Preoperative and postoperative full-process optimization nursing system for thoracic surgery department

Abstract

The invention discloses a preoperative and postoperative full-process optimization nursing system for the thoracic surgery department, relates to the technical field of medical nursing, and aims to solve the problem that physical execution nursing and information management decision are separated in the prior art; and the nursing intervention is lack of real-time response and closed-loop regulation and control capability to the highly dynamic physiological state of the thoracic surgery patient in the perioperative period. The nursing system is characterized by comprising a multi-mode physiological and environmental perception layer, a dynamic nursing decision engine and a closed-loop physical intervention execution layer. Wherein the sensing layer is responsible for data acquisition, state model construction, risk prediction and optimal intervention planning by a decision engine, and the execution layer accurately implements intervention and feeds back. By the adoption of the technical scheme, individualized, dynamic and prospective management of perioperative nursing of the patient in the thoracic surgery department can be achieved, a refined, individualized and intelligent management means is provided, and the clinical value is remarkable.

Description

Technical Field

[0001] The present invention relates to the field of medical care technology, and in particular to a full-process optimized nursing system for pre- and post-operative thoracic surgery. Background Art

[0002] In modern clinical medical practice, thoracic surgery places extremely stringent demands on the quality of perioperative care due to its high technical complexity, high traumatic stress, and high risk of postoperative complications. A successful thoracic surgery treatment plan not only relies on the accuracy of the surgical operation itself, but also depends to a large extent on the adequacy of preoperative preparation and the refinement of postoperative rehabilitation management. Therefore, building a nursing system that can cover the entire process from admission assessment to discharge rehabilitation and provide efficient, accurate, and personalized support plays a vital role in improving patient prognosis, shortening hospitalization periods, and optimizing the allocation of medical resources. Against this background, a modern nursing system that integrates automation and information technology has emerged, and to a certain extent has improved the efficiency and standardization of nursing work.

[0003] In the existing technology, automation and information solutions for clinical nursing needs have mainly developed along two different technical paths. One is automated nursing equipment with physical execution as the core. For example, the Chinese patent document with publication number CN112716714B discloses a nursing robot system, which realizes the automated transfer, position change and even cleaning and bathing of patients in different physical spaces through the coordinated action of precise mechanical arms and multi-functional cabins (such as nursing beds, wheelchairs, and sanitary cabins). From a technical principle point of view, this solution is based on solving the most labor-intensive and physically risky links in nursing work. Its core contribution is to introduce mechatronics technology into physical nursing scenarios, replace heavy manual operations with standardized mechanical processes, significantly improve the safety and consistency of operations, and effectively reduce the risk of occupational injuries to nursing staff.

[0004] Another technical path is an auxiliary nursing system with information management as the core. For example, the Chinese patent document with publication number CN103690284B discloses an ICU auxiliary nursing system. The system constructs a local area network with a nurse station server as the center and bedside terminals as nodes, and realizes data docking with the hospital information system (HIS), so as to present standardized nursing procedures, medical order information, etc. to nursing staff in a digital and structured manner. Its original design intention was to solve the problems of complicated clinical information and error-prone nursing processes. It realizes standardized guidance and process management of nursing tasks through information technology, effectively avoids nursing errors caused by information asymmetry or human negligence, and plays an important role in improving the standardization and quality of nursing work.

[0005] However, as clinical medicine places unprecedented demands on the precision and dynamic responsiveness of nursing interventions, particularly in high-risk, complex perioperative care such as thoracic surgery, a deep-seated contradiction between the two aforementioned technical approaches has gradually emerged at the system architecture level. Specifically, there is a significant disconnect between automation at the physical execution level and standardization at the information management level. The aforementioned nursing robotic system is essentially an open-loop physical execution unit. While it can accurately execute preset instructions, it lacks the ability to perceive and provide feedback on the patient's real-time physiological state. For example, it cannot dynamically adjust its movement trajectory or pause operations based on immediate changes in physiological parameters (such as heart rate or blood oxygen saturation) caused by pain or dyspnea during transfer. Conversely, the aforementioned ICU-assisted nursing system, while achieving standardized information transmission, has a one-way and relatively static information flow. The system generates nursing tasks based on medical orders stored in the HIS, but is unable to deeply integrate and analyze information such as the patient's real-time physiological data, subjective experience, and dynamic recovery trends during the execution of nursing tasks to optimize or dynamically adjust existing nursing standards. This separation of information management from physical execution creates a barrier between "data" and "action."

[0006] For thoracic surgery patients, the perioperative state is highly dynamic and varies significantly from person to person. Preoperative pulmonary function and psychological status, postoperative pain levels, respiratory patterns, expectoration ability, and potential risks of complications such as atelectasis and infection constitute a complex, multidimensional information space. An effective nursing system must be able to capture the dynamic changes within this space in real time and provide immediate, precise intervention responses. Existing technical solutions are inherently limited in that they either remain at the physical level of "using machines to replace hands and feet" or the information level of "using screens to replace pen and paper," failing to establish an integrated nursing logic encompassing multidimensional data perception, intelligent analysis and decision-making, and closed-loop feedback control. This architectural gap prevents the system from truly providing continuous, dynamic, and personalized support for the interconnected care needs of thoracic surgery patients, from preoperative pulmonary function training and psychological counseling to postoperative positioning management, respiratory rehabilitation, and pain intervention. As a result, nursing decisions remain heavily reliant on the experience of medical staff and limited rounds of observation, making it difficult to implement proactive risk warnings and optimal intervention pathway planning based on massive amounts of real-time data.

[0007] Therefore, how to design a system that can deeply integrate real-time physiological data perception, multi-dimensional information intelligent analysis and closed-loop nursing execution and control to break the barriers between physical nursing and information management, and realize dynamic optimization of the entire process from data collection to decision support to precise intervention, so as to provide thoracic surgery patients with truly personalized, forward-looking and integrated high-quality care, has become a key challenge currently faced by technical personnel in this field and a technical problem that needs to be solved urgently. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to overcome the inherent defect of the existing technology that physical execution of nursing care is separated from information management decision-making. This defect leads to the lack of real-time response and closed-loop control capabilities of nursing intervention to the highly dynamic physiological state of thoracic surgery patients during the perioperative period. In view of this, the purpose of the present invention is to provide a full-process optimization nursing system for preoperative and postoperative thoracic surgery. The system deeply integrates multimodal physiological perception, model-based predictive decision-making and closed-loop physical intervention, and constructs an integrated technical system from real-time data to precise action to achieve individualized, dynamic and forward-looking management of perioperative nursing for thoracic surgery patients.

[0009] To achieve the above-mentioned purpose of the invention, the present invention provides a full-process optimized nursing system for pre- and post-operative thoracic surgery, which includes: a multimodal physiological and environmental perception layer, a dynamic nursing decision engine, and a closed-loop physical intervention execution layer.

[0010] The multimodal physiological and environmental perception layer is configured to continuously collect multi-dimensional data related to the patient's perioperative status in real time and perform standardized preprocessing on the data. The multimodal physiological and environmental perception layer includes a non-contact vital sign monitoring module, a wearable multi-parameter physiological sensing module, a patient behavior and environmental interaction monitoring module, and a data preprocessing and fusion gateway.

[0011] Specifically, the non-contact vital signs monitoring module is configured to collect the patient's respiratory and body temperature information without disturbance. This module includes a millimeter-wave radar sensor and a long-wave infrared thermal imaging sensor. The millimeter-wave radar sensor uses a frequency-modulated continuous wave (FMCW) system with an operating frequency range of 60 GHz to 64 GHz. It is fixedly mounted on a cantilever bracket 1.5 to 2.0 meters above the patient's bed, with its main beam axis aligned with the center of the patient's thorax. By analyzing the phase changes in the radar echo signal caused by the rise and fall of the thorax, the radar sensor calculates the patient's respiratory rate, respiratory amplitude, and respiratory rhythm stability parameters in real time, with a respiratory rate measurement accuracy of ±0.5 breaths per minute. The long-wave infrared thermal imaging sensor operates in the wavelength range of 8 to 14 microns, has a detector resolution of 320 x 240 pixels, and a temperature resolution better than 50 millikelvin. This sensor, mounted side by side with the millimeter-wave radar sensor, captures the patient's body surface temperature distribution and uses image processing algorithms to identify areas of abnormal body temperature as early screening indicators for potential infection or inflammation.

[0012] Furthermore, the wearable multi-parameter physiological sensing module is configured to be attached to the patient's body surface for obtaining core circulation and peripheral perfusion parameters. The module is embodied as a flexible, breathable, integrated physiological sensing patch with a biocompatible silicone base, which integrates a multi-channel photoplethysmography (PPG) sensing unit, a single-lead electrocardiogram (ECG) acquisition unit, and a three-axis micro-electromechanical system (MEMS) accelerometer. The PPG sensing unit includes a red light emitting diode with a central wavelength of 660 nanometers, an infrared light emitting diode with a central wavelength of 940 nanometers, and a photodiode. By alternately driving the red and infrared light emitting diodes and detecting the reflected light intensity, the unit synchronously measures blood oxygen saturation (SpO2), heart rate (HR), and pulse wave waveform. The ECG acquisition unit includes two Ag / AgCl dry electrodes for collecting lead II ECG signals. The three-axis MEMS accelerometer, with a range of ±4g and a sampling frequency of 100 Hz, is used to monitor the patient's body position, turning movements, and chest vibration caused by coughing. The physiological sensing patch communicates wirelessly with the data preprocessing and fusion gateway via the Bluetooth 5.2 low-power protocol.

[0013] Furthermore, the patient behavior and environment interaction monitoring module is configured to quantify the patient's active rehabilitation behavior and response to intervention instructions. This module includes an array microphone, a mattress-integrated pressure sensor matrix, and a handheld respiratory rehabilitation training terminal. The array microphone consists of four MEMS microphones spaced 10 cm apart and is mounted on the headboard. It uses a beamforming algorithm to directionally pick up sound from the patient area. The audio signals collected by these microphones are used to analyze cough frequency and intensity and distinguish dry coughs from effective expectoration through acoustic characteristics. The mattress-integrated pressure sensor matrix consists of a 16x32 flexible piezoresistive sensor array covering the main area of ​​the mattress, which is used to accurately identify the patient's bed position (supine, side-lying), the time the position is maintained, and the intention to leave the bed. The handheld respiratory rehabilitation training terminal is integrated with a differential pressure flow sensor. Patients use this terminal to perform instructed deep breathing and effective coughing training. The terminal measures the patient's lung function parameters such as forced vital capacity (FVC) and forced expiratory volume in one second (FEV1) in real time.

[0014] The data preprocessing and fusion gateway receives raw data streams from each of the aforementioned sensor modules and performs timestamp alignment, digital filtering, artifact removal, and formatting. For example, it applies a 0.5 Hz to 40 Hz digital bandpass filter to the ECG signal to remove baseline drift and power frequency interference, and separates the gravity component of the accelerometer signal to extract dynamic acceleration. The processed structured data packets are sent to the dynamic care decision engine via encrypted Transmission Control Protocol (TCP).

[0015] The dynamic nursing decision engine is the computing core of the system, which is configured to receive and process multidimensional data streams from the perception layer, build a patient dynamic state model, perform risk prediction, and generate an optimized individualized nursing intervention protocol. The dynamic nursing decision engine runs on a central server with graphics processing unit (GPU) acceleration capability. The hardware specifications of the server include a central processing unit with at least 16 cores, 128 gigabytes of error correction code memory (ECC RAM), and at least two graphics processing units equipped with tensor cores and video memory of not less than 24 gigabytes. The engine includes: a multidimensional patient state vector (PSV) generation module, a complication risk prediction module based on deep learning, and an optimal intervention protocol planning module based on model predictive control (MPC).

[0016] Specifically, the multidimensional patient state vector (PSV) generation module is used to fuse discrete, heterogeneous sensor data into a unified, high-dimensional mathematical representation to comprehensively characterize the patient's physiological and behavioral state at any time t. The generated patient state vector S(t) is defined as:

[0017] S(t)=[s resp (t),s circ (t),s act (t),s pain (t),s pulm (t)] T

[0018] Among them, S resp (t) is the respiratory state subvector, which includes the respiratory frequency and respiratory depth index calculated by the millimeter wave radar, and the degree of respiratory muscle auxiliary force analyzed by the MEMS accelerometer. circ (t)) is a cyclic state subvector, including heart rate, time domain and frequency domain indicators of heart rate variability (HRV), and blood oxygen saturation calculated from PPG and ECG data. act (t) is the activity and body position sub-vector, which includes the body position information identified by the pressure matrix, the frequency and amplitude of turning over monitored by the accelerometer, and the quantitative parameters of cough events (frequency, peak acceleration). pain (t) is an indirect representation subvector of pain and stress state, which is comprehensively evaluated by analyzing the ultra-low frequency component in HRV, sudden changes in respiratory rate, and the subjective pain score (VAS) input by the patient through the interactive interface. pulm (t) is the pulmonary function rehabilitation status subvector, which contains the latest FEV1 / FVC ratio measured by the handheld respiratory rehabilitation training terminal.

[0019] This module uses an extended Kalman filter (EKF) framework to fuse sensor data of different frequencies and noise characteristics to generate a smooth and robust estimate of the state vector S(t).

[0020] Furthermore, the complication risk prediction module based on deep learning is configured to dynamically predict the probability of key postoperative complications (such as atelectasis, lower respiratory tract infection, and pulmonary embolism) occurring within a specific future time window (e.g., 6 hours, 12 hours, or 24 hours) based on the time series of the patient's state vector. The core of this module is a long short-term memory (LSTM) recurrent neural network model. The LSTM network architecture includes three hidden layers, each containing 256 LSTM units, and uses dropout and layer normalization techniques to prevent overfitting. The input of the model is the patient state vector sequence S(t-Δt), S(t-2Δt),..., S(t) in the past hour, and the output is a probability vector P r (t+τ)=[P a (t+τ),P i (t+τ),P e (t+τ)], where τ is the prediction time window. The model is pre-trained on a desensitized dataset containing complete perioperative data from tens of thousands of previous thoracic surgery patients and can be fine-tuned online and personalized using the patient's own data after admission.

[0021] Furthermore, the optimal intervention protocol planning module, based on model predictive control (MPC), is the core of the entire decision-making engine. Its function is to solve a constrained optimization problem to calculate a series of future nursing interventions that maximize the patient's recovery benefits while satisfying safety constraints. This module formalizes the nursing process as a dynamic control problem.

[0022] The control input vector U(t) of the system is defined as:

[0023] U(t)=[θ b (t),θ l (t),P m (x,y,t),f v (t),I v (t),G r (t)] T

[0024] Among them, θ b (t) and θ l (t) is the angle of the backboard and legboard of the intelligent nursing bed; P m (x, y, t) is the pressure distribution function of each airbag unit of the mattress, which is used to achieve active posture guidance and pressure redistribution; f v (t) and Iv (t) is the frequency and intensity of the vibration expectoration aid applied to the patient's back; G r (t) The next set of training targets (such as target inspiratory volume and flow rate curve) set for the respiratory rehabilitation training terminal.

[0025] At the beginning of each control period (e.g., every 15 minutes), this module solves a finite-horizon optimal control problem of the following form to determine the optimal control sequence U(k), U(k+1), ..., U(k+N-1) for the next N time steps:

[0026] Minimize the objective function:

[0027]

[0028] Where S(i) is the future patient state vector predicted by the system dynamic model under the control input U(i). r (i) is the ideal rehabilitation trajectory reference vector set according to the standard clinical pathway and the individual characteristics of the patient. r (i) is the reference control input, which is usually zero or maintains the current state. Q, R, and P are positive definite weight matrices used to balance tracking error, control energy consumption, and terminal state deviation. r is a scalar weight factor used to convert the complication risk P predicted by the LSTM module r (i+τ) is included in the optimization objective, and its value is dynamically adjusted according to the severity of the risk.

[0029] The optimization problem is subject to a series of strict constraints:

[0030] Physical constraints: U min ≤U(i)≤U max , ensure that all control inputs are within the safe operating range of the actuator.

[0031] Physiological safety restraints: S min ≤S(i)≤S max , ensuring that the predicted future physiological state (such as heart rate, blood oxygen) is always within the preset safety threshold. For example, any control sequence that causes the predicted SpO2 to fall below 92% will be rejected.

[0032] Clinical logic constraints: For example, vibration expectoration intervention (f v >0) only when the patient is in an effective drainage position (such as θ b >30°) and no acute pain attack (s pain is allowed only when the

[0033] This module solves the nonlinear MPC problem using numerical optimization algorithms such as the interior-point method or sequential quadratic programming (SQP) to determine the optimal control sequence. Only the first control input, U(k), is then sent to the physical intervention execution layer for execution. In the next control cycle, the system repeats the entire process based on the latest measured values, forming a rolling optimization closed-loop control system.

[0034] The closed-loop physical intervention execution layer is configured to accurately execute the nursing intervention protocol generated by the dynamic nursing decision engine and provide real-time feedback on the execution status to the decision engine. The execution layer includes an intelligent nursing bed unit, an integrated environmental control unit, and an information exchange and education terminal.

[0035] Specifically, the intelligent nursing bed unit is the main carrier of physical intervention. Its bed frame is made of high-strength aluminum alloy, and the drive system uses multiple silent DC linear push rods with built-in position encoders to achieve precise control of the backboard (0-75°), leg board (0-40°) and the height of the entire bed, with an angle control accuracy of ±0.2 degrees. Its mattress consists of 20 independently inflatable and deflated fiber-reinforced polyurethane (TPU) airbags, arranged in a 4x5 matrix. Each airbag is connected to an air circuit control module consisting of a micro air pump and a solenoid valve, which can adjust the pressure according to the pressure function P output by the decision engine. m (x, y, t) The airbag pressure can be independently adjusted within the range of 1-5kPa, achieving a range from conventional pressure alternation to prevent pressure sores, to precise body position guidance for assisting patients to turn to a specific angle. Four low-frequency vibration tactile transducers are symmetrically implanted under the chest and back area of ​​the mattress. Their operating frequency range is 20-150 Hz and they are driven by a Class D power amplifier. They can respond to commands f v (t) and I v (t) Produce mechanical vibrations of specific frequency and intensity to assist in loosening lung secretions.

[0036] The integrated environmental control unit interfaces with the hospital's building automation system (BAS) to adjust the local environment based on the patient's stress level. When the decision engine's PSV analysis indicates a high stress level (e.g., an elevated LF / HF ratio of HRV), the unit can automatically dim the room lights, play pre-set soothing music, or adjust the temperature and humidity of the ventilation system to create a conducive environment for physiological relaxation and recovery.

[0037] The information interaction and education terminal is typically a touchscreen mounted on a bedside cantilever. It not only allows patients to input their subjective feelings (such as VAS pain scores), but its primary function is to visually present the care plan, rehabilitation progress, and training feedback generated by the respiratory rehabilitation training terminal to the patient or nurse. For example, it will display the time and target for the next position adjustment planned by the MPC module, or provide animation guidance on how to cooperate with respiratory training.

[0038] The present invention also provides a method for optimizing the whole process of pre- and post-operative nursing care for thoracic surgery based on the above system, which comprises the following steps:

[0039] Step S1: Data is collected from thoracic surgery patients in the perioperative period through the multimodal physiological and environmental perception layer. The data includes respiratory and body temperature data obtained by the non-contact vital signs monitoring module, heart rate, blood oxygen, electrocardiogram and body movement data obtained by the wearable multi-parameter physiological sensing module, and cough sounds, in-bed body pressure distribution and active lung function training parameters obtained by the patient behavior and environment interaction monitoring module.

[0040] Step S2: In the data preprocessing and fusion gateway, the heterogeneous multi-source data collected in step S1 are time-aligned, filtered, denoised, and structured packaged to form a standardized data stream.

[0041] Step S3: The standardized data stream is transmitted to the dynamic nursing decision engine, and the multi-dimensional patient state vector (PSV) generation module uses the extended Kalman filter algorithm to fuse the multi-source data into a unified high-dimensional vector S(t) that can fully characterize the patient's immediate status.

[0042] Step S4: Input the time series of the patient state vector S(t) into the deep learning-based complication risk prediction module. This module uses a pre-trained LSTM neural network model to calculate and output in real time the probability vector P of the patient developing key complications such as atelectasis and lower respiratory tract infection within the preset time window τ in the future. r (t+τ).

[0043] Step S5: Start the optimal intervention protocol planning module based on model predictive control (MPC). This module is based on the current patient state vector S(t), complication risk probability vector P r (t+τ) and the preset individualized ideal rehabilitation trajectory S r(t), a constrained optimal control problem is constructed and solved to minimize the deviation of the rehabilitation trajectory, control energy consumption, and weighted complication risk. The objective function J and constraints of this problem are described in the system section above. By solving this optimization problem, an optimal nursing intervention control sequence {U(k), U(k+1), ..., U(k+N-1)} covering the next N time steps is generated.

[0044] Step S6: Send the first control instruction U(k) in the optimal control sequence to the closed-loop physical intervention execution layer.

[0045] Step S7: The closed-loop physical intervention execution layer accurately drives its subordinate intelligent nursing bed unit, integrated environmental control unit and other actuators according to the received control instruction U(k) to complete physical intervention actions such as adjusting the bed angle, applying a specific pattern of back vibration, guiding the patient to perform breathing training or adjusting indoor environmental parameters.

[0046] Step S8: While performing the intervention, the multimodal physiological and environmental perception layer continuously monitors the patient's immediate physiological response to the intervention and feeds the new data back to the decision engine. The system returns to step S1, thereby forming a closed-loop nursing process with continuous rolling optimization and dynamic adjustment.

[0047] The beneficial effects of the present invention are:

[0048] The present invention constructs a closed-loop control system of perception-decision-execution, taking the patient's real-time multi-dimensional physiological state as the core driving variable of the system operation. It not only solves the problem of separation between information management and physical execution in the existing technology, but more importantly, by introducing model-based predictive decision-making and optimal control theory, it transforms nursing intervention from a passive, rule-based response mode to an active, forward-looking, intelligent regulation mode aimed at optimizing the patient's long-term recovery outcome. This provides an unprecedented refined, individualized and intelligent management method for the high-risk and high-complexity clinical scenario of perioperative thoracic surgery, which has significant clinical application value and technological advancement significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a system block diagram of a full-process optimized nursing system for pre- and post-operative thoracic surgery according to the present invention.

[0050] The accompanying drawings are numbered as follows: 100, multimodal physiological and environmental perception layer; 110, non-contact vital signs monitoring module; 120, wearable multi-parameter physiological sensing module; 130, patient behavior and environment interaction monitoring module; 140, data preprocessing and fusion gateway; 200, dynamic nursing decision engine; 210, multi-dimensional patient state vector (PSV) generation module; 220, complication risk prediction module based on deep learning; 230, optimal intervention protocol planning module based on model predictive control (MPC); 300, closed-loop physical intervention execution layer; 310, intelligent nursing bed unit; 320, integrated environmental control unit; 330, information interaction and education terminal. DETAILED DESCRIPTION

[0051] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The exemplary embodiments and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0052] See Figure 1 , which shows the system architecture of a full-process optimized nursing system for pre- and post-operative thoracic surgery provided by the present invention. The core design concept of the present invention is to build a closed-loop control technology system integrating perception, decision-making, and execution. This system can use the patient's real-time multi-dimensional physiological and behavioral status as the core driving variable, and through model-based predictive decision-making and optimal control theory, it can achieve active, forward-looking, and optimized regulation of the nursing intervention process, thereby overcoming the fundamental defects of the existing technology in which information management and physical execution are separated and nursing intervention lags behind changes in patient status.

[0053] In a specific example, the system was deployed in the postoperative intensive care unit of the Department of Thoracic Surgery to provide comprehensive and optimized care for a 68-year-old male patient who underwent a right upper lobectomy. The system's deployment and operation are detailed below.

[0054] The system as a whole includes three core layers that are coupled and work together: a multimodal physiological and environmental perception layer 100, a dynamic nursing decision engine 200, and a closed-loop physical intervention execution layer 300.

[0055] First, a detailed engineering description of the multimodal physiological and environmental perception layer 100 is provided. The fundamental task of this perception layer 100 is to construct a comprehensive, high-fidelity digital twin data stream, continuously mapping the patient's dynamic physiological and behavioral states during the perioperative period in real time. To achieve this goal, this perception layer 100 integrates a variety of heterogeneous sensors and performs preliminary data integration through a data preprocessing and fusion gateway 140.

[0056] Specifically, the sensing layer 100 includes a non-contact vital signs monitoring module 110. This module 110 is designed to provide non-interfering, long-term, continuous monitoring of a patient's core vital signs, particularly respiratory status. Physically, this module 110 is a composite sensing unit integrated on a cantilever bracket directly above the bed. Mounted 1.75 meters above the mattress surface, it ensures the sensor's field of view fully covers the patient's primary chest and abdominal areas. This composite sensing unit integrates a millimeter-wave radar sensor and a long-wave infrared thermal imaging sensor. The millimeter-wave radar sensor is the Texas Instruments IWR6843AOP, which utilizes a frequency-modulated continuous wave (FMCW) operating system within the 60 GHz to 64 GHz Industrial, Scientific, and Medical (ISM) band. The radar front end transmits a linear frequency-modulated signal via an antenna-on-package (AoP) integrated into the package (AoP). The signal has a 40 MHz / μs frequency modulation slope and a 50 millisecond frame period. The radar echo signal is processed on-chip, and a two-dimensional fast Fourier transform (FFT) is performed on the intermediate frequency signal to obtain a range-Doppler map. Furthermore, by extracting the phase information located on the patient's chest distance unit, the inverse tangent demodulation algorithm is used to calculate the tiny displacement caused by the chest rise and fall. The time series of the displacement signal directly reflects the respiratory movement. By performing spectral analysis on the respiratory waveform, the respiratory rate can be accurately calculated, and its measurement accuracy has been verified to be ±0.5 times / minute through clinical comparison. At the same time, by analyzing the envelope and peak-to-valley values ​​of the respiratory waveform, the respiratory amplitude index and respiratory rhythm stability coefficient can be obtained to judge changes in the respiratory pattern. The long-wave infrared thermal imaging sensor uses the Lepton 3.5 module of FLIR, whose detector is an uncooled vanadium oxide (VOx) microbolometer, the focal plane array resolution is 320x240 pixels, and the pixel pitch is 12 microns. Its operating wavelength range covers 8 microns to 14 microns, and the thermal sensitivity (NETD) is better than 50 millikelvin. The sensor uses a germanium lens with a field of view of 57 degrees, which can capture a complete thermal map of the patient's upper body. The collected thermal radiation data is processed by non-uniformity correction (NUC) and digital detail enhancement (DDE) algorithms to generate a pseudo-color temperature distribution map. By setting a threshold for the normal body temperature range (for example, 36.5 to 37.2 degrees Celsius) and applying an image segmentation algorithm, it can automatically identify and highlight areas with abnormal body temperature, such as localized skin temperature increases caused by inflammatory responses, providing visual clues for early detection of potential infections.

[0057] Furthermore, the sensing layer 100 also includes a wearable multi-parameter physiological sensing module 120. The design goal of this module 120 is to obtain core parameters reflecting the state of the circulatory system, such as heart rate, blood oxygen saturation, etc. in close contact with the body. In terms of physical form, the module 120 is a disposable flexible physiological sensing patch with a size of 12 cm x 5 cm and a thickness of 3 mm. Its base material is medical-grade platinum vulcanized silicone rubber (such as Wacker Silpuran 2400), which has excellent biocompatibility and breathability, and is attached to the patient's left chest wall with medical-grade pressure-sensitive adhesive. The patch integrates a multi-channel photoplethysmography (PPG) sensor unit, a single-lead electrocardiogram (ECG) acquisition unit, and a three-axis micro-electromechanical system (MEMS) accelerometer through a flexible printed circuit board (FPC). The PPG sensing unit uses Maxim Integrated's MAX86150 biosensor front-end, which integrates a red light-emitting diode (LED) with a central wavelength of 660 nanometers, an infrared light-emitting diode (ID) with a central wavelength of 940 nanometers, and a high-sensitivity photodiode. The red and infrared LEDs are alternately driven using time-division multiplexing, and the reflected light intensity is synchronously collected. After conversion by the built-in 20-bit ADC, arterial oxygen saturation (SpO2) and heart rate (HR) are calculated according to the Lambert-Beer law. The ECG acquisition unit uses Analog Devices' AD8233 ECG front-end chip, connected to two screen-printed Ag / AgCl dry electrodes at each end of the patch to collect lead II ECG signals. Its common-mode rejection ratio (CMRR) is as high as 110 decibels, effectively suppressing motion artifacts and power frequency interference. The three-axis MEMS accelerometer uses the Bosch BMI270, with a range of ±4g and a sampling frequency of 100 Hz. It is used to accurately monitor the patient's body tilt angle, the frequency and amplitude of turning movements, and high-frequency chest vibrations caused by coughing and expectoration. The entire physiological sensing patch is controlled by an onboard Nordic nRF52840 system-on-chip (SoC). The SoC transmits the collected PPG waveforms, ECG waveforms, and three-axis acceleration data in an encrypted manner to the data preprocessing and fusion gateway 140 once per second via the Bluetooth Low Energy 5.2 (BLE 5.2) protocol.

[0058] Furthermore, the perception layer 100 includes a patient behavior and environment interaction monitoring module 130, whose function is to quantify the patient's active rehabilitation behavior and its responsiveness to nursing intervention instructions. This module 130 is composed of multiple sub-devices. One is an array microphone consisting of four Knowles SPH0645LM4H-B MEMS microphones with a spacing of 10 cm, arranged linearly and integrated on the upper edge of the headboard. Through the delay-and-sum (DSS) beamforming algorithm running in the data preprocessing and fusion gateway 140, the microphone array can form a pickup beam directed at the patient's mouth and nose area, effectively suppressing ambient noise. The collected audio signal is digitized at a sampling rate of 48 kHz and 24-bit accuracy, and then analyzed by a pre-trained convolutional neural network (CNN) model to identify cough events and quantify their frequency, duration, and peak sound pressure level. Furthermore, by analyzing the Mel-Frequency Cepstral Coefficient (MFCC) features of cough sounds, the model can accurately distinguish dry, ineffective coughs from effective expectoration accompanied by sputum sounds. The second component is a mattress-integrated pressure sensing matrix embedded within the mattress. This matrix consists of a 16x32 flexible piezoresistive sensor array, using Interlink Electronics' FSR 402 sensors. The physical dimensions cover the 1.8m x 0.8m main support area of ​​the mattress. Using a row-column scanning drive circuit, the system acquires a complete pressure distribution map at a frequency of 5 Hz. By analyzing the centroid position, peak area, and overall shape of the pressure map, a support vector machine (SVM) classifier can accurately identify the patient's bed position (supine, left lateral decubitus, right lateral decubitus, semi-recumbent) and intention to leave the bed. The third component is a handheld respiratory rehabilitation training terminal. This terminal features an ergonomic design and integrates a high-precision differential pressure flow sensor (such as the Honeywell Zephyr HAF series) with a flow measurement range of ±250 standard liters per minute (SLPM). Patients perform deep breathing and forced exhalation exercises using the information interaction and education terminal 330, following the instructions displayed on the terminal. The microcontroller (MCU) within the terminal integrates flow data in real time to derive respiratory capacity, thereby accurately measuring key lung function rehabilitation indicators such as forced vital capacity (FVC), forced expiratory volume in one second (FEV1), and peak expiratory flow (PEF). This data is uploaded to the system wirelessly (e.g., via Wi-Fi).

[0059] Finally, the data streams from the perception layer 100 converge at the data preprocessing and fusion gateway 140. This gateway's hardware entity is an industrial-grade embedded computer deployed in the ward, equipped with an NXP i.MX 8MPlus processor and running a customized Yocto Linux operating system. Its core function is to receive heterogeneous raw data streams from the aforementioned sensor modules 110, 120, and 130 and perform a series of key preprocessing steps. First, all incoming data packets are assigned a unified, high-precision timestamp using the Network Time Protocol (NTP) to achieve precise temporal alignment. Second, digital filtering is applied to various signals. For example, a fourth-order Butterworth digital bandpass filter with a cutoff frequency of 0.5 Hz to 40 Hz is applied to the raw ECG signal to remove baseline drift and high-frequency noise. A low-pass filter is applied to the accelerometer signal to separate the gravity component from the dynamic acceleration component. Finally, an artifact removal algorithm, such as one based on independent component analysis (ICA), is executed to identify and remove motion artifacts from the PPG signal. Finally, all processed data is encapsulated into standardized structured data packets in JSON format and reliably sent to the dynamic nursing decision engine 200 located in the hospital data center via a Transmission Control Protocol (TCP) connection encrypted based on TLS1.3.

[0060] Next, the dynamic nursing decision engine 200 is described in detail. The engine 200 is the "brain" of the entire system and runs on a high-performance central server. The hardware configuration of the server includes two Intel Xeon Gold 6248R processors (each with 24 cores), 256 gigabytes of DDR4 error correction code memory (ECC RAM) and two NVIDIA A100 graphics processors bridged by NVLink (each with 40 gigabytes of HBM2e video memory). The software architecture of the engine 200 includes three core modules: a multidimensional patient state vector (PSV) generation module 210, a deep learning-based complication risk prediction module 220, and an optimal intervention protocol planning module 230 based on model predictive control (MPC) as the decision core.

[0061] Specifically, the function of the multi-dimensional patient state vector (PSV) generation module 210 is to fuse the discrete, multi-rate, noisy sensor data transmitted from the perception layer 100 into a unified, smooth, and robust mathematical representation, namely the patient state vector S(t). This vector comprehensively describes the patient's physiological and behavioral state at any time t. In a preferred embodiment, S(t) is defined as a high-dimensional vector containing 25 elements. This vector is organized into multiple logical sub-vectors: s resp(t) is the respiratory state subvector, including the respiratory rate (times / minute), respiratory depth index (in arbitrary units au), and respiratory rhythm standard deviation calculated by the millimeter wave radar, as well as the cough frequency (times / hour) and cough intensity (peak sound pressure level dB) analyzed by the MEMS accelerometer and microphone. circ (t) is the cyclic state subvector, which includes heart rate (beats / minute), multiple time-domain indices (such as SDNN, RMSSD) and frequency-domain indices (such as LF, HF, LF / HF ratio) of heart rate variability (HRV), and blood oxygen saturation (SpO2%) calculated from PPG and ECG data. act (t) is the activity and position subvector, which includes the current position (coded value) identified by the pressure matrix, the duration (minutes) since the last position change, and the number of turns in the past hour. pain (t) is an indirect representation of the pain and stress state. This vector is generated by integrating multiple information sources: the 0-10 point visual analog scale (VAS) entered by the patient through the information interaction and education terminal 330 as direct input, combined with the LF / HF ratio in HRV (reflecting sympathetic nerve excitability) and the sudden increase in respiratory rate as objective physiological indicators of stress, and a weighted fusion algorithm is used to derive a comprehensive pain stress index. pulm (t) is the pulmonary function rehabilitation status subvector, which includes the most recent FEV1 / FVC ratio (%) measured by the handheld respiratory rehabilitation training terminal.

[0062] To account for the data refresh rates and noise characteristics of different sensors, module 210 employs an extended Kalman filter (EKF) framework. The EKF's state equations predict the next value of the state vector based on a simplified physiological model, while the measurement equations define the relationship between each sensor reading and the state vector elements. By recursively performing predictions and updates, the EKF generates an optimal estimate of the true state vector S(t), effectively smoothing out noise and filling in temporary gaps in sensor data.

[0063] Furthermore, the deep learning-based complication risk prediction module 220 has the task of using historical information on the patient's status to prospectively predict the risk of critical postoperative complications in the future. The core of this module is a long short-term memory (LSTM) recurrent neural network model. The LSTM network architecture has been carefully designed and includes three stacked hidden layers, each layer contains 256 LSTM units, and a Dropout rate of 0.3 and layer normalization (LayerNormalization) are applied after each layer to prevent overfitting and improve the generalization ability of the model. The input of the model is a time series, specifically the patient state vector sequence {S(t-59),...,S(t)} for the past 60 minutes, where each S(t) is the state vector output by the PSV generation module 210 at the tth minute. The output of the model is a three-dimensional probability vector P r (i + τ), corresponding to the predicted probabilities of atelectasis, lower respiratory tract infection, and deep vein thrombosis / pulmonary embolism within a future time window of τ = 12 hours. The model was pre-trained offline on a desensitized dataset containing complete perioperative multimodal data from over 30,000 previous thoracic surgery patients. After admission, the model is fine-tuned through online learning using the patient's own ongoing data, enabling highly personalized risk prediction.

[0064] Furthermore, the Model Predictive Control (MPC)-based Optimal Intervention Protocol Planning Module 230 forms the core of the entire decision-making system. Its function is to solve a strictly constrained, finite-horizon optimal control problem at the beginning of each control decision cycle (e.g., every 15 minutes) to calculate a series of future nursing intervention actions that maximize the patient's long-term recovery benefits. This process transforms nursing intervention from an experience-driven, passive response model to a model-driven, proactive optimization model.

[0065] The module first defines the controllable input vector U(t) of the system, which represents the set of actions that can be performed by the physical intervention execution layer 300. In a specific implementation, U(t) includes: the backboard angle θ of the smart nursing bed b (t) and leg plate angle θ l (t); Target pressure distribution function P of each airbag unit of the mattress m (x, y, t); frequency f of the vibration expectoration aid applied to the patient's back v (t) and intensity I v (t); and the target parameters G for the next set of respiratory rehabilitation training issued to the patient through the information interaction terminal r (t).

[0066] At each decision time k, the MPC module solves an optimization problem of the following form to determine the optimal control sequence U(k), U(k+1), ..., U(k+N-1) for the next N time steps (e.g., N = 24, corresponding to the next 6 hours, with each step length of 15 minutes):

[0067] The optimization objective function J is designed to minimize a comprehensive cost:

[0068]

[0069] Where S(i) is the future patient state vector predicted by a simplified, online identifiable patient physiological dynamic model (in the form of S(i+1)=f(S(i),U(i))). r (i) is the ideal recovery trajectory reference vector pre-set based on the standard clinical pathway and individual patient characteristics (such as age and surgery type). r (i) is the reference control input, which is usually set to maintain the current state to penalize unnecessary control actions. Q, R, P are positive-definite weight matrices that are used to finely balance the error in tracking the ideal trajectory, the energy consumption of the control action, and the state deviation at the end of the prediction period. For example, the weight corresponding to SpO2 in the Q matrix is ​​set very high to ensure stable blood oxygen saturation. r is a dynamically adjusted scalar weight factor that converts the complication risk probability vector P predicted by the LSTM module r The L1 norm of (i+τ) (i.e., the sum of all risk probabilities) is introduced into the optimization objective, so that the controller can actively avoid control strategies that may lead to increased risks.

[0070] This optimization problem is subject to a series of stringent constraints that reflect clinical safety and logistics:

[0071] Physical actuator constraints: e.g., 0°≤θ b (t)≤75°,20Hz≤f v (t)≤60Hz, etc., ensure that all control inputs are within the physical operating range of the actuator.

[0072] Physiological safety hard constraints: For example, the predicted SpO2(i) must always be greater than 92%, and the predicted heart rate HR(i) must be between 50 and 120 beats / min. Any control sequence that violates these hard constraints will be considered an infeasible solution and discarded during the optimization process.

[0073] Clinical logic soft constraints and rules: For example, only when the patient is in an effective drainage position (such as θ b Vibratory expectoration intervention (i.e., allowing f v>0).

[0074] This nonlinear, constrained MPC problem is solved using efficient numerical optimization algorithms, such as the Interior Point Method. After the solution is complete, the optimal control sequence {U*(k), U*(k+1), ...} is obtained, but the system only sends the first control input U*(k) in the sequence to the physical intervention execution layer 300 for execution. In the next control cycle (after 15 minutes), the system obtains the latest patient status measurements and repeats the entire "perception-prediction-optimization" process described above. This rolling optimization approach enables the system to continuously adapt to changes in patient status and external disturbances, achieving true closed-loop adaptive control.

[0075] Finally, the closed-loop physical intervention execution layer 300 is described in detail. The execution layer 300 is the physical implementation carrier of the nursing intervention measures, and its task is to accurately and reliably execute the control instructions U(k) from the dynamic nursing decision engine 200.

[0076] The execution layer 300 includes a core intelligent nursing bed unit 310. The bed frame is welded from high-strength 6061-T6 aluminum alloy profiles, and the drive system uses a variety of silent DC linear actuators from Denmark's LINAK company. Each actuator has a built-in Hall effect sensor as a position encoder, which achieves smooth and precise control of the backboard (0-75°), leg board (0-40°) and the lifting of the entire bed (40-80 cm), with an angle control accuracy of ±0.2 degrees. The mattress part is the technical core of the unit. It consists of 20 independently inflatable and deflated fiber-reinforced polyurethane (TPU) airbag units, arranged in a 4x5 matrix. Each airbag unit is connected to an integrated air path control module through an independent micro air tube. The module contains a micro diaphragm air pump and two three-way solenoid valves. By controlling the air pump speed and accurately controlling the switching time of the solenoid valve through PWM signals, the system can output the target pressure function P according to the decision engine. m (x, y, t), the internal pressure of each airbag is independently and quickly adjusted in the range of 1 to 5 kilopascals (kPa), with a pressure control accuracy of ±0.1kPa. This capability not only enables traditional anti-pressure ulcer pressure alternation (i.e., periodically reducing pressure in different areas), but also enables fine, active posture guidance, such as gently assisting the patient from supine to lateral position by sequentially pressurizing the airbags on one side. In addition, four EX 60R low-frequency vibration tactile transducers from Visaton, Germany, are symmetrically implanted under the chest and back area of ​​the mattress. These transducers are driven by a four-channel Class D power amplifier and can output instructions f according to the decision engine. v (t) and I v(t) generates sinusoidal mechanical vibrations of specific frequency and intensity in the frequency range of 20 to 150 Hz, which are transmitted to the patient's back through the mattress to assist in the loosening and discharge of lung secretions.

[0077] The execution layer 300 also includes an integrated environmental control unit 320. This unit communicates bidirectionally with the hospital's building automation system (BAS) via the standard BACnet / IP protocol. When the dynamic care decision engine 200 analyzes PSV and indicates a patient is in a high-stress state (e.g., if the LF / HF ratio of HRV is consistently above 3.5), it generates corresponding environmental control instructions. Upon receiving these instructions, the unit automatically dims the brightness of the ward's intelligent lighting controller (e.g., the DALI protocol) to 150 lux, plays pre-set, clinically proven relaxation music (e.g., white noise or light music at a specific frequency) through the network speakers, and sends instructions to the VAV (variable air volume) terminal of the air conditioning system to adjust the supply air temperature to 24 degrees Celsius and maintain the humidity at 55%, thereby creating a microenvironment that is both physiologically and psychologically conducive to the patient's rest and recovery.

[0078] Finally, the execution layer 300 includes an information interaction and education terminal 330. This terminal is a 15.6-inch capacitive touchscreen with a resolution of 1920x1080, mounted on an adjustable bedside arm. It serves not only as an interface for patients to input their subjective experiences (such as VAS pain scores and nausea levels), but more importantly, as a window for human-computer interaction, clearly presenting the care plan to the patient or nurse. For example, it will display a countdown message, "In 10 minutes, the bed will automatically adjust to a 30-degree right lateral position to promote left lung drainage," to encourage patient engagement and proactive cooperation. During respiratory rehabilitation training, it uses gamified animations (such as blowing a balloon or raising a sailboat) to guide patients on how to effectively inhale and exhale using the handheld respiratory rehabilitation training terminal. The terminal also displays training results in real time (such as a comparison of the achieved inhalation volume with the target value), providing immediate, positive feedback and significantly improving patient engagement and compliance.

[0079] The present invention also accordingly provides a method for optimizing nursing care for the entire process before and after thoracic surgery based on the above-mentioned system. The method realizes a complete closed-loop nursing process through the collaborative work of the above-mentioned hardware and software. The method includes steps S1 to S8, and its specific implementation has been elaborated in detail in the description of the various components of the system, namely, from multimodal data acquisition (S1), to data preprocessing (S2), to state vector generation (S3), risk prediction (S4), optimal intervention planning (S5), and finally to intervention execution (S6, S7) and the formation of a new closed-loop feedback (S8). This continuously rolling cycle ensures that the nursing measures are always closely coupled with the patient's most subtle physiological changes, and are always guided by the global optimal rehabilitation path.

[0080] Example

[0081] In order to further verify the beneficial effects of the technical solution of the present invention, a specific implementation case is now provided.

[0082] Example 1:

[0083] A 68-year-old male patient was admitted to the hospital for peripheral lung cancer in the right upper lobe and underwent thoracoscopic right upper lobectomy with systematic lymph node dissection. After the surgery, he returned to the intensive care unit and applied the full-process optimized nursing system described in the present invention.

[0084] One hour after surgery, the system began full operation. The multimodal physiological and environmental sensing layer 100 began collecting data: the non-contact monitoring module 110 measured the patient's respiratory rate at 22 breaths / minute, with a regular rhythm; the wearable physiological sensor patch 120 measured a heart rate of 98 beats / minute, SpO2 of 97% (with nasal cannula oxygen at 2 L / min), and an ECG showing sinus tachycardia; the mattress pressure matrix 130 indicated the patient was in a supine position. The PSV generation module 210 of the dynamic care decision engine 200 fused this data into the initial state vector S(t).

[0085] Eight hours after surgery, the LSTM model of the risk prediction module 220 detects a slow downward trend in the respiratory depth index based on the S(t) sequence of the past few hours, and the cough sound captured by the microphone array is identified as a dry, weak cough. The model outputs the probability of atelectasis risk in the next 12 hours, P a (t+12h) increased from baseline 0.15 to 0.45.

[0086] The MPC optimal intervention protocol planning module 230 receives this risk warning. In its next 15-minute decision cycle, the weight w associated with the atelectasis risk in the objective function J is r The optimal control sequence U*(k) calculated by the MPC solver is as follows:

[0087] Intelligent nursing bed unit 310: adjust the backboard angle θ b Slowly increase from 15° to 45° in 2 minutes.

[0088] Mattress airbag unit: Adjust P m (x, y, t), form a support on the patient's left side, and gently guide them to lie on their left side 15°.

[0089] Vibration Expectoration Module: Applies 40 Hz, medium-intensity vibration to the lower right side of the back (corresponding to the surgical area) for 10 minutes.

[0090] Information interaction and education terminal 330: displays animation to guide patients to perform three sets of deep breathing and effective coughing exercises in conjunction with vibration.

[0091] The physical intervention execution layer 300 accurately executed the above instructions. During the intervention, the sensing layer 100 continuously monitored the patient's SpO2, which remained stable at 96-97%, and his heart rate, which rose slightly to 105 beats per minute before falling back. After the intervention, the handheld respiratory rehabilitation training terminal measured the patient's FVC, which improved by 15% compared to before the intervention. Over the following hours, the MPC module continued to plan multiple similar postural adjustments and vibration sputum removal interventions.

[0092] Comparative Example 1:

[0093] A 69-year-old male patient whose demographic characteristics, disease diagnosis and surgical method completely matched those of Example 1 received traditional standard nursing procedures.

[0094] The nurse measured and recorded vital signs every four hours. Eight hours after surgery, during the nurse's ward rounds, the patient recorded respiratory rate of 24 breaths / minute, heart rate of 102 beats / minute, and SpO2 of 95%. The patient complained of pain at the incision site and was reluctant to take deep breaths or cough. The nurse administered pain medication as directed and verbally encouraged the patient to cough and expectorate, but no quantitative monitoring or physical assistance was provided.

[0095] Sixteen hours after the operation, the patient became agitated, with shortness of breath reaching 30 breaths per minute and SpO2 dropping to 90%. A bedside chest X-ray confirmed large areas of right lung atelectasis. Emergency treatments, including high-flow oxygen, bronchoscopic suctioning, and intensive back massage, were initiated. The patient remained in the ICU for an additional 48 hours.

[0096] Data comparison

[0097] The following table shows a quantitative comparison of the key clinical results in Example 1 and Comparative Example 1:

[0098]

[0099]

[0100] The detailed descriptions and data comparisons of the above examples and comparative examples clearly demonstrate that the present invention's optimized pre- and postoperative nursing system for thoracic surgery, through its unique closed-loop perception-decision-execution architecture, significantly enhances the ability to predict and proactively intervene in postoperative complications, transforming the nursing model from a passive, delayed response to a proactive, forward-looking, and optimized management. This system not only effectively prevents serious complications such as atelectasis and improves patients' clinical outcomes, but also significantly reduces monitoring time and nurse workload, possessing significant technological advancement significance and clinical application value.

[0101] The technical solution of the present invention is not limited to the above-mentioned specific embodiments. Any technical variations made according to the technical solution of the present invention fall within the protection scope of the present invention.

Claims

1. A full-process optimization nursing system for pre- and post-operative thoracic surgery, characterized by: The system comprises: a multimodal physiological and environmental sensing layer (100) configured to continuously collect multidimensional physiological and behavioral data related to the patient's perioperative status in real time; a dynamic nursing decision engine (200) configured to receive and process the multi-dimensional data collected by the multimodal physiological and environmental perception layer (100), perform risk prediction and dynamic model construction based on the data, and generate an optimized individualized nursing intervention protocol accordingly; and a closed-loop physical intervention execution layer (300) configured to execute precise physical intervention actions according to the nursing intervention protocol generated by the dynamic nursing decision engine (200); The multimodal physiological and environmental perception layer (100) is further configured to continuously monitor the patient's immediate physiological response to the physical intervention action, and feed back new data containing the physiological response to the dynamic nursing decision engine (200) for rolling optimization in the next decision cycle, thereby forming a closed-loop adaptive control loop of perception-decision-execution.

2. The full-process optimization nursing system for pre- and post-operative thoracic surgery according to claim 1 is characterized in that: The multimodal physiological and environmental perception layer (100) comprises: a non-contact vital sign monitoring module (110) configured to collect the patient's respiratory and body temperature information without disturbance; A wearable multi-parameter physiological sensing module (120) is configured to be attached to the patient's body surface and is used to obtain core circulation and peripheral perfusion parameters; a patient behavior and environment interaction monitoring module (130) configured to quantify the patient's active rehabilitation behavior and response to intervention instructions; and A data preprocessing and fusion gateway (140) is configured to receive raw data streams from the non-contact vital sign monitoring module (110), the wearable multi-parameter physiological sensing module (120) and the patient behavior and environment interaction monitoring module (130), and perform timestamp alignment, digital filtering, artifact removal and formatting and packaging on the raw data streams to form structured data packets and transmit them to the dynamic nursing decision engine (200).

3. The full-process optimization nursing system for pre- and post-operative thoracic surgery according to claim 2 is characterized in that: The non-contact vital signs monitoring module (110) comprises: A millimeter-wave radar sensor, operating in a frequency-modulated continuous wave (FMCW) regime between 60 GHz and 64 GHz, is mounted on a cantilever 1.5 to 2 meters above the bed and aimed at the center of the patient's thorax. It analyzes phase shifts in the radar echo signal caused by chest movement to calculate the patient's respiratory rate, amplitude, and rhythm stability in real time. A long-wave infrared thermal imaging sensor with an operating wavelength range of 8 microns to 14 microns, a detector resolution of 320x240 pixels, and a temperature resolution better than 50 millikelvin is arranged to be installed side by side with the millimeter-wave radar sensor to capture the patient's body surface temperature distribution map and identify abnormal body temperature areas through image processing algorithms as an early screening indicator for potential infection or inflammation.

4. The full-process optimization nursing system for pre- and post-operative thoracic surgery according to claim 2 is characterized in that: The wearable multi-parameter physiological sensing module (120) is embodied as a flexible, breathable, integrated physiological sensing patch with a biocompatible silicone base, wherein the patch internally integrates: a multi-channel photoplethysmography (PPG) sensing unit comprising a red light emitting diode (LED) with a central wavelength of 660 nanometers, an infrared light emitting diode (IRLED) with a central wavelength of 940 nanometers, and a photodiode, configured to synchronously measure a patient's blood oxygen saturation (SpO2), heart rate (HR), and pulse waveform by alternately driving the red and IR LEDs and detecting reflected light intensity; a single-lead electrocardiogram (ECG) acquisition unit comprising two Ag / AgCl dry electrodes configured to acquire a lead II ECG signal; as well as A three-axis microelectromechanical system (MEMS) accelerometer with a range of ±4g and a sampling frequency of 100 Hz was set up to monitor the patient's body angle, turning movements, and chest vibration caused by coughing.

5. The full-process optimization nursing system for pre- and post-operative thoracic surgery according to claim 1 is characterized in that: The dynamic nursing decision engine (200) comprises: a multi-dimensional patient state vector (PSV) generation module (210) configured to fuse the discrete, heterogeneous sensory data from the perception layer (100) into a unified, high-dimensional mathematical representation, namely the patient state vector S(t), to comprehensively characterize the physiological and behavioral state of the patient at any time; a deep learning-based complication risk prediction module (220), configured to dynamically predict the probability of a critical postoperative complication occurring within a specific future time window based on the time series of the patient state vector S(t); and An optimal intervention protocol planning module (230) based on model predictive control (MPC) is configured to calculate a series of future nursing intervention actions that can maximize the patient's recovery benefits while satisfying safety constraints by solving a constrained optimal control problem based on the current patient state vector S(t) and the complication risk probability.

6. The full-process optimization nursing system for pre- and post-operative thoracic surgery according to claim 5 is characterized in that: The multidimensional patient state vector (PSV) generation module (210) is configured to fuse the sensor data using an extended Kalman filter (EKF) framework to generate a smooth, robust estimate of the patient state vector S(t); the patient state vector S(t) is defined as S(t)=[s resp (t),s circ (t),s act (t),s pain (t),s pulm (t)] T Where: s resp (t) is the respiratory state subvector, including respiratory rate, respiratory depth index and respiratory muscle auxiliary effort; s circ (t) is the circulatory state subvector, including heart rate, heart rate variability (HRV) index and blood oxygen saturation; s act (t) is the activity and body position sub-vector, including body position information, turning frequency and amplitude, and cough event quantification parameters; s pain (t) is an indirect representation subvector of pain and stress state, which is generated by comprehensively analyzing HRV index, respiratory rate changes and subjective pain score input by patients; and s pulm (t) is the pulmonary function rehabilitation state subvector, which includes the pulmonary function parameters measured by the handheld respiratory rehabilitation training terminal.

7. The full-process optimization nursing system for pre- and post-operative thoracic surgery according to claim 5 is characterized in that: The core of the deep learning-based complication risk prediction module (220) is a long short-term memory (LSTM) recurrent neural network model; the LSTM model is configured to receive a time series of patient state vectors S(t) within a specific time period in the past as input and output a probability vector P r (t+τ) is the prediction result, the probability vector P r (t+τ) represents the probability that the patient will develop one or more complications, including at least atelectasis, lower respiratory tract infection, and pulmonary embolism, within a preset time window τ in the future; the LSTM model is pre-trained on a desensitized dataset containing complete perioperative data of previous patients, and can be fine-tuned online using the patient's own data after admission.

8. The full-process optimization nursing system for pre- and post-operative thoracic surgery according to claim 1 is characterized in that: The closed-loop physical intervention execution layer (300) includes: An intelligent nursing bed unit (310) is provided as a main carrier of physical intervention, and is used to perform bed posture adjustment, active posture guidance, and vibration expectoration assistance; an integrated environmental control unit (320) configured to interface with the hospital's building automation system to control the lighting, sound, temperature, and humidity of the ward based on the patient's stress status; and An information interaction and education terminal (330) is configured to allow patients to input their subjective feelings and visually present nursing plans, rehabilitation progress, and rehabilitation training guidance and feedback to patients or nurses.

9. The full-process optimization nursing system for pre- and post-operative thoracic surgery according to claim 8 is characterized in that: The intelligent nursing bed unit (310) comprises: Multiple silent DC linear actuators with built-in position encoders are configured to precisely control the backrest angle, legrest angle, and overall bed height. A mattress composed of a plurality of independently inflatable and deflable fiber-reinforced polyurethane (TPU) airbags arranged in a matrix form, wherein the airbags are connected to an air circuit control module composed of a micro air pump and a solenoid valve, which is set to adjust the pressure according to the pressure function P included in the nursing intervention protocol. m (x, y, t), independently adjust the pressure of each airbag within the preset pressure range to achieve pressure alternation to prevent pressure sores and assist the patient in turning to a specific angle of posture guidance; and A plurality of low-frequency vibration tactile transducers symmetrically implanted below the chest and back area of ​​the mattress are configured to respond to the frequency instruction f in the nursing intervention protocol. v (t) and intensity instruction I v (t), generates specific mechanical vibrations in the frequency range of 20-150 Hz to assist in loosening lung secretions.

10. The system according to claim 5, wherein: The optimal intervention protocol planning module (230) based on model predictive control (MPC) is configured to generate the nursing intervention action by solving a finite-time optimal control problem in each control cycle; The control input vector U(t) of the optimal control problem includes at least the bed board angle of the intelligent nursing bed, the pressure distribution function of each airbag unit of the mattress, the frequency and intensity of the vibration expectoration assistance, and the target parameters of the respiratory rehabilitation training; The objective function J(S k ,U k ) is set to minimize a comprehensive cost, which includes the future patient state vector S(i) predicted by the system dynamic model and the ideal rehabilitation trajectory reference vector S in the prediction time domain. r (i), the tracking error between the control input U(i), the energy consumption of the control input U(i), and the weighted complication risk P output by the complication risk prediction module (220) r The sum of (i+τ); The optimal control problem is subject to at least three types of constraints: physical constraints that ensure that the control input is within the safe operating range of the actuator, physiological safety constraints that ensure that the predicted future physiological state indicators are within the preset safety thresholds, and clinical logic constraints that ensure that the intervention measures are consistent with the clinical care logic.

Images