluate the parsing quality.

[0044] Optionally, the core steps' operational points list, applicable condition logical judgment statements, required resource list, and expected target quantitative indicators can be automatically extracted from the converted structured solution template. The core steps' operational points list is directly obtained from the "Core Steps List" field. The applicable condition logical judgment statements are generated into executable code snippets from the "Applicable Condition Logic" field. The required resource list is organized into a material list from the "Required Resource List" field. The expected target quantitative indicators are parsed into specific values ​​from the "Expected Target Indicators" field. For example, "Inflammation Reduction" is parsed as "The proportion of the inflamed area is reduced to below 5%".

[0045] Understanding this approach, the structured information of all candidate intervention protocols is stored in a relational database to form a nursing intervention protocol library. The structured information of each candidate protocol is stored as a record in a database table. The database table fields include protocol identifier, risk level, recovery stage, core steps, applicable conditions, required resources, expected goals, and extraction completeness score. The system establishes indexes according to multiple dimensions such as protocol identifier, risk level, and recovery stage. For example, a B-tree index is created for the risk level field to accelerate queries. The nursing intervention protocol library supports programmatic queries and invocations using a structured query language, such as querying all protocols with a risk level of "intermediate" and a recovery stage of "intermediate recovery stage."

[0046] In one embodiment of the invention, the core content and options of the selected nursing intervention program library are presented to the patient through an interactive interface. A combined structured and unstructured feedback collection method is designed. The structured method includes questionnaire ratings with preset options, while the unstructured method includes open-ended voice or text input boxes. Patient feedback input is received and analyzed in real time. The structured scoring results are statistically analyzed, and the unstructured input undergoes sentiment analysis and key intent extraction. By combining the statistical results of the structured scoring and the analysis results of the unstructured input, the patient's acceptance of different programs, tolerance to different nursing procedures, and preference for different follow-up frequencies are quantified to form a quantitative real-time preference feedback. The patient's personal historical data is retrieved, including past medical history, drug allergy history, surgical details, and previous recovery patterns. Each program in the nursing intervention program library is compared for compatibility with the patient's personal historical data, and programs with contraindications or high-risk incompatibility are eliminated. Based on the obtained real-time preference feedback, the implementation details of the remaining programs are adjusted, including medication reminder times, priority of follow-up examination items, and the format of health education materials. By integrating the compatibility comparison results with the adjusted execution details, a personalized care plan is customized for the patient, including a specific execution schedule, key points of operation, and precautions.

[0047] In practice, the implementation involves a process of personalized adaptation and adjustment of protocols in the nursing intervention protocol library based on the patient's personal historical data and real-time feedback to generate a personalized nursing plan. The system retrieves personal historical data from the patient's electronic health record, including past medical history such as diabetes or hypertension, drug allergy history such as allergy records to specific antibiotics, surgical details such as the date and specific type of cataract surgery, and past recovery patterns such as a quantitative description of the wound healing speed after previous surgeries. The nursing intervention protocol library contains multiple candidate protocols retrieved and matched from the nursing knowledge graph. Each candidate protocol has structured core steps, applicable conditions, required resources, and expected goal fields. The system first performs a compatibility comparison.

[0048] In some embodiments, each intervention protocol in the nursing intervention protocol library is compared for compatibility with the patient's personal history data. The system iterates through each protocol record in the nursing intervention protocol library, extracts the applicable conditions and contraindications list of the protocol, and performs a logical comparison item by item with the corresponding fields in the patient's personal history data. The compatibility comparison operation is performed based on a predefined rule engine. The rule engine matches personal history data entries with protocol contraindications. For example, if a protocol contains the core step of "using glucocorticoid eye drops" and the patient's personal history data records "history of steroid glaucoma," the rule engine determines that a contraindication exists. The system calculates a quantified incompatibility risk value for each protocol to assist in decision-making. The formula for calculating the incompatibility risk value is:

[0049] in: This indicates that the proposed solution is incompatible with risk values. This represents the total number of taboo conditions defined in the current scheme; Indicates the first The severity of each contraindication condition is weighted, and this weight is preset by medical experts; It is a binary indicator function that triggers the first [function name] when the patient's personal history data is [data type]. The function value is 1 if the condition is contraindicated, and 0 otherwise. The system sets an incompatibility risk threshold and removes solutions with incompatibility risk values ​​exceeding the threshold. For example, a solution for high-risk patients would be removed if it contains a drug component that the patient is allergic to. The remaining solutions are considered as initially compatible solutions and enter the subsequent adjustment phase.

[0050] Understandably, obtaining real-time feedback on patients' preferences for initial treatment options is achieved through an interactive interface. This interface, available on mobile applications or web pages, presents patients with the core content and available options of treatment plans selected through compatibility filtering from a nursing intervention plan library. The content is presented in a structured summary format, including key elements such as medication frequency, follow-up appointment times, and daily nursing actions. A combination of structured and unstructured feedback collection methods is designed. The structured method includes questionnaire ratings for pre-set options, such as asking patients to select their acceptance level for "four eye drops daily" on a scale of 1 to 5. The unstructured method includes open-ended voice or text input boxes allowing patients to input free text or voice descriptions of their preferences. The system receives and analyzes patient feedback in real time, statistically calculating the average score and distribution of each option for the structured ratings. For the unstructured input, sentiment analysis and key intent extraction are performed. Sentiment analysis uses a pre-trained model to output an emotion polarity score, and key intent extraction uses natural language processing to identify specific intentions such as "desire to reduce follow-up frequency" or "preference for video guidance."

[0051] In practice, the execution details of the remaining plans are adjusted based on real-time preference feedback. These adjustments directly modify the plan parameters based on the feedback analysis results. For example, if a patient has a low structured score for high-frequency medications, the system may reduce the daily dosage but extend the total treatment course. If a patient reports inconvenience during weekdays via text input, the system will adjust the medication reminder time to the early morning and evening. Adjustments also include prioritizing follow-up examinations; for example, if a patient reports difficulty traveling, the system will replace some non-urgent hospital check-up items with monitoring items that can be done at home. Furthermore, the format of health education materials may be adjusted; for example, if a patient expresses a preference for video materials, the system will provide short demonstration videos instead of purely text-based manuals. The specific changes before and after these adjustments can be illustrated through data comparison; see Table 1, which illustrates a hypothetical scenario.

[0052] Table 1: Comparison of Nursing Plan Before and After Adjustment

[0053] In some embodiments, the analysis of combined structured scoring results and unstructured input quantifies patients' acceptance of different options, tolerance to different nursing procedures, and preference for different follow-up frequencies. The quantification process uses a weighted scoring model to transform multivariate feedback into numerical indicators. For each option, the system calculates a comprehensive preference score, which is a fusion of the average structured score and the sentiment score obtained from the analysis of unstructured input, weighted according to preset criteria. This comprehensive preference score is used to rank and select different option options, forming a quantified real-time preference feedback dataset that directly drives subsequent adjustment logic.

[0054] Optionally, the system integrates compatibility comparison results with adjusted execution details to create a personalized care plan for the patient. All plan elements, after compatibility screening and feedback-based adjustments, are integrated into a unified structured document. The personalized care plan includes a detailed execution schedule, listing the specific times for daily medication administration, the dates and times of follow-up appointments, and operational guidelines outlining the detailed steps and standards for each nursing action. The precautions section specifically highlights warnings based on the patient's personal history, such as "Sulfonamides are contraindicated." The final personalized care plan generated by the system is presented to the patient in the form of a visual timeline and an interactive task list, and its traceability and integrity are ensured through digital signatures and timestamps.

[0055] Understandably, the personalized adaptation and adjustment process supports multiple iterations. Patients can provide feedback on the initial personalized care plan generated by the system, and the system further refines the plan details based on the new feedback data until the patient confirms its implementation. All adjustment records, original feedback data, and the final version of the personalized care plan are encrypted and stored in the patient's personal health record, forming the data foundation for continuous optimization of care strategies.

[0056] See Figure 4 This is a comparative chart showing the adjustments made to the elements of a personalized post-operative ophthalmic care plan. It visually demonstrates the quantitative changes resulting from adjustments to the core elements of the care plan based on patient feedback. All adjustments are derived from patients' structured scores and unstructured feedback, reflecting the core principle of personalized care. While reducing medication and follow-up frequency, online follow-up was added, reducing the burden on patients and ensuring medical safety through home data collection. The chart uses intuitive numerical comparisons to allow both medical staff and patients to quickly understand the scope and significance of the plan adjustments, improving the transparency and feasibility of the care plan. It demonstrates the complementary relationship between "reducing the frequency of hospital follow-ups" and "adding home monitoring," proving that the plan adjustments are not simply about reducing medical intervention, but about optimizing resource allocation while ensuring safety.

[0057] In one embodiment of the present invention, during the execution cycle of the personalized care plan, the system continuously receives and processes the latest monitoring data stream. The processed latest data is input into a complication risk assessment model to obtain updated recovery progress scores, potential complication probability values, and risk level labels. The updated data is compared with the data used when the original plan was formulated. If the risk level label is upgraded or the potential complication probability value exceeds a new threshold, an emergency revision process for the care plan is triggered. This process includes re-retrieving high-risk intervention plans from the nursing knowledge graph based on the new risk indication and quickly integrating them into the original plan. If the risk has not changed significantly, the intensity and frequency of subsequent nursing activities in the original plan are fine-tuned based on the updated recovery progress score. The finalized care plan is parsed and broken down into atomic tasks with clearly defined execution times, content, and recipients. Corresponding natural language descriptions or graphical guidance content are generated for each atomic task, forming nursing guidance instructions. Reminder messages containing specific times, locations, and precautions are generated for tasks requiring patient participation, such as follow-up examinations and medication administration, forming follow-up reminders. Based on the patient's preset channel preferences, different instructions and reminders are distributed to corresponding terminals, including smartphone application push notifications, SMS messages, telephone voice notifications, or emails. A confirmation feedback mechanism for instructions and reminders is established to track patient reception and confirmation, and to provide repeated reminders or escalation notifications for unconfirmed tasks.

[0058] In practice, the implementation involves a complete process of dynamically optimizing the personalized care plan by incorporating the latest monitoring data streams and generating executable nursing instructions and follow-up reminders. The personalized care plan is a structured care program customized for the patient, including a specific implementation schedule, key operational points, and precautions. Throughout the entire cycle of the personalized care plan's execution, the system continuously receives and processes the latest monitoring data streams from the patient's terminal. These latest monitoring data streams include newly acquired eye images, updated physiological parameter time series, and newly submitted self-symptom assessment reports. The system performs the same feature extraction and standardization processes on these latest monitoring data streams as it did on the initial data processing.

[0059] In some embodiments, the processed latest data is input into the complication risk assessment model to obtain updated recovery progress scores, potential complication probability values, and risk level labels. The complication risk assessment model uses a standardized set of postoperative recovery features corresponding to the new data for forward computation. The system compares the updated data with the data used when the original plan was developed. The comparison focuses on the risk level labels and potential complication probability values. A change in the risk level label from "intermediate" to "high" is considered an upgrade. A potential complication probability value exceeding a new threshold refers to any specific complication probability value exceeding a preset dynamic threshold, which may be set based on the patient's initial risk assessment results. If the risk level label is upgraded or the potential complication probability value exceeds the new threshold, the system immediately triggers an emergency revision process for the care plan.

[0060] Understandably, the emergency revision process for nursing care plans involves re-retrieving high-risk intervention plans from the nursing knowledge graph based on the new risk indicators and quickly integrating them into the original plan. The new risk indicators are identified by the most significantly changed probability values ​​of potential complications. The system re-queries the nursing knowledge graph using the new, higher risk level as the primary search key, locating the nursing strategy branch with the higher risk level. From there, it retrieves intervention plans targeting the currently dominant high-risk risk. After rapid structured analysis, the core intervention steps of the retrieved plans are extracted and integrated into the corresponding parts of the original personalized nursing care plan, forming a revised nursing care plan. The revised nursing care plan will prioritize covering nursing measures in the original plan that conflict with or are insufficiently effective in addressing the new high-risk factors.

[0061] In practice, if the risk does not change significantly after the latest monitoring data stream is input (i.e., the risk level label remains unchanged and the probability values ​​of all potential complications are within the original threshold range), the system will fine-tune the intensity and frequency of subsequent nursing activities in the original plan based on the updated recovery progress score. The fine-tuning operation is based on a preset score-intensity mapping function, which translates the numerical change in the recovery progress score into adjustments to nursing parameters. For example, when the recovery progress score increases from 65 to 80, the system may, based on the mapping relationship, reduce the daily eye cleaning frequency from four times to three times, or adjust the follow-up interval from once a week to once every two weeks. The fine-tuning decision function that determines whether an intensity adjustment is needed is expressed as:

[0062] in: The suggested adjustment amount indicates the intensity or frequency of nursing activities; a positive value represents an increase, and a negative value represents a decrease. It is a predefined adjustment function; This indicates the updated recovery progress score; This indicates the recovery progress score at the time the original plan was formulated. The system uses the calculated adjustment amount... Automatically modify task parameters for future cycles in a personalized care plan.

[0063] In some embodiments, the finalized care plan is parsed and broken down into atomic tasks with clearly defined execution times, content, and recipients. The finalized care plan may be a revised plan that has undergone urgent revisions or a slightly adjusted original plan. The breakdown process is based on the structured fields of the care plan. Each atomic task corresponds to an indivisible independent nursing action. For example, "Apply one drop of levofloxacin eye drops to the right eye at 8:00 AM on October 27, 2023" is an atomic task, and "Schedule an intraocular pressure check at the Central Hospital at 10:00 AM on November 3, 2023" is another atomic task. The system generates corresponding natural language descriptions or graphical guidance for each atomic task to form nursing guidance instructions. For medication tasks, the instruction text details the drug name, dosage, level, and steps; for nursing operations, graphical guidance may be provided, such as a diagram of handwashing procedures.

[0064] Optionally, reminder messages containing specific times, locations, and precautions can be generated for tasks requiring active patient participation, such as follow-up examinations and medication administration, forming follow-up reminders. The reminder message generation template is bound to the atomic task type. The system distributes different instructions and reminders to corresponding terminals based on the patient's preset channel preferences. These channel preferences are selected by the patient in the system settings, and the terminals include smartphone app push notifications, SMS, telephone voice notifications, or email. For routine tasks such as medication administration, push notifications are usually sent to the smartphone app; for important follow-up appointments, both SMS and app push notifications may be sent simultaneously; for high-risk reminders or escalation notifications of unconfirmed tasks, telephone voice notifications may be initiated.

[0065] Understandably, establishing a confirmation and feedback mechanism for instructions and reminders tracks patient reception and confirmation. Each issued nursing guidance instruction and follow-up reminder is accompanied by a unique task identifier, requiring the patient to confirm it on their terminal. The system monitors the confirmation status and confirmation time for each task. If a task is not confirmed within a preset time, such as a medication reminder not being confirmed within 30 minutes of being issued, the system automatically triggers a repeat reminder. The repeat reminder may be sent again through the same channel or escalated to a higher priority channel. All records of instruction issuance, reception, confirmation, and repeat reminders are persistently stored, forming a quantitative assessment basis for patient compliance and serving as feedback input for subsequent nursing plan optimization.

[0066] See Figure 5 This is a postoperative ophthalmic sign recovery trend monitoring chart, which visually presents the dynamic changes of three core ocular signs within 28 days post-surgery, possessing multiple professional values ​​in clinical and research settings. The chart clearly delineates the "rapid regression period" (0-10 days) and the "stable recovery period" (10-28 days), providing data support for developing phased nursing strategies in clinical practice. The synchronized decline of the three indicators aligns with a normal recovery trajectory; an abnormally high or persistently fluctuating indicator may indicate potential risks of complications such as infection or delayed healing. Long-term monitoring can verify the effectiveness of surgical and nursing protocols, providing objective evidence for optimizing treatment strategies. The intuitive decline trend allows patients to clearly perceive their recovery progress, enhancing their confidence and adherence to treatment.

[0067] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A smart follow-up and nursing method for ophthalmic patients undergoing postoperative recovery, characterized in that, Includes the following steps: Collect multimodal continuous monitoring data of ophthalmic patients during postoperative recovery; The multimodal continuous monitoring data is subjected to spatiotemporal synchronization and feature extraction processing to generate a standardized postoperative recovery feature set, which includes quantitative indicators of ocular signs, physiological state trend indicators, and subjective symptom feature vectors. The standardized set of postoperative recovery features is input into a preset complication risk assessment model, which outputs a recovery progress score, a probability value of potential complications, and a risk level label. Based on the recovery progress score, the probability value of potential complications, and the risk level marker, a matching nursing intervention program library is retrieved from the nursing knowledge graph; Based on the patient's personal historical data and real-time feedback, the plans in the nursing intervention plan library are personalized and adjusted to generate a personalized nursing plan. The personalized care plan is dynamically optimized by combining the latest monitoring data stream, generating executable nursing instructions and follow-up reminders.

2. The intelligent follow-up and nursing method for ophthalmic patients oriented towards postoperative recovery as described in claim 1, characterized in that, The process of performing spatiotemporal synchronization and feature extraction on the multimodal continuous monitoring data to generate a standardized set of postoperative recovery features includes: The multimodal continuous monitoring data includes eye image sequences taken by the patient, time series of physiological parameters recorded by wearable devices, and self-symptom assessment reports entered by the patient through an interactive interface. The eye image sequence is subjected to image quality assessment and screening, and images that are blurry or do not meet the angle requirements are removed, while qualified eye images are retained; The qualified ocular images are segmented into lesion areas, analyzed for hyperemia, and feature extracted from healing areas. The quantitative indicators of ocular signs are calculated, including the proportion of inflammatory area, neovascularization density, and corneal edema index. The time series of physiological parameters recorded by the wearable device is aligned with the shooting time of each qualified eye image, and the statistical features of physiological parameters within the corresponding time window are extracted to form the physiological state trend index, which includes intraocular pressure fluctuation curve, heart rate variability rate and sleep quality index. Natural language processing is performed on the self-symptom assessment report to extract keywords and quantify symptom intensity, constructing the subjective symptom feature vector, which includes numerical representations of pain level, photophobia degree, and blurred vision score.

3. The intelligent follow-up and nursing method for ophthalmic patients oriented towards postoperative recovery according to claim 1, characterized in that, The standardized set of postoperative recovery features is input into a pre-defined complication risk assessment model, which outputs a recovery progress score, a probability value of potential complications, and a risk level label, including: The quantitative indicators of eye signs are input into the visual feature analysis sub-network of the model. The visual feature analysis sub-network is based on a convolutional neural network structure and is used to assess the degree of abnormality of the signs and output the abnormality score of the signs. The physiological state trend indicators are input into the time series analysis submodule of the model. The time series analysis submodule analyzes the changing trend and stability of the indicators and outputs the physiological risk coefficient. The subjective symptom feature vector is input into the subjective feedback fusion unit of the model. The subjective feedback fusion unit combines the patient's historical baseline data to assess the degree to which the symptoms deviate from the normal recovery trajectory and outputs the symptom deviation degree. The recovery process score is calculated by integrating the abnormal vital signs score, the physiological risk coefficient, and the symptom deviation score through an ensemble learning algorithm. Based on the abnormal physical signs score, physiological risk coefficient, and symptom deviation, the probability of infection risk, the probability of abnormal intraocular pressure, and the probability of delayed healing are obtained by querying a preset probability mapping table, and the highest value is taken as the probability value of potential complications. Based on the recovery progress score and the numerical range of the potential complication probability value, the corresponding risk level marker is matched from a preset rule base.

4. The intelligent follow-up and nursing method for ophthalmic patients oriented towards postoperative recovery as described in claim 1, characterized in that, The step of retrieving a matching nursing intervention protocol from the nursing knowledge graph based on the recovery progress score, potential complication probability value, and risk level marker includes: Using the risk level marker as the primary search key, the corresponding nursing strategy branch is located in the nursing knowledge graph; Using the recovery process score as a secondary screening criterion, a set of nursing plans applicable to the recovery stage is selected under the identified nursing strategy branch. Using the specific risk type indicated by the probability value of the potential complication as the guide for optimizing the plan, candidate plans containing targeted interventions are further screened from the set of nursing plans. The selected candidate solutions are analyzed in a structured manner to extract the core steps, applicable conditions, required resources and expected goals of the solutions, forming a preliminary nursing intervention solution library.

5. The intelligent follow-up and nursing method for ophthalmic patients oriented towards postoperative recovery according to claim 1, characterized in that, The process of adapting and adjusting the nursing intervention plans in the database based on the patient's personal historical data and real-time feedback to generate a personalized nursing plan includes: Retrieve the patient's personal history data, which includes past medical history, drug allergy history, surgical details, and past recovery patterns; Each plan in the nursing intervention plan library is compared with the patient's personal historical data for compatibility, and plans that are contraindicated or have high risk of incompatibility are eliminated. Obtain real-time feedback from patients regarding their preferences for initial treatment options. This real-time preference feedback includes preferences for medication timing, convenience of follow-up visits, and acceptance of nursing procedures. Based on the real-time preference feedback, the execution details of the remaining plans are adjusted, including adjusting the medication reminder time, adjusting the priority of follow-up examination items, and adjusting the form of health education materials; By integrating the compatibility comparison results with the adjusted execution details, a personalized care plan is customized for the patient, including a specific execution schedule, key points of operation, and precautions.

6. The intelligent follow-up and nursing method for ophthalmic patients oriented towards postoperative recovery according to claim 5, characterized in that, The process of obtaining real-time feedback from patients regarding their preferences for initial treatment options includes: The core content and options of the selected nursing intervention plan library are presented to the patient through an interactive interface. Design a feedback collection method that combines structured and unstructured approaches. The structured approach includes questionnaire ratings with pre-defined options, while the unstructured approach includes open-ended voice or text input. It receives and analyzes patient feedback in real time, performs statistical analysis on structured scores, and performs sentiment analysis and key intent extraction on unstructured input. By combining the results of structured scoring with the analysis of unstructured inputs, the patient's acceptance of different treatment plans, tolerance to different nursing procedures, and preference for different follow-up frequencies are quantified to form the aforementioned real-time preference feedback.

7. The intelligent follow-up and nursing method for ophthalmic patients oriented towards postoperative recovery according to claim 3, characterized in that, The subjective symptom feature vector is input into the subjective feedback fusion unit of the model. This unit, combined with the patient's historical baseline data, assesses the degree to which the symptoms deviate from the normal recovery trajectory and outputs the symptom deviation score, including: Standardized change curves of various subjective symptoms during the typical postoperative recovery period corresponding to the type of ophthalmic surgery were extracted from the patient's historical baseline data. The scores of each item in the subjective symptom feature vector at the current time point are compared with the standard values ​​at the corresponding time points on the standard change curve, and the instantaneous deviation value of each symptom is calculated. Based on the weight of each symptom in the complication warning, the weighted sum of each instantaneous deviation value is calculated to obtain the comprehensive instantaneous deviation score; Analyze the changing trend of the subjective symptom feature vector over a period of time, calculate the angle or distance between the trend and the standard recovery trend, and obtain the trend deviation score; The comprehensive instantaneous deviation score and the trend deviation score are combined according to a preset ratio to obtain the final symptom deviation degree.

8. The intelligent follow-up and nursing method for ophthalmic patients oriented towards postoperative recovery according to claim 1, characterized in that, The personalized care plan is dynamically optimized by incorporating the latest monitoring data stream, generating executable nursing instructions and follow-up reminders, including: During the execution cycle of the personalized care plan, the latest monitoring data stream is continuously received and processed; The latest processed data is input into the complication risk assessment model to obtain an updated recovery progress score, potential complication probability value, and risk level label. Compare the updated data with the data used when the original plan was developed. If the risk level marker is upgraded or the probability value of potential complications exceeds the new threshold, the emergency revision process of the care plan is triggered. The emergency revision process includes: retrieving high-risk intervention plans from the nursing knowledge graph based on the new risk indications, and quickly integrating them into the original plan to form a revised nursing plan; If the risk does not change significantly, the intensity and frequency of subsequent care activities in the original plan will be fine-tuned based on the updated recovery progress score. The finalized nursing plan is translated into specific daily or weekly nursing instructions and follow-up reminders, and distributed through designated communication channels.

9. The intelligent follow-up and nursing method for ophthalmic patients oriented towards postoperative recovery according to claim 4, characterized in that, The selected candidate solutions are subjected to structured analysis to extract the core steps, applicable conditions, required resources, and expected goals of each solution, forming a preliminary nursing intervention solution library, including: Natural language processing is performed on the text description of each candidate solution to identify and label the key entities and relationships describing nursing actions, medication information, examination items, frequency and cycle; Based on the annotation results, the unstructured text solution is converted into a structured solution template. The solution template contains fixed fields, which are used to fill in the core steps, applicable conditions, required resources and expected goals. From the converted structured solution template, the list of key operation points of the core steps, the logical judgment statements of the applicable conditions, the list of required resources, and the quantitative indicators of the expected goals are automatically extracted. All candidate solutions are stored in a relational database with structured information and indexed according to dimensions such as solution identifier, risk level, and recovery stage, forming a nursing intervention solution library that can be queried and invoked programmatically.

10. The intelligent follow-up and nursing method for ophthalmic patients oriented towards postoperative recovery according to claim 8, characterized in that, The process of translating the finalized nursing plan into specific daily or weekly nursing guidance instructions and follow-up reminders, and distributing them through designated communication channels, includes: The finalized care plan is analyzed and broken down into atomic tasks with clearly defined execution time, content, and target. For each atomic task, generate corresponding natural language descriptions or graphical guidance content to form the nursing guidance instructions; For tasks requiring active patient participation, such as follow-up examinations and medication administration, a reminder message containing specific time, location, and precautions is generated, forming the aforementioned follow-up reminder; Based on the patient's preset channel preferences, different instructions and reminders are distributed to corresponding terminals, including smartphone application push notifications, SMS messages, telephone voice notifications, or emails; Establish a confirmation and feedback mechanism for instructions and reminders, track patients' receipt and confirmation status, and provide repeated reminders or escalation notifications for unconfirmed tasks.

Images