Intelligent medical data self-supervision deep processing analysis system

Abstract

The invention relates to the technical field of wisdom medical treatment, in particular to a wisdom medical data self-supervision deep processing analysis system, which comprises a multi-department data sensing layer for acquiring multi-department data and postoperative follow-up visit data of a patient; the operation period feature processing center extracts data features of all departments, correlates data before and after an operation through an attention mechanism, and generates a patient perioperative period unified feature map; the full-cycle dynamic analysis unit analyzes the state of the patient in stages based on the patient perioperative period unified characteristic spectrum, and outputs a risk index in combination with a self-supervised training prediction model; the adaptive clinical decision module generates customized suggestions for different departments according to the analysis result and the risk index; the cross-department data collaboration unit is used for constructing a visual collaboration platform, so that doctors of all departments share and interact with a whole-cycle analysis result of an operation; and the dynamic feedback optimization unit collects clinical diagnosis and treatment results and performs reverse fine tuning on parameters of the self-supervised model. Therefore, the problems of strategy solidification, insufficient energy utilization efficiency and the like in the prior art are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart medical treatment, in particular to a smart medical treatment data self-supervised deep processing analysis system. BACKGROUND

[0002] With the acceleration of the transformation of the medical industry towards digitization and precision, multi-source heterogeneous medical data such as electronic medical records, medical images, test reports, and real-time physiological signals are growing explosively, providing core support for early disease diagnosis, treatment plan optimization, and clinical research. However, in the long-term service process of the smart medical treatment data processing system, due to factors such as high data annotation cost, strict cross-institution privacy protection, and complex and variable clinical conditions (such as emergency data surges), it is prone to problems such as low data utilization, weak model generalization ability, high privacy leakage risk, and delayed fault response, which leads to fluctuations in intelligent diagnosis accuracy, obstacles to cross-institution data sharing, and even affects the timeliness and reliability of clinical decision-making. With the development of smart medical treatment towards multi-modal fusion, cross-regional collaboration, and real-time dynamic decision-making, higher requirements are put forward for data processing systems in terms of unannotated dependence, privacy security protection, wide-condition adaptability, and processing efficiency, and there is an urgent need for a self-supervised deep processing analysis system that can real-time perceive multi-dimensional data state, dynamically optimize feature extraction strategy, and efficiently balance privacy and precision, to ensure the safe and efficient use of medical data in diversified clinical scenarios.

[0003] However, traditional smart medical treatment data processing technology has inherent defects: the processing strategy is fixed, relying mainly on structured data annotated by humans or simple threshold screening rules, without integrating multi-dimensional information such as clinical ontology knowledge, environmental parameters, and load prediction, making it difficult to adapt to the complex features of unstructured data and the heterogeneous characteristics of cross-institution data distribution, resulting in low processing accuracy, high data redundancy, and even feature mis-extraction; privacy protection methods are single, traditional centralized model training requires transmission of original medical data, which violates HIPAA, GDPR, and other privacy regulations, and existing federated learning frameworks mostly use static privacy budget allocation, without dynamically adjusting protection strategies according to data sensitivity, which easily causes core privacy leakage or model precision loss due to excessive encryption of non-core data; data processing resource management is extensive, passive data reception and batch processing mode consumes high energy, and active processing lacks dynamic adaptability to clinical needs (such as emergency priority processing), not only reducing data conversion efficiency, but also easily exacerbating hardware device wear and tear; and the system lacks modular redundancy design, a single data preprocessing module or feature extraction module failure leads to the failure of the overall processing system, far exceeding the stringent requirements of clinical data processing continuity, and the overall technology faces multiple challenges such as fixed processing strategy, privacy and precision imbalance, and insufficient energy utilization efficiency. SUMMARY

[0004] The application provides a smart medical data self-supervised deep processing analysis system to solve the problems of strategy solidification and insufficient energy utilization efficiency in the prior art.

[0005] The first aspect of the application provides a smart medical data self-supervised deep processing analysis system, comprising: a multi-department data perception layer, a surgery period feature processing hub, a whole cycle dynamic analysis unit, a self-adaptive clinical decision module, a cross-department data collaboration unit, and a dynamic feedback optimization unit; wherein the multi-department data perception layer is configured to collect multi-department data and postoperative follow-up data of a patient wearable device; the surgery period feature processing hub is configured to extract features of each department data, associate preoperative and postoperative data through an attention mechanism, and generate a patient perioperative unified feature map; the whole cycle dynamic analysis unit is configured to analyze the patient state in stages based on the patient perioperative unified feature map, and output a risk index combined with a self-supervised training prediction model; the self-adaptive clinical decision module is configured to generate customized recommendations for different departments according to the analysis results and the risk index; the cross-department data collaboration unit is configured to construct a visual collaboration platform to enable doctors of each department to share and interact with the surgery whole cycle analysis results; and the dynamic feedback optimization unit is configured to collect clinical diagnosis and treatment results and fine-tune the self-supervised model parameters in reverse.

[0006] Preferably, the multi-department data perception layer comprises a data acquisition module and a follow-up data integration unit, wherein the data acquisition module is configured to collect diagnosis and treatment data of internal medicine, surgery, and anesthesiology departments; and the follow-up data integration unit is configured to receive postoperative heart rate, activity level, and sleep quality data transmitted by a patient wearable device and align the data according to timestamps.

[0007] Preferably, the surgery period feature processing hub comprises a feature extraction unit, a cross-temporal data association module, and a feature map generation engine, wherein the feature extraction unit is configured to extract physiological indicator features, image features, and surgery operation features from department data; the cross-temporal data association module is configured to associate preoperative baseline data and postoperative recovery data through a multi-head attention mechanism and mine temporal association relationships; and the feature map generation engine is configured to associate multi-dimensional features with a patient individual and construct a perioperative unified feature map containing time and indicator dimensions.

[0008] Preferably, the whole-cycle dynamic analysis unit comprises a staged state evaluation module, a self-supervised prediction model training unit, and a risk index calculation engine, wherein the staged state evaluation module is configured to divide the whole cycle of a patient into a preoperative evaluation period, an intraoperative monitoring period, and a postoperative recovery period, and analyze the physiological state and recovery progress in each stage; the self-supervised prediction model training unit is configured to perform self-supervised training based on unannotated clinical data, and predict the state evolution process of the patient through a Perceiver model; and the risk index calculation engine is configured to generate a quantitative risk index by fusing the complication occurrence probability and the recovery delay risk indicator according to the state evolution process of the patient.

[0009] Preferably, the adaptive clinical decision module comprises a department demand analysis unit, a decision rule generation module, and a suggestion optimization unit, wherein the department demand analysis unit is configured to identify the clinical demands of different departments such as internal medicine, surgery, and anesthesiology; the decision rule generation module is configured to generate preliminary diagnosis and treatment suggestion rules according to the patient risk index and the department demands; and the suggestion optimization unit is configured to iteratively optimize the preliminary suggestions in combination with historical diagnosis and treatment effect data, and generate customized clinical suggestions.

[0010] Preferably, the cross-department data collaboration unit comprises a visual platform construction module, a data sharing permission management unit, and a real-time interaction module, wherein the visual platform construction module constructs a three-dimensional visual collaboration interface using WebGL to display the patient whole-cycle feature map and analysis results; the data sharing permission management unit is configured to assign data viewing and editing permissions according to the department and responsibilities; and the real-time interaction module is configured to perform online labeling of key data, addition of diagnosis and treatment opinions, and real-time interactive collaboration.

[0011] Preferably, the dynamic feedback optimization unit comprises a diagnosis and treatment result collection module, a feedback data matching unit, and a model parameter fine-tuning engine, wherein the diagnosis and treatment result collection module is configured to collect the postoperative complication occurrence, recovery period, diagnosis and treatment effect score, and clinical result data of a patient; the feedback data matching unit is configured to associate and match the diagnosis and treatment results with the model input data and prediction results of the corresponding patient; and the model parameter fine-tuning engine is configured to use a gradient descent algorithm to reversely fine-tune the weight parameters based on the matched feedback data, and optimize the prediction accuracy.

[0012] The second aspect embodiment of the application provides a smart medical data self-supervised deep processing analysis method, comprising: acquiring multi-department diagnosis and treatment data and postoperative follow-up data; analyzing and processing the multi-department diagnosis and treatment data and the postoperative follow-up data, extracting physiological index features, image features and surgical operation features, associating preoperative and postoperative data through an attention mechanism, and generating a unified feature map of a patient perioperative period; based on the unified feature map of the patient perioperative period, dividing a patient whole cycle into a preoperative evaluation period, an intraoperative monitoring period and a postoperative recovery period, analyzing physiological states and recovery progress in each stage, performing self-supervised training based on unannotated clinical data, predicting a patient state evolution process through a Perceiver model, fusing complication occurrence probability and recovery delay risk indicators according to the patient state evolution process, generating a risk index, identifying clinical needs of different departments according to analysis results and the risk index, generating preliminary diagnosis and treatment suggestion rules, iteratively optimizing the preliminary diagnosis and treatment suggestion rules in combination with historical diagnosis and treatment effect data, and generating customized clinical suggestions; based on a visual collaborative platform, enabling doctors of each department to share and interact with surgical whole cycle analysis results and the customized suggestions, collecting clinical diagnosis and treatment results in real time, and reversely fine-tuning parameters.

[0013] The third aspect embodiment of the application provides an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, the processor executes the program to implement a smart medical data self-supervised deep processing analysis method according to the above-mentioned embodiments.

[0014] The fourth aspect embodiment of the application provides a computer readable storage medium having a computer program stored thereon, the program being executed by a processor to implement a smart medical data self-supervised deep processing analysis method according to the above-mentioned embodiments.

[0015] Therefore, the application has the following beneficial effects: The embodiments of the present application integrate department data and postoperative follow-up data through a multi-department data perception layer, provide a complete data basis for whole-cycle analysis; the surgery period feature processing hub correlates preoperative and postoperative data by means of an attention mechanism, generates a unified feature atlas, and improves the relevance and accuracy of feature extraction; the whole-cycle dynamic analysis unit combines a self-supervised prediction model to output a risk index in stages, dynamically monitors and predicts the risk of the patient's state; the adaptive clinical decision module generates customized recommendations for different departments, improving the personalization and practicality of clinical decision-making; the cross-department data collaboration unit promotes data sharing and interaction through a visualization platform, breaking down the barriers to cross-department collaboration; the dynamic feedback optimization unit relies on clinical results to fine-tune the model in reverse, ensuring the continuous iterative optimization of system performance, improving the utilization efficiency and analysis accuracy of medical data, and strengthening the cross-department collaboration capability and the scientific nature of clinical decision-making, effectively supporting efficient diagnosis and treatment and patient safety in the context of smart medical care. Thus, the problems of strategy solidification and insufficient energy utilization efficiency in the prior art are solved.

[0016] Additional aspects and advantages of the present application will be made apparent by the following description and the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0017] The above and / or additional aspects and advantages of the present application will become apparent and be readily understood from the following description, taken in conjunction with the accompanying drawings, in which: Figure 1 A structural schematic diagram of a smart medical data self-supervised deep processing analysis system according to an embodiment of the present application is shown in FIG. 1; Figure 2 A schematic diagram of a multi-department data perception layer according to an embodiment of the present application is shown in FIG. 2; Figure 3 A schematic diagram of a surgery period feature processing hub according to an embodiment of the present application is shown in FIG. 3; Figure 4 A schematic diagram of a whole-cycle dynamic analysis unit according to an embodiment of the present application is shown in FIG. 4; Figure 5 A schematic diagram of an adaptive clinical decision module according to an embodiment of the present application is shown in FIG. 5; Figure 6 A schematic diagram of a cross-department data collaboration unit according to an embodiment of the present application is shown in FIG. 6; Figure 7 A schematic diagram of a dynamic feedback optimization unit according to an embodiment of the present application is shown in FIG. 7; Figure 8 A structural schematic diagram of a smart medical data self-supervised deep processing analysis system according to an embodiment of the present application is shown in FIG. 8; Figure 9A flowchart of a smart medical data self-supervised deep processing analysis method according to an embodiment of the present application is provided. Figure 10 A schematic diagram of a smart medical data self-supervised deep processing analysis method according to an embodiment of the present application is provided. Figure 11 A structural schematic diagram of an electronic device according to an embodiment of the present application is provided. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0019] The smart medical data self-supervised deep processing analysis system according to the embodiments of the present application will be described below with reference to the drawings. In view of the problem of strategy solidification mentioned in the above background art, the present application provides a smart medical data self-supervised deep processing analysis system. In the system, the department data and postoperative follow-up data are integrated through a multi-department data perception layer to provide a complete data basis for full-cycle analysis; a surgery period feature processing hub correlates preoperative and postoperative data by means of an attention mechanism to generate a unified feature map, thereby improving the relevance and accuracy of feature extraction; a full-cycle dynamic analysis unit combines a self-supervised prediction model to output risk indexes in stages, dynamically monitors and predicts the risk of the patient state; an adaptive clinical decision module generates customized recommendations for different departments to improve the personalization and practicality of clinical decision-making; a cross-department data collaboration unit promotes data sharing and interaction through a visualization platform to break down the barriers to cross-department collaboration; and a dynamic feedback optimization unit relies on clinical results to fine-tune the model in reverse, ensuring continuous iterative optimization of system performance, improving the utilization efficiency and analysis accuracy of medical data, and strengthening the cross-department collaboration capability and scientific nature of clinical decision-making, effectively supporting efficient diagnosis and treatment and patient safety in the context of smart medical care. Thus, the problems of strategy solidification and insufficient energy utilization efficiency in the prior art are solved.

[0020] Figure 1 A structural schematic diagram of a smart medical data self-supervised deep processing analysis system according to an embodiment of the present application is provided.

[0021] The present application provides a smart medical data self-supervised deep processing analysis system, which comprises: a multi-department data perception layer 100, a surgery period feature processing hub 200, a full-cycle dynamic analysis unit 300, an adaptive clinical decision module 400, a cross-department data collaboration unit 500, and a dynamic feedback optimization unit 600.

[0022] The multi-department data perception layer 100 is configured to collect multi-department data and postoperative follow-up data of the patient wearable device; the surgery period feature processing hub 200 is configured to extract features of the department data, associate preoperative and postoperative data through an attention mechanism, and generate a unified feature map of the patient perioperative period; the whole-cycle dynamic analysis unit 300 is configured to analyze the patient state in stages based on the unified feature map of the patient perioperative period, and output a risk index in combination with a prediction model of self-supervised training; the adaptive clinical decision module 400 is configured to generate customized recommendations for different departments according to the analysis results and the risk index; the cross-department data collaboration unit 500 is configured to construct a visual collaboration platform, so that doctors of different departments share and interact with the analysis results of the whole-cycle surgery; and the dynamic feedback optimization unit 600 is configured to collect clinical diagnosis and treatment results, and fine-tune the parameters of the self-supervised model in reverse.

[0023] It can be understood that, in the embodiments of the present application, the multi-department data perception layer integrates department data and postoperative follow-up data to provide a complete data basis for whole-cycle analysis; the surgery period feature processing hub associates preoperative and postoperative data through an attention mechanism to generate a unified feature map, thereby improving the relevance and accuracy of feature extraction; the whole-cycle dynamic analysis unit outputs a risk index in stages in combination with a self-supervised prediction model, and dynamically monitors and predicts the risk of the patient state; the adaptive clinical decision module generates customized recommendations for different departments, thereby improving the personalization and practicality of clinical decision-making; the cross-department data collaboration unit promotes data sharing and interaction through a visual platform, thereby breaking down the barriers to cross-department collaboration; and the dynamic feedback optimization unit fine-tunes the model in reverse based on clinical results, thereby ensuring continuous iterative optimization of system performance, improving the utilization efficiency and analysis accuracy of medical data, strengthening the cross-department collaboration capability and the scientific nature of clinical decision-making, and effectively supporting efficient diagnosis and treatment and patient safety in the context of smart medical care. Thus, the problems of strategy solidification and insufficient energy utilization efficiency in the prior art are solved.

[0024] In the embodiments of the present application, the multi-department data perception layer 100 includes: as shown in Figure 2 the data acquisition module and the follow-up data integration unit.

[0025] The data acquisition module is configured to collect diagnosis and treatment data of internal medicine, surgery, and anesthesiology departments; and the follow-up data integration unit is configured to receive postoperative heart rate, activity, and sleep quality data transmitted by the patient wearable device, and align the data according to timestamps.

[0026] It is understood that the embodiments of this application centrally collect core data from multiple departments through a data acquisition module, avoiding information gaps caused by scattered data storage. This provides comprehensive clinical data support for subsequent multidisciplinary joint diagnosis and treatment plan formulation. The follow-up data integration unit accurately receives dynamic data on heart rate, activity level, and sleep quality transmitted by patients through wearable devices after surgery, and associates them with previous medical data in a timestamp-aligned manner, constructing a patient data chain covering the entire cycle from pre-operative to intra-operative to post-operative. This not only reduces the shortcomings of untimely and discontinuous data collection in traditional post-operative follow-up, allowing doctors to grasp the dynamics of post-operative recovery in real time, promptly identify abnormal indicators and intervene, but also helps doctors more accurately judge treatment effects and recovery patterns through time-dimensional data correlation analysis, improving the scientific nature of treatment decisions. At the same time, it also reduces the workload of medical staff in manually integrating follow-up data and improves the efficiency of follow-up management.

[0027] For example, in the follow-up management of patients after orthopedic knee replacement surgery, patients wear a smart bracelet with heart rate monitoring, step counting, and sleep analysis functions. This bracelet will record the patient's heart rate (e.g., an average of 82 beats / min on day 2 and 88 beats / min on day 4), activity level (1800 steps on day 2 and 2500 steps on day 4), and sleep quality data (2.1 hours of deep sleep duration and a sleep score of 72 on day 2 and day 4, respectively) in real time from 6:00 AM to 10:00 PM daily. The data (2.5 hours in duration, 78 points in sleep) was transmitted to the follow-up data integration unit. After receiving the data, the unit automatically aligned the various indicators according to minute-level timestamps. For example, it linked and integrated the heart rate of 86 beats / min at 9:30 on the 4th day after surgery, the activity level of 320 steps, and the sleep segment data (light sleep from 8:00 to 8:30) with the patient's preoperative knee joint imaging data, intraoperative prosthesis model, and postoperative analgesia medication information recorded by the anesthesiology department to generate a visualized "time-indicator" dynamic curve. Through this integrated result, doctors can clearly observe the coordinated changes in postoperative activity levels, heart rate, and sleep quality. When a patient's activity level suddenly drops to 1200 steps and heart rate rises to 95 beats per minute while sleep score drops to 65 on the 6th postoperative day, doctors can quickly combine the fluctuation points of the indicators corresponding to the timestamps to determine the possible problem of limited activity caused by postoperative joint swelling. They can then adjust the rehabilitation training plan in a timely manner and remotely remind the patient to return for a follow-up visit. This effectively avoids the recovery risks caused by fragmented and delayed follow-up data, while reducing the number of offline follow-up visits for patients and improving the efficiency of follow-up management and patient compliance.

[0028] In this embodiment of the application, the surgical feature processing center 200 includes: such as Figure 3 As shown, the system includes a feature extraction unit, a cross-time series data association module, and a feature map generation engine.

[0029] The feature extraction unit is used to extract physiological indicators, imaging features and surgical operation features from departmental data; the cross-temporal data association module is used to associate preoperative baseline data and postoperative recovery data through a multi-head attention mechanism to explore temporal correlations; and the feature map generation engine is used to associate multi-dimensional features with individual patients to construct a unified perioperative feature map that includes time and indicator dimensions.

[0030] Understandably, the feature extraction unit in this application accurately extracts physiological indicators, imaging features, and surgical operation features from departmental data, breaking down the isolation barriers between different types of data and providing comprehensive and crucial basic data support for subsequent analysis, avoiding omissions of diagnostic and treatment information due to data fragmentation. The cross-temporal data association module, through a multi-head attention mechanism, dynamically focuses on the core correlation points between preoperative baseline data and postoperative recovery data, efficiently mining the temporal dependency relationship between preoperative state, intraoperative operation, and postoperative recovery, making up for the shortcomings of traditional static data analysis in capturing dynamic correlations. The feature map generation engine precisely binds multi-dimensional features to individual patients, constructing a unified perioperative feature map, intuitively presenting the trajectory of patient feature changes throughout the entire cycle, providing doctors with a visual decision-making basis for developing personalized surgical plans, risk assessments, and rehabilitation plans, and helping to trace the correlation between surgical effects and specific features, improving the utilization efficiency of perioperative data and the scientific and accurate nature of diagnostic and treatment decisions.

[0031] It should be noted that the feature map generation engine uses the individual patient as the unique anchor (identified by hospital ID, surgery number, etc.) to accurately bind multi-dimensional features throughout the perioperative period—covering physiological indicators from preoperative to postoperative time (such as blood routine, liver and kidney function, pain score), imaging features (such as preoperative lesion CT parameters, intraoperative ultrasound signals, postoperative healing imaging scores), and surgical operation features (such as surgery type, duration, and anastomosis method). Simultaneously, it constructs a two-dimensional "time-indicator" architecture: the time axis covers key nodes such as 1 week before surgery, 1 day before surgery, various stages during surgery, and 6 hours to 7 days postoperatively, ensuring continuity throughout the entire cycle; the indicator axis is divided into three modules according to diagnostic and treatment logic: physiological function, imaging diagnosis, and surgical operation, enabling the classification of multiple types of indicators. Based on this, the engine uses a feature association strength quantification algorithm to mine the correlations between features (such as the strong correlation between preoperative coagulation function and postoperative wound healing speed), and presents this in a visual form (color marking of abnormal features, line thickness reflecting correlation strength), ultimately generating a unified perioperative feature map. This atlas is not only a panoramic view of the entire surgical process for patients, helping doctors trace the complete trajectory of "preoperative status - intraoperative intervention - postoperative recovery" and quickly locate abnormal features and related factors, but also provides quantitative and intuitive decision support for personalized surgical planning, postoperative risk warning, and review of treatment effects, effectively solving the problems of scattered data and difficulty in identifying temporal correlations in traditional diagnosis and treatment.

[0032] Feature association strength quantification formula:

[0033] in, The Pearson correlation coefficient; Let X be the i-th observation value of variable X; Let X be the sample mean of variable X; Let Y be the i-th observation value; Let Y be the sample mean of the variable Y; The number of samples; For index symbols.

[0034] Multi-head attention mechanism formula:

[0035] in, To output the projection matrix; This is the query vector, corresponding to the features of postoperative recovery data; For key vectors, corresponding to preoperative baseline data features; This is a value vector, corresponding to the association weight vector of the preoperative data; For the overall output of the multi-head attention mechanism; This is the output of the first attention head; For the number of heads; For splicing operations; This is the output of the second attention head.

[0036] In this embodiment of the application, the full-cycle dynamic analysis unit 300 includes: as follows Figure 4 As shown, the system includes a phased state assessment module, a self-supervised prediction model training unit, and a risk index calculation engine.

[0037] The phased status assessment module is used to divide the patient's entire life cycle into the preoperative assessment period, the intraoperative monitoring period, and the postoperative recovery period, and analyze the physiological status and recovery progress at each stage; the self-supervised prediction model training unit is used to perform self-supervised training based on unlabeled clinical data, and predict the evolution of the patient's status through the Perceiver model; the risk index calculation engine is used to generate a quantitative risk index by integrating the probability of complication and the risk index of recovery delay based on the evolution of the patient's status.

[0038] It is understandable that the embodiments of this application break down the entire patient lifecycle into three core stages—preoperative assessment, intraoperative monitoring, and postoperative recovery—through a phased status assessment module, and analyze the differences in treatment focus for each stage: the preoperative assessment focuses on the patient's basic physiological state and surgical suitability; the intraoperative monitoring stage tracks vital signs and surgical procedure suitability in real time; and the postoperative recovery stage focuses on wound healing, inflammation resolution, and other recovery progress, laying a precise phased data foundation for subsequent prediction and risk assessment; the self-supervised prediction model training unit relies on unlabeled clinical data for training, which can fully explore the data without a large amount of manual annotation. Based on underlying patterns, it significantly reduces data processing costs and can accurately predict patient status evolution trends based on historical data, enabling proactive forecasting of changes in patient status. The risk index calculation engine, with status evolution prediction results as its core, integrates complication probability and recovery delay risk indicators, generating an intuitive risk index through quantitative algorithms. This helps doctors quickly identify high-risk patients and accurately adjust treatment plans, while also providing quantitative evidence for medical resource allocation and treatment outcome review. This improves the efficiency of the entire surgical cycle, reduces the aggravation of complications or recovery delays caused by insufficient risk prediction, and improves the quality of patient prognosis.

[0039] It should be noted that the entire patient lifecycle is divided into the preoperative assessment period, the intraoperative monitoring period, and the postoperative recovery period. The physiological status and recovery progress at each stage are analyzed. From the perspective of the preoperative assessment period, this stage usually covers one week to one day before surgery and is a key link in the early prevention and control of surgical risks. The analysis focuses on the patient's basic physiological status and surgical suitability. This requires reviewing the patient's static basic data, such as age, weight, history of underlying diseases, and key preoperative physiological indicators to determine whether the patient has any surgical contraindications. It also requires assessing the patient's physiological suitability potential in conjunction with the surgical plan. For example, for patients undergoing lung surgery, lung function tests are used to analyze their postoperative lung function recovery ability. For patients undergoing joint replacement surgery, lower limb muscle strength and range of motion are assessed to predict the starting point of postoperative rehabilitation training, screen high-risk suitable patients in advance, and provide a preoperative baseline for key intraoperative monitoring and postoperative recovery goals. This avoids sudden intraoperative risks or unclear postoperative recovery starting points due to insufficient preoperative assessment. During intraoperative monitoring, which spans the entire surgical procedure, this stage serves as the core window for real-time risk intervention. The focus shifts to the synergistic fit between dynamic physiological states and surgical procedures. This requires relying on intraoperative monitoring equipment to collect and analyze the patient's dynamic physiological indicators in real time: during anesthesia induction, the focus is on the fluctuations in heart rate and blood pressure; during the surgical procedure, the focus is on the correlation between vital signs and surgical procedures; and during suturing, the focus is on the patient's recovery of spontaneous breathing, level of consciousness, etc., to assess anesthesia recovery suitability. Simultaneously, it is necessary to analyze the rationality of the physiological stress response in conjunction with the details of the surgical procedure. For example, if the surgery lasts longer than expected by one hour and the patient's heart rate remains above 120 beats / min and blood pressure is 20% higher than baseline, it is necessary to determine whether there is excessive intraoperative stress and adjust the anesthesia plan or fluid resuscitation strategy accordingly. Through real-time dynamic analysis, abnormal physiological signals during surgery can be quickly captured to prevent the accumulation of risks postoperatively, while also providing an intraoperative stress baseline for postoperative recovery assessment. The postoperative recovery period, typically extending from 6 hours after surgery until patient discharge, is a crucial stage for tracking recovery progress and managing subsequent risks. The analysis focuses on the alignment of physiological recovery trends with recovery goals. In the early postoperative period (6 hours to 1 day), the focus is on acute recovery indicators such as vital sign stability, wound condition, drainage tube flow and nature, and the first postoperative blood routine test and inflammatory markers to determine if the patient has passed the acute postoperative risk period. In the mid-postoperative period (2-7 days), the focus shifts to functional recovery indicators such as independent movement ability, organ function recovery, and changes in pain scores. In the long-term postoperative period (follow-up), the focus is on rehabilitation integrity indicators such as wound healing quality, basic functional recovery, and the presence of residual complications. Simultaneously, the recovery progress at each stage is compared with the standard recovery cycle for the same disease. If deviations in recovery indicators are observed, the causes are promptly investigated and the intervention plan adjusted.

[0040] Self-supervised training based on unlabeled clinical data predicts the evolution of patient conditions. By designing targeted pre-training tasks, the Perceiver model is trained and autonomously mines supervisory signals from massive amounts of unlabeled clinical data. It learns both common evolutionary patterns of the population and captures individual differences. Only current and historical unlabeled patient data need to be input to accurately output short-term (e.g., predicting whether there will be signs of infection 12 hours after 6 hours post-surgery) and long-term (e.g., predicting 1 week of activity ability and 1 month of recovery completeness after 3 days post-surgery) state evolution results. This fully releases the value of more than 80% of unlabeled clinical data, significantly reduces data processing costs, and can predict abnormal trends in patient conditions in advance, allowing doctors to adjust treatment plans ahead of time.

[0041] Perceiver model formula:

[0042]

[0043]

[0044] in, The symbol for cross-attention operation; For query vector; The key vector; It is a value vector; Use the Softmax activation function; This is the transpose of the key matrix K; This is the transpose of the query matrix and the key matrix; Key vector The dimension; This is the scaling factor; For self-attention operations; The input feature matrix; To query the weight matrix; This is the key weight matrix; Value weight matrix; for Transpose of; This is a symbol for a feedforward neural network. It is a Gaussian error linear unit; This is the first layer weight matrix; For the first layer bias; This is the weight matrix for the second layer; This is the second layer bias.

[0045] Based on the patient's evolving condition, a quantitative risk index is generated by integrating the probability of complication occurrence and the risk indicators of delayed recovery. By matching similar historical cases, the probability of various complications such as postoperative infection and bleeding is calculated using Bayesian probability formulas. By comparing with the standard recovery cycle, the progress deviation of the patient at key points is converted into a risk value of delayed recovery. Finally, by combining the weights of the severity of complications, a quantitative risk index with a scale of 0-100 is generated through a weighted fusion formula, which comprehensively and intuitively reflects the patient's risk level throughout the entire cycle, helping doctors to adjust treatment plans in advance.

[0046] Bayesian probability formula:

[0047] in, Presenting the patient's evolving condition When complications occur The posterior probability; For historical complications In the cases, the current state has evolved. The conditional probability; Complications in historical databases The prior probability; State evolution appears in historical database The marginal probability.

[0048] Weighted fusion formula:

[0049] in, This serves as the final quantitative risk index. Number of complication categories; For the first Weights for complications; For the first The probability of such complications occurring; Weighting for the risk of recovery delay; This represents the total risk value for recovery delay. : Normalization coefficient, ensuring the result is mapped to the 0-100 range; i is the index symbol.

[0050] For example, taking a 58-year-old patient undergoing laparoscopic cholecystectomy (with type 2 diabetes) as an example, the phased status assessment module focuses on analyzing the patient's basic physiological status and surgical suitability during the preoperative assessment period (3 days to 1 day before surgery): reviewing the patient's fasting blood glucose fluctuations (7.9-9.1 mmol / L) and glycated hemoglobin (6.8%) over the past 3 days, combined with liver and kidney function (ALT 38 U / L, creatinine 76 μmol / L) and cardiopulmonary function (normal ECG, oxygen saturation 98%), determining that the preoperative blood glucose lowering regimen needs to be adjusted (changing oral medication to short-acting insulin to lower fasting blood glucose to 7.2 mmol / L 1 day before surgery) to avoid intraoperative blood glucose fluctuations or postoperative infection risks; during the intraoperative monitoring period (from anesthesia induction to the end of surgery, lasting 75 minutes), physiological indicators at different stages are tracked in real time: during anesthesia induction period Monitor heart rate (maintained at 60-70 beats / min) and blood pressure (stable at 120-130 / 70-80 mmHg, avoiding a drop below 15% of baseline). During the gallbladder removal procedure, focus on monitoring blood loss (cumulative <30ml) and end-tidal carbon dioxide partial pressure (35-45 mmHg). If the patient's blood pressure briefly drops to 105 / 65 mmHg, immediately coordinate with the anesthesia team to adjust the fluid resuscitation rate to ensure physiological status and surgical procedure coordination. During the postoperative recovery period (6 hours to 4 days post-surgery), track recovery progress in stages: 6 hours post-surgery, monitor body temperature (36.5-37.1℃) and wound exudate (no redness, swelling, or bleeding); 2 days post-surgery, assess independent activity ability (able to walk 20 meters) and dietary tolerance (no abdominal distension on semi-liquid diet); 3 days post-surgery, repeat blood routine test (white blood cell count 6.5×10⁻⁶). 9 / L, no inflammation elevation), compared with the standard recovery period for laparoscopic cholecystectomy (getting out of bed 1 day after surgery and being discharged 3-4 days after surgery), it was found that the patient's wound healing speed was slightly slower than the standard due to diabetes (the epidermal scab formation was delayed by about 12 hours). The frequency of local dressing changes was increased in time, and the patient was finally discharged smoothly 4 days after surgery without complications. This fully demonstrates that the module achieves closed-loop management of preoperative risk prevention and control, real-time intervention during surgery, and postoperative progress tracking through phased and targeted assessment, avoiding risk omissions caused by general analysis of the whole cycle.

[0051] In this embodiment, the adaptive clinical decision module 400 includes, as follows: Figure 5 As shown, the module includes a departmental needs analysis unit, a decision rule generation module, and a suggestion optimization unit.

[0052] The departmental needs analysis unit is used to identify the clinical needs of different departments such as internal medicine, surgery, and anesthesiology; the decision rule generation module is used to generate preliminary treatment suggestion rules based on the patient risk index and departmental needs; and the suggestion optimization unit is used to iteratively optimize the preliminary suggestions by combining historical treatment effect data to generate customized clinical suggestions.

[0053] Understandably, the embodiments of this application accurately identify the differentiated needs of internal medicine, surgery, and anesthesiology through the departmental needs analysis unit, avoiding the disconnect between recommendations and clinical reality due to neglecting the professional characteristics of each department; the decision rule generation module uses the patient's quantitative risk index as the core basis and combines the needs of each department to generate preliminary treatment recommendation rules, ensuring that the recommendations are both in line with the individual risk of the patient and adapted to the department's treatment focus; the recommendation optimization unit relies on historical treatment effect data to iteratively optimize the preliminary recommendations, and finally generate customized clinical recommendations. By verifying the risk correlation and historical data, the scientificity and suitability of the recommendations are improved, helping doctors in each department to efficiently obtain treatment plans that fit their own needs and the patient's condition, reducing decision-making bias and promoting the improvement of multi-departmental collaboration efficiency.

[0054] It should be noted that identifying the clinical needs of different departments such as internal medicine, surgery, and anesthesiology begins by anchoring the core functional needs of each department—internal medicine focuses on preoperative management of underlying diseases (e.g., achieving target blood glucose levels in diabetic patients) and diagnosis and treatment of postoperative complications (e.g., drug intervention for infections); surgery emphasizes surgical plan adaptation (e.g., laparoscopic / open surgery choice), intraoperative coordination, and postoperative functional recovery (e.g., wound healing and activity training); and anesthesiology focuses on preoperative anesthesia tolerance assessment (e.g., airway patency and the impact of liver and kidney function on anesthetic metabolism), intraoperative vital sign stabilization, and postoperative recovery management (e.g., pain control and antiemetics); then... The system details the needs of each stage of surgery (preoperative, intraoperative, and postoperative) (e.g., preoperatively, internal medicine sets target indicators, surgery prepares the skin, and anesthesiology checks the completeness of examinations; intraoperatively, internal medicine responds to sudden onset of underlying diseases, surgery requires operational cooperation, and anesthesiology controls vital sign thresholds; postoperatively, internal medicine prevents complications, surgery tracks recovery progress, and anesthesiology addresses residual anesthesia). Simultaneously, it uncovers implicit collaborative needs (e.g., internal medicine needs surgery to clarify the operation time to control the cycle, and surgery needs anesthesiology to estimate the anesthesia duration to plan the process). This comprehensive approach covers both explicit and implicit needs, avoiding confusion or disconnect between requirements and laying the foundation for generating appropriate treatment recommendations for each department.

[0055] By combining historical treatment outcome data, the initial recommendations are iteratively optimized to generate customized clinical recommendations. These recommendations are based on individual patient characteristics (such as age, underlying diseases, and risk index), core elements of the initial recommendations (such as medication, monitoring, and rehabilitation plans), and departmental needs. Historical case data of patients with similar surgeries and conditions are selected to extract outcome indicators such as complication rates, recovery periods, and changes in indicators. Then, through horizontal comparison (such as differences in infection rates among different medication regimens) and longitudinal tracking (such as the correlation between dressing change frequency and wound healing), the shortcomings of the initial recommendations are assessed (e.g., high infection rates with single antibiotics, insufficient dressing change frequency in diabetic patients). This is followed by multiple rounds of adjustment, verification, and further adjustment (e.g., changing to combination therapy, adding local care, verifying effects with detailed data, and then optimizing dosage or adding skin testing). Finally, a personalized recommendation is generated that integrates individual characteristics, historical verification, and departmental needs (including precise medication dosage, monitoring points, and key points for departmental collaboration), ensuring that the recommendations are scientific and aligned with clinical practice, thus helping to improve treatment outcomes.

[0056] For example, taking a 62-year-old patient undergoing radical gastrectomy for gastric cancer (with hypertension, mild renal insufficiency, and a quantitative risk index of 72, including a 28% probability of postoperative bleeding and a 35-point risk value for delayed recovery) as an example, the decision rule generation module generates preliminary treatment recommendations based on the patient's risk index and the needs of internal medicine, surgery, and anesthesiology: For the internal medicine department's need for "preoperative management of underlying diseases + prevention of postoperative complications," combined with the patient's hypertension (daily blood pressure 150 / 95 mmHg) and bleeding risk, the module generates the rule: "Three days before surgery, adjust the antihypertensive medication to a long-acting calcium channel blocker, targeting blood pressure at 130-140 / 80-85 mmHg; monitor renal function (serum creatinine, blood urea nitrogen) and coagulation function (PT, APTT) daily postoperatively; if coagulation indicators are abnormal, promptly administer low molecular weight heparin to prevent thrombosis." For the surgical department's need for "…", the module generates the following recommendations: Based on the needs for "surgical procedure adaptation and postoperative recovery management," and considering the probability of bleeding and the type of surgery, the following rules were generated: "Use ultrasonic scalpel for precise hemostasis during surgery; choose a midline abdominal incision to shorten operation time; observe incision bleeding every 6 hours within 24 hours postoperatively; and begin progressive enteral nutrition support on the 3rd postoperative day (initial dose 20ml / h, increasing gradually if there is no discomfort)." For the anesthesiology department's needs for "preoperative anesthesia assessment and intraoperative vital sign stability," and considering mild renal insufficiency and bleeding risk, the following rules were generated: "Use propofol + remifentanil anesthesia with minimal impact on renal function preoperatively; continuously monitor invasive arterial blood pressure (recorded every 5 minutes) and urine output (target ≥0.5ml / kg per hour) intraoperatively; if blood pressure falls below 20% of baseline, immediately initiate fluid resuscitation and vasoactive drug intervention." These rules not only closely address the patient's high risk of bleeding and delayed recovery but also precisely adapt to the treatment priorities of each department, forming a multi-dimensional collaborative preliminary treatment framework, laying the foundation for subsequent optimization based on historical data.

[0057] In this embodiment of the application, the cross-departmental data collaboration unit 500 includes, for example: Figure 6 As shown, the visualization platform includes a construction module, a data sharing and permission management unit, and a real-time interaction module.

[0058] The visualization platform module utilizes WebGL to build a 3D visualization collaborative interface, displaying the patient's full-cycle feature map and analysis results; the data sharing permission management unit is used to allocate data viewing and editing permissions according to the department and responsibilities; and the real-time interaction module is used to annotate key data online, add diagnostic and treatment opinions, and conduct real-time interactive collaboration.

[0059] It is understood that the visualization platform construction module in this application uses WebGL to create a three-dimensional visualization collaborative interface, presenting the patient's preoperative basic disease indicators, intraoperative vital sign fluctuations, postoperative recovery progress and other full-cycle characteristic maps and risk analysis results in intuitive three-dimensional charts, avoiding data comprehension bias caused by traditional text or two-dimensional tables, and helping various departments quickly grasp key information; the data sharing permission management unit accurately allocates permissions according to departmental functions and medical staff responsibilities, ensuring patient data privacy and security while avoiding irrelevant data interference or missing key data permissions, ensuring that each department can obtain and operate securely as needed; the real-time interaction module supports medical staff from multiple departments to annotate key data online and add treatment opinions in real time, eliminating the need for offline meetings or repeated file transfers, greatly shortening the communication cycle, reducing information transmission errors, improving the consistency and efficiency of treatment decisions in various departments through data collaboration, helping multiple departments form a unified understanding of treatment around the patient's full-cycle condition, reducing treatment bias caused by information asymmetry, and improving the patient's treatment experience and prognosis.

[0060] For example, taking a 70-year-old patient undergoing unilateral hip replacement surgery (with hypertension and type 2 diabetes, and a preoperative risk index of 68) as an example, the visualization platform module uses WebGL technology to create a 3D visualization collaborative interface, transforming the patient's full-cycle data into an intuitive dynamic map: In the preoperative stage, the interface displays the trend of underlying disease indicators using a 3D line graph overlaid with a risk heat map—the blue line shows the patient's blood pressure fluctuations three times a day (morning, noon, and evening) in the two weeks before surgery (gradually decreasing from 155 / 92 mmHg to 138 / 85 mmHg), and the orange line simultaneously shows the changes in fasting blood glucose (adjusted from 8.3 mmol / L to 7.1 mmol / L). The risk heat map uses yellow to mark abnormal periods of "blood pressure fluctuations > 10 mmHg in the 3 days before surgery," facilitating the internal medicine department to quickly locate key nodes in the control of underlying diseases; In the intraoperative stage, the interface switches to a 3D dynamic monitoring model—the surgical operation scene is rendered in real time in the center (hip prosthesis implantation position, dynamic statistics of intraoperative blood loss, and cumulative blood loss expressed in red bars). The image updates in real time (currently displaying 45ml). The right-hand 3D panel synchronously displays indicators such as heart rate (green waveform, stable at 65-75 beats / min), depth of anesthesia (purple value, maintained at 40-50%), and blood oxygen saturation (blue value, 98%-99%). Anesthesiologists can drag the model to view the correlation between key intraoperative operations and physiological indicators, quickly identifying subtle fluctuations such as "blood pressure briefly rising to 142 / 88mmHg during prosthesis implantation." In the postoperative stage, the interface presents a 3D recovery progress radar chart—using different colored sectors to show the current recovery status from four dimensions: "wound healing," "mobility," "inflammation control," and "underlying disease stability" (the wound healing sector reaches 80%, corresponding to the visualization effect of the incision redness and swelling subsiding in the 3D model; the mobility sector reaches 60%, synchronously displaying the dynamic simulation that the patient can walk 30 meters with the help of a walker 3 days after surgery). Surgeons can intuitively compare with the standard recovery cycle (mobility should reach 75% 1 week after surgery) and adjust the rehabilitation training plan in a timely manner. This interface avoids the problems of scattered data and time-consuming interpretation in traditional text reports, allowing medical staff in internal medicine, surgery, and anesthesiology to quickly grasp key information about patients throughout their entire life cycle without repeated communication, providing intuitive data support for cross-departmental collaborative decision-making.

[0061] In this embodiment of the application, the dynamic feedback optimization unit 600 includes, as follows: Figure 7 As shown, the diagnostic results acquisition module, feedback data matching unit, and model parameter fine-tuning engine are included.

[0062] The diagnostic and treatment result acquisition module is used to collect data on postoperative complications, recovery period, treatment effect scores, and clinical outcomes of patients; the feedback data matching unit is used to associate and match the diagnostic and treatment results with the corresponding patient's model input data and prediction results; and the model parameter fine-tuning engine is used to fine-tune the weight parameters in reverse using the gradient descent algorithm based on the matched feedback data to optimize the prediction accuracy.

[0063] It is understood that the treatment result acquisition module in this application focuses on core postoperative clinical data, comprehensively collecting information on patient complications, actual recovery periods, and treatment effect scores. This provides real and vivid clinical feedback for model optimization, preventing the model from gradually becoming ineffective due to its disconnect from actual treatment data. The feedback data matching unit accurately correlates the collected actual treatment results with the corresponding patient's model input data and initial prediction results, constructing a data chain, clearly locating model prediction biases, and providing a clear direction for subsequent parameter adjustments. Based on the correlated feedback data, the model parameter fine-tuning engine uses a gradient descent algorithm to backtrack and fine-tune the model weight parameters, gradually correcting prediction biases. It continuously learns based on accumulated clinical data, continuously improving prediction accuracy to help doctors more accurately assess patient risks, formulate treatment plans, reduce treatment errors caused by model biases, and improve overall treatment quality and patient prognosis.

[0064] It's important to note that the process of linking treatment outcomes with the corresponding patient's model input data and prediction results involves several steps. First, using the patient's unique ID as the core link, the actual postoperative treatment outcome is bound to the patient's preoperative / intraoperative model input data and the model's output prediction results, ensuring data ownership remains consistent. Second, the linking dimensions are refined according to the treatment timeline. For example, a diagnosis of infection on the 3rd postoperative day precisely corresponds to the physiological indicator input data from 3 days before surgery to 6 hours after surgery, as well as the infection risk prediction results generated based on this data for days 1-7 postoperatively, clearly identifying the specific stage at which the deviation occurred. Furthermore, point-to-point matching is performed on prediction items. For instance, a model prediction of a 7-day postoperative recovery period corresponds to an actual discharge on the 9th day, and a model prediction of no bleeding risk corresponds to a small amount of bleeding on the 2nd postoperative day. This precise matching clearly identifies deviations in specific prediction dimensions, such as the model's assessment of bleeding risk during the recovery period, avoiding vague problem localization caused by general associations. Finally, the matched data intuitively presents which predictions the model made based on which inputs, and what the actual treatment outcome was, providing a clear direction for deviations in subsequent model parameter fine-tuning, ensuring more targeted optimization.

[0065] Patient Unique ID Binding Formula:

[0066] in, The name of the data association set; This serves as a unique identifier for the i-th patient; A unique identifier for the patient; For the number of patients; For patients Preoperative multimodal data; For real-time monitoring data during surgery; This represents the actual postoperative diagnosis and treatment results.

[0067] Timeline refinement correlation formula:

[0068] in, For time mapping functions; This is a mapping symbol with a time offset; For time difference; The timestamp of the actual event; The timestamp for the event predicted by the model; is the input data vector for the i-th patient from the preoperative to the intraoperative combined stage.

[0069] Predicted item point-to-point matching formula:

[0070] in, For the prediction-actual matching matrix; This is the predicted recovery period for the i-th patient; This is the predicted risk of complications for the i-th patient; This represents the actual recovery period for the i-th patient. Mark the actual occurrence of complications for the i-th patient.

[0071] Gradient descent algorithm formula:

[0072] in, These are learnable weight parameters; For the first The parameter values ​​at the next iteration; For the first The parameter values ​​after the next iteration; The learning rate; For gradient operators; This is the loss function.

[0073] For example, taking a 70-year-old patient undergoing unilateral hip replacement surgery (with hypertension and type 2 diabetes, and a preoperative risk index of 68) as an example, the visualization platform module uses WebGL technology to create a 3D visualization collaborative interface, transforming the patient's full-cycle data into an intuitive dynamic map: In the preoperative stage, the interface displays the trend of underlying disease indicators using a 3D line graph overlaid with a risk heat map—the blue line shows the patient's blood pressure fluctuations three times a day (morning, noon, and evening) in the two weeks before surgery (gradually decreasing from 155 / 92 mmHg to 138 / 85 mmHg), and the orange line simultaneously shows the changes in fasting blood glucose (adjusted from 8.3 mmol / L to 7.1 mmol / L). The risk heat map uses yellow to mark abnormal periods of "blood pressure fluctuations > 10 mmHg in the 3 days before surgery," facilitating the internal medicine department to quickly locate key nodes in the control of underlying diseases; In the intraoperative stage, the interface switches to a 3D dynamic monitoring model—the surgical operation scene is rendered in real time in the center (hip prosthesis implantation position, dynamic statistics of intraoperative blood loss, and cumulative blood loss expressed in red bars). The image updates in real time (currently displaying 45ml). The right-hand 3D panel synchronously displays indicators such as heart rate (green waveform, stable at 65-75 beats / min), depth of anesthesia (purple value, maintained at 40-50%), and blood oxygen saturation (blue value, 98%-99%). Anesthesiologists can drag the model to view the correlation between key intraoperative operations and physiological indicators, quickly identifying subtle fluctuations such as "blood pressure briefly rising to 142 / 88mmHg during prosthesis implantation." In the postoperative stage, the interface presents a 3D recovery progress radar chart—using different colored sectors to show the current recovery status from four dimensions: "wound healing," "mobility," "inflammation control," and "underlying disease stability" (the wound healing sector reaches 80%, corresponding to the visualization effect of the incision redness and swelling subsiding in the 3D model; the mobility sector reaches 60%, synchronously displaying the dynamic simulation that the patient can walk 30 meters with the help of a walker 3 days after surgery). Surgeons can intuitively compare with the standard recovery cycle (mobility should reach 75% 1 week after surgery) and adjust the rehabilitation training plan in a timely manner. This interface avoids the problems of scattered data and time-consuming interpretation in traditional text reports, allowing medical staff in internal medicine, surgery, and anesthesiology to quickly grasp key information about patients throughout their entire life cycle without repeated communication, providing intuitive data support for cross-departmental collaborative decision-making.

[0074] This application proposes a smart healthcare data self-supervised deep processing and analysis system. It integrates departmental data and postoperative follow-up data through a multi-departmental data perception layer, providing a complete data foundation for full-cycle analysis. The surgical feature processing center uses an attention mechanism to correlate pre- and post-operative data, generating a unified feature map to improve the relevance and accuracy of feature extraction. A full-cycle dynamic analysis unit, combined with a self-supervised prediction model, outputs risk indices in stages, dynamically monitoring patient status and predicting risks. An adaptive clinical decision-making module generates customized suggestions for different departments, enhancing the personalization and practicality of clinical decisions. A cross-departmental data collaboration unit promotes data sharing and interaction through a visualization platform, breaking down barriers to cross-departmental collaboration. A dynamic feedback optimization unit relies on a clinical outcome-based reverse fine-tuning model to ensure continuous iterative optimization of system performance, improving the utilization efficiency and analytical accuracy of medical data, while also strengthening cross-departmental collaboration and the scientific nature of clinical decision-making, effectively supporting efficient diagnosis and treatment and patient safety in smart healthcare scenarios. This solves the problems of rigid strategies and insufficient energy utilization efficiency in existing technologies.

[0075] The following will illustrate a smart healthcare data self-supervised deep processing and analysis system through a specific embodiment, such as... Figure 8 As shown, it includes: Taking the perioperative management of laparoscopic cholecystectomy (LC) in the hepatobiliary surgery department of a primary-level tertiary hospital as the application scenario, this system addresses the characteristics of this surgery, which involves multi-departmental collaboration involving internal medicine (preoperative underlying disease assessment), surgery (surgical operation), and anesthesiology (intraoperative monitoring), and is prone to problems such as bile leakage, infection, and delayed recovery after surgery. The system deeply integrates each module with the actual medical process, with the multi-departmental data perception layer serving as the core link for system data input. It focuses on the collection and standardized integration of medical data in all dimensions and with high precision, providing a solid data foundation for subsequent feature processing and analysis. The data acquisition module is not simply data scraping. Based on the hospital's existing information system architecture, it connects to the Electronic Medical Record (EMR) system and the Orthopaedic Surgical Information System (ORIS) via the HL7FHIR international medical data exchange standard. This enables automated, non-manual collection of core diagnostic and treatment data from multiple departments. For internal medicine, the module focuses on acquiring fasting blood glucose (collected daily at 8:00 AM, reference range 3.9-6.1 mmol / L, this frequency is based on the clinical need for continuous monitoring of stable preoperative baseline blood glucose), blood pressure (collected three times a day at 7:00 AM, 12:00 PM, and 7:00 PM, target <140 / 90 mmHg, covering the patient's diurnal blood pressure fluctuations), and liver and kidney function indicators (ALT, AST, bilirubin, automatically synchronized to the system after detection using a biochemical analyzer). For surgery, the collected data includes preoperative gallbladder ultrasound images (using GE LogiqE). 9. Ultrasound imaging, 512×512 resolution, including key information such as gallbladder wall thickness and stone size / number; surgical operation data (operation time accurate to the minute, intraoperative blood loss calculated by suction capacity + gauze weight method, hemostasis number recorded in real time by the surgical nurse); and postoperative incision score (based on a 5-level standard of "red-swelling-heat-pain-exudation", which is checked and marked by the surgeon in the system daily, with level 1 being normal and level 5 being severe infection); for the anesthesiology department, real-time collection of intraoperative circulatory monitoring data (heart rate every 10 seconds, blood pressure every 30 seconds, and end-tidal carbon dioxide every 5 seconds, transmitted via a 5G private network with low latency through the anesthesia monitor to ensure data real-time performance); anesthetic drug dosage (propofol calculated in mg / kg, remifentanil calculated in μg / kg / min, automatically calculated by the anesthesia machine); and postoperative analgesia plan (background dose, single press dose, lockout time, and other parameters of the intravenous PCA pump).The follow-up data integration unit improves the data chain from the perspective of "continued monitoring outside the hospital." Postoperative patients are uniformly provided with Huawei Sports Health Watches (GT5). Before discharge, nurses bind the devices to the hospital system (connecting via Bluetooth to the patient's mobile app, which then interfaces with the hospital's follow-up system) and provide 15 minutes of one-on-one training (including device charging, viewing abnormal pop-up windows, and interpreting activity targets). Data collection focuses on the critical recovery period of 1-7 days post-surgery, during which heart rate is automatically recorded every 30 minutes. When the resting heart rate is >100 bpm or <50 bpm... Upon arrival, the device immediately pops up a reminder to the patient and simultaneously pushes an alert to the attending physician's mobile phone. Activity levels are tracked by daily cumulative steps (a target of ≥2000 steps is set, as clinical studies have shown this number can promote gastrointestinal motility and reduce the risk of postoperative intestinal adhesions) and exercise intensity (determined by the device's built-in accelerometer; moderate-intensity exercise sessions should last <10 minutes to avoid excessive activity affecting incision healing). Sleep quality is recorded by daily deep sleep duration (target ≥2 hours, insufficient deep sleep can lead to decreased immunity) and nighttime awakenings (≤2 times, frequent awakenings often indicate postoperative pain or discomfort). To address the issue of time misalignment in multi-source data, all follow-up data is synchronized to a "UTC+8" timestamp via the hospital's intranet NTP time server and automatically associated with the "postoperative days" in the hospital system. For example, heart rate data collected at 14:00 on the 3rd postoperative day will be automatically marked as "72h postoperatively," ensuring consistency with preoperative and intraoperative data in terms of time dimension.

[0076] The perioperative feature processing center undertakes the core function of transforming "data → features → atlas," constructing a unified perioperative feature atlas that can be intuitively interpreted through professional feature extraction and intelligent correlation analysis. In the feature extraction unit, a dual-drive model of "domain knowledge guidance + deep learning empowerment" is adopted to ensure the medical professionalism and quantitative accuracy of the features: For physiological indicators, raw data is not used directly, but rather time-series trend analysis of internal medicine blood glucose and blood pressure data is performed using Python's Pandas library. For example, the fluctuation range of blood glucose in the 3 days before surgery is calculated (i.e., the difference between the maximum and minimum blood glucose values ​​within 3 days; ≤1.5mmol / L is defined as stable, and this threshold refers to the clinical standard for preoperative management of underlying diseases). For intraoperative blood pressure data in the anesthesiology department, the "percentage of abnormal duration" is calculated (the time when blood pressure exceeds the range of 50-100 beats / min divided by the total surgical time; ≤5% is acceptable, and an excessively high percentage of abnormalities indicates intraoperative circulatory instability). For imaging features, a ResNet50 deep learning model is used for processing. This model is based on 2000 gallbladder ultrasound imaging data from the past 5 years (analyzed by 3 associate chief physicians). The pre-training weights were fine-tuned using labeled data. The training process utilized a hospital GPU cluster (4 NVIDIA Tesla V100 graphics cards, 50 iterations, and cross-entropy as the loss function). Ultimately, the system can automatically segment the gallbladder region in ultrasound images and extract quantitative features such as "mean gallbladder wall thickness" (normal <3mm, >3mm suggests chronic cholecystitis) and "maximum stone diameter" (>2cm increases the risk of stone impaction during surgery). The feature extraction accuracy rate reached over 92%. For surgical operation features, the raw surgical data was transformed into more clinically meaningful derived features, such as "operation efficiency features" (intraoperative blood loss divided by operation time, ≤5ml / min is excellent; for example, if a patient's intraoperative blood loss was 50ml and the operation time was 10 minutes, the efficiency feature is 5ml / min, which meets the excellent standard) and "operation safety features" (the number of hemostasis attempts ≤2 is excellent; excessive hemostasis attempts indicate that the surgical area's blood vessels are difficult to manage).The cross-time series data association module breaks down the data fragmentation between preoperative, intraoperative, and postoperative periods through an 8-head attention mechanism (based on the Transformer architecture, with each attention head having 64 dimensions, a total of 512 dimensions, and using the ScaledDot-ProductAttention calculation method). The input data covers the 7 days before surgery (time steps T1-T7, one time step per day, including blood glucose and blood pressure data) and the 60 minutes during surgery (time steps T8-T13, one time step every 10 minutes, including heart rate, blood pressure, and end-tidal carbon dioxide). Carbon data), 7 days post-surgery (time steps T14-T20, one time step per day, including heart rate and activity data), and hidden correlations were uncovered through attention weight allocation. For example, after training on 1000 patient data, the model automatically found that patients with "preoperative blood glucose fluctuation >2mmol / L in 3 days" and "abnormal blood pressure duration >10% during surgery" had a postoperative infection rate of 35%, which is much higher than the 8% of ordinary patients. Therefore, a high attention weight of 0.8 was assigned to this type of temporal correlation (compared to only 0.2-0.3 for ordinary correlation), highlighting its clinical importance. The feature map generation engine uses the Neo4j graph database to construct individual patient feature maps. The database is deployed on the hospital's private cloud (using a master-slave architecture, with the master node responsible for data writing and the slave nodes responsible for querying, capable of handling 150+ query requests per second to ensure high availability). The map contains three types of nodes: time nodes (accurate to the day, such as 7 days before surgery, during surgery, and 1-7 days after surgery), indicator nodes (with specific values ​​and normal ranges, such as "blood glucose 5.8 mmol / L (3.9-6.1)"), and risk nodes (marking the risk level, such as "infection risk - medium"), as well as two types of relationships: "time-indicator" association (recording indicator collection). The time frame (e.g., "3 days post-surgery → heart rate 105 bpm") is linked to "indicators-risks" (the direction of the indicator's impact on risk is indicated, e.g., "gallbladder wall thickness 5mm → infection risk ↑"). In the doctor's workstation's visual interface, the graph supports zooming, filtering (e.g., displaying only abnormal indicators), and hovering to view details (hovering the mouse over an indicator node displays the normal range, and abnormal data is highlighted in red). Clicking on a risk node expands all associated indicator bases. For example, clicking "high infection risk" will display "large fluctuations in preoperative blood glucose (4.2-6.8 mmol / L) + insufficient postoperative activity (1200 steps / day)," helping doctors quickly understand the source of risk.

[0077] The full-cycle dynamic analysis unit, with "phased assessment + precise prediction" as its core, enables dynamic monitoring and risk quantification of the perioperative status of patients undergoing LC surgery, providing a basis for clinical intervention decisions. The phased status assessment module, based on the clinical treatment pathway of LC surgery, scientifically divides the patient's entire life cycle into three key stages. The division and assessment criteria for each stage refer to the clinical routine of perioperative management for laparoscopic cholecystectomy: The preoperative assessment period is set from T-7d to T-1d (the 7-day duration is because medication adjustments for underlying diseases such as hypertension and diabetes require 5-7 days to achieve a stable effect), assessing "underlying disease control." The target criteria are "blood glucose fluctuation ≤1.5mmol / L in the 3 days before surgery" and "blood pressure control days ≥5 days" (control means systolic blood pressure <140mmHg and diastolic blood pressure <90mmHg). Those who meet the criteria are classified as "low risk," and those who do not meet either criterion are classified as "medium risk." The assessment results are updated daily and pushed to the internist's workstation. Medium-risk patients need to increase the follow-up frequency (twice daily). The intraoperative monitoring period is set from T0 to T+1h (covering LC surgery). The average operation time is 40-60 minutes (with a 20-minute buffer time). "Intraoperative stability" is assessed, with the criteria being "abnormal blood pressure duration ≤ 5%" and "intraoperative blood loss / operation time ≤ 5 ml / min". Meeting both criteria indicates "low risk", failing one indicates "medium risk", and failing two indicates "high risk". High-risk patients will trigger real-time alerts from the anesthesiology and surgery departments, requiring immediate adjustments to the surgical or anesthesia plan. The postoperative recovery period is set from T+1 to T+7 days (this stage is the peak period for postoperative complications and the recovery period for physical function). "Recovery progress" is assessed, with the criteria being "activity level ≥ 1500 steps 3 days post-operation" (below this number indicates slow gastrointestinal function recovery) and "no redness, swelling, or effusion at the incision site 5 days post-operation". Meeting both criteria indicates "normal recovery", and failing either indicates "delayed recovery". Patients with delayed recovery require a personalized rehabilitation plan developed by the surgeon.To address the high cost of medical data annotation, the self-supervised prediction model training unit employs an unsupervised training mode. Training data is derived from perioperative data of 5000 patients undergoing LC surgery at the hospital over the past three years. Differential privacy technology (adding Gaussian noise with an intensity ε=0.1) and patient identity de-identification (removing names, ID numbers, and contact information, retaining only the hospital number as a unique identifier) ​​are used. The training task is designed as a "time-series data completion task"—randomly masking 10% of the time-series segments in the data (e.g., masking blood pressure data at the 20th-30th minute of surgery, or activity data on the second postoperative day). The Perceiver model (input dimension 512, encoder layers 12, decoder layers 6) predicts the values ​​of the masked segments using related data from other time periods. The model training... The training used a hospital AI training platform (equipped with 8 NVIDIA A100 GPUs and employing mixed precision training). The mean absolute error (MAE) was used as the evaluation metric. After 80 training rounds, the MAE decreased from an initial 15.6 to 3.2, indicating that the model had fully learned the temporal patterns of perioperative data. Upon completion of training, the model outputs a "patient status evolution curve," which displays the predicted trends of daily activity levels, heart rate, sleep quality, and other indicators from 1 to 7 days post-surgery in a line graph format. In the graph, blue represents the predicted curve, and red represents the normal reference range. When the predicted curve deviates from the normal range by more than 30% (e.g., predicting 1200 steps of activity on day 3 post-surgery, compared to the normal reference value of 2000 steps, a deviation rate of 40%), the system automatically marks it as a "high-risk evolution" and sends an alert to the attending physician. The risk index calculation engine integrates "stage assessment results" and "model prediction results" to construct a quantitative risk index system ranging from 0 to 100 points: the base score is 60 points (set as the baseline score, representing the ideal state with no significant risk). If the preoperative, intraoperative, and postoperative stages all meet the standards, 60 points are retained; for each stage that fails to meet the standards, 10 points are deducted (e.g., if the preoperative or intraoperative stages fail to meet the standards, 20 points are deducted, and the base score becomes 40 points); the prediction deviation score has a maximum of 40 points and is calculated as "the deviation rate between the model's predicted state and the normal curve × 40", where the deviation rate = (predicted...). The deviation score is calculated as follows: (Value - Normal Reference Value) / Normal Reference Value (Absolute Value). The higher the deviation rate, the more points are added. For example, if a patient's predicted activity level 3 days after surgery is 1800 steps, the normal reference value is 2000 steps, and the deviation rate is 10%, then the deviation score = 10% × 40 = 4 points. If the deviation rate is 50%, then the deviation score = 20 points. The final risk index = base score + predicted deviation score, and is divided into low risk (<40 points, managed according to the standard procedure), medium risk (40-79 points, followed by multidisciplinary visits twice a week), and high risk (≥80 points, followed by multidisciplinary ward rounds once a day).Taking patient A as an example, their preoperative blood glucose fluctuation was 3 mmol / L (not up to standard), the intraoperative blood pressure abnormality rate was 8% (not up to standard), and the postoperative activity level was 1000 steps (not up to standard). The baseline score was 60 - 3 × 10 = 30 points. The model prediction deviation rate was 40%, and the prediction deviation score was 40 × 40% = 16 points. The total risk index was 46 points, which is classified as medium risk. Intervention should be carried out according to the medium risk management process.

[0078] The adaptive clinical decision-making module focuses on "demand orientation + dynamic optimization". By accurately identifying departmental needs, constructing personalized decision-making rules, and iteratively optimizing clinical recommendations, it provides customized guidance that fits the actual working scenarios for internal medicine, surgery, and anesthesiology departments. The departmental needs analysis unit uses a two-dimensional approach of "qualitative questionnaire + quantitative data mining" to clarify the core needs of each department: at the qualitative level, a needs questionnaire is distributed to doctors with a title of attending physician or above in each department (a total of 100 questionnaires are distributed, 92 valid questionnaires are recovered, and the effective recovery rate is 92%). The questionnaire contains 15 questions such as "current pain points in diagnosis and treatment", "the most critical环节 that requires system support", and "suggested output form". After statistics, it is found that internal medicine departments are most concerned about the "precise adjustment of preoperative underlying disease control plans" (mentioned by 85% of doctors), surgery departments are most concerned about the "early intervention timing of postoperative complications" (mentioned by 90% of doctors), and anesthesiology departments are most concerned about the "drug adjustment plan for intraoperative circulatory stability" (mentioned by 88% of doctors); at the quantitative level, the SPSS statistical software is used to conduct a correlation analysis on the diagnosis and treatment data of 3000 LC surgery patients in the hospital in the past two years. It is found that the correlation coefficient r between "preoperative blood glucose fluctuation > 2 mmol / L" and "postoperative infection rate" is 0.63 (P < 0.01), and the correlation coefficient r between "the proportion of abnormal intraoperative blood pressure duration > 10%" and "the incidence of postoperative cardiovascular and cerebrovascular adverse events" is 0.58 (P < 0.01). These data further verify the reasonableness and urgency of departmental needs.The decision rule generation module constructs a dynamically updated decision rule base based on "risk index + departmental needs." The rule base adopts an "if-else" logic framework, with each rule containing three parts: "trigger condition," "execution suggestion," and "reference basis." It is updated monthly based on the latest clinical practice. For internal medicine, the trigger condition is set as "patient risk index ≥ 40 points and preoperative blood glucose fluctuation > 2 mmol / L," the execution suggestion is "adjust the metformin dose from 0.5g bid (twice daily, 0.5g each time) to 0.85g bid, and monitor blood glucose four times daily (fasting + 2 hours after each meal)," and the reference basis is marked as "based on data from 300 patients, this plan improves the blood glucose target achievement rate by 40%." For surgery, the trigger condition is "patient risk index ≥ 80 points and postoperative activity level < 1000 steps in 3 days," and the execution suggestion is "perform 1 step each at 9:00 AM and 3:00 PM daily." The bedside rehabilitation guidance includes abdominal breathing training (10 minutes each time, 8-10 breaths per minute) and ankle pump exercises (15 minutes each time, one set per hour), with a target activity level of 1500 steps / day. The reference is that "clinical studies show this rehabilitation program can shorten postoperative gastrointestinal function recovery time by 12 hours." For the anesthesiology department, the trigger condition is "the proportion of abnormal blood pressure during surgery >10%," and the recommended implementation is to "adjust the norepinephrine dose from 0.05 μg / (kg・min) to 0.08 μg / (kg・min), monitor blood pressure every 5 minutes, and if blood pressure remains unstable, reduce the propofol infusion rate by 10%." The reference is that "anesthesiology department quality control data shows that this adjustment program has an 85% effective blood pressure control rate." All recommendations are pushed through a pop-up window on the doctor's workstation, which includes screenshots of abnormal patient indicators and rule references for easy understanding and decision-making by doctors.The optimization unit focuses on "clinical efficacy" and continuously iterates and optimizes initial recommendations. First, "efficacy rate" is defined as "the proportion of patients whose relevant indicators reach the target after adopting the recommendation out of the total number of users." For example, if an internal medicine recommendation is applied to 50 patients, and 35 patients' blood glucose fluctuations decrease to ≤1.5 mmol / L, the efficacy rate is 35 / 50 = 70%. If the efficacy rate is <60%, the optimization process is initiated—taking the original internal medicine recommendation "adjust the metformin dose to 1.0 g bid when blood glucose is not at the target" as an example, this recommendation was applied to 60 patients, with only 33 achieving the target, an efficacy rate of 55%, and 12 patients experiencing gastrointestinal side effects such as nausea and diarrhea (incidence rate 20%). Through analysis... The reasons for side effects (excessive dosage leading to gastrointestinal irritation) and low target achievement rate (limited effect of single drug adjustment) led to the optimization of the regimen to "metformin dose reduced to 0.85g bid + combined with acarbose 50mg tid (3 times daily, 50mg each time, taken with meals)". The optimized recommendation was applied to 80 patients, with 66 achieving the target (effective rate 82%), the incidence of gastrointestinal side effects reduced to 5% (only 4 cases), and the stability of patients' blood glucose control significantly improved (average fluctuation range reduced from 2.8mmol / L to 1.2mmol / L). The optimized recommendation will replace the content in the original rule base and will be marked "optimized version" with a comparison of the effects before and after optimization, making it easier for doctors to understand the value of the improvement.

[0079] The cross-departmental data collaboration unit aims to break down "information silos" by building a visual collaboration platform, improving access control, and supporting real-time interaction to achieve efficient collaboration and data sharing among internal medicine, surgery, and anesthesiology departments. The visual platform module is developed based on WebGL technology. WebGL's advantages include support for real-time 3D model rendering (map loading time <3 seconds), adaptation to different screen resolutions (from 13-inch laptops to 27-inch doctor workstation monitors), and support for offline caching (data from the past 24 hours can be viewed even during network interruptions). The platform interface is functionally divided into three core areas: the left-hand "Patient Profile Area" uses a card-style design, prioritizing the display of core patient information (age, gender, underlying disease type) and risk index (displayed using red / yellow / green circular dashboards for intuitive differentiation). The dashboard displays risk levels and key abnormal indicators (such as "large preoperative blood glucose fluctuations" and "insufficient postoperative activity"). Abnormal indicators are highlighted in red, while normal indicators are collapsed. Clicking "expand" will show the complete list of indicators. A detailed risk index calculation is attached below the dashboard (e.g., "base score 40 + deviation score 16 = 56") to help doctors quickly grasp the patient's overall condition. The central "3D Feature Atlas Area" uses a time axis as the vertical axis (extending from 7 days preoperatively to 7 days postoperatively, with key events marked at each time point, such as "3-day preoperative blood glucose monitoring" and "1-day postoperative incision examination") and indicator types as the horizontal axis (divided into physiological indicators, imaging indicators, and operational indicators). (Three categories, each distinguished by a different color) to construct a 3D atlas. The atlas supports mouse dragging and rotation (360° view), zooming (zoom in to view specific values, zoom out to observe the overall trend), and time interval filtering (e.g., viewing only "intraoperative" data). Clicking on any time point will pop up the original data table for that period (e.g., heart rate data and activity data 3 days post-surgery) and the model prediction curve (blue prediction line compared to red normal range line). For example, clicking "3 days post-surgery" will display "average heart rate 95 beats / min (normal 60-100)", "activity 1200 steps (target 1500)" and the predicted postoperative 4 days. The activity level is 1400 steps, helping doctors predict recovery trends. The "Suggestion Area" on the right displays customized suggestions categorized by department (blue label for internal medicine, green label for surgery, and orange label for anesthesiology). Each suggestion includes three parts: "Suggestion Content," "Historical Effectiveness Rate," and "Number of Application Cases." For example, an internal medicine suggestion might be labeled "Effectiveness Rate 82% (80 cases applied)." Clicking on the suggestion expands to "Application Cases" (anonymously showcasing the treatment effects of 3 similar patients), enhancing doctors' trust in the suggestions. The platform also supports data export (in Excel or PDF format) and printing, facilitating case discussions or consultation reports.The data sharing permission management unit is based on a dual control logic of "role-department" and, combined with hospital information security management standards, constructs a strict permission system: First, it clarifies the permission scope of each role—internal medicine doctors can only view "preoperative basic disease data (blood glucose, blood pressure, liver and kidney function)" and "customized internal medicine suggestions," but cannot edit surgical data or anesthesiology monitoring data; surgeons can view "patient full-cycle data (preoperative-intraoperative-postoperative)" and "surgical suggestions," and can edit "postoperative incision score" and "rehabilitation guidance record" (after editing, the system automatically records the modification history, including the modifier, modification time, and content before and after modification); anesthesiologists can only view "intraoperative monitoring data (heart rate, blood pressure, end-tidal carbon dioxide)" and "anesthesiology suggestions," but cannot view postoperative follow-up data (such as patient activity level and sleep quality); the department head has "full authority." The system is "limited," allowing users to view data from all departments, edit departmental suggestions, and approve permission change requests. Permission configuration is centrally managed by the hospital's information technology department. If a doctor needs to change permissions (e.g., the head of internal medicine needs to view surgical data for a consultation), they must submit a "Permission Change Request Form" (stating the reason for the request, the required permissions, and the duration of use). After the department head signs and approves the request, the information technology department will complete the configuration within 24 hours. To ensure data security, the platform employs multiple protection measures: data transmission uses the TLS 1.3 encryption protocol (to prevent theft during transmission), data storage uses the AES-256 encryption algorithm (even if the storage device is stolen, the data cannot be decrypted), operation logs are retained for 6 months (recording all data viewing, editing, and export operations for easy traceability), and copying and pasting data to external systems is prohibited (achieved by disabling the browser's copy function).The real-time interaction module provides a convenient tool for multi-departmental collaboration, solving the problems of "communication lag" and "information asymmetry" in traditional collaboration: the online annotation function allows doctors to add annotations in real time while viewing data. For example, after a surgeon checks the patient's incision at 10:00 AM every day post-surgery, they can annotate on the platform, "2 days post-surgery, mild redness and swelling of the incision, no exudate, sterile dressing changed." The annotation content is synchronized in real time through the hospital's intranet (synchronization delay <10 seconds). After viewing, internal medicine doctors can immediately adjust the medication plan for underlying diseases based on the annotation (e.g., when annotating "incision redness and swelling," the internal medicine doctor can adjust the medication plan for underlying diseases). Doctors can leave messages such as, "Do not adjust the dosage of hypoglycemic drugs for now. It is recommended to closely monitor the incision for any worsening. If oozing occurs, infection should be investigated." The comment function supports online communication among doctors from multiple departments. Comments are arranged in chronological order, and each comment indicates the sender, department, and time of sending. Doctors can "reply" or "like" comments. For example, an anesthesiologist might comment to a patient with "large fluctuations in blood pressure during surgery," "It is recommended to strengthen postoperative analgesia to avoid pain causing a further increase in blood pressure. The background dose of the PCA pump can be increased by 20%." The surgeon could then reply, "The PCA has been adjusted as recommended." Pump parameters were adjusted, and the patient's postoperative pain score decreased from 4 to 2 (VAS scoring standard). All messages and replies were automatically saved to the "Collaboration Records" section of the patient's electronic medical record for easy access later. The online consultation function supports initiating real-time multi-departmental consultations. Doctors can click "Initiate Consultation" on the platform, select participating departments (such as internal medicine, surgery, and anesthesiology) and consultation time, and the system will automatically send email and SMS reminders to participating doctors. During the consultation, the platform has a built-in screen sharing function (supporting the sharing of patient feature maps, risk indices, and raw data) and multi-person voice calls (supporting up to [number missing]). The system allows for simultaneous speaking by up to 10 people with clear, lag-free audio and video (optional, facilitating face-to-face communication between doctors). The consultation process is automatically recorded (saved to the hospital's consultation system for one year) and generates a written summary (automatically extracting key decisions, such as "unanimously agreeing to continue postoperative use of metformin 0.85g bid + acarbose 50mg tid, enhanced surgical incision dressing changes, and follow-up analgesia by the anesthesiology department"). After confirmation by the consultation moderator, the summary is automatically synchronized to the workstations of all participating doctors and the patient's electronic medical record, ensuring that consultation decisions are executable and traceable.

[0080] The dynamic feedback optimization unit constructs a closed-loop optimization mechanism of "clinical feedback → data matching → parameter fine-tuning." By collecting real treatment results, it adjusts the parameters of the self-supervised model in reverse, continuously improving the model's predictive accuracy and clinical applicability. The treatment result collection module focuses on the "final treatment result" of patients within one month after surgery. The collected content covers three dimensions: complications, recovery efficiency, and treatment satisfaction, ensuring the comprehensiveness and clinical value of the feedback data. The occurrence of complications is entered daily by surgical nurses through the electronic medical record system, including whether bile leakage occurred (yes / no, if so, record the time of occurrence, such as "5 days post-surgery," and the treatment method, such as "endoscopic nasobiliary drainage"), and whether there was incision infection (yes / no, record the degree of infection, such as "mild infection, only redness and swelling" or "severe infection, with purulent exudate"). After entry, it must be reviewed and confirmed by the attending physician (if the review fails, it will be returned for modification) to avoid data errors. The recovery cycle data is entered by the discharge nurse at the time of patient discharge, including the actual number of days discharged (such as "actually discharged after 8 days, compared with the target of 5-7 days to calculate the delay days), and the first flatus after surgery. Timeframes (e.g., "passing gas 36 hours post-surgery," with a target of 24-48 hours to assess gastrointestinal function recovery speed) are used, along with all examination reports from the patient's hospitalization period (e.g., post-operative blood routine and liver function reports) as supplementary evidence for assessing the recovery period. Treatment effectiveness scoring employs a dual-perspective evaluation from both the patient and doctor. Upon discharge, patients complete a "satisfaction rating" (1-5 points, 1 point indicating "very dissatisfied," 5 points indicating "very satisfied") via the hospital's app, and can also provide written feedback (e.g., "Very satisfied with the post-operative rehabilitation guidance; the nurse's explanation was very detailed"). Doctors, within one week of patient discharge, complete a "treatment effectiveness rating" (1-5 points, based on complication rate, whether the recovery period is met, and patient compliance), for example, "Patient had no post-operative complications, recovery period of 7 days (met the target), satisfaction score 5 points, treatment effectiveness rating 5 points." To ensure data quality, the module employs a three-tiered quality control mechanism: Level 1 quality control involves self-checking by the nurses entering data (checking data completeness, such as whether the time of complication occurrence was omitted); Level 2 quality control involves daily random checks by the department's quality control physicians (20% of patients are checked to verify data accuracy, such as comparing temperature charts with infection records); and Level 3 quality control involves weekly review by the hospital's quality control department (reviewing the logical consistency of the data, such as whether "exhausting gas 3 days post-surgery" and "discharge 5 days post-surgery" conform to clinical norms), promptly contacting relevant doctors to correct any issues found, ensuring that the accuracy of the feedback data is ≥98%.The core task of the feedback data matching unit is to accurately correlate treatment results with the corresponding patient's model input data and prediction results, avoiding optimization bias caused by misattribution. The matching process unfolds based on a three-dimensional framework of "patient unique ID - timeline - prediction item": The patient unique ID uses the "hospital number" as the core link (e.g., hospital number "202405001"). Through SQL statements, it associates the "model input data" (e.g., preoperative blood glucose, intraoperative blood pressure, postoperative activity level), "model prediction results" (e.g., predicted recovery period of 6 days, predicted infection risk of 35 points) and "actual treatment results" (e.g., actual recovery period of 8 days, actual infection) corresponding to this hospital number in the system database, ensuring complete data consistency. Timeline matching aligns the occurrence time of the actual treatment result with the time interval of the model input and prediction. For example, the result "incision infection 5 days postoperatively" needs to correspond to "preoperative 3 days postoperatively". The model input data (physiological indicators from day d to 6 hours post-surgery, as this period is the core basis for the model to predict the risk of infection 1-7 days post-surgery) and the model's predicted risk of infection 1-7 days post-surgery (prediction results) are compared using timestamps to clarify the specific stage at which the deviation occurred (e.g., "3-5 days post-surgery"), avoiding errors in deviation positioning due to time misalignment. Prediction item matching achieves a "point-to-point" correspondence, linking each prediction dimension of the model with the actual result one by one. For example, the model's prediction of "7-day recovery period" corresponds to the actual "8-day discharge", the model's prediction of "no bleeding risk" corresponds to the actual "no bleeding 2 days post-surgery", and the model's prediction of "low infection risk (35 points)" corresponds to the actual "infection 5 days post-surgery". Through this precise correspondence, the deviations of the model in specific dimensions such as "recovery period prediction", "bleeding risk judgment", and "infection risk prediction" are clearly identified, rather than vaguely locating "the model prediction is inaccurate". After matching is completed, a feedback data matching report is generated. The report includes the patient's hospital number, key indicators of model input, prediction results, actual results, deviation type (such as "recovery cycle deviation of 1 day" or "infection risk prediction deviation"), and deviation degree (such as "deviation rate of 14.3%)", providing a clear direction of deviation for subsequent model fine-tuning.The model parameter fine-tuning engine, based on the matched feedback data, uses a mini-batch gradient descent algorithm to back-tune the weight parameters of the self-supervised model (Perceiver model), ensuring the targeting and effectiveness of the optimization. First, data preparation is performed: for every 100 patients' feedback data points accumulated, a "fine-tuning dataset" is constructed, divided in a 7:3 ratio into a training set (70 cases, used for parameter updates) and a validation set (30 cases, used to evaluate the optimization effect). Data preprocessing includes standardization (converting index values ​​to a 0-1 range to avoid the influence of units) and outlier handling (using the 3σ principle to remove extreme outliers, such as data that does not conform to clinical reality, like "postoperative activity level 10,000 steps"). The loss function chosen is mean squared error (MSE). Since risk index prediction is a regression task, MSE can effectively measure the deviation between the "model-predicted risk index" and the "actual risk label" (marking "complications occurred" as 1 and "no complications occurred" as 0, then mapping to a risk label of 0-100). The SE calculation formula is "MSE=1 / n×∑(predicted value-actual label value)²", where n is the number of samples. The gradient descent parameter settings have been verified through multiple experiments: learning rate η=0.001 (this value can avoid parameter oscillation and non-convergence due to excessive step size, or excessive optimization speed due to excessively small step size), batch size b=32 (balancing computational efficiency and gradient estimation accuracy; the hospital GPU server (NVIDIA A100, 32GB VRAM) can process 32 samples at a time, with each batch taking about 2 seconds to compute), and number of iterations = 100 rounds (the model loss tends to stabilize after 50 rounds, and 100 rounds can ensure convergence to the optimal solution). After each round of training, the "risk identification accuracy" (i.e., the proportion of high-risk, medium-risk, and low-risk patients correctly identified by the model) is evaluated using the validation set. If the accuracy improves by ≥2% compared to before fine-tuning, the current parameters are saved as the new model parameters, and the original model is overwritten. If the accuracy does not improve or decreases, the parameters of the previous round are rolled back, the reasons are analyzed (such as sample bias, improper parameter settings), and then fine-tuned again. Taking a batch of fine-tuning as an example, it was found that the model had a large prediction bias for patients with "preoperative gallbladder wall thickness > 5 mm and postoperative activity level < 1000 steps"—the actual infection rate of this type of patient was 30%, but the model predicted a risk index of only 45 points (lower limit of medium risk), which is considered "lower predicted risk." By calculating the weight gradient corresponding to the features of this type of patient using the gradient descent algorithm, it was found that the weight of the "gallbladder wall thickness" feature was only 0.2 (lower than the average weight of 0.35 for other risk features). Therefore, its weight was adjusted to 0.3, and the weight of the "postoperative activity level" feature was also fine-tuned (from 0.25 to 0.3). After the adjustment, the predicted risk index value of this type of patient increased to 65 points (upper limit of medium risk), which is more consistent with the actual infection rate of 30%. The risk identification accuracy of the validation set increased from 88% before fine-tuning to 92%, and the MSE decreased from 8.5 to 4.3, indicating a significant improvement in the model's prediction accuracy.In addition, the engine has established a parameter version management mechanism, retaining parameter files for nearly 10 fine-tuning versions, supporting one-click rollback (if new parameters are found to increase the prediction bias for other types of patients, rollback to the previous version is possible), and generating model fine-tuning reports that record the amount of fine-tuning data, parameter adjustment details, and optimization effect comparisons, providing complete documentation support for model iteration.

[0081] In summary, this application's embodiments utilize a multi-departmental data perception layer with automated data acquisition and off-site wearable device monitoring to achieve comprehensive coverage and temporal alignment of preoperative underlying diseases, intraoperative procedures, and postoperative recovery data. This reduces manual data entry errors and fills gaps in off-site monitoring. The surgical feature processing center combines deep learning and attention mechanisms to accurately extract quantitative features and uncover hidden correlations. The visualized atlas allows doctors to intuitively grasp the sources of risk, significantly improving the efficiency of clinical judgment. The full-cycle dynamic analysis unit generates a quantitative risk index through phased assessment and a self-supervised model, enabling early warning of complications such as bile leakage and infection, and facilitating timely intervention for high-risk patients. The adaptive clinical decision-making module customizes and iteratively optimizes treatment recommendations based on departmental needs, improving the control rate of underlying diseases while reducing drug side effects; the cross-departmental data collaboration unit breaks down information silos with a visualization platform and real-time interactive functions, coupled with strict access control, ensuring data security while shortening the time for multi-departmental collaboration; the dynamic feedback optimization unit continuously improves model accuracy through closed-loop fine-tuning, further enhancing the accuracy of risk prediction, reducing the incidence of postoperative complications after LC, shortening the patient recovery period, and improving the efficiency of medical team collaboration and patient satisfaction, providing solid support for high-quality and precise perioperative management of LC in primary hospitals.

[0082] Next, referring to the accompanying drawings, a self-supervised deep processing and analysis method for smart medical data is proposed according to an embodiment of this application.

[0083] like Figure 9 As shown, this self-supervised deep processing and analysis method for smart healthcare data includes the following steps: In step S101, multidisciplinary diagnosis and treatment data and postoperative follow-up data are acquired.

[0084] It is understood that the embodiments of this application, by acquiring multidisciplinary diagnostic and treatment data and integrating key information from the entire perioperative process, provide a complete diagnostic and treatment perspective for subsequent feature extraction and risk analysis, avoiding judgment bias caused by data fragmentation. Postoperative follow-up data fills the gap in out-of-hospital recovery monitoring, compensates for the shortcomings of traditional medical care that relies solely on inpatient data and struggles to track the dynamics of key postoperative rehabilitation stages, reduces manual input errors, ensures data accuracy and consistency, provides high-quality, multi-dimensional data source support for risk prediction, and captures postoperative abnormalities early through real-time follow-up data, laying a solid foundation for the generation of subsequent precise diagnostic and treatment recommendations and multidisciplinary collaboration, thereby improving perioperative management efficiency and patient prognosis.

[0085] In step S102, the multidisciplinary diagnosis and treatment data and postoperative follow-up data are analyzed and processed to extract physiological indicators, imaging features and surgical operation features. The data before and after surgery are linked through the attention mechanism to generate a unified feature map of the patient in the perioperative period.

[0086] Among them, the unified perioperative feature map of patients refers to a visual data carrier that is built based on graph databases, integrates multi-department and multi-dimensional data of patients throughout the entire cycle of preoperative, intraoperative and postoperative periods, clearly presents the data attribution and time sequence patterns through time-series correlation and risk correlation, and can intuitively show the correspondence between indicators and risks, and help doctors quickly locate the source of risks.

[0087] It is understood that the embodiments of this application generate a unified perioperative feature map of patients, integrating multi-dimensional data (such as physiological indicators, imaging features, surgical operation data, etc.) scattered across various departments before, during, and after surgery through a graph database. This allows doctors to intuitively obtain the full-cycle data context of patients during the perioperative period, avoiding the limitations of the diagnostic and treatment perspective caused by data fragmentation. By using attention mechanisms to link pre- and post-operative data, and through temporal and risk correlations, it clearly presents data attribution, temporal patterns, and the correspondence between indicators and risks. This not only helps doctors quickly uncover hidden clinical correlations but also visualizes the sources of risk, significantly shortening the time for doctors to locate risks and assess conditions, and improving the efficiency of clinical decision-making. At the same time, it provides structured and standardized data support for subsequent full-cycle dynamic analysis and adaptive clinical decision-making, laying a key foundation for precise perioperative management and reducing the risk of postoperative complications.

[0088] In step S103, based on the patient's perioperative unified feature map, the patient's entire cycle is divided into the preoperative assessment period, the intraoperative monitoring period, and the postoperative recovery period. The physiological status and recovery progress at each stage are analyzed. Self-supervised training is performed based on unlabeled clinical data. The patient's status evolution process is predicted through the Perceiver model. Based on the patient's status evolution process, the probability of complication and the risk index of recovery delay are integrated to generate a risk index. Based on the analysis results and the risk index, the clinical needs of different departments are identified, and preliminary treatment recommendation rules are generated. Combined with historical treatment effect data, the preliminary treatment recommendation rules are iteratively optimized to generate customized clinical recommendations.

[0089] Among them, the Perceiver model is a general perception model based on the Transformer architecture proposed by the DeepMind team. It encodes high-dimensional multimodal inputs into low-dimensional latent representations through an iterative cross-attention mechanism, achieving efficient processing and feature fusion of cross-modal data while maintaining computational efficiency.

[0090] It is understood that the embodiments of this application utilize the Perceiver model to efficiently process multimodal high-dimensional data, such as physiological indicators, images, and surgical procedures, contained in the unified feature map of patients during the perioperative period. This data is encoded into low-dimensional latent representations, overcoming the limitations of traditional models in integrating multi-source heterogeneous data. While compressing data dimensions, core information is retained, ensuring computational efficiency to meet the actual computing power needs of hospitals. Based on the characteristics of self-supervised training using unlabeled clinical data, it perfectly addresses the pain points of high cost and scarcity of labeled samples in medical data annotation. Perioperative data patterns can be learned through tasks such as time-series data completion without extensive manual annotation, significantly reducing the model training threshold. It accurately captures the temporal correlation and evolution logic of data at each stage before, during, and after surgery, reliably predicting the evolution of patient status. This provides a high-precision predictive basis for the subsequent fusion of complication probability and recovery delay risk indicators to generate a quantitative risk index, thereby helping to accurately identify departmental clinical needs, optimize customized treatment recommendations, and improve the accuracy of perioperative risk prediction and the scientific nature of treatment decisions.

[0091] It should be noted that, based on the patient's condition evolution, a risk index is generated by integrating the probability of complication and the risk indicators of delayed recovery. First, using the condition evolution curve output by the Perceiver model as a basis, the trend deviation rate from the normal recovery curve is calculated, reflecting the degree of deviation from the overall recovery trajectory. Then, combined with historical data of similar cases and the patient's current indicators, the probability of complications such as infection and bile leakage is determined. Next, a recovery delay risk score is calculated based on dimensions such as postoperative activity level achievement rate, first flatus time, and incision healing status. Finally, these scores are weighted according to clinical priority and integrated into a risk index of 0-100 points, categorized as low / medium / high, while also providing a detailed breakdown of risk sources, allowing doctors to quickly grasp the core risk points.

[0092] Formula for calculating trend deviation rate:

[0093] in, For the patient's first The actual state value of the day; The normal recovery curve Reference value for the day; For the evaluation period; This is an index for the time period.

[0094] Formula for calculating the probability of complications:

[0095] in, The probability of a patient developing a certain type of complication; is the base of the natural logarithm; This is the intercept term of the logistic regression model; For the i-th input feature The corresponding coefficients; Let i be the i-th input feature.

[0096] Formula for calculating recovery delay risk score:

[0097] in, The total number of indicators participating in the recovery assessment; This is the index for the k-th recovery indicator; The weight of the k-th recovery indicator; The actual completion time of the k-th recovery indicator; Let be the expected completion time for the k-th recovery indicator.

[0098] For example, in the perioperative management of laparoscopic cholecystectomy in a primary-level tertiary hospital, the Perceiver model achieves accurate risk prediction by processing multimodal time-series data. The model's input includes multi-source information such as intraoperative blood glucose and blood pressure (e.g., fasting blood glucose at 8:00 am and blood pressure monitoring in the morning, noon and evening), intraoperative anesthesiology data on heart rate every 10 seconds and blood pressure every 30 seconds, and postoperative wearable device records of heart rate every 30 minutes and daily activity. It adopts a self-supervised training mode—randomly masking 10% of the time-series segments (e.g., blood pressure data at 20-30 minutes during surgery or activity records at 2 days postoperatively) and learning the data association rules through an iterative cross-attention mechanism, encoding high-dimensional heterogeneous data into low-dimensional latent representations. After 80 rounds of training on a hospital GPU cluster, the mean absolute error (MAE) decreased from 15.6 to 3.2. In practical applications, the model can generate postoperative status evolution curves from 1 to 7 days postoperatively. For example, it can accurately predict that a patient's activity level is only 1200 steps on postoperative day 3 (below the target value of 2000 steps) and resting heart rate is consistently >100 beats / min. Combined with the preoperative imaging features of gallbladder wall thickness >5mm and the operation data of more than 2 hemostasis attempts during surgery, it is automatically marked as "high-risk evolution" and a quantitative risk index is generated. Its calculation efficiency is 9 times higher than that of the traditional Transformer and does not require a large amount of labeled data. It provides a reliable decision-making basis for clinical adjustment of rehabilitation programs (such as increasing the frequency of bedside rehabilitation guidance) and optimization of underlying disease management, effectively improving the accuracy of perioperative risk prediction and the timeliness of intervention.

[0099] In step S104, based on the visual collaboration platform, doctors from various departments can share and interact on the analysis results and customized suggestions of the entire surgical cycle, collect clinical diagnosis and treatment results in real time, and fine-tune parameters in reverse.

[0100] Among them, the visual collaboration platform integrates information through an intuitive visual interface, supports real-time interaction and data sharing among multiple people or roles, and helps teams to efficiently advance their work.

[0101] It is understood that, by utilizing this visual collaboration platform, the embodiments of this application enable doctors from various departments to intuitively share the analysis results and customized suggestions throughout the entire surgical cycle. Through real-time interaction, they can quickly align their diagnostic and treatment approaches and reduce information discrepancies. At the same time, they can collect clinical diagnostic and treatment results in a timely manner and, based on the platform's efficient data flow capabilities, precisely fine-tune parameters in reverse to dynamically adapt to actual diagnostic and treatment needs. This not only ensures the scientific and timely nature of diagnostic and treatment decisions throughout the entire surgical cycle but also improves the efficiency of cross-departmental collaboration, reduces the diagnostic and treatment risks caused by information delays or poor communication, and further optimizes the overall diagnostic and treatment process.

[0102] According to the embodiments of this application, a self-supervised deep processing and analysis method for smart medical data is proposed. This method integrates departmental data and postoperative follow-up data through a multi-departmental data perception layer, providing a complete data foundation for full-cycle analysis. The surgical feature processing center uses an attention mechanism to correlate pre- and post-operative data, generating a unified feature map to improve the relevance and accuracy of feature extraction. A full-cycle dynamic analysis unit, combined with a self-supervised prediction model, outputs risk indices in stages to dynamically monitor patient status and predict risks. An adaptive clinical decision-making module generates customized suggestions for different departments, improving the personalization and practicality of clinical decisions. A cross-departmental data collaboration unit promotes data sharing and interaction through a visualization platform, breaking down barriers to cross-departmental collaboration. A dynamic feedback optimization unit relies on a clinical outcome-based reverse fine-tuning model to ensure continuous iterative optimization of system performance, improving the utilization efficiency and analytical accuracy of medical data, while also strengthening cross-departmental collaboration capabilities and the scientific nature of clinical decision-making, effectively supporting efficient diagnosis and treatment and patient safety in smart medical scenarios. This solves the problems of strategy rigidity and insufficient energy utilization efficiency in existing technologies.

[0103] The following will illustrate a self-supervised deep processing and analysis method for smart healthcare data through a specific embodiment, such as... Figure 10 As shown, it includes: This study targets patients undergoing laparoscopic cholecystectomy (LC) at the hospital, involving departments including: hepatobiliary surgery (the lead surgeon), anesthesiology (intraoperative monitoring and risk management), laboratory medicine (physiological indicator testing), radiology (preoperative imaging assessment), rehabilitation medicine (postoperative functional recovery), and nursing (postoperative follow-up and data collection). It covers the collaborative needs of the entire surgical cycle, acquires multi-departmental data and postoperative follow-up data, and relies on the hospital's existing information system to achieve "no blind spots" in data collection without additional hardware costs, maximizing the use of existing hospital resources. The collection of multidisciplinary diagnostic and treatment data must cover key information throughout the entire process from patient admission to the end of surgery. Among the basic patient information extracted through the hospital's electronic medical record (EMR) system, age (patients >65 years old have a higher risk of delayed postoperative recovery), BMI (≥28 kg / m² indicates obesity, which may increase the difficulty of intraoperative procedures and the risk of postoperative infection), and underlying diseases (such as hypertension requiring recording the preoperative blood pressure control range, and diabetes requiring recording glycated hemoglobin levels) are important foundations for preoperative risk assessment. In addition to gallstone size (stones >2 cm are prone to causing cystic duct obstruction) and gallbladder wall thickness (>3 mm indicates chronic cholecystitis, possibly accompanied by gallbladder adhesions to surrounding tissues), preoperative diagnostic data for hepatobiliary surgery must also include preoperative gallbladder contraction function test results (such as ultrasound findings). A gallbladder contraction rate <30% indicates gallbladder dysfunction and is one of the indications for surgery. In the preoperative anesthesiology assessment, the ASA classification (Class I for healthy patients, Class II for patients with mild systemic disease, and Class III for patients with severe systemic disease) directly determines the anesthesia plan. Cardiopulmonary function indicators must include preoperative electrocardiogram (to check for myocardial ischemia) and pulmonary function tests (whether the FEV1 / FVC ratio is >70% to assess pulmonary ventilation). Laboratory data must differentiate between preoperative baseline values ​​and postoperative dynamic changes. Preoperative liver function indicators (ALT normal reference range 7-40 U / L; if >80 U / L, liver protection treatment is required first), coagulation function indicators (PT normal 11-13 seconds; prolonged >3 seconds requires coagulation factor supplementation), and complete blood count (WBC normal 4-10 × 10⁻⁶). 9A C-reactive protein level (CRP) of >50 mg / L (indicating potential infection) is crucial for determining the timing of surgery. Liver function and inflammatory markers (such as CRP, which indicates a severe inflammatory response) should be checked daily for 3 days post-surgery. Preoperative abdominal ultrasound / CT images retrieved from the radiology department's PACS system should include the number of gallstones, the degree of gallbladder atrophy (gallbladder volume <10 ml indicates severe atrophy, increasing the difficulty of gallbladder bed dissection during surgery), the diameter of intrahepatic and extrahepatic bile ducts (to rule out complications of bile duct stones), and the positional relationship between the gallbladder and portal vein (to avoid vascular damage during surgery). Intraoperative data acquisition relies on anesthesia monitoring equipment and... The surgical anesthesia system requires real-time linkage. The anesthesia monitor needs to record heart rate (normal 60-100 beats / min), systolic / diastolic blood pressure (normal 90-139 / 60-89 mmHg), blood oxygen saturation (SpO2 should be ≥95% continuously; if it is below 92%, ventilation parameters need to be adjusted), and end-tidal carbon dioxide (normal 35-45 mmHg; abnormally high levels may indicate excessive pneumoperitoneum pressure) every 5 minutes. The surgical anesthesia system also needs to record operational data including the type of laparoscopic instrument used during the operation (such as a 30° laparoscope or a 0° laparoscope), details such as whether hemostatic materials (such as hemostatic gauze) were used during the operation. Postoperative follow-up data is collected through the Nursing Department's "Postoperative Follow-up APP," which connects in real time with the hospital's EMR system and can automatically link patient hospitalization information to avoid duplicate entry. Within one week after surgery, the patient's NRS pain score (0 points for no pain, 10 points for severe pain; a score >4 points requires adjustment of the analgesia regimen), time to first flatus (no flatus after >48 hours indicates delayed bowel function recovery), and recovery of eating (time to transition from liquid to semi-liquid diet) must be recorded daily. At the same time, the patient's nausea, vomiting, fever (temperature >38.5℃ requires investigation for infection) and other adverse reactions must be recorded. One month after surgery, ultrasound examination should focus on whether there is fluid accumulation in the gallbladder bed area (depth >2cm requires puncture and drainage) and whether the intrahepatic and extrahepatic bile ducts are dilated (to rule out bile leakage or bile duct stenosis). Three months after surgery, the quality of life score (SF-36 scale) should cover eight dimensions, including physiological function (such as the ability to bend over normally and lift heavy objects), bodily pain, and social function (such as the ability to return to work), to comprehensively assess the impact of surgery on the patient's long-term life.

[0104] Data processing and perioperative unified feature map generation, using Python data stacks (such as Pandas for data cleaning and Matplotlib for feature visualization) and attention mechanism models as the core technologies, can be implemented using the existing server resources of the hospital's AI laboratory without the need for additional complex hardware environments. In the feature extraction stage, the raw data must first be standardized, and then core features are extracted according to three dimensions: "physiological," "imaging," and "operational." Physiological indicators must clearly distinguish between the three stages: "preoperative baseline," "intraoperative dynamics," and "postoperative recovery." Preoperative ALT (>40 U / L should be marked as "abnormal liver function") and PT (>13 seconds should be marked as "abnormal coagulation function") are key to assessing surgical tolerance. Intraoperative mean arterial pressure (MAP should be maintained between 65-85 mmHg; fluctuations >15 mmHg are marked as "hemodynamic instability") and intraoperative lowest oxygen saturation (SpO2 <95% should have its duration recorded) directly reflect anesthesia safety. Postoperative day 1 WBC (>10×10⁶) is also important. 9 / L indicates possible postoperative infection), and NRS score (>4 points indicates insufficient analgesia) are core indicators for assessing postoperative recovery status. Each indicator needs to be labeled with its normal reference range and abnormal threshold for subsequent model identification. Image feature extraction utilizes a ResNet50 deep learning model (pre-trained weights use the ImageNet dataset, then fine-tuned using CT images from nearly 2000 LC patients in the hospital). The model can automatically identify and quantify gallbladder wall thickness (≥3mm for "thickened", >5mm for "significantly thickened"), maximum stone diameter (≥2cm for "large stone", >3cm for "giant stone"), and the degree of adhesion between the gallbladder and surrounding tissues (0 for no adhesion, 1 for no adhesion, 1 for no adhesion). Grade 1 adhesions are classified as localized, Grade 2 as extensive, and Grade 3 as complete adhesions. The confidence level of the imaging features (e.g., the accuracy of adhesion identification must be ≥90%) is also output to ensure feature reliability. Surgical operation features need to be cross-validated with surgical and anesthesia records. Features such as operation duration (≥60 minutes is considered "long surgery," and the reasons for the extension, such as time spent separating adhesions), intraoperative pneumoperitoneum pressure (routinely maintained at 12-14 mmHg; if >15 mmHg, the duration must be recorded to avoid excessive pressure leading to subcutaneous emphysema), and whether conversion to open surgery was required (0 = no, 1 = yes; the reason for conversion must be noted, such as massive bleeding or bile duct injury). Each feature must correspond to specific clinical significance to lay the foundation for subsequent correlation analysis. In the attention mechanism model building phase, the TensorFlow framework was used to construct a model containing an input layer, a self-attention layer, and a fully connected layer. The input layer needs to convert the three types of features into vectors of a uniform dimension (e.g., physiological indicator vector dimension is 12, image feature vector dimension is 8, and surgical operation feature vector dimension is 5). The self-attention layer calculates the attention weights between features, focusing on strengthening the data associations that significantly affect postoperative recovery. For example, in the association of "preoperative gallbladder wall thickness - intraoperative MAP fluctuation amplitude - postoperative WBC increase amplitude", the attention weight of gallbladder wall thickness is set to 0.4 (higher than the 0.2-0.3 of other features). After model training, it can be concluded that "for every 1mm increase in gallbladder wall thickness, the probability of intraoperative MAP fluctuation > 15mmHg increases by 8%, and the postoperative WBC increase > 2×10 9The quantitative relationship is "a 10% increase in the probability of / L"; in the correlation of "preoperative ASA classification - intraoperative minimum SpO2 value - postoperative flatus time", the attention weight of ASA III patients is higher than that of I / II patients. The model can output the conclusion that "when ASA III patients have an intraoperative SpO2 < 95% for more than 10 minutes, the probability of postoperative flatus time being prolonged by more than 12 hours increases by 25%". The final generated "patient perioperative unified feature map" should be presented in the form of a structured table, including the patient's unique ID (e.g., LC20240501), preoperative features (gallbladder wall 4mm + ASA II + ALT 35U / L + gallbladder contraction rate 40%), intraoperative features (MAP fluctuation 12mmHg + operation time 55 minutes + pneumoperitoneum pressure 12mmHg + no conversion to open surgery), and postoperative features (WBC 8×10 9 The results include: L+NRS score 3+18 hours postoperative flatus and no adverse reactions. At the same time, abnormal indications for each feature are marked (if there are no abnormalities, mark "normal"; if there are abnormalities, mark "needs attention") to facilitate subsequent diagnosis and treatment decisions.

[0105] The process of cycle segmentation, status prediction, and customized clinical suggestion generation is implemented through a hospital AI platform (such as a medical AI deployment platform built on Huawei Cloud or Alibaba Cloud). The core is to form a closed loop of "cycle management - AI prediction - suggestion optimization". In the full cycle segmentation stage, it is necessary to combine the clinical process characteristics of LC surgery to clarify the time range and core management objectives of each stage: the preoperative assessment period is set at 1-7 days before surgery. The first 3 days focus on completing basic examinations (liver function, coagulation function, imaging examination). If ALT > 80U / L is found, liver protection treatment should be initiated (such as intravenous infusion of diammonium glycyrrhizinate). After 2-3 days of treatment, a follow-up examination should be conducted. Surgery can be scheduled only if the indicators drop to the normal range. One day before surgery, a second assessment by the anesthesiology department should be completed to confirm that there is no change in ASA classification (if the patient has a cold before surgery, the ASA classification may rise from level II to level III, and surgery should be delayed). At the same time, surgical risks should be explained and preoperative preparations should be made (such as fasting and water restriction time, bowel preparation plan). Intraoperative monitoring is conducted on the day of surgery, typically lasting 40-90 minutes, and should be performed in three sub-phases: anesthesia induction, surgical procedure, and anesthesia recovery. During anesthesia induction (approximately 10 minutes), focus on monitoring heart rate and blood pressure changes to avoid sudden drops in blood pressure caused by induction drugs. During surgical procedure (approximately 30-60 minutes), record MAP, SpO2, and intraoperative blood loss every 5 minutes (for blood loss > 50 ml, record the bleeding site, such as gallbladder artery bleeding). If MAP < 65 mmHg, rapid fluid resuscitation is necessary (e.g., infusion of 200-300 ml of balanced electrolyte solution). If SpO2 < 95%, adjust ventilator parameters (e.g., increase tidal volume). During anesthesia recovery (approximately 10 minutes), monitor the patient's recovery of consciousness to avoid delayed awakening (if awakening is not achieved after > 30 minutes, investigate the cause). Postoperative recovery lasts 1-30 days post-surgery, with days 1-3 considered the "acute recovery period." Monitor body temperature, NRS score, WBC count, and liver function daily. If body temperature > 38.5℃ and WBC > 12 × 10⁻⁶, further monitoring is required. 9 / L, an ultrasound examination is needed to rule out bile leakage or abdominal infection; 4-7 days after surgery is the "functional recovery period", focusing on assessing bowel function recovery (such as whether the patient can eat semi-liquid food normally) and wound healing (whether there is redness, swelling and oozing); 8-30 days after surgery is the "discharge follow-up period", during which the patient's recovery is monitored remotely by phone or APP, and a follow-up ultrasound is reminded one month after surgery. In the self-supervised training and state prediction phase, preprocessing is required for 5,000 unlabeled clinical data of LC cases from the hospital over the past 3 years: removing data with more than 20% missing key indicators (such as intraoperative MAP and postoperative flatus time), standardizing physiological indicators using Z-score (eliminating the influence of dimensions), and normalizing image features (adjusting pixel values ​​to the 0-1 range); the Perceiver model is implemented using the PyTorch framework. This model processes high-dimensional clinical data through the "attention pooling" mechanism. The input layer contains a 128-dimensional feature vector (integrating physiological, imaging, and operational data), the hidden layer is set to 6 layers, the activation function is ReLU, the batch size is set to 32, the learning rate is set to 1e-4, the training epochs are 50, and the model performance is validated every 10 epochs (the MAE of the predicted NRS score on postoperative day 3 is <0.8 and the AUC of the predicted flatus probability on postoperative day 7 is >0.9 as the standard). After the model training is complete, inputting the patient's perioperative feature map will output a "patient status evolution curve". Taking patient LC20240501 as an example, the curve can intuitively show: the NRS score dropped from 4 to 3 at 6 hours postoperatively, remained at 3 at 12 hours postoperatively, and dropped to 2 at 2 days postoperatively; the patient passed gas at 18 hours postoperatively, and could eat a semi-liquid diet at 2 days postoperatively; the WBC count increased from 8×10 at 3 days postoperatively. 9 / L decreased to 7×10 9 / L, 7 days after surgery, ALT dropped from 35U / L to 28U / L, clearly showing the patient's recovery trend. During the risk index and customized suggestion generation stage, the risk index calculation needs to integrate "core risk + secondary risk": the core risk is the probability of bile leakage (based on the degree of preoperative gallbladder adhesion + intraoperative conversion to open surgery; patients with grade 3 adhesion and conversion to open surgery have a bile leakage probability of up to 15%; this patient has grade 1 adhesion and no conversion to open surgery, so the bile leakage probability is 2%) and the risk of delayed recovery (based on postoperative flatus time prediction + age factor; age > 65 years can increase the risk of delayed recovery by 5%; this patient is 55 years old and predicted to have flatus 18 hours postoperatively, so the risk of delayed recovery is 8%). The secondary risk is the risk of infection (based on postoperative WBC + body temperature; this patient has no abnormalities, so the risk of infection is 3%, with a weight of 0.1). The final risk index formula is adjusted to "(probability of bile leakage × 0.6 + risk of delayed recovery × 0.4 + risk of infection × 0.1) × 100", which calculates the patient's risk index to be 4.7 (still low risk; 0-20 is low risk, 21-50 is medium risk, and 51-100 is high risk). Preliminary treatment recommendations should be formulated according to risk stratification: Low-risk patients do not require extended liver protection treatment preoperatively (if ALT is normal), should maintain pneumoperitoneum pressure at 12-14 mmHg during surgery, and can get out of bed 24 hours postoperatively; Intermediate-risk patients (e.g., ASA III) require extended liver protection treatment preoperatively for 5 days, with pneumoperitoneum pressure controlled below 12 mmHg during surgery, and can get out of bed 48 hours postoperatively; High-risk patients (e.g., with cirrhosis) require multidisciplinary consultation to determine the surgical plan. Preliminary recommendations need to be iteratively optimized based on historical treatment data. Data from 500 low-risk LC patients in the hospital over the past two years revealed that: patients assisted with ambulation 12 hours postoperatively had an average time to first flatus of 16 hours, 12 hours shorter than those who got out of bed 24 hours postoperatively (average 28 hours), and the incidence of postoperative pulmonary infection decreased from 5% to 2%; simultaneously, patients using ibuprofen + acetaminophen for postoperative analgesia had an 8% incidence of nausea and vomiting, lower than the 20% incidence with opioids alone. Therefore, the initial recommendation was optimized to "assist the patient in getting out of bed 12 hours after surgery, and use ibuprofen (0.3g each time, 3 times a day) + acetaminophen (0.5g each time, 3 times a day) for postoperative analgesia". For the patient (with mild hypertension, and blood pressure controlled at around 130 / 80 mmHg before surgery), the additional recommendation was "monitor blood pressure after surgery to avoid blood pressure fluctuations caused by analgesics", thus forming the final customized recommendation.

[0106] The visual collaboration platform application relies on the hospital's existing "Clinical Multidisciplinary Team (MDT) Collaboration Platform." Based on DingTalk Enterprise Edition, the platform requires no initial setup and only 3-5 days to complete medical data interface integration and permission configuration, enabling rapid deployment. The platform employs a three-tiered permission management system: Level 1 permissions (e.g., department heads) allow viewing all patient data and modifying treatment plans; Level 2 permissions (e.g., attending physicians) allow viewing patient data within their department and submitting suggestions; and Level 3 permissions (e.g., nurses) only allow data entry and viewing of basic information. Furthermore, the national standard SM4 encryption algorithm is used to encrypt data transmission and storage, complying with medical data security standards. In terms of information sharing, the platform features a "patient-specific treatment space." Within 24 hours of patient admission, attending physicians in hepatobiliary surgery can upload perioperative characteristic atlases (including image annotations), risk index calculation processes, and preliminary recommendations. Anesthesiologists can log in to directly view intraoperative MAP and SpO2 monitoring recommendations (e.g., for ASA II patients, intravenous general anesthesia with laryngeal mask airway is recommended to avoid endotracheal intubation injury) and note anesthesia precautions (e.g., if the patient is sensitive to propofol, the induction dose should be reduced). Rehabilitation physicians can obtain detailed postoperative activity guidance (e.g., specific steps for getting out of bed 12 hours post-surgery: sit up for 30 minutes, stand by the bedside for 10 minutes, and finally walk slowly for 50 meters) and upload rehabilitation assessment forms. Laboratory and radiology departments can upload examination reports in real time, and the platform automatically links them to the patient's treatment space, avoiding time wasted by doctors manually searching for reports. Regarding real-time interactive functions, the platform supports a hybrid "online + offline" MDT meeting mode: the preoperative MDT meeting (held one day before surgery) is chaired by the director of hepatobiliary surgery, with offline participants (hepatobiliary surgeons, anesthesiologists, and radiologists) in the hospital's MDT meeting room, and out-of-town experts (such as professors of hepatobiliary surgery from provincial hospitals) participating remotely through the platform; during the meeting, radiologists use the platform's "image annotation tool" to mark the adhesion sites between the gallbladder and duodenum on preoperative CT images with a red pen, and simultaneously provide a voice explanation that "the adhesion area has a rich blood supply, and careful separation is required during surgery"; anesthesiologists, based on ASA classification and cardiopulmonary function indicators, submit feedback on the platform such as "it is recommended to use laryngeal mask ventilation, and the dosage of anesthesia induction drugs is propofol 1.5mg / kg and remifentanil 1μg / kg"; doctors from various departments can leave messages on the platform in real time (such as "the patient's preoperative ALT is normal, and there is no need to delay surgery"); after the meeting, the platform automatically generates meeting minutes, which are archived to the patient's treatment space after confirmation by all participants to avoid subsequent disputes.Regarding parameter fine-tuning, the platform includes an "abnormal data early warning module": On the first postoperative day, a nurse from the nursing department enters a patient's NRS score of 4 (higher than the model's predicted score of 2) through the app. The platform automatically triggers an early warning (notifying the attending physician in hepatobiliary surgery and the anesthesiologist via SMS and platform pop-up), and displays an analysis of the abnormal cause (such as insufficient analgesic dosage or patient insensitivity to the drug). After logging into the platform, the hepatobiliary surgeon checks the patient's postoperative blood pressure (135 / 85 mmHg, no significant fluctuation) and body temperature (37.2℃, normal). Based on the anesthesiologist's feedback that "a single intravenous injection of flurbiprofen ester 50mg is recommended," the analgesia plan is modified on the platform, and the reason for the adjustment is noted. After the adjustment, the nurse records the NRS score every 4 hours. On the second day, the score is entered as 2 (returning to the predicted range). The platform automatically records the adjustment process (score before adjustment, adjustment plan, and effect after adjustment) and synchronizes the data to the hospital's AI laboratory for iterative training of the Perceiver model (such as adding samples of "NRS score > 4 points requires flurbiprofen ester supplementation" to improve the model's prediction accuracy). In addition, the platform also has a "treatment effect statistics module" that automatically calculates the adoption rate of doctors' suggestions in each department (e.g., the adoption rate of anesthesiology department suggestions reached 92%) and the incidence of postoperative complications in patients (e.g., the incidence of postoperative bile leakage in low-risk patients was 0.5%), providing data support for departmental performance evaluation. At the same time, it regularly generates "LC surgery treatment quality reports" to help hospitals optimize the overall treatment process.

[0107] In summary, this application's embodiments collect multi-department data through the hospital's existing information system, eliminating the need for additional hardware costs and reducing hospital operating costs. Regarding data value, core features are extracted through technical models, transforming fragmented data into structured graphs and visualized curves, thus improving the scientific nature of medical decision-making. In terms of diagnostic and treatment accuracy, management priorities are divided according to surgical cycles, and personalized plans are generated by combining risk grading and historical data, reducing postoperative complications and accelerating patient recovery. Regarding multi-departmental collaboration, information barriers are broken down using an MDT collaborative platform, enabling data sharing, real-time interaction, and anomaly fine-tuning while ensuring medical data security. In terms of hospital management, it provides a basis for performance evaluation, and quality reports help optimize treatment processes, thereby improving both the level of medical services and the patient's medical experience and long-term quality of life.

[0108] Figure 11 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: The memory 1101, the processor 1102, and the computer program stored on the memory 1101 and executable on the processor 1102.

[0109] When the processor 1102 executes the program, it implements a smart medical data self-supervised deep processing and analysis method provided in the above embodiments.

[0110] Furthermore, electronic devices also include: Communication interface 1103 is used for communication between memory 1101 and processor 1102.

[0111] The memory 1101 is used to store computer programs that can run on the processor 1102.

[0112] The memory 1101 may include high-speed RAM (Random Access Memory) and may also include non-volatile memory, such as at least one disk storage.

[0113] If the memory 1101, processor 1102, and communication interface 1103 are implemented independently, then the communication interface 1103, memory 1101, and processor 1102 can be interconnected via a bus to complete communication between them. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 11 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0114] Optionally, in a specific implementation, if the memory 1101, processor 1102, and communication interface 1103 are integrated on a single chip, then the memory 1101, processor 1102, and communication interface 1103 can communicate with each other through an internal interface.

[0115] The processor 1102 may be a CPU (Central Processing Unit), an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of this application.

[0116] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method for self-supervised deep processing and analysis of smart medical data.

[0117] In the description of this specification, the references to "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0118] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0119] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0120] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0121] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0122] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A self-supervised deep processing and analysis system for intelligent medical data, characterized in that, include: The system comprises a multi-departmental data perception layer, a surgical feature processing center, a full-cycle dynamic analysis unit, an adaptive clinical decision-making module, a cross-departmental data collaboration unit, and a dynamic feedback optimization unit; among these components... The multi-departmental data sensing layer is used to collect multi-departmental data and postoperative follow-up data from patients' wearable devices; The surgical feature processing center is used to extract features from data from various departments, and to associate pre- and post-operative data through an attention mechanism to generate a unified perioperative feature map of the patient. The full-cycle dynamic analysis unit is used to analyze the patient's status in stages based on the patient's perioperative unified feature map, and output a risk index in combination with the self-supervised training prediction model. The adaptive clinical decision module is used to generate customized suggestions for different departments based on the analysis results and the risk index. The cross-departmental data collaboration unit is used to build a visual collaboration platform, enabling doctors from different departments to share and interact on the analysis results of the entire surgical cycle; The dynamic feedback optimization unit is used to collect clinical diagnosis and treatment results and fine-tune the parameters of the self-supervised model in reverse.

2. The intelligent medical data self-supervised deep processing and analysis system according to claim 1, characterized in that, The multi-departmental data perception layer includes a data acquisition module and a follow-up data integration unit. The data acquisition module is used to collect diagnosis and treatment data from multiple departments, including internal medicine, surgery, and anesthesiology. The follow-up data integration unit is used to receive postoperative heart rate, activity level, and sleep quality data transmitted from the patient's wearable device and align them by timestamp.

3. The intelligent medical data self-supervised deep processing and analysis system according to claim 1, characterized in that, The perioperative feature processing center includes a feature extraction unit, a cross-temporal data association module, and a feature map generation engine. The feature extraction unit is used to extract physiological indicator features, imaging features, and surgical operation features from departmental data. The cross-temporal data association module is used to associate preoperative baseline data and postoperative recovery data through a multi-head attention mechanism to explore temporal correlations. The feature map generation engine is used to associate multi-dimensional features with individual patients and construct a unified perioperative feature map that includes time and indicator dimensions.

4. The intelligent medical data self-supervised deep processing and analysis system according to claim 1, characterized in that, The full-cycle dynamic analysis unit includes a phased state assessment module, a self-supervised prediction model training unit, and a risk index calculation engine. The phased state assessment module divides the patient's entire lifecycle into preoperative assessment, intraoperative monitoring, and postoperative recovery phases, analyzing the physiological state and recovery progress at each stage. The self-supervised prediction model training unit performs self-supervised training based on unlabeled clinical data, using the Perceiver model to predict the evolution of the patient's condition. The risk index calculation engine generates a quantitative risk index based on the patient's condition evolution process, integrating complication probability and recovery delay risk indicators.

5. The intelligent medical data self-supervised deep processing and analysis system according to claim 1, characterized in that, The adaptive clinical decision-making module includes a departmental needs analysis unit, a decision rule generation module, and a suggestion optimization unit. The departmental needs analysis unit is used to identify the clinical needs of different departments such as internal medicine, surgery, and anesthesiology. The decision rule generation module is used to generate preliminary treatment suggestion rules based on the patient's risk index and departmental needs. The suggestion optimization unit is used to iteratively optimize the preliminary suggestions by combining historical treatment effect data to generate customized clinical suggestions.

6. The intelligent medical data self-supervised deep processing and analysis system according to claim 1, characterized in that, The cross-departmental data collaboration unit includes a visualization platform construction module, a data sharing permission management unit, and a real-time interaction module. The visualization platform construction module uses WebGL to build a three-dimensional visualization collaboration interface to display the patient's full-cycle feature map and analysis results. The data sharing permission management unit is used to allocate data viewing and editing permissions according to the department and responsibilities. The real-time interaction module is used to annotate key data online, add diagnostic and treatment opinions, and conduct real-time interactive collaboration.

7. The intelligent medical data self-supervised deep processing and analysis system according to claim 1, characterized in that, The dynamic feedback optimization unit includes a treatment result acquisition module, a feedback data matching unit, and a model parameter fine-tuning engine. The treatment result acquisition module is used to collect data on postoperative complications, recovery period, treatment effect scores, and clinical outcomes of patients. The feedback data matching unit is used to associate and match the treatment results with the corresponding patient's model input data and prediction results. The model parameter fine-tuning engine is used to fine-tune the weight parameters using a gradient descent algorithm based on the matched feedback data to optimize prediction accuracy.

8. A method for applying to a smart medical data self-supervised deep processing and analysis system according to any one of claims 1-7, characterized in that, The method includes: Obtain multidisciplinary diagnosis and treatment data as well as postoperative follow-up data; The multidisciplinary diagnosis and treatment data and postoperative follow-up data are analyzed and processed to extract physiological indicators, imaging features and surgical operation features. The data before and after surgery are linked through the attention mechanism to generate a unified feature map of the patient in the perioperative period. Based on the unified perioperative feature map of the patients, the entire patient cycle is divided into the preoperative assessment period, the intraoperative monitoring period, and the postoperative recovery period. The physiological status and recovery progress at each stage are analyzed. Self-supervised training is performed based on unlabeled clinical data. The Perceiver model is used to predict the evolution of the patient's condition. Based on the evolution of the patient's condition, the probability of complication and the risk index of delayed recovery are integrated to generate a risk index. Based on the analysis results and the risk index, the clinical needs of different departments are identified, and preliminary treatment recommendation rules are generated. Combined with historical treatment effect data, the preliminary treatment recommendation rules are iteratively optimized to generate customized clinical recommendations. Based on the visual collaboration platform, doctors from various departments can share and interact with the analysis results of the entire surgical cycle and the customized suggestions, collect clinical diagnosis and treatment results in real time, and fine-tune parameters in reverse.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the program to implement the self-supervised deep processing and analysis method for smart medical data as claimed in claim 8.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When a computer program or instruction is executed, it implements the self-supervised deep processing and analysis method for smart medical data as claimed in claim 8.

Images