Intelligent management system for thoracic surgery intensive care unit based on multi-modal data fusion

Abstract

The invention discloses a multi-modal data fusion-based intelligent management system for a thoracic surgery monitoring unit, belongs to the technical field of medical information and artificial intelligence, and aims to solve the limitation of an existing thoracic surgery monitoring system in the aspects of multi-modal data fusion, heterogeneous data semantic alignment and intelligent deep analysis and prediction decision. The system is characterized by comprising a multi-modal data acquisition unit, a heterogeneous data fusion and semantic alignment module, an intelligent analysis and prediction decision module, a man-machine interaction and visual presentation module and a secure storage and management module. By the adoption of the technical scheme, comprehensive multi-modal data fusion, high real-time performance, deep intelligent analysis and prospective prediction can be achieved, intelligent decision support, resource optimization, continuous learning and self-adaptive optimization are provided, and the intelligent level and patient management efficiency of the thoracic surgery intensive care unit are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of medical information technology and artificial intelligence technology, and in particular to an intelligent management system for thoracic surgery monitoring rooms based on multimodal data fusion. Background Technology

[0002] With the continuous development of global healthcare, especially in the field of surgery, the safety of patients during the perioperative period and the quality of postoperative recovery are receiving increasing attention. Thoracic surgery, due to its involvement of vital organs within the thoracic cavity and its high complexity, often requires patients to undergo close and continuous vital sign monitoring and rehabilitation management in the intensive care unit (ICU) postoperatively. Against this backdrop, building an efficient, precise, and intelligent monitoring and management system has become a key technological direction for improving the safety of thoracic surgery, optimizing the allocation of medical resources, and improving patient outcomes. Such a system aims to provide timely and comprehensive decision support to medical staff by acquiring and deeply analyzing multi-dimensional information such as patient physiological data, environmental parameters, and the status of medical equipment in real time, thereby effectively reducing the risk of complications and accelerating the patient's recovery process.

[0003] In the existing technology, various technical solutions have been developed for patient monitoring or medical device management. For example, Chinese Patent Publication No. CN108471957B discloses a patient monitor. The core of this solution lies in achieving unified management of time information by setting a master clock and multiple operating systems, and enabling data interaction with external measuring devices to improve the accuracy and real-time performance of the monitoring system. Specifically, it ensures the consistency of time sequences across different data sources through a standardized time synchronization mechanism, which is of great significance for accurately recording and analyzing patients' physiological parameters. Correspondingly, Chinese Patent Publication No. CN116157053B proposes a patient monitor and its device information management method. This technology mainly achieves interconnection between the patient monitor and various medical devices and device management servers through a communication interface, thereby automatically collecting and effectively managing device information, significantly improving the operating efficiency and maintenance convenience of medical devices. Within their respective specific application scopes, the above technologies have indeed effectively solved some of the technical problems faced at the time, such as real-time monitoring of a single patient's vital signs or automated management of medical devices, laying the foundation for the development of the medical monitoring field.

[0004] However, with the increasing complexity of thoracic surgery and the growing demand for refined and personalized postoperative monitoring, some inherent characteristics of the aforementioned existing technologies at the principle level are gradually revealing their limitations in addressing current challenges. Specifically, CN108471957B focuses on monitoring the vital signs of a single patient, and its data processing mode mainly revolves around structured physiological parameters, such as heart rate, blood pressure, and blood oxygen saturation. This single-modal, point-based monitoring paradigm is inadequate when dealing with the complex and multi-factor-related pathological changes that may occur in patients after thoracic surgery. The real-world scenario in a thoracic surgical intensive care unit goes far beyond this, involving a large amount of unstructured or semi-structured data, such as chest imaging data (X-rays, CT images), dynamic changes in drainage volume, ambient temperature and humidity, air quality, and operating parameters of various life support equipment (such as ventilator mode and infusion pump speed). The aforementioned patented solutions lack the ability to comprehensively process and deeply integrate this heterogeneous, multimodal data. The reason for this is that its architecture design did not fully consider the standardized collection, efficient transmission, and semantic alignment of multi-source data from the source, resulting in each data source being like an "information island" and making it difficult to form a synergistic effect.

[0005] Furthermore, the equipment information management method proposed in CN116157053B primarily focuses on the operational status, usage records, and maintenance information of medical equipment, aiming to optimize equipment management processes. While this has a positive effect on improving hospital operational efficiency, its core value does not lie in the direct, dynamic assessment and early warning of patients' conditions. The data processing approach of this solution is relatively simplistic, failing to fully utilize cutting-edge technologies such as artificial intelligence, machine learning, and big data analytics for in-depth real-time analysis and prediction of postoperative patient monitoring data in thoracic surgery. For example, a single equipment malfunction warning cannot replace the intelligent identification and correlation analysis of comprehensive indicators such as subtle changes in patient breathing patterns, abnormal drainage fluid characteristics, or imaging progression. In the high-risk, time-sensitive environment of thoracic surgery monitoring, single-dimensional data insights are far from sufficient. Therefore, although existing systems can accumulate massive amounts of data, the lack of efficient multimodal data fusion mechanisms and intelligent analysis capabilities prevents this data from being fully transformed into valuable clinical insights and intelligent decision support. This creates a deep-seated technological contradiction: a significant gap exists between the need for comprehensive, accurate, real-time, and forward-looking patient status assessment and management in thoracic surgery intensive care units and the current monitoring systems' ability to provide fragmented, single-modal data presentation lacking intelligent correlation analysis capabilities. Under the existing model, data effectiveness is not simply linearly positively correlated with the increase in data volume. On the contrary, without efficient fusion and intelligent parsing mechanisms, the influx of heterogeneous data may lead to information overload, reducing the efficiency of medical staff in capturing key information, thus causing the overall decision support capability of the system to decrease rather than increase, creating an "information paradox." This limitation—the inability to integrate fragmented data into a unified, comprehensive patient profile and, based on this, conduct intelligent risk assessment and early warning—severely restricts the intelligence level of the intensive care unit and its ability to respond to emergencies.

[0006] Therefore, how to construct an intelligent management system for thoracic surgery intensive care units that can effectively integrate multimodal patient data, achieve high real-time performance and high compatibility of data, and utilize artificial intelligence technology for in-depth analysis and intelligent decision support has become a key challenge and an urgent technical problem for those skilled in the art. Summary of the Invention

[0007] This invention aims to overcome the significant limitations of existing thoracic surgery monitoring systems in multimodal data fusion, heterogeneous data semantic alignment, and intelligent deep analysis and predictive decision-making. It provides an intelligent management system for thoracic surgery monitoring rooms based on multimodal data fusion. This system can achieve comprehensive, real-time acquisition and deep fusion of patient physiological status, environmental parameters, medical equipment operating status, and other clinical data. By integrating advanced artificial intelligence algorithms, it performs efficient intelligent analysis and forward-looking prediction of the fused data, thereby providing medical staff with accurate clinical decision support and early risk warnings, significantly improving the intelligence level of the thoracic surgery monitoring room and patient management efficiency.

[0008] To achieve the above-mentioned objectives, this invention provides an intelligent management system for a thoracic surgery intensive care unit based on multimodal data fusion. The system mainly includes: a multimodal data acquisition unit, a heterogeneous data fusion and semantic alignment module, an intelligent analysis and predictive decision-making module, a human-computer interaction and visualization module, and a secure storage and management module. The system acquires patient physiological data, medical imaging data, environmental parameter data, medical equipment operating parameter data, and manually entered and semi-structured data in real time and in parallel through the multimodal data acquisition unit, and transmits these heterogeneous data to the heterogeneous data fusion and semantic alignment module. The heterogeneous data fusion and semantic alignment module preprocesses, standardizes, semantically aligns, and deeply fuses the received heterogeneous data to generate a unified, high-dimensional patient state feature vector. The intelligent analysis and predictive decision-making module receives the patient state feature vector and uses various artificial intelligence algorithms to perform pattern recognition, anomaly detection, risk assessment, and trend prediction on the feature vector. The human-computer interaction and visualization module presents the analysis results, early warning information, and decision suggestions from the intelligent analysis and predictive decision-making module to medical staff in an intuitive and operable form. The secure storage and management module is responsible for the secure storage, retrieval, backup, and access control of all data to ensure data integrity and security.

[0009] In a preferred embodiment of the present invention, the multimodal data acquisition unit is used to acquire multi-source heterogeneous data in the thoracic surgery monitoring room in real time and in parallel, and includes the following sub-units:

[0010] First, the physiological parameter acquisition subunit is used to collect various physiological vital signs data of patients, including but not limited to electrocardiogram (ECG), pulse oxygen saturation (SpO2), non-invasive blood pressure (NIBP), invasive blood pressure (IBP), body temperature, respiratory rate, end-tidal carbon dioxide (EtCO2), and pulmonary function parameters. The physiological parameter acquisition subunit is implemented through integrated medical-grade sensors, such as: an ECG sensor array with at least twelve leads and a sampling frequency of no less than 1000Hz for real-time acquisition of cardiac electrophysiological activity; a blood oxygen sensor based on transmission photoplethysmography with measurement wavelengths including 660nm and 940nm and a response time of less than 2 seconds for continuous monitoring of blood oxygen saturation and pulse rate; a non-invasive blood pressure module using oscillometric or Korotkoff sound principles with a measurement range of 0-300mmHg and a measurement accuracy of ±3mmHg, supporting cuff switching for adults, children, and newborns; an invasive blood pressure module using piezoresistive or strain gauge sensors with a measurement range of -50 to 300mmHg and an accuracy of ±2mmHg for monitoring arterial pressure, central venous pressure, or pulmonary artery pressure; a body temperature module using a thermistor or infrared sensor array with a measurement accuracy of ±0.1℃; and a respiratory sensor using impedance or respiratory carbon dioxide gas analysis to acquire respiratory rate and EtCO2 values. All physiological parameter data are output in a uniform data format, such as HL7 or DICOM-RS standard, with high-precision timestamps to ensure temporal consistency for subsequent data fusion.

[0011] Second, the medical image acquisition subunit is used to acquire the patient's chest medical image data, including but not limited to chest X-rays, computed tomography (CT) images, and ultrasound images. This subunit interfaces with the hospital's Picture Archiving and Communication System (PACS) to acquire the latest image data in real-time or near real-time using the DICOM standard protocol and extract relevant image diagnostic reports. The interface uses a DICOM Query / Retrieve service provider (SCP) mode to achieve accurate retrieval and high-speed transmission of patient images from the PACS server, and supports image compression formats such as JPEG 2000 and RLE to optimize transmission efficiency. This unit can automatically parse DICOM metadata, including patient ID, examination date, image modality, slice thickness, pixel spacing, etc., providing structured information for subsequent image analysis.

[0012] Third, the environmental parameter acquisition subunit is used to monitor environmental indicators within the monitoring room, including but not limited to room temperature, humidity, air quality (such as PM2.5 concentration and CO2 concentration), noise level, and light intensity. This environmental parameter acquisition subunit consists of a distributed micro-sensor network, for example: using thermocouples or platinum resistance sensors for temperature measurement with an accuracy of ±0.5℃; using capacitive or resistive humidity sensors for humidity measurement with an accuracy of ±3%RH; and using a PM2.5 sensor based on the laser scattering principle with a resolution of 1 μg / m³. 3 The system employs a CO2 sensor based on the non-dispersive infrared (NDIR) principle, with a measurement range of 0-5000ppm; a sound pressure level meter is used for noise measurement, with a response frequency range of 20Hz-20kHz; and a photodiode array is used for light intensity measurement, with a resolution of 1 lux. The sensor network transmits data in real-time to a data aggregation point via low-power wireless communication protocols such as Wi-Fi or ZigBee, and integrates a high-precision real-time clock (RTC) module to ensure timestamp synchronization of all environmental data.

[0013] Fourth, the medical equipment operation parameter acquisition subunit is used to acquire the real-time operating status and parameters of various life support and treatment devices in the intensive care unit, including but not limited to ventilator settings (such as ventilation mode, tidal volume, respiratory rate, positive end-expiratory pressure PEEP, and inhaled oxygen concentration FiO2), infusion pump rate, infusion pump dosage, extracorporeal membrane oxygenation (ECMO) equipment flow and pressure, and alarm information from other monitoring devices. This medical equipment operation parameter acquisition subunit performs protocol conversion and data acquisition through dedicated communication interfaces with medical devices (such as RS-232, USB, Ethernet) or a Medical Device Integration (MDI) platform. For example, for ventilators, the output parameter frames are parsed using standard serial communication protocols (such as RS-232C) to extract the current ventilation mode, preset tidal volume, respiratory rate, actual PEEP value, and FiO2 value; for infusion pumps and infusion pumps, the current infusion / infusion rate, total infusion / infusion volume, remaining volume, and alarm status are obtained through the corresponding communication modules. All collected device operating parameters are standardized into a unified JSON or XML format and include a timestamp.

[0014] Fifth, the manual data entry and semi-structured data acquisition subunit is used to collect patient clinical data manually entered by medical staff through the system interface, as well as other semi-structured information that is difficult to obtain through automated equipment, including but not limited to the amount, color, and characteristics of drainage fluid (such as thoracic drainage and pericardial drainage), wound condition, pain scores (such as VAS scores), consciousness status assessments (such as GCS scores), nutritional intake, excretion, medication records, and specific nursing operation records. This subunit provides a structured data entry interface, supporting text input, drop-down selection, numeric input, and image upload functions, and achieves semantic constraints and standardization of data through a preset clinical terminology dictionary (such as a subset of SNOMED CT). This subunit has a verification mechanism to ensure the validity and completeness of the entered data and automatically adds the information of the person entering the data and a timestamp.

[0015] In a preferred embodiment of the present invention, the heterogeneous data fusion and semantic alignment module is used to preprocess, standardize, semantically align, and deeply fusion the heterogeneous data collected by the multimodal data acquisition unit, thereby generating a unified, high-dimensional patient state feature vector. The module includes the following components:

[0016] First, the data parsing and preprocessing unit performs preliminary parsing, format conversion, missing value imputation, and noise removal on data from different sources. For physiological signal data, methods such as adaptive filtering, wavelet transform, or empirical mode decomposition (EMD) are used to remove baseline drift, power frequency interference, and motion artifacts, and interpolation algorithms (such as linear interpolation and cubic spline interpolation) are used to handle transient missing values ​​in sensor data. For medical image data, pixel value normalization, image enhancement (such as contrast stretching and histogram equalization), and region of interest (ROI) extraction are performed. For text and semi-structured data, natural language processing (NLP) techniques are applied for named entity recognition, sentiment analysis (for patient descriptions), and key information extraction. All data is converted into a unified numerical or vector representation after preprocessing.

[0017] Second, the timestamp synchronization and sequence reconstruction unit ensures precise temporal alignment of data from different modalities and acquisition frequencies. This unit employs high-precision time synchronization protocols, such as Network Time Protocol (NTP) or Precise Time Protocol (PTP), to maintain nanosecond-level synchronization between all data acquisition devices and the central processing system's time source. For data with different sampling frequencies, resampling techniques (such as linear interpolation, zero-order hold, or spline interpolation) are used to unify them to a common time reference or a specific sampling frequency. Through a sliding time window method, real-time data streams from different modalities are aggregated into a time-synchronized multivariate time-series dataset to reflect the patient's overall condition at a specific point in time or within a given time period.

[0018] Third, the ontology and knowledge graph construction unit is used to define semantic relationships between different data modalities and construct a domain ontology model that includes medical concepts, clinical events, and equipment status. This unit is based on ontology engineering methods and uses the Web Ontology Language (OWL) to define core concepts, attributes, and relationships, such as the association between "patient" and "vital signs," "imaging examinations," and "medication"; and the parameter association between "ventilator" and "ventilation mode" and "tidal volume." This ontology model combines internationally standardized medical terminology (such as SNOMED CT and LOINC) with clinical knowledge specific to the field of thoracic surgery. Through knowledge graph technology (such as RDF triple storage), heterogeneous data is mapped to a unified semantic space, solving the semantic gap problem caused by data heterogeneity. For example, "heart rate" data from different devices is mapped to a unique "heart rate" concept in the ontology, while recording its source, measurement method, and unit.

[0019] Fourth, feature-level fusion algorithms are used to fuse preprocessed and semantically aligned multimodal data at the feature level using deep learning or machine learning methods. These algorithms employ a multi-branch neural network architecture, where each branch processes raw or low-level features of one modality. The high-level features extracted from different branches are then integrated through connection layers, attention mechanisms, or gated recurrent units (GRU) layers. For example, for physiological signal sequences and medical images, convolutional neural networks (CNNs) are used to extract image features, and recurrent neural networks (RNNs) are used to extract time-series features. Then, through a cross-attention mechanism, the model learns the dependencies and contribution weights between features of different modalities, generating a joint feature vector containing information from all modalities. This joint feature vector is high-dimensional and information-rich, comprehensively depicting the patient's current physiological and pathological state.

[0020] In a preferred embodiment of the present invention, the intelligent analysis and predictive decision-making module is used to receive the patient state feature vector generated by the heterogeneous data fusion and semantic alignment module, and to use various artificial intelligence algorithms to perform pattern recognition, anomaly detection, risk assessment, and trend prediction on the feature vector. The module includes the following sub-modules:

[0021] First, the physiological state abnormality identification and early warning submodule is used to monitor the changing trends of patients' physiological parameters in real time and identify potential abnormal patterns. This submodule employs a deep learning-based time-series analysis model, such as a Long Short-Term Memory (LSTM) network or a Gated Recurrent Unit (GRU), to model continuous physiological time-series data. By learning the time dependence and fluctuation range under normal physiological conditions, the model can detect abnormal changes exceeding the normal baseline in real time, such as arrhythmias, sudden drops in blood pressure, persistent decreases in blood oxygen saturation, and abnormal breathing patterns. The model is trained on a large amount of historical normal and abnormal physiological data and can identify subtle, non-obvious early abnormal patterns. When an abnormality is detected, the submodule generates different levels of early warning information based on preset clinical thresholds and the degree of abnormality.

[0022] Second, the intelligent assessment submodule for imaging progression is used to automatically analyze patients' chest medical images and assess changes in lesions and the progression of complications. This submodule employs a deep learning model based on a three-dimensional convolutional neural network (3D CNN) or Vision Transformer to perform image segmentation, target detection, and lesion quantification on CT or X-ray sequences. For example, for pleural effusion, the model can automatically segment the effusion area and quantify its volume changes; for pulmonary infection or consolidation, the model can identify the extent, density, and distribution of lesions and assess their evolution trend. The model can automatically compare images at different time points, identify and quantify subtle progressions or improvements, and provide auxiliary diagnosis and risk assessment for potential complications such as pneumothorax, atelectasis, and pulmonary embolism.

[0023] Third, the comprehensive risk assessment and complication prediction submodule is one of the core intelligent decision support functions of this invention. It is used to dynamically and prospectively predict the risk of patient complications based on all fused multimodal features. This submodule employs a machine learning model within a multi-task learning or ensemble learning framework, such as XGBoost, LightGBM, or Stacking Ensemble models. The model takes features from all modalities, including physiological, imaging, environmental, device, and manual input, as input to predict the probability of a patient developing specific complications (such as acute respiratory distress syndrome (ARDS), pulmonary infection, heart failure, coagulation dysfunction, sepsis, or renal insufficiency) within a specific time window (e.g., 24 hours, 48 ​​hours, or 72 hours).

[0024] The core algorithm of the prediction model can be expressed as:

[0025] Let the fused patient state feature vector be x = [x1, x2, ..., xn].n ], where x1 represents a feature extracted from multimodal data.

[0026] For each complication C j This invention trains a binary classification prediction model, such as a logistic regression model, whose prediction probability P(C j The formula for calculating (1 / x) is:

[0027]

[0028] Among them, w j Is it related to complication C j The relevant feature weight vector, b j These are bias terms. These parameters are obtained by training the model on a large amount of historical patient data and are designed to maximize predictive accuracy.

[0029] Furthermore, to combine the correlation between different complications with the physician's attention to specific complications, a comprehensive risk score can be introduced.

[0030]

[0031] Where M is the total number of complications, α j Complication C j Clinical weight (determined by expert knowledge or clinical data), S j Complication C j The severity score is used to assess the patient's risk. This comprehensive risk score provides healthcare professionals with a macro-level overview of patient risk and triggers customized alerts based on the type of high-risk complication.

[0032] Fourth, the rehabilitation trend prediction and intervention suggestion submodule is used to assess the dynamic trajectory of patient rehabilitation and provide personalized rehabilitation suggestions. This submodule utilizes time series prediction models (such as ARIMA, Prophet, or deep learning sequence models) to predict short- and medium-term trends in key rehabilitation indicators (such as changes in drainage volume, respiratory function recovery indicators, time to ambulation, pain scores, etc.). Based on the predicted rehabilitation trajectory and a pre-set rehabilitation path model, this submodule can identify deviations in the patient's rehabilitation process and, in conjunction with clinical guidelines and historical treatment experience, provide targeted intervention suggestions, such as adjusting drainage strategies, optimizing respiratory rehabilitation training programs, adjusting analgesia programs, or providing early mobilization guidance.

[0033] Fifth, the dynamic resource optimization suggestion submodule is used to dynamically optimize the allocation of medical resources in the intensive care unit based on the patient's current status and predicted needs. This submodule analyzes data such as the severity of the patient's condition, risk of complications, required level of care, and utilization rate of medical equipment. Combined with information on intensive care unit beds, human resources (doctor-to-nurse ratio), and equipment availability, it generates resource allocation suggestions through optimization algorithms (such as linear programming and heuristic algorithms). For example, when the system predicts that a patient's risk of complications will significantly increase in the next few hours, it can suggest increasing the patient's level of care or allocating medical staff with specific skills for intensive monitoring; when it predicts that a piece of equipment (such as a ventilator) will soon be idle, its status can be updated and recommended to other patients who may need it.

[0034] In a preferred embodiment of the present invention, the human-computer interaction and visualization module is used to present the analysis results, early warning information, and decision suggestions of the intelligent analysis and predictive decision-making module to medical staff in an intuitive and operable form, and to support medical staff in data querying and feedback. The module includes:

[0035] First, a unified view building unit is used to integrate data and analysis results from different modalities to construct a unified, multi-dimensional patient monitoring dashboard. This dashboard can display real-time trends in vital signs, image comparisons, laboratory test results, drainage status, medication records, predicted probabilities of complication risks, and recovery progress curves. This unit supports custom layouts and view switching, allowing medical staff to choose different information display modes according to their needs, such as a detailed view for a single patient or an overview view for multiple patients. All data is presented in visual formats such as graphs, tables, or heatmaps to ensure clear and easy-to-understand information.

[0036] Second, the intelligent early warning and alert unit generates real-time, tiered alarm information based on the warning level and content output by the intelligent analysis and prediction decision-making module. The alarm information includes visual cues (such as screen flashing, color coding), auditory cues (such as alarm sounds of different tones and rhythms), and message push notifications (such as notifications sent via mobile device apps, SMS, or the hospital information system). This unit supports priority management and intelligent noise reduction of alarm information to avoid invalid or excessive alarm information interfering with medical staff. Medical staff can confirm, mark, or process alarm information; these operations will be recorded and used for subsequent system learning and optimization.

[0037] Third, the decision support information presentation unit displays the decision suggestions provided by the intelligent analysis and predictive decision-making module in a structured and interpretable format. These suggestions include preventative interventions for high-risk complications, specific adjustment plans for abnormal physiological parameters, clinical interpretation of imaging abnormalities, and personalized rehabilitation plans. This unit provides a confidence assessment of the suggestions and allows for tracing back to key data and analytical models supporting the suggestions, enhancing healthcare professionals' trust in the system's recommendations. For complex suggestions, this unit can provide links to relevant clinical guidelines or knowledge bases to facilitate in-depth understanding by healthcare professionals.

[0038] Fourth, a feedback and learning interface is provided to collect feedback from medical staff on the system's analysis results, prediction accuracy, and decision-making suggestions. This interface allows medical staff to provide positive or negative evaluations of anomalies identified by the system, predicted risks, or provided suggestions, and to input detailed clinical observations or intervention measures. The feedback data, after being anonymized, is used as part of the training data for the continuous learning and iterative optimization of the artificial intelligence model in the intelligent analysis and prediction decision-making module, thereby continuously improving the system's prediction accuracy and decision support capabilities, and achieving system self-adaptation and evolution.

[0039] In a preferred embodiment of the present invention, the secure storage and management module is used to securely and efficiently store, retrieve, back up, and manage access permissions for all data generated during system operation, ensuring data integrity, confidentiality, and traceability. The module includes:

[0040] First, a distributed database system is used to store massive amounts of multimodal data. This database system employs a distributed architecture, such as the Hadoop Distributed File System (HDFS) combined with NoSQL databases (e.g., MongoDB, Cassandra) or NewSQL databases (e.g., TiDB), to support high-concurrency data write and read operations and to elastically expand storage capacity. Structured data (e.g., physiological parameters, device parameters) is stored in relational databases (e.g., PostgreSQL), while unstructured data (e.g., images, waveforms, logs) is stored in object storage or document databases to accommodate the storage needs of different data types.

[0041] Second, a data encryption and access control unit is used to protect the security of patient privacy data and sensitive system information. This unit encrypts all data in transmission using Transport Layer Security (TLS) and encrypts data stored in the database at rest using Advanced Encryption Standard (AES-256). The unit implements a Role-Based Access Control (RBAC) mechanism, setting granular data access permissions for different user groups such as medical staff, administrators, and researchers, ensuring that only authorized users can access specific data or perform specific operations. All data access and operations are recorded in detail in the audit log to meet compliance requirements.

[0042] Third, a data backup and recovery mechanism ensures the continuity and availability of system data. This mechanism supports periodic full backups and real-time incremental backups, storing data copies in off-site data centers to prevent data loss due to single points of failure. In the event of data corruption or system failure, the mechanism can quickly initiate a data recovery process, restoring the system to its most recent valid state, minimizing service interruption time and the risk of data loss. The mechanism also includes a data version management function, allowing for rollback to historical data versions.

[0043] The beneficial effects of this invention are:

[0044] This invention effectively solves the bottleneck of existing monitoring systems when processing heterogeneous and multimodal data by constructing a highly integrated, multi-layered intelligent management architecture, and overcomes the limitations of lacking deep intelligent analysis and forward-looking prediction capabilities. Specifically, this invention achieves:

[0045] First, comprehensive multimodal data fusion: Through a specially designed multimodal data acquisition unit and a heterogeneous data fusion and semantic alignment module, the system can seamlessly integrate various heterogeneous data from patient physiology, medical imaging, environment, medical equipment, and manual input. This full-dimensional data acquisition and refined fusion ensures a comprehensive understanding of the patient's condition, avoids the "information silo" effect caused by traditional single-modal monitoring, and allows all relevant information to converge into a unified, high-dimensional patient profile.

[0046] Secondly, high real-time performance and high compatibility: From the initial design stage, the system fully considered the stringent real-time requirements of medical scenarios. Through high-precision timestamp synchronization, efficient data transmission protocols, and optimized data processing workflows, it ensures low latency throughout the entire chain from data acquisition to intelligent analysis and decision support. Simultaneously, by supporting multiple industry-standard communication protocols and a flexible interface design, the system exhibits extremely high compatibility with medical devices, enabling seamless integration with various existing intensive care unit equipment.

[0047] Third, deep intelligent analysis and forward-looking prediction: Leveraging advanced artificial intelligence (AI) algorithms, particularly deep learning and ensemble learning models, the system can go beyond simple threshold alarms, achieving pattern recognition of patient physiological abnormalities, quantitative assessment of imaging lesions, and accurate prediction of the risk of complications. This forward-looking predictive capability allows medical staff to receive early warnings before the condition worsens, thus gaining valuable intervention time, significantly improving the timeliness and accuracy of clinical decision-making, effectively reducing the incidence of complications, and improving patient prognosis.

[0048] Fourth, intelligent decision support and resource optimization: The system not only provides data and early warnings, but also offers personalized and structured clinical decision recommendations based on big data analysis and AI models, covering all aspects from preventative intervention and treatment plan adjustments to rehabilitation pathway optimization. Furthermore, the system can dynamically recommend medical resource allocation plans based on changes in the patient's condition, optimize intensive care unit operational efficiency, reduce medical costs, and achieve refined management of medical resources.

[0049] Fifth, continuous learning and adaptive optimization: Through embedded feedback and learning interfaces, the system can continuously collect clinical feedback from medical staff and use this real-world experience to continuously train and optimize the built-in AI model. This closed-loop learning mechanism enables the system to continuously improve its predictive accuracy and decision support capabilities over time, thereby achieving system adaptation and evolution, and ensuring that its intelligence level can continuously adapt to ever-changing clinical needs.

[0050] In summary, this invention completely resolves the fundamental technical contradictions in existing monitoring systems regarding multimodal data fusion, intelligent analysis, and forward-looking prediction. It provides a comprehensive, accurate, efficient, and intelligent management solution for thoracic surgery intensive care units, significantly improving the life safety of perioperative patients, enhancing postoperative recovery quality, and optimizing the allocation of medical resources. Attached Figure Description

[0051] Figure 1 This is a system block diagram of the present invention;

[0052] In the diagram: 1. Multimodal data acquisition unit; 2. Heterogeneous data fusion and semantic alignment module; 3. Intelligent analysis and predictive decision-making module; 4. Human-computer interaction and visualization module; 5. Secure storage and management module; 101. Physiological parameter acquisition subunit; 102. Medical image acquisition subunit; 103. Environmental parameter acquisition subunit; 104. Medical equipment operation parameter acquisition subunit; 105. Manual data entry and semi-structured data acquisition subunit; 201. Data parsing and preprocessing unit; 202. Timestamp synchronization and sequence reconstruction unit; 203. Ontology and knowledge graph construction unit. ; 204. Feature-level fusion algorithm; 301. Physiological state abnormality identification and early warning submodule; 302. Intelligent assessment of imaging progress submodule; 303. Comprehensive risk assessment and complication prediction submodule; 304. Rehabilitation trend prediction and intervention suggestion submodule; 305. Dynamic resource optimization suggestion submodule; 401. Unified view construction unit; 402. Intelligent early warning and prompt unit; 403. Decision support information presentation unit; 404. Feedback and learning interface; 501. Distributed database system; 502. Data encryption and access control unit; 503. Data backup and recovery mechanism. Detailed Implementation

[0053] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0054] This invention provides an intelligent management system for thoracic surgery intensive care units based on multimodal data fusion. By integrating advanced data acquisition technology, a sophisticated heterogeneous data fusion mechanism, and cutting-edge artificial intelligence algorithms, it aims to comprehensively improve the quality of perioperative patient monitoring and the efficiency of clinical decision-making in thoracic surgery. The system has a rigorous architecture and clear logic, mainly comprising a multimodal data acquisition unit, a heterogeneous data fusion and semantic alignment module, an intelligent analysis and predictive decision-making module, a human-computer interaction and visualization module, and a secure storage and management module. These core modules work together to construct a highly intelligent, data-driven patient monitoring and management platform.

[0055] In one specific embodiment, the multimodal data acquisition unit (1) serves as the system's perception layer, responsible for capturing multi-source heterogeneous data streams related to patients, the environment, and medical equipment within the thoracic surgery monitoring room in real time and in parallel. The unit is internally composed of multiple highly specialized sub-units, including a physiological parameter acquisition sub-unit (101), a medical image acquisition sub-unit (102), an environmental parameter acquisition sub-unit (103), a medical equipment operation parameter acquisition sub-unit (104), and a manual input and semi-structured data acquisition sub-unit (105).

[0056] Specifically, the physiological parameter acquisition subunit (101) integrates a high-precision medical-grade sensor array for continuous acquisition of various vital signs of the patient. For example, electrocardiogram (ECG) data is acquired through a twelve-lead ECG sensor array with a sampling frequency set to 2000Hz to ensure the capture of minute changes in cardiac electrophysiological activity. The sensor array uses Ag / AgCl electrodes and a bioelectric amplifier with a high common-mode rejection ratio (CMRR ≥ 100dB) to effectively suppress power frequency interference and motion artifacts. Pulse oxygen saturation (SpO2) and pulse rate monitoring utilizes transmission photoplethysmography (PPG) technology, employing a dual-wavelength LED light source of 660nm (red light) and 940nm (near-infrared light) and a photosensitive detector, with a response time of less than 1.5 seconds. The sensor probe is designed to prevent detachment and resist ambient light interference. The non-invasive blood pressure (NIBP) module is based on the oscillometric principle, using a miniature air pump and a precision pressure sensor to control cuff inflation and deflation. It measures blood pressure from 0 to 300 mmHg with an accuracy within ±2 mmHg and intelligently adapts to different cuff sizes for adults, children, and newborns. The invasive blood pressure (IBP) module uses a high-sensitivity piezoresistive sensor with a measurement range of -50 to 350 mmHg and an accuracy of ±1 mmHg. It is used to accurately monitor arterial pressure, central venous pressure, or pulmonary artery pressure. The sensor is connected to a disposable pressure sensor kit to ensure sterility. The body temperature module is equipped with multiple thermistor sensors or an infrared thermometer array, achieving a measurement accuracy of up to ±0.05℃, enabling multi-point monitoring of core body temperature and surface temperature. Respiratory rate and end-tidal carbon dioxide (EtCO2) are measured using an impedance sensor in conjunction with a bypass infrared CO2 gas analyzer. The infrared analyzer has a fast response time (<100 ms) and high selectivity, and can display respiratory waveforms and EtCO2 values ​​in real time. All of these physiological parameter data are standardized into data streams conforming to the HL7 v2.x or DICOM-RS standard after acquisition, and are timestamped with nanosecond-level high precision by a hardware real-time clock (RTC) module to ensure time synchronization and consistency between different physiological parameters.

[0057] Furthermore, the medical image acquisition subunit (102) achieves deep integration with the hospital's existing Picture Archiving and Communication System (PACS). Its core lies in adhering to the DICOM standard protocol and proactively initiating requests to the PACS server in DICOM Query / Retrieve SCU (Service Class User) mode to acquire the patient's latest chest medical image data in real-time or near real-time, including high-resolution chest X-rays, multi-slice spiral CT images, and bedside ultrasound images. The interface mechanism supports lossless or near-lossless image compression formats such as JPEG 2000 and RLE to optimize data transmission bandwidth and efficiency. This subunit can automatically parse DICOM metadata, extracting key information such as the patient's unique identifier (Patient ID), study instance UID, examination date and time, image modality, slice thickness, pixel spacing, and window level / width. This structured metadata is crucial for subsequent image analysis and lesion localization.

[0058] Specifically, the environmental parameter acquisition subunit (103) constructs a distributed micro-sensor network for comprehensive monitoring of environmental indicators within the monitoring room. Room temperature measurement utilizes a high-precision platinum resistance temperature sensor (Pt1000), with a measurement range of 0℃ to 50℃, an accuracy of ±0.2℃, and a response time of less than 5 seconds. Humidity measurement employs a polymer capacitive humidity sensor, with a measurement range of 0-100%RH and an accuracy of ±2%RH. For air quality monitoring, the PM2.5 concentration sensor, based on the laser scattering principle, can accurately detect particulate matter in the range of 0.3μm to 10μm, with a resolution of 1μg / m³. 3 The maximum range is 1000 μg / m 3The CO2 concentration sensor employs the non-dispersive infrared (NDIR) principle, with a measurement range of 0-5000ppm and an accuracy of ±50ppm, used for real-time assessment of ventilation in the intensive care unit. Noise level monitoring is achieved through a high-precision sound pressure level meter, with a frequency response range of 20Hz to 20kHz and a measurement accuracy of ±0.5dB, effectively identifying continuous noise and transient high-noise events. Light intensity measurement utilizes a silicon photodiode array, whose spectral response range closely matches the human visual curve, with a measurement resolution of 1 lux, used to assess the impact of light on patients' circadian rhythms. The sensor network uses a low-power Wi-Fi Mesh network or ZigBee protocol for data transmission, ensuring real-time data aggregation to the central processing unit. Each sensor node integrates a temperature-compensated crystal oscillator (TCXO) real-time clock (RTC) module, ensuring that the timestamps of all environmental data are synchronized with the system's master clock at the microsecond level.

[0059] Furthermore, the medical equipment operating parameter acquisition subunit (104) acquires real-time parameters of various life support and treatment devices in the monitoring room through a standardized communication interface protocol. For example, for ventilators (such as... The Evita V500 system, via its RS-232C or Ethernet interface, parses the output data frames according to the manufacturer's communication protocol (such as PDMS) to extract key parameters, including but not limited to the current ventilation mode (such as PCV, VCV, PSV), preset tidal volume, actual inhaled / exhaled tidal volume, respiratory rate, positive end-expiratory pressure (PEEP), inhaled oxygen concentration (FiO2), peak airway pressure, plateau pressure, mean airway pressure, and compliance / resistance curves. For infusion pumps and syringe pumps (such as Braun PerfusorCompact plus), the system obtains the current infusion / injection rate (ml / h), total infusion / injection volume, remaining volume, and various alarm statuses (such as tubing blockage, air bubbles, and fluid depletion) via its USB or RS-232 interface. For extracorporeal membrane oxygenation (ECMO) devices, the system obtains the pump speed, blood flow rate, pre- and post-membrane pressure, transmembrane pressure differential, gas flow rate, and oxygen concentration via the Ethernet interface. All collected device operating parameters are converted into data packets in a unified JSON or XML format, with an accurate timestamp to ensure consistency with the patient's physiological data in time, thereby enabling real-time monitoring and evaluation of treatment plans such as respiratory support and fluid management.

[0060] Furthermore, the manual data entry and semi-structured data acquisition subunit (105) provides healthcare professionals with an intuitive and structured data entry interface to capture clinical information that is difficult to obtain directly through automated equipment. The interface supports multiple input methods, including text input (such as wound descriptions and skin conditions), drop-down selections (such as drainage fluid color and characteristics), numerical input (such as precise drainage fluid volume and VAS pain score of 0-10), and image uploads (such as wound photographs and close-ups of abnormal signs). To ensure semantic consistency and standardization of the data, the subunit integrates a pre-defined clinical terminology dictionary, such as a thoracic surgery-related subset of SNOMED CT (Systematized Nomenclature of Medicine—Clinical Terms), forcing healthcare professionals to select standard terms during data entry to avoid ambiguity. For example, when recording drainage fluid, standardized descriptions such as "slightly bloody," "serous," and "chylous" can be selected, and precise milliliters can be entered. When assessing consciousness, a structured entry option for the Glasgow Coma Scale (GCS) score is provided. The sub-unit also has a strict data verification mechanism, such as numerical range verification, date format verification, and mandatory field checks, to ensure the validity and integrity of the entered data. At the same time, it automatically adds the user ID of the person entering the data and the operation timestamp to ensure the traceability of the data.

[0061] In a preferred embodiment of the present invention, the heterogeneous data fusion and semantic alignment module (2), as the core data processing hub of the system, performs in-depth processing on the heterogeneous data from the multimodal data acquisition unit (1) to ultimately generate a unified, high-dimensional patient state feature vector. The module is composed of a data parsing and preprocessing unit (201), a timestamp synchronization and sequence reconstruction unit (202), an ontology and knowledge graph construction unit (203), and a feature-level fusion algorithm (204), achieving refined data processing through a pipelined operation.

[0062] Specifically, the data parsing and preprocessing unit (201) is responsible for the initial cleaning and format conversion of the original heterogeneous data. For physiological signal data, this unit uses a multi-scale denoising algorithm based on wavelet transform, combined with empirical mode decomposition (EMD) or variational mode decomposition (VMD) techniques to remove baseline drift, power frequency interference (50 / 60Hz), and motion artifacts. At the same time, it processes instantaneous missing values ​​in sensor data through algorithms based on Kalman filtering or cubic spline interpolation to ensure the continuity and smoothness of the signal. For medical image data, it performs pixel value normalization (mapping pixel intensity to the range of 0-1 or -1 to 1), image enhancement (such as adaptive histogram equalization, nonlocal mean denoising), and combines deep learning models such as U-Net or Mask R-CNN to automatically extract regions of interest (ROIs), such as lung parenchyma, pleural effusion areas, or lesion areas. For text and semi-structured data, Natural Language Processing (NLP) techniques are applied, including Chinese word segmentation, part-of-speech tagging, named entity recognition (such as identifying drug names, disease diagnoses, and surgical names), and sentiment analysis (such as quantifying patients' self-reported pain levels). Unstructured text is then transformed into structured feature vectors (e.g., through Word2Vec or BERT encoding). All preprocessed data is ultimately converted into a unified numerical or fixed-length vector representation, laying the foundation for subsequent fusion.

[0063] Furthermore, the timestamp synchronization and sequence reconstruction unit (202) undertakes the crucial task of accurately aligning multimodal data in the time dimension. This unit employs a high-precision time synchronization protocol, such as Precise Time Protocol (PTP, IEEE 1588) based on hardware time synchronization or Network Time Protocol (NTP), to ensure that all data acquisition devices (including sensors, imaging equipment, and medical devices) maintain sub-microsecond or even nanosecond synchronization with the time source of the central processing system. For data streams with different sampling frequencies (e.g., ECG 2000Hz, SpO2 100Hz, blood pressure 1Hz), the unit uses adaptive resampling techniques, such as upsampling based on cubic spline interpolation or downsampling based on moving average, to unify them to a common time reference or a specific output frequency (e.g., generating one comprehensive data point per second), to address the problem of inconsistent time resolution. By setting a fixed-length sliding time window (e.g., 5 minutes or 10 minutes), the unit aggregates real-time data streams from different modalities into a time-synchronized multivariate time-series dataset, which comprehensively reflects the patient's overall physiological, environmental, and treatment status at a specific time point or time period.

[0064] Furthermore, the ontology and knowledge graph construction unit (203) is the core mechanism for achieving semantic alignment of heterogeneous data. Based on ontology engineering methods, the unit defines a domain ontology model for the thoracic surgery monitoring field using the Web Ontology Language (OWL). This model defines in detail core concepts (e.g., "patient," "vital signs," "medical images," "medical equipment," "complications," "drugs"), attributes (e.g., "heart rate value," "lung imaging features," "ventilator ventilation mode"), and the complex relationships between them (e.g., "patient" has "vital signs," "medical images" show "lung lesions," "ventilator" acts on "patient" through "ventilation mode"). The ontology model deeply integrates international standard medical terminology, such as SNOMED CT (for clinical concepts) and LOINC (for laboratory tests and observation results), as well as clinical knowledge specific to the thoracic surgery field (e.g., specific signs of different chest diseases, types of common complications). Through knowledge graph technology, such as using RDF (Resource Description Framework) triple storage and SPARQL query language, heterogeneous data from different data sources are mapped into a unified semantic space. For example, "heart rate" data from different brands of monitors, although their internal representations may differ, are all associated with the unique "heart rate" concept in the ontology through ontology mapping, while retaining meta-information such as their original source, measurement method, and unit, thereby completely eliminating the semantic gap caused by data heterogeneity.

[0065] Finally, the feature-level fusion algorithm (204) performs deep and intelligent fusion of the preprocessed and semantically aligned multimodal data at the feature level. The algorithm employs a multi-branch deep neural network architecture, with each branch specifically responsible for processing the raw features of a particular modality or low-level features extracted through modality-specific preprocessing. For example, for physiological signal sequences (such as ECG waveforms and blood pressure trends), one-dimensional convolutional neural networks (1D CNN) and long short-term memory networks (LSTM) are used to capture their temporal features and dynamic patterns; for medical image data (such as CT slice sequences), three-dimensional convolutional neural networks (3D CNN) or Vision Transformer (ViT) are used to extract spatial and high-dimensional visual features; and for structured numerical data (such as laboratory test results and drainage volume), a fully connected network is used. The high-level feature vectors output by these modality-specific feature extractors are then concatenated through a unified concatenation layer. To better learn the dependencies and contribution weights between different modal features, the algorithm further introduces a cross-attention mechanism, allowing the model to dynamically assign importance to different modal features during the fusion process. Furthermore, gated recurrent unit (GRU) layers or Transformer encoder layers can be used to further abstract and integrate the fused sequence features, generating a high-dimensional, highly information-dense joint feature vector. This joint feature vector not only contains information from all modalities but also, through the nonlinear transformation of the deep learning model, can capture complex interactions and potential correlation patterns between modalities, thus comprehensively and accurately characterizing the patient's current physiological and pathological state and clinical context.

[0066] In a preferred embodiment of the present invention, the intelligent analysis and prediction decision module (3) is the intelligent core of the system. It receives the patient state feature vector generated by the heterogeneous data fusion and semantic alignment module (2) and uses various artificial intelligence algorithms to perform pattern recognition, anomaly detection, risk assessment and trend prediction on the feature vector. The module is composed of a physiological state anomaly identification and early warning submodule (301), an imaging progress intelligent assessment submodule (302), a comprehensive risk assessment and complication prediction submodule (303), a rehabilitation trend prediction and intervention suggestion submodule (304), and a dynamic resource optimization suggestion submodule (305).

[0067] Specifically, the physiological state abnormality identification and early warning submodule (301) is dedicated to real-time monitoring of the dynamic changes in the patient's physiological parameters and accurately identifying potential abnormal patterns. The core of this submodule is a deep learning-based time-series analysis model, such as a gated recurrent unit (GRU) enhanced with a bidirectional long short-term memory network (Bi-LSTM) or an attention mechanism, to model continuous physiological time-series data. This model is trained on a large amount of historical normal physiological state data (e.g., from MIMIC-IV of over 10,000 post-thoracic surgery patients or a local hospital database), learning the range of physiological parameter fluctuations, time dependence, and cross-parameter correlations under healthy and stable conditions. Through online real-time inference, the model can detect abnormal changes deviating from the normal baseline, such as arrhythmias (e.g., new onset or worsening of ventricular premature beats, atrial fibrillation), sudden drops in blood pressure (e.g., systolic blood pressure consistently below 90 mmHg with increased heart rate), persistent oxygen saturation (SpO2) below 90%, abnormal breathing patterns (e.g., Cheyne-Stokes breathing, sighing breathing, or apnea), and abnormal fluctuations in end-tidal carbon dioxide (EtCO2). The model can identify subtle, non-obvious early abnormal patterns and generate different levels of warning information based on preset clinical thresholds and the degree of abnormality (e.g., based on duration and degree of deviation), ranging from low-priority trend alerts to high-priority emergency alarms.

[0068] Furthermore, the intelligent assessment submodule for imaging progression (302) utilizes advanced computer vision technology to automatically analyze the patient's chest medical images to assess changes in lesions and the progression of complications. This submodule employs a three-dimensional convolutional neural network (3D CNN), such as 3D U-Net or V-Net, to segment the lungs in chest CT images, identify lesions (such as pleural effusion, pulmonary consolidation, and pulmonary nodules), and perform precise quantification. For pleural effusion, the model can automatically segment the effusion area and calculate its precise volume (ml) change, as well as its accumulation or reduction rate over time (e.g., 24 hours, 48 ​​hours). For pulmonary infection or consolidation, the model can identify the extent, density (Hounsfield units), and distribution (e.g., lobular distribution) of lesions, and assess their infiltration, absorption, or progression trends. For chest X-rays, the model uses a Vision Transformer or ResNet-based detector to assist in identifying complications such as pneumothorax, atelectasis, and pleural thickening. The model can automatically compare image sequences from different time points (e.g., preoperative, postoperative day 1, and postoperative day 3) to identify and quantify subtle progressions or improvements, such as the volume of new pneumothorax or the degree of lung re-expansion after drainage. When imaging abnormalities or progression are detected, the submodule can provide auxiliary diagnostic suggestions and risk assessments based on their severity.

[0069] The comprehensive risk assessment and complication prediction submodule (303) is one of the core intelligent decision support functions of this invention. Its purpose is to dynamically and prospectively predict the risk of a patient developing specific complications within a specific time window (e.g., 24 hours, 48 ​​hours, 72 hours) based on all fused multimodal features. This submodule employs a machine learning model within an ensemble learning framework, specifically a Stacking Ensemble model. This model uses the prediction results of various basic learners (such as Gradient Boosting Trees, XGBoost, LightGBM; Support Vector Machines (SVM); and Deep Neural Networks (DNN)) as input to a meta-learner (such as logistic regression or random forest), thereby combining the advantages of different models to improve the robustness and accuracy of the prediction. The model uses features from all modalities, including physiological parameters, medical imaging features, environmental parameters, medical equipment operating status, and manually entered structured and semi-structured data, as input.

[0070] The core algorithm of the prediction model can be expressed as:

[0071] Let the fused patient state feature vector be x = [x1, x2, ..., xn]. n ], where x1 represents a feature extracted from multimodal data.

[0072] For each complication C j (For example, acute respiratory distress syndrome (ARDS), lung infection, heart failure, coagulation disorders, sepsis, renal insufficiency), this invention trains a binary classification prediction model, such as a logistic regression model, whose prediction probability P(C j The formula for calculating (1 / x) is:

[0073]

[0074] Among them, w j Is it related to complication C j The relevant feature weight vector, b j These are bias terms. These parameters are obtained by training the model on a large amount of historical patient data and are designed to maximize predictive accuracy.

[0075] Furthermore, to combine the correlation between different complications with the physician's attention to specific complications, a comprehensive risk score R can be introduced. total

[0076]

[0077] Where M is the total number of complications (in this embodiment, M = 6, corresponding to the six complications mentioned above), α j Complication C j Clinical weighting (determined by a multidisciplinary panel of experts based on the incidence of complications, mortality, and impact on patient prognosis; for example, the α value for ARDS is higher than that for ordinary lung infections), S j Complication C j The severity score (usually an integer from 1 to 5, representing the degree of impact on the patient's health) provides healthcare professionals with a comprehensive overview of the patient's risk and triggers customized alerts based on the type of high-risk complication. For example, in cases of increased ARDS risk, the system may recommend immediate adjustments to lung-protective ventilation strategies or preparation for prone positioning ventilation.

[0078] In one specific embodiment, a 65-year-old patient admitted to the intensive care unit after thymoma resection was monitored.

[0079] Example (System of the Invention):

[0080] The patient, a 65-year-old male, was admitted to the intensive care unit 24 hours after thymoma resection.

[0081] The multimodal data acquisition unit acquires data in real time:

[0082] Physiological parameters: heart rate 85 bpm, blood pressure 125 / 78 mmHg, SpO2 96%, EtCO2 40 mmHg, respiratory rate 18 breaths / min. However, the system detected occasional premature ventricular contractions (PVCs) on the ECG signal, with a frequency of 2 PVCs / min. In the past 2 hours, SpO2 decreased from 98% to 96%, respiratory rate gradually increased from 16 breaths / min to 18 breaths / min, and tidal volume decreased slightly.

[0083] Imaging data: Postoperative chest X-ray showed mild atelectasis in the left lower lung. The system, through the intelligent assessment submodule of imaging progress, compared the chest X-ray taken 24 hours prior and identified a 15% increase in the area of ​​the atelectasis region. At the same time, AI detected a small amount of new pleural effusion in the left pleural cavity, with a volume of approximately 50 ml.

[0084] Equipment data: Ventilator ventilation mode is PSV, PEEP 5cmH2O, FiO2 35%. The system detected a 10% deviation between the actual tidal volume and the preset tidal volume.

[0085] Manually entered data: The drainage fluid was serum sample, with a volume of 120ml / 4 hours, but the patient's VAS pain score was 7, which was higher than the target value (<4), indicating insufficient analgesia.

[0086] The heterogeneous data fusion and semantic alignment module fuses the above data to generate a patient state feature vector.

[0087] The comprehensive risk assessment and complication prediction submodule of the intelligent analysis and predictive decision module immediately runs. Based on model calculations, the system predicts that the patient has a 15% probability of developing acute respiratory distress syndrome (ARDS) within the next 24 hours (high-risk threshold is 10%), an 8% probability of lung infection, and a 5% probability of heart failure. The main contributing factors to the increased risk of ARDS are: progression of atelectasis, a trend of decreasing SpO2, decreased tidal volume, and occasional premature ventricular contractions.

[0088] Based on this, the human-computer interaction and visualization module immediately issued a high-level warning (visual flashing and mid-frequency alarm sound), indicating that "the patient has a high risk of ARDS; it is recommended to repeat bedside ultrasound to assess the progression of pleural effusion, and consider adjusting the ventilation strategy, optimizing analgesia, and encouraging deep breathing and coughing." Following the warning and recommendations, medical staff promptly repeated the ultrasound, confirmed the increase in pleural effusion, adjusted the analgesia plan, and strengthened airway management. Over the next 48 hours, the patient's respiratory function gradually improved, and the condition did not develop into ARDS.

[0089] Comparative example (traditional monitoring system):

[0090] The same patient was monitored using a traditional monitoring system.

[0091] Traditional monitoring systems typically rely on threshold alarms based on a single physiological parameter.

[0092] Physiological parameters: Heart rate, blood pressure, SpO2, and respiratory rate are all within the "normal range" (e.g., SpO2 > 90%). Occasional premature ventricular contractions may be recorded simply as part of routine ECG monitoring, but their association with deteriorating respiratory function is not identified. Decreased tidal volume and progression of atelectasis do not directly trigger alarms because no trend and multimodal correlation analysis was performed.

[0093] Imaging data: Postoperative chest X-rays are routinely reviewed by radiologists, but due to the workload of doctors, comparative analysis may be delayed. Small amounts of newly developed effusion may not be identified as potential risks in the first instance, or may not be correlated with physiological data.

[0094] Equipment data: The ventilator parameters are displayed on the traditional monitor interface, but the warning mechanism for the deviation from the actual tidal volume is relatively simple and may not be detected in time.

[0095] Manually entered data: Pain scores are recorded by nurses, but there is a lack of systematic correlation analysis with other data to assess their impact on respiratory function or rehabilitation.

[0096] In traditional systems, due to the lack of multimodal data fusion and deep intelligent analysis capabilities, the system does not proactively identify early increases in a patient's risk of ARDS. Doctors and nurses may only discover the problem when the patient experiences significant shortness of breath, a persistent SpO2 below 90%, or a chest X-ray showing significant worsening of lung lesions (at which point the condition may have progressed to the middle or late stages). This can lead to delayed intervention and increase the patient's risk of developing severe ARDS.

[0097] The following table compares the differences between the system of this invention and the traditional monitoring system in the identification and intervention of ARDS risk in this patient:

[0098]

[0099]

[0100] The rehabilitation trend prediction and intervention suggestion submodule (304) utilizes a multivariate time series prediction model, such as a Transformer model based on the attention mechanism or a multivariate gated recurrent unit (GRU) network, to predict the short-term and medium-term trends of key rehabilitation indicators for patients. These key rehabilitation indicators include, but are not limited to, changes in the amount and characteristics of pleural drainage fluid, lung function recovery indicators (such as vital capacity and maximum expiratory flow rate), pain score (VAS), time to ambulation, and nutritional intake. By learning the dynamic trajectories and interrelationships of patient indicators from a large amount of historical rehabilitation data, the model can predict the trends of these indicators over the next 24 to 72 hours. Based on the predicted rehabilitation trajectory and a pre-defined post-thoracic surgery rehabilitation pathway model (e.g., a standardized rehabilitation process for lobectomy, pneumonectomy, or thoracoscopic surgery), the submodule can identify deviations in the patient's rehabilitation process (e.g., persistently higher-than-expected drainage fluid volume, stagnation in lung function recovery) and, in conjunction with clinical guidelines and historical treatment experience, provide targeted and personalized intervention suggestions. These recommendations cover adjusting drainage strategies (such as recommendations on the timing of extubation), optimizing respiratory rehabilitation training programs (such as recommending increasing nebulization frequency and adjusting postural drainage), adjusting analgesia programs (such as recommending assessing analgesia pump dosage and increasing oral analgesics), or providing early mobilization guidance, thereby ensuring the smooth progress of the patient's rehabilitation process.

[0101] Furthermore, the dynamic resource optimization suggestion submodule (305) aims to dynamically optimize the allocation of medical resources in the intensive care unit (ICU) based on the patient's current status and predicted needs. This submodule comprehensively analyzes data such as the severity of the patient's condition (e.g., through APACHE II or SOFA scores), the predicted probability of complication risks, the required level of care, and the real-time utilization rate of medical equipment (e.g., ventilators, ECMO, infusion pumps). Combining ICU bed occupancy, real-time scheduling and skill distribution of medical staff (doctors, nurses), and equipment accessibility, the submodule uses optimization algorithms, such as mixed integer linear programming (MILP) or reinforcement learning-based heuristic algorithms, to generate optimal resource allocation suggestions. For example, when the system predicts that a high-risk patient will require more intensive monitoring in the next few hours, it may suggest deploying nurses or specialists with advanced life support skills to the patient's bedside for focused monitoring; when ICU bed shortages are predicted, it may suggest early transfer of low-risk, well-recovering patients based on patient recovery trend predictions to free up beds. Furthermore, when a medical device (such as a portable ultrasound machine) is about to be released, the system can update its status and recommend it to other patients who may need it, thereby improving the utilization efficiency of medical resources and reducing operating costs.

[0102] In a preferred embodiment of the present invention, the human-computer interaction and visualization module (4) serves as a bridge between the system and medical staff, responsible for presenting the complex analysis results, high-priority early warning information, and structured decision suggestions of the intelligent analysis and predictive decision module (3) to medical staff in an intuitive, efficient, and operable manner. It also supports data querying and feedback by medical staff, forming a closed loop of human-computer interaction. The module internally includes a unified view construction unit (401), an intelligent early warning and prompting unit (402), a decision support information presentation unit (403), and a feedback and learning interface (404).

[0103] Specifically, the unified view construction unit (401) is responsible for integrating raw data and intelligent analysis results from different modalities to construct a unified, multi-dimensional patient monitoring dashboard. The dashboard can display the patient's vital signs trend graph in real time (e.g., dynamic curves of 24-hour ECG, blood pressure, SpO2, and respiratory rate, and can overlay normal ranges or alarm thresholds), medical image comparison graphs (e.g., comparison views of preoperative and postoperative CT images, which can overlay AI-segmented lesion areas), laboratory test results (e.g., tables and trend graphs of blood routine and biochemical indicators), precise drainage status (e.g., real-time volume, cumulative volume, color, and characteristics of pleural drainage fluid), medication records (including drug name, dosage, route of administration, and time), complication risk prediction probability (intuitively displaying the predicted probability and comprehensive risk score of various complications in the form of pie charts, bar charts, or heat maps), and rehabilitation progress curves (e.g., lung capacity recovery trend and pain score decline trend). The unit supports highly customizable layouts and view switching, allowing healthcare professionals to flexibly choose different information display modes based on their clinical needs and concerns. Examples include a detailed panoramic view for a single patient, an overview view of key indicators for multiple patients (for quick rounds), or a list view of patients at risk of specific complications. All data is presented in highly visual formats such as graphs, tables, heatmaps, or 3D renderings, ensuring clear and easy-to-understand information and reducing cognitive load.

[0104] Furthermore, the intelligent early warning and prompting unit (402) generates real-time, graded alarm information based on the warning level and specific content output by the intelligent analysis and prediction decision module (3). The alarm information includes multi-sensory prompts: visual prompts (e.g., color coding changes in specific areas of the screen, such as green for safety, yellow for attention, and red for emergency; flashing alarm values; or pop-up warning boxes), auditory prompts (e.g., playing alarm sounds of different tones, rhythms, and volumes according to the alarm level, with high-frequency rapid alarm sounds used in emergency situations), and message pushes (e.g., sending real-time notifications to responsible medical staff through the hospital information system HIS / LIS, mobile device App, SMS, or dedicated voice communication system). The unit has priority management and intelligent noise reduction functions for alarm information. By filtering or downgrading short-term transient signals, known benign physiological fluctuations, or abnormal events with low repeatability, it avoids invalid or excessive alarm information from interfering with the normal work of medical staff and reduces alarm fatigue. Medical staff can confirm alarm information, mark it as processed, or transfer it for further processing. These operations, along with the processing time and the information of the personnel involved, will be recorded in the system log and used for subsequent system learning and optimization.

[0105] Furthermore, the decision support information presentation unit (403) presents the decision suggestions provided by the intelligent analysis and predictive decision module (3) to medical staff in a structured and interpretable form. These suggestions cover a range of topics, from preventative interventions (e.g., early respiratory rehabilitation training for patients at high risk of ARDS), specific adjustments to abnormal physiological parameters (e.g., fluid resuscitation or vasopressors based on the cause of a sudden drop in blood pressure), clinical interpretation of imaging abnormalities (e.g., suggesting potential causes and further examinations for newly diagnosed small pleural effusions), and personalized rehabilitation plans (e.g., adjusting rehabilitation intensity and frequency based on the patient's ability to get out of bed and pain levels). The unit provides a confidence assessment for each suggestion, for example, indicating "This suggestion has a 92% confidence level," and allows for tracing back to key data points supporting the suggestion (e.g., trends in high-risk physiological parameters, specific imaging findings) and the analytical model used (e.g., key features identified by the XGBoost model), thereby enhancing medical staff's trust in and willingness to adopt the system's recommendations. For complex recommendations or rare clinical situations, the unit can provide links to relevant clinical guidelines, best practice pathways, or the system's built-in medical knowledge base, enabling healthcare professionals to gain a deeper understanding of the medical principles and evidence behind their decisions and assisting them in making final clinical decisions.

[0106] Finally, the feedback and learning interface (404) constructs a closed-loop learning mechanism for the system, allowing medical staff to provide real-time feedback on the system's analysis results, prediction accuracy, and decision-making suggestions. The interface provides a user-friendly interface where medical staff can give positive or negative evaluations of system-identified anomalies (e.g., whether an alarm is a true positive), predicted risks (e.g., the accuracy of a complication prediction), or provided suggestions. They can also selectively input detailed clinical observations, actual interventions, or textual comments on the system's performance. After rigorous anonymization and de-identification, the feedback data automatically serves as new training data for the continuous learning and iterative optimization of the artificial intelligence model in the intelligent analysis and prediction decision-making module. This continuous training mechanism, based on real clinical feedback, enables the system to continuously improve its prediction accuracy and decision support capabilities, achieving system self-adaptation and evolution, and ensuring that its intelligence level continuously adapts to changing clinical needs and medical advancements.

[0107] In a preferred embodiment of the present invention, the secure storage and management module (5) is the cornerstone of the entire system. It is used to securely and efficiently store, retrieve, back up, and manage permissions for all massive multimodal data generated during system operation, thereby ensuring the integrity, confidentiality, availability, and traceability of the data, and complying with various regulatory requirements for medical data security and privacy protection. The module consists of a distributed database system (501), a data encryption and permission control unit (502), and a data backup and recovery mechanism (503).

[0108] Specifically, the distributed database system (501) is designed to store massive amounts of multimodal data at the TB or even PB level. Its architecture employs a hybrid storage strategy to adapt to the storage needs of different data types. Structured data, such as patient basic information, physiological parameters, equipment operating parameters, and laboratory test results, is stored in a high-performance relational database management system (RDBMS), such as a PostgreSQL or MySQL cluster. These databases support ACID transaction characteristics, ensuring data consistency and integrity. Unstructured or semi-structured data, such as medical images (DICOM files), raw physiological waveform data, system logs, and unstructured clinical notes, are stored in a distributed file system (such as HDFS) combined with a NoSQL database (such as MongoDB or Cassana) or an object storage service (such as MinIO). For scenarios requiring high throughput and elastic scaling, a NewSQL database, such as TiDB, can also be used, which combines the SQL interface of a relational database with the horizontal scaling capabilities of NoSQL. This hybrid distributed architecture can support high-concurrency data write and read operations and can elastically scale storage capacity according to the growth of data volume.

[0109] Furthermore, the data encryption and access control unit (502) is the core component for ensuring the security of patient privacy data and sensitive system information. This unit enforces Transport Layer Security (TLS) encryption (e.g., TLS 1.2 or TLS 1.3) on all data transmitted within the system, ensuring that data is not eavesdropped on or tampered with during network transmission. For static data stored in databases and file systems, the unit uses Advanced Encryption Standard (AES-256) for encryption; all data is encrypted before being written to the storage medium and automatically decrypted upon reading. The unit implements a strict Role-Based Access Control (RBAC) mechanism, setting granular data access permissions for different user groups such as medical staff (doctors, nurses), system administrators, data analysts, and researchers. For example, a nurse may only have write access to the nursing records of the patients under their care and read-only access to all patients' vital signs; a doctor may have read and write access to all medical records of the patients under their care; and a researcher may only be able to access anonymized and desensitized data. All data access and operations, including login, query, modification, and deletion, are recorded in detail in an immutable audit log to meet relevant compliance requirements in the healthcare industry (such as GDPR, HIPAA, or China's Personal Information Protection Law).

[0110] Finally, the data backup and recovery mechanism (503) ensures the continuity and availability of system data in the face of various failures. The mechanism supports a strategy combining periodic full backups and real-time incremental backups. Full backups are typically performed weekly or monthly, completely replicating all data to independent storage media; incremental backups are performed hourly or daily, backing up only data that has changed since the last full or incremental backup. All data copies are stored in off-site data centers to prevent data loss due to geographical disasters or single points of failure. In the event of data corruption, primary storage device failure, or system-level disaster, the mechanism can quickly initiate a data recovery process, restoring system data to its most recent valid state by selecting the most recent valid backup version, minimizing service interruption time and the risk of data loss. The mechanism also includes granular data version management capabilities, allowing for regression to a specific historical point in time when necessary to meet the needs of compliance reviews or data analysis.

[0111] In summary, the intelligent management system for thoracic surgery intensive care units proposed in this invention, based on multimodal data fusion, overcomes the inherent limitations of traditional monitoring systems in terms of data dimensionality, real-time performance, depth of intelligent analysis, and decision support capabilities by constructing a highly integrated and technologically advanced system architecture that extends from bottom-level data acquisition to upper-level intelligent decision support. Through comprehensive multimodal data integration, high-precision semantic alignment and time synchronization, and deep learning-driven predictive models, the system enables refined insights into patient status and proactive warnings of potential risks. Its proactive and personalized clinical decision suggestions and dynamic resource optimization capabilities not only significantly improve the intelligence level and patient management efficiency of thoracic surgery intensive care units, but more importantly, through early intervention and precision care, effectively improve perioperative patient safety and postoperative recovery quality. The system's continuous learning mechanism further ensures the continuous evolution of its predictive accuracy and decision support capabilities, enabling it to continuously adapt to ever-changing clinical needs, reflecting the deep integration and innovation of medicine and cutting-edge artificial intelligence technologies.

[0112] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. A smart management system for thoracic surgery intensive care unit based on multimodal data fusion, characterized in that, The system includes: The multimodal data acquisition unit (1) is used to acquire multi-source heterogeneous data in the thoracic surgery monitoring room in real time and in parallel, including patient physiological data, medical imaging data, environmental parameter data, medical equipment operating parameter data, as well as manually entered and semi-structured data. The heterogeneous data fusion and semantic alignment module (2) is connected to the multimodal data acquisition unit (1) and is used to preprocess, standardize, semantically align and deeply fuse the heterogeneous data received by the multimodal data acquisition unit (1) to generate a unified, high-dimensional patient state feature vector. The intelligent analysis and prediction decision module (3) is connected to the heterogeneous data fusion and semantic alignment module (2) and is used to receive the patient state feature vector and use a variety of artificial intelligence algorithms to perform pattern recognition, anomaly detection, risk assessment and trend prediction on the feature vector. The human-computer interaction and visualization module (4), connected to the intelligent analysis and predictive decision-making module (3), is used to present the analysis results, early warning information, and decision suggestions of the intelligent analysis and predictive decision-making module (3) to medical staff in an intuitive and operable manner; and The secure storage and management module (5) is connected to the multimodal data acquisition unit (1), the heterogeneous data fusion and semantic alignment module (2), the intelligent analysis and prediction decision module (3), and the human-computer interaction and visualization presentation module (4). It is responsible for the secure storage, retrieval, backup, and access control of all data generated by the multimodal data acquisition unit (1), the heterogeneous data fusion and semantic alignment module (2), and the intelligent analysis and prediction decision module (3), so as to ensure the integrity and security of the data.

2. The intelligent management system for thoracic surgery monitoring rooms based on multimodal data fusion according to claim 1, characterized in that, The multimodal data acquisition unit (1) includes: The physiological parameter acquisition subunit (101) is used to collect various physiological vital signs data of patients, including electrocardiogram, pulse oxygen saturation, non-invasive blood pressure, invasive blood pressure, body temperature, respiratory rate and end-tidal carbon dioxide. The medical image acquisition subunit (102) is used to acquire the patient's chest medical image data, including chest X-ray, computed tomography images and ultrasound images; The environmental parameter acquisition subunit (103) is used to monitor environmental indicators in the monitoring room, including room temperature, humidity, air quality, noise level and light intensity. The medical equipment operation parameter acquisition subunit (104) is used to acquire the real-time operating status and parameters of various life support and treatment equipment in the intensive care unit, including ventilator settings, infusion pump rate, syringe pump dosage, and extracorporeal membrane oxygenation (ECMO) equipment flow and pressure; and The manual data entry and semi-structured data acquisition subunit (105) is used to collect patient clinical data manually entered by medical staff through the system interface, including: the amount, color and characteristics of drainage fluid, wound condition, pain score, consciousness status assessment, nutritional intake, excretion, medication use records and specific nursing operation records.

3. The intelligent management system for thoracic surgery monitoring rooms based on multimodal data fusion according to claim 2, characterized in that, The physiological parameter acquisition subunit (101) is implemented by integrating a medical-grade sensor, which includes: An ECG sensor array with at least twelve leads and a sampling frequency of no less than 1000Hz, using Ag / AgCl electrodes and a bioelectric amplifier with a high common-mode rejection ratio (CMRR≥100dB); The pulse oximeter sensor based on transmission photoplethysmography has measurement wavelengths including 660nm and 940nm, and a response time of less than 2 seconds. The probe of the pulse oximeter sensor has the ability to prevent detachment and resist ambient light interference. The non-invasive blood pressure module, which uses the oscillometric or Korotkoff sound method, has a measurement range of 0-300 mmHg and a measurement accuracy of ±3 mmHg. It also supports cuff switching for adults, children, and newborns. Invasive blood pressure modules employing piezoresistive or strain gauge sensors have a measurement range of -50 to 300 mmHg and an accuracy of ±2 mmHg, and are used to accurately monitor arterial pressure, central venous pressure, or pulmonary artery pressure. A body temperature module employing a thermistor or infrared sensor array has a measurement accuracy of ±0.1℃ and can achieve multi-point monitoring of core body temperature and body surface temperature; and A respiratory sensor employing impedance or respiratory carbon dioxide gas analysis is used to acquire respiratory rate and end-tidal carbon dioxide value, and the respiratory sensor has a fast response time of less than 100ms. The physiological parameter acquisition subunit (101) outputs all physiological parameter data in HL7 or DICOM-RS standard data format, along with a high-precision timestamp.

4. The intelligent management system for thoracic surgery monitoring room based on multimodal data fusion according to claim 2, characterized in that: The medical image acquisition subunit (102) interfaces with the hospital's Picture Archiving and Communication System (PACS) to acquire the latest medical image data in real time or near real time using the DICOM standard protocol, including chest X-rays, multi-slice spiral CT images, and bedside ultrasound images, and extract relevant image diagnostic reports. The interface interface adopts the DICOM Query / Retrieve service provider (SCP) mode and supports JPEG 2000 and RLE image compression formats. The medical image acquisition subunit (102) can automatically parse DICOM metadata, including patient unique identifier, examination sequence number, examination date and time, image modality, slice thickness, pixel spacing, and window width / window level. The environmental parameter acquisition subunit (103) is composed of a distributed micro-sensor network, which includes: a platinum resistance temperature sensor with an accuracy of ±0.2℃; a polymer capacitive humidity sensor with an accuracy of ±2%RH; and a PM2.5 sensor using the laser scattering principle with a resolution of 1μg / m³. 3 A CO2 sensor based on the non-dispersive infrared (NDIR) principle is used, with an accuracy of ±50ppm; a sound pressure level meter is used for noise measurement, with an accuracy of ±0.5dB; and a silicon photodiode array is used for light intensity measurement, with a resolution of 1 lux. The distributed micro-sensor network transmits data to the data aggregation point in real time via a low-power wireless communication protocol. Each sensor node integrates a real-time clock module with a temperature-compensated crystal oscillator to ensure that the timestamps of all environmental data are synchronized. The low-power wireless communication protocol includes Wi-Fi or ZigBee.

5. The intelligent management system for thoracic surgery monitoring rooms based on multimodal data fusion according to claim 1, characterized in that, The heterogeneous data fusion and semantic alignment module (2) includes: The data parsing and preprocessing unit (201) is used to perform preliminary parsing, format conversion, missing value imputation and noise removal on data from different sources; The timestamp synchronization and sequence reconstruction unit (202) is used to ensure that data from different modalities and different acquisition frequencies are accurately aligned in the time dimension; The ontology and knowledge graph construction unit (203) is used to define semantic relationships between different data modalities and construct a domain ontology model that includes medical concepts, clinical events, and equipment status; and Feature-level fusion algorithm (204) is used to fuse preprocessed and semantically aligned multimodal data at the feature level through deep learning or machine learning methods.

6. The intelligent management system for thoracic surgery monitoring room based on multimodal data fusion according to claim 5, characterized in that: The data parsing and preprocessing unit (201) uses a wavelet transform-based multi-scale denoising algorithm combined with empirical mode decomposition or variational mode decomposition techniques to process physiological signal data, and uses Kalman filtering or cubic spline interpolation algorithms to process instantaneous missing data in sensor data. For medical image data, the data parsing and preprocessing unit (201) performs pixel value normalization and image enhancement, and combines deep learning models such as U-Net or Mask R-CNN to automatically extract regions of interest. For text and semi-structured data, the data parsing and preprocessing unit (201) applies natural language processing methods to transform unstructured text into structured feature vectors. Natural language processing methods include Chinese word segmentation, part-of-speech tagging, named entity recognition, and sentiment analysis. The timestamp synchronization and sequence reconstruction unit (202) adopts a high-precision time synchronization protocol, including Precise Time Protocol (PTP) or Network Time Protocol (NTP), to ensure that all data acquisition devices and the time source of the central processing system are synchronized at the sub-microsecond level. The timestamp synchronization and sequence reconstruction unit (202) adopts adaptive resampling, including upsampling based on cubic spline interpolation or downsampling based on moving average, to unify data with different sampling frequencies to a common time base or a specific output frequency. The timestamp synchronization and sequence reconstruction unit (202) aggregates real-time data streams of different modalities into a time-synchronized multivariate time series dataset by setting a sliding time window of fixed length.

7. The intelligent management system for thoracic surgery monitoring room based on multimodal data fusion according to claim 5, characterized in that: The ontology and knowledge graph construction unit (203) is based on ontology engineering methods and uses Web ontology language to define a domain ontology model for the field of thoracic surgery monitoring. This model integrates international standard medical terminology, including SNOMED CT and LOINC, as well as clinical knowledge specific to the field of thoracic surgery. The ontology and knowledge graph construction unit (203) maps heterogeneous data into a unified semantic space while retaining its original source, measurement method and unit information. The feature-level fusion algorithm (204) adopts a multi-branch deep neural network architecture, with each branch responsible for processing the original features or low-level features of a modality. This includes using a one-dimensional convolutional neural network and a long short-term memory network to process physiological signal sequences, using a three-dimensional convolutional neural network or a Vision Transformer to process medical image data, and using a fully connected network to process structured numerical data. The feature-level fusion algorithm (204) integrates the high-level features extracted from different branches through a unified connection layer, a cross-attention mechanism, or a gated recurrent unit layer to generate a high-dimensional joint feature vector.

8. The intelligent management system for thoracic surgery monitoring rooms based on multimodal data fusion according to claim 1, characterized in that, The intelligent analysis and predictive decision-making module (3) includes: The physiological state abnormality identification and early warning submodule (301) is used to monitor the changes in the patient's physiological parameters in real time and identify potential abnormal patterns. It uses a deep learning-based time series analysis model to detect abnormal changes beyond the normal baseline in real time. The deep learning-based time series analysis model includes a long short-term memory network (LSTM) or a gated recurrent unit (GRU). The intelligent assessment submodule for imaging progress (302) is used to automatically analyze the patient's chest medical images, assess the changes in lesions and the progression of complications. It uses a deep learning model based on a three-dimensional convolutional neural network or Vision Transformer to perform image segmentation, target detection and lesion quantification on CT or X-ray sequence images. The integrated risk assessment and complication prediction submodule (303) is used to dynamically and prospectively predict the risk of patient complications based on all fused multimodal features; The rehabilitation trend prediction and intervention suggestion submodule (304) is used to assess the dynamic trajectory of patient rehabilitation and provide personalized rehabilitation suggestions, utilizing time series prediction models to predict short- and medium-term trends of key rehabilitation indicators for patients; and The dynamic resource optimization suggestion submodule (305) is used to dynamically optimize the allocation of medical resources in the intensive care unit based on the patient's current status and predicted needs. It analyzes data such as the severity of the patient's condition, risk of complications, nursing level requirements and medical equipment utilization rate, and generates resource allocation suggestions through optimization algorithms in combination with intensive care unit resource information.

9. The intelligent management system for thoracic surgery monitoring rooms based on multimodal data fusion according to claim 8, characterized in that, The comprehensive risk assessment and complication prediction submodule (303) adopts a machine learning model under the ensemble learning framework. The machine learning model includes a Stacking Ensemble model. The machine learning model uses the prediction results of multiple basic learners as the input of a meta-learner. The basic learners include Gradient Boosting Trees, Support Vector Machines and Deep Neural Networks. The meta-learner includes Logistic Regression or Random Forest. The comprehensive risk assessment and complication prediction submodule (303) takes features from physiological parameters, medical imaging features, environmental parameters, medical equipment operating status, and manually entered structured and semi-structured data as input to predict the probability of a patient developing a specific complication within a specific time window in the future. The prediction model in the submodule (303) calculates the predicted probability of complications using the following formula: Where x = [x1, x2, ..., x n [] is the fused patient state feature vector, where x1 represents a feature extracted from multimodal data, and C j It is a complication, P(C) j =1|x) is complication C j The predicted probability of occurrence, w j Is it related to complication C j The relevant feature weight vector, b j It is a bias term; The submodule (303) also calculates the comprehensive risk score using the following formula: Where M is the total number of complications, α j Complication C j Clinical weight, S j Complication C j The severity score.

10. The intelligent management system for thoracic surgery monitoring room based on multimodal data fusion according to claim 1, characterized in that: The human-computer interaction and visualization module (4) includes a unified view construction unit (401), an intelligent early warning and prompting unit (402), a decision support information presentation unit (403), and a feedback and learning interface (404); The unified view construction unit (401) is used to integrate data and analysis results from different modalities to construct a unified, multi-dimensional patient monitoring dashboard. The dashboard can display the patient's vital signs trend graph, image comparison graph, laboratory test results, drainage status, medication records, complication risk prediction probability and rehabilitation progress curve in real time, and supports custom layout and view switching. The intelligent early warning and prompting unit (402) is used to generate real-time, graded alarm information based on the early warning level and content output by the intelligent analysis and prediction decision module (3). The alarm information includes visual prompts, auditory prompts and message pushes, and has alarm information priority management and intelligent noise reduction functions. The secure storage and management module (5) includes a distributed database system (501), a data encryption and access control unit (502), and a data backup and recovery mechanism (503); The distributed database system (501) is used to store massive amounts of multimodal data. Its architecture adopts a hybrid storage strategy, with structured data stored in a relational database and unstructured data stored in a distributed file system, NoSQL database, or object storage service. The data encryption and access control unit (502) encrypts all data in transmission using transport layer security protocols, statically encrypts data stored in the database using advanced encryption standards, and implements a role-based access control mechanism to set fine-grained data access permissions for different user groups. All data access and operations are recorded in detail in the audit log. as well as The data backup and recovery mechanism (503) supports periodic full backup and real-time incremental backup, stores data copies in a remote data center, and can quickly start the data recovery process in the event of data corruption or system failure. The data backup and recovery mechanism (503) also includes a data version management function.

Images