Home-based intelligent follow-up assessment system and method for cardiopulmonary function in post-cardiac surgery patients

Abstract

This invention discloses a home-based intelligent follow-up assessment system and method for cardiopulmonary function in post-cardiac surgery patients, relating to the field of data analysis technology. It includes: an initial dynamic physiological baseline module, used to establish a personalized initial dynamic physiological baseline for each patient based on multimodal physiological time-series data from the patient's stable post-operative period using a Gaussian mixture model and a hidden Markov model; a deviation time-series evolution trajectory module, used to acquire the patient's physiological feature vector in real time, and obtain the deviation index between the actual and initial baseline data states by matching it with the initial dynamic physiological baseline and calculating the Mahalanobis distance, thereby generating a deviation time-series evolution trajectory; and a graded early warning module, used to dynamically time-warp and match the deviation time-series evolution trajectory with a preset clinical pathophysiological pattern library, calculate the posterior probability through a multi-level Bayesian network, and trigger graded early warnings based on the posterior probability. This invention improves the sensitivity of early risk identification.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to a home-based intelligent follow-up assessment system and method for cardiopulmonary function in post-cardiac surgery patients. Background Technology

[0002] In existing technologies, home follow-up for post-cardiac surgery patients largely relies on patients' subjective symptom reporting, regular outpatient check-ups, or simple remote single-parameter monitoring. This lacks dynamic modeling of patients' individual physiological baselines, making it impossible to effectively quantify the subtle evolution of cardiopulmonary function over time. Furthermore, early warning mechanisms are often based on general thresholds for the population, resulting in low specificity and high false alarm rates in identifying early compensatory functional decline and occult clinical deterioration patterns. This makes it difficult to provide medical staff with decision support with clear clinical pathophysiological significance, and suffers from delays in intervention and inefficient resource allocation. Summary of the Invention

[0003] To address the aforementioned technical issues, this invention provides a home-based intelligent follow-up assessment system and method for cardiopulmonary function in post-cardiac surgery patients. This technical solution resolves the problems of difficulty in providing medical staff with decision support that has clear clinical pathophysiological significance, as well as the issues of delayed intervention and insufficient resource allocation efficiency.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A smart home-based follow-up assessment system for cardiopulmonary function in post-cardiac surgery patients, including: Initial dynamic physiological baseline module, deviation time-series evolution trajectory module, graded early warning module; Among them, the deviation time-series evolution trajectory module is electrically connected to the initial dynamic physiological baseline module, and the graded early warning module is electrically connected to the deviation time-series evolution trajectory module; The initial dynamic physiological baseline module is used to establish a personalized initial dynamic physiological baseline for patients based on their postoperative stable period multimodal physiological time series data through Gaussian mixture model and hidden Markov model. The deviation time-series evolution trajectory module is used to acquire the patient's physiological feature vector in real time, and obtain the deviation index of the actual-initial baseline data state by matching the initial dynamic physiological baseline and calculating the Mahalanobis distance, thereby generating the deviation time-series evolution trajectory. The graded early warning module is used to dynamically time-normalize and match the deviation time evolution trajectory with a preset clinical pathophysiological pattern library, calculate the posterior probability through a multi-level Bayesian network, and trigger graded early warning based on the posterior probability.

[0005] Preferably, the initial dynamic physiological baseline module specifically includes: The patient physiological data acquisition unit is based on a medical-grade smart bracelet worn by the patient. The bracelet contains a PPG photoelectric heart rate sensor, a three-axis accelerometer, and a body temperature sensor. The device is calibrated and data is collected for more than 30 minutes at a fixed time every day for 3-7 days after the patient's postoperative discharge in a resting state. Multimodal physiological time-series data of the patient in a stable state are obtained, including the patient's ECG HRV data, blood oxygen saturation and PI data, respiratory RR data, activity intensity and steps, and sleep quality data. The ECG signal is denoised and motion artifacts are removed and unified to the same timestamp. The temporal feature extraction unit calculates the mean, standard deviation, extreme values, and interquartile range of the multimodal physiological time series data under the patient's stable state using a 5-minute sliding time window. It then obtains the linear regression slope and autocorrelation coefficient of the multimodal physiological time series data under the patient's stable state and extracts the temporal features of the multimodal physiological time series data under the patient's stable state. The frequency domain feature extraction unit uses Fast Fourier Transform to convert the time domain features of the patient's multimodal physiological time series data under steady-state conditions to the frequency domain, calculates its power spectral density, main frequency component, frequency band energy ratio and spectral entropy, and extracts the frequency domain features of the patient's multimodal physiological time series data under steady-state conditions.

[0006] Preferably, the initial dynamic physiological baseline module also includes: The initial physiological pattern unit integrates the temporal and frequency domain features of the multimodal physiological time series data of the patient in a stable state, performs normalization processing, and splices them to obtain the comprehensive feature vector of the multimodal physiological time series data of the patient in a stable state, and extracts the initial physiological pattern of the patient in the early stage of rehabilitation. The mean vector and covariance matrix establishment unit takes the comprehensive feature vector of multimodal physiological time series data under the patient's stable state as input, and uses the Bayesian information criterion to automatically select the optimal component of the comprehensive feature vector of multimodal physiological time series data under the patient's stable state, and randomly initializes the mean vector and covariance matrix of K optimal components of the comprehensive feature vector of multimodal physiological time series data under the patient's stable state. The static physiological feature baseline distribution unit is iteratively optimized using the expectation-maximization algorithm to calculate the log-likelihood function of the comprehensive feature vector of the multimodal physiological time series data under the patient's steady state. Its mean vector and covariance matrix are continuously updated until the log-likelihood function converges, thereby obtaining the Gaussian mixture function model parameters of the patient's multimodal physiological time series data. The probability density distribution of the patient's multimodal physiological time series data in the multidimensional space under the patient's steady state is determined as the patient's static physiological feature baseline distribution.

[0007] Preferably, the initial dynamic physiological baseline module also includes: The temporal evolution law unit introduces a hidden Markov model. The comprehensive feature vector of multimodal physiological temporal data under the steady state of the patient is used as the observation sequence. The number of hidden states of the model is set, and the parameters of the hidden Markov model are randomly initialized, including the initial state probability distribution, the state transition probability matrix and the probability density distribution of multimodal physiological temporal data under the steady state of the patient in the multidimensional space. The Baum-Welch algorithm is used to estimate the maximum likelihood of the model parameters. The model is iteratively optimized until convergence, and a hidden Markov model of multimodal physiological temporal data under the steady state of the patient is established to capture the temporal evolution law of its physiological state. The physiological baseline establishment unit establishes a personalized initial dynamic physiological baseline for each patient based on the patient's static physiological characteristic baseline distribution and the temporal evolution of physiological state.

[0008] Preferably, the deviation time-series evolution trajectory module specifically includes: The real-time data acquisition unit sets a fixed sliding time window of 5 minutes, monitors the patient's physiological data in real time within the time window, performs normalization processing, and extracts the patient's physiological feature vector within the time window. The matching baseline calculation unit compares the patient's physiological feature vector within the time window with the patient's personalized initial dynamic physiological baseline. It calculates the Euclidean distance between the patient's physiological feature vector within the current time window and each baseline feature vector in the patient's personalized initial dynamic physiological baseline. The baseline with the shortest Euclidean distance is selected as the matching baseline. The mean vector and covariance matrix of the patient's multimodal physiological time series data under the steady state on the baseline with the shortest Euclidean distance are calculated to obtain its baseline Gaussian distribution parameters. The deviation index calculation unit calculates the Mahalanobis distance between the patient's physiological feature vector and the matching baseline within the current time window, which serves as the deviation index of the patient's actual physiological state from the initial baseline data.

[0009] Preferably, the deviation time-series evolution trajectory module also includes: The short-term fluctuation extraction unit uses time as the horizontal axis and the deviation index of the patient's actual physiological state from the initial baseline data as the vertical axis to generate a continuous time series of deviation index. Using the exponential weighted moving average method, it calculates the deviation trend of the patient's actual physiological state from the initial baseline data within 1 hour and captures its short-term fluctuations. The long-term trend identification unit uses the autoregressive integral moving average method to calculate the deviation trend of the patient's actual physiological state from the initial baseline data within 24 hours, and identifies its long-term trend and periodic changes.

[0010] Preferably, the deviation time-series evolution trajectory module also includes: The deviation time-series evolution trajectory unit calculates the average deviation of the patient's actual physiological state from the initial baseline data within a time window, using it as the target value. It calculates the difference between the deviation index of the patient's actual physiological state from the initial baseline data and the target value at each moment, and performs cumulative calculation to obtain the cumulative deviation value of the patient's actual physiological state from the initial baseline data. A decision threshold b is set. If the cumulative deviation value of the patient's actual physiological state from the initial baseline data exceeds the decision threshold b, an alarm is triggered. The point at which the cumulative deviation value of the patient's actual physiological state from the initial baseline data begins to rise monotonically to exceed the decision threshold is determined as the mutation point where the physiological state has deviated adversely. The time-series evolution trajectory of the patient's current cardiopulmonary function relative to its initial physiological baseline is generated.

[0011] Preferably, the tiered early warning module specifically includes: The mode definition unit, based on clinical guidelines and expert knowledge, is defined as follows: if a patient's ECG HRV data decreases by more than 30% in standard deviation and is accompanied by an increase in heart rate, the patient is defined as being in the arrhythmia risk mode; if nocturnal respiratory RR data increases by more than 25% and is accompanied by a trend of decreasing blood oxygen saturation, the patient is defined as being in the heart failure exacerbation mode; if the resting heart rate continues to increase and exceeds 20% of the patient's personalized initial dynamic physiological baseline within the time window and is accompanied by a slight increase in body temperature, the patient is defined as being in the infection risk mode; if the recovery time after activity is prolonged and the time for the heart rate to return to the patient's personalized initial dynamic physiological baseline within the time window exceeds 10 minutes, the patient is defined as being in the activity intolerance mode. The patient clinical pathophysiology model database establishment unit integrates arrhythmia risk model, heart failure aggravation model, infection risk model and activity intolerance model to establish a patient clinical pathophysiology model database. The optimal matching pattern acquisition unit compares and analyzes each pattern in the patient's clinical pathophysiological pattern library based on the temporal evolution trajectory of the patient's current cardiopulmonary function deviation. Using the Sakoe-Chiba bandwidth-constrained search path, it calculates the minimum regular distance between the patient's current cardiopulmonary function deviation temporal evolution trajectory and any pattern in the patient's clinical pathophysiological pattern library, and selects the pattern with the shortest distance as the optimal matching pattern. The model building unit uses any pattern in the patient clinical pathophysiology pattern library as the root node, the patient physiological subsystem as the intermediate node, and the specific observed variables in the pattern as the leaf nodes to build a multi-level Bayesian network model. The posterior probability acquisition unit uses the current real-time observation data of the patient as evidence, inputs it into the leaf nodes of the multi-level Bayesian network model, performs probabilistic inference, and calculates the posterior probability of each pathophysiological pattern of the root node based on the current real-time observation data of the patient. The graded early warning unit sets a probability threshold greater than 0.7 as the pattern matching benchmark. If the posterior probability exceeds the probability threshold, it determines that the current patient's cardiopulmonary function status matches the matching pattern, automatically identifies abnormal evolution trajectories that conform to specific clinical pathophysiological patterns, and triggers graded early warnings associated with the abnormal patterns.

[0012] Preferably, the graded early warning module also includes: If the processing unit matches a patient in a heart failure aggravation mode or a high-risk arrhythmia mode, and the blood oxygen saturation drops by more than 5% within 1 hour, it is determined as a Level 1 warning. The system automatically pushes an emergency alert to the mobile terminals of the attending physician and on-duty nurse, along with a real-time snapshot of the patient's data and suggestions. If the matching patient is in an infection risk mode or an activity intolerance mode, and the posterior probability of each pathophysiological mode at the root node based on the current real-time observation data of the patient is greater than 0.85, it is determined as a Level 2 warning. The system generates a prominent prompt on the doctor's workstation and suggests arranging a telephone follow-up on the same day or scheduling an outpatient follow-up appointment for the next day. If any mode in the patient's clinical pathophysiological mode library is matched, but the posterior probability of each pathophysiological mode at the root node based on the current real-time observation data of the patient is between 0.7 and 0.85, it is determined as a Level 3 warning. The system marks the patient in the patient management list and generates a summary report the next day for the doctor's morning meeting to review.

[0013] Furthermore, home-based intelligent follow-up assessment methods for cardiopulmonary function in post-cardiac surgery patients include: S1 is used to establish a personalized initial dynamic physiological baseline for patients based on multimodal physiological time-series data of patients during the postoperative stable period through Gaussian mixture model and hidden Markov model. S2 is used to acquire the patient's physiological feature vector in real time, and obtain the deviation index of the actual-initial baseline data by matching the initial dynamic physiological baseline and calculating the Mahalanobis distance, thereby generating the deviation time series evolution trajectory. S3 is used to dynamically time-normalize and match the deviation time-series evolution trajectory with a preset clinical pathophysiological model library, calculate the posterior probability through a multi-level Bayesian network, and trigger a graded early warning based on the posterior probability.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a home-based intelligent follow-up assessment scheme for cardiopulmonary function in post-cardiac surgery patients. This system constructs a personalized initial dynamic physiological baseline and dynamically calculates the temporal trajectory of cardiopulmonary function deviation, achieving a shift from static threshold-based early warning to individualized, dynamic, and patterned risk assessment. It can accurately capture subtle evolutionary trends of cardiopulmonary function during home rehabilitation for post-cardiac surgery patients. Based on a multi-level Bayesian network and a clinical pathophysiological pattern database, it performs intelligent matching and hierarchical early warning, improving the sensitivity and specificity of early risk identification. This effectively assists medical staff in accurately screening and stratifying high-risk patients. By simultaneously pushing personalized intervention suggestions, it enhances patient self-management compliance and the timeliness of rehabilitation guidance. Attached Figure Description

[0015] Figure 1 Framework diagram of a home-based intelligent follow-up assessment system for cardiopulmonary function in post-cardiac surgery patients; Figure 2 Flowchart of a home-based intelligent follow-up assessment method for cardiopulmonary function in post-cardiac surgery patients. Detailed Implementation

[0016] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0017] Reference Figure 1 As shown, the home-based intelligent follow-up assessment system for cardiopulmonary function in post-cardiac surgery patients includes: Initial dynamic physiological baseline module, deviation time-series evolution trajectory module, graded early warning module; Among them, the deviation time-series evolution trajectory module is electrically connected to the initial dynamic physiological baseline module, and the graded early warning module is electrically connected to the deviation time-series evolution trajectory module; The initial dynamic physiological baseline module is used to establish a personalized initial dynamic physiological baseline for patients based on their postoperative stable period multimodal physiological time series data through Gaussian mixture model and hidden Markov model. The deviation time-series evolution trajectory module is used to acquire the patient's physiological feature vector in real time, and obtain the deviation index of the actual-initial baseline data state by matching the initial dynamic physiological baseline and calculating the Mahalanobis distance, thereby generating the deviation time-series evolution trajectory. The graded early warning module is used to dynamically time-normalize and match the deviation time evolution trajectory with a preset clinical pathophysiological pattern library, calculate the posterior probability through a multi-level Bayesian network, and trigger graded early warning based on the posterior probability.

[0018] The initial dynamic physiological baseline module specifically includes: The patient physiological data acquisition unit is based on a medical-grade smart bracelet worn by the patient. The bracelet contains a PPG photoelectric heart rate sensor, a three-axis accelerometer, and a body temperature sensor. The device is calibrated and data is collected for more than 30 minutes at a fixed time every day for 3-7 days after the patient's postoperative discharge in a resting state. Multimodal physiological time-series data of the patient in a stable state are obtained, including the patient's ECG HRV data, blood oxygen saturation and PI data, respiratory RR data, activity intensity and steps, and sleep quality data. The ECG signal is denoised and motion artifacts are removed and unified to the same timestamp. The temporal feature extraction unit calculates the mean, standard deviation, extreme values, and interquartile range of the multimodal physiological time series data under the patient's stable state using a 5-minute sliding time window. It then obtains the linear regression slope and autocorrelation coefficient of the multimodal physiological time series data under the patient's stable state and extracts the temporal features of the multimodal physiological time series data under the patient's stable state. The frequency domain feature extraction unit uses Fast Fourier Transform to convert the time domain features of the patient's multimodal physiological time series data under steady-state conditions to the frequency domain, calculates its power spectral density, main frequency component, frequency band energy ratio and spectral entropy, and extracts the frequency domain features of the patient's multimodal physiological time series data under steady-state conditions.

[0019] The initial dynamic physiological baseline module also includes: The initial physiological pattern unit integrates the temporal and frequency domain features of the multimodal physiological time series data of the patient in a stable state, performs normalization processing, and splices them to obtain the comprehensive feature vector of the multimodal physiological time series data of the patient in a stable state, and extracts the initial physiological pattern of the patient in the early stage of rehabilitation. The mean vector and covariance matrix establishment unit takes the comprehensive feature vector of multimodal physiological time series data under the patient's stable state as input, and uses the Bayesian information criterion to automatically select the optimal component of the comprehensive feature vector of multimodal physiological time series data under the patient's stable state, and randomly initializes the mean vector and covariance matrix of K optimal components of the comprehensive feature vector of multimodal physiological time series data under the patient's stable state. The static physiological feature baseline distribution unit is iteratively optimized using the expectation-maximization algorithm to calculate the log-likelihood function of the comprehensive feature vector of the multimodal physiological time series data under the patient's steady state. Its mean vector and covariance matrix are continuously updated until the log-likelihood function converges, thereby obtaining the Gaussian mixture function model parameters of the patient's multimodal physiological time series data. The probability density distribution of the patient's multimodal physiological time series data in the multidimensional space under the patient's steady state is determined as the patient's static physiological feature baseline distribution.

[0020] The initial dynamic physiological baseline module also includes: The temporal evolution law unit introduces a hidden Markov model. The comprehensive feature vector of multimodal physiological temporal data under the steady state of the patient is used as the observation sequence. The number of hidden states of the model is set, and the parameters of the hidden Markov model are randomly initialized, including the initial state probability distribution, the state transition probability matrix and the probability density distribution of multimodal physiological temporal data under the steady state of the patient in the multidimensional space. The Baum-Welch algorithm is used to estimate the maximum likelihood of the model parameters. The model is iteratively optimized until convergence, and a hidden Markov model of multimodal physiological temporal data under the steady state of the patient is established to capture the temporal evolution law of its physiological state. The physiological baseline establishment unit establishes a personalized initial dynamic physiological baseline for each patient based on the patient's static physiological characteristic baseline distribution and the temporal evolution of physiological state.

[0021] When using it, combine the content of the above modules: Traditional postoperative rehabilitation monitoring relies heavily on discrete clinical follow-ups or single physiological parameter threshold alarms, lacking systematic fusion analysis of continuous, multimodal physiological time-series data from patients' home environments. Existing methods often employ general thresholds or fixed baselines, making it difficult to capture the dynamics and temporal dependence of individual physiological patterns. They also lack robustness to data noise and motion artifacts, resulting in limited sensitivity and specificity for detecting early abnormal physiological changes. The beneficial effect of this approach lies in constructing a comprehensive physiological feature vector covering both the time and frequency domains through multi-sensor data fusion and spatiotemporal feature extraction, enabling a refined characterization of multidimensional physiological patterns in patients under stable conditions. By combining Gaussian mixture models and hidden Markov models, not only is a probability distribution benchmark consistent with individual static physiological characteristics established, but the temporal evolution of physiological states is also captured, forming a personalized and dynamic physiological baseline. This enhances the ability to detect subtle physiological deviations in the early stages of postoperative rehabilitation, providing crucial data-driven evidence for personalized rehabilitation early warning and intervention.

[0022] The deviation time series evolution trajectory module specifically includes: The real-time data acquisition unit sets a fixed sliding time window of 5 minutes, monitors the patient's physiological data in real time within the time window, performs normalization processing, and extracts the patient's physiological feature vector within the time window. The matching baseline calculation unit compares the patient's physiological feature vector within the time window with the patient's personalized initial dynamic physiological baseline. It calculates the Euclidean distance between the patient's physiological feature vector within the current time window and each baseline feature vector in the patient's personalized initial dynamic physiological baseline. The baseline with the shortest Euclidean distance is selected as the matching baseline. The mean vector and covariance matrix of the patient's multimodal physiological time series data under the steady state on the baseline with the shortest Euclidean distance are calculated to obtain its baseline Gaussian distribution parameters. The deviation index calculation unit calculates the Mahalanobis distance between the patient's physiological feature vector and the matching baseline within the current time window, which serves as the deviation index of the patient's actual physiological state from the initial baseline data.

[0023] The deviation time series evolution trajectory module also includes: The short-term fluctuation extraction unit uses time as the horizontal axis and the deviation index of the patient's actual physiological state from the initial baseline data as the vertical axis to generate a continuous time series of deviation index. Using the exponential weighted moving average method, it calculates the deviation trend of the patient's actual physiological state from the initial baseline data within 1 hour and captures its short-term fluctuations. The long-term trend identification unit uses the autoregressive integral moving average method to calculate the deviation trend of the patient's actual physiological state from the initial baseline data within 24 hours, and identifies its long-term trend and periodic changes.

[0024] The deviation time series evolution trajectory module also includes: The deviation time-series evolution trajectory unit calculates the average deviation of the patient's actual physiological state from the initial baseline data within a time window, using it as the target value. It calculates the difference between the deviation index of the patient's actual physiological state from the initial baseline data and the target value at each moment, and performs cumulative calculation to obtain the cumulative deviation value of the patient's actual physiological state from the initial baseline data. A decision threshold b is set. If the cumulative deviation value of the patient's actual physiological state from the initial baseline data exceeds the decision threshold b, an alarm is triggered. The point at which the cumulative deviation value of the patient's actual physiological state from the initial baseline data begins to rise monotonically to exceed the decision threshold is determined as the mutation point where the physiological state has deviated adversely. The time-series evolution trajectory of the patient's current cardiopulmonary function relative to its initial physiological baseline is generated.

[0025] When using it, combine the content of the above modules: Traditional patient physiological monitoring methods often rely on fixed universal thresholds or static baselines, which are difficult to adapt to individual physiological dynamic changes and fluctuations in status at different times. This leads to high false alarm rates, delayed early abnormality identification, a lack of time-series quantification and multi-timescale analysis of deviation trends, an inability to effectively distinguish between short-term fluctuations and long-term deterioration trends, and a failure to achieve dynamic early warning and abrupt change point capture based on cumulative deviation. The beneficial effects of this approach are that it achieves individualized and precise monitoring of patients' physiological status through real-time normalization processing via sliding window and personalized dynamic baseline matching, combined with Mahalanobis distance to quantify deviation. By using exponentially weighted moving average and autoregressive integral moving average methods, it captures short-term fluctuations and long-term trends respectively, improving the ability to analyze the time-series trajectory of cardiopulmonary function evolution. Through cumulative deviation calculation and threshold triggering mechanisms, it achieves automatic detection of early abnormal changes and abrupt change point location, enhancing the timeliness and accuracy of early warning and providing dynamic and continuous decision support for clinical intervention.

[0026] The tiered early warning module specifically includes: The mode definition unit, based on clinical guidelines and expert knowledge, is defined as follows: if a patient's ECG HRV data decreases by more than 30% in standard deviation and is accompanied by an increase in heart rate, the patient is defined as being in the arrhythmia risk mode; if nocturnal respiratory RR data increases by more than 25% and is accompanied by a trend of decreasing blood oxygen saturation, the patient is defined as being in the heart failure exacerbation mode; if the resting heart rate continues to increase and exceeds 20% of the patient's personalized initial dynamic physiological baseline within the time window and is accompanied by a slight increase in body temperature, the patient is defined as being in the infection risk mode; if the recovery time after activity is prolonged and the time for the heart rate to return to the patient's personalized initial dynamic physiological baseline within the time window exceeds 10 minutes, the patient is defined as being in the activity intolerance mode. The patient clinical pathophysiology model database establishment unit integrates arrhythmia risk model, heart failure aggravation model, infection risk model and activity intolerance model to establish a patient clinical pathophysiology model database. The optimal matching pattern acquisition unit compares and analyzes each pattern in the patient's clinical pathophysiological pattern library based on the temporal evolution trajectory of the patient's current cardiopulmonary function deviation. Using the Sakoe-Chiba bandwidth-constrained search path, it calculates the minimum regular distance between the patient's current cardiopulmonary function deviation temporal evolution trajectory and any pattern in the patient's clinical pathophysiological pattern library, and selects the pattern with the shortest distance as the optimal matching pattern. The model building unit uses any pattern in the patient clinical pathophysiology pattern library as the root node, the patient physiological subsystem as the intermediate node, and the specific observed variables in the pattern as the leaf nodes to build a multi-level Bayesian network model. The posterior probability acquisition unit uses the current real-time observation data of the patient as evidence, inputs it into the leaf nodes of the multi-level Bayesian network model, performs probabilistic inference, and calculates the posterior probability of each pathophysiological pattern of the root node based on the current real-time observation data of the patient. The graded early warning unit sets a probability threshold greater than 0.7 as the pattern matching benchmark. If the posterior probability exceeds the probability threshold, it determines that the current patient's cardiopulmonary function status matches the matching pattern, automatically identifies abnormal evolution trajectories that conform to specific clinical pathophysiological patterns, and triggers graded early warnings associated with the abnormal patterns.

[0027] The tiered early warning module also includes: If the processing unit matches a patient in a heart failure aggravation mode or a high-risk arrhythmia mode, and the blood oxygen saturation drops by more than 5% within 1 hour, it is determined as a Level 1 warning. The system automatically pushes an emergency alert to the mobile terminals of the attending physician and on-duty nurse, along with a real-time snapshot of the patient's data and suggestions. If the matching patient is in an infection risk mode or an activity intolerance mode, and the posterior probability of each pathophysiological mode at the root node based on the current real-time observation data of the patient is greater than 0.85, it is determined as a Level 2 warning. The system generates a prominent prompt on the doctor's workstation and suggests arranging a telephone follow-up on the same day or scheduling an outpatient follow-up appointment for the next day. If any mode in the patient's clinical pathophysiological mode library is matched, but the posterior probability of each pathophysiological mode at the root node based on the current real-time observation data of the patient is between 0.7 and 0.85, it is determined as a Level 3 warning. The system marks the patient in the patient management list and generates a summary report the next day for the doctor's morning meeting to review.

[0028] When using it, combine the content of the above modules: Existing technologies for remote early warning of cardiovascular diseases largely rely on static thresholds or single-parameter alarms, lacking the ability to dynamically match the temporal evolution trajectory of patients' physiological parameters with clinical pathophysiological patterns. They often overlook the synergistic changes among multiple parameters and the patient's personalized baseline, resulting in low specificity, high false alarm rate, and difficulty in timely identifying the evolution of hidden risks. This step's early warning module, by integrating dynamic time warping and multi-level Bayesian networks, achieves intelligent matching and probabilistic inference of the temporal trajectory of patients' cardiopulmonary function deviation with clinical pathophysiological patterns. Based on personalized dynamic baselines and multi-parameter association rules, it can accurately identify abnormal patterns such as arrhythmia, worsening heart failure, infection risk, and activity intolerance. It triggers graded early warnings based on posterior probability and risk level, improving the accuracy and clinical applicability of early warnings and assisting medical staff in achieving early intervention and tiered management.

[0029] Reference Figure 2 As shown, the home-based intelligent follow-up assessment method for cardiopulmonary function in post-cardiac surgery patients includes: S1 is used to establish a personalized initial dynamic physiological baseline for patients based on multimodal physiological time-series data of patients during the postoperative stable period through Gaussian mixture model and hidden Markov model. S2 is used to acquire the patient's physiological feature vector in real time, and obtain the deviation index of the actual-initial baseline data by matching the initial dynamic physiological baseline and calculating the Mahalanobis distance, thereby generating the deviation time series evolution trajectory. S3 is used to dynamically time-normalize and match the deviation time-series evolution trajectory with a preset clinical pathophysiological model library, calculate the posterior probability through a multi-level Bayesian network, and trigger a graded early warning based on the posterior probability.

[0030] Based on the above module content, the specific implementation method is as follows: Mr. Zhang, a 65-year-old patient, was discharged 5 days after undergoing coronary artery bypass grafting. As instructed by his doctor, he wore a medical-grade smart bracelet. After the system was activated, the initial dynamic physiological baseline module began to work. From the 3rd to the 7th day after discharge, at 10:00 AM every day, when Mr. Zhang was resting quietly, the bracelet automatically collected more than 30 minutes of his PPG, acceleration, and body temperature data. The raw signals were denoised and motion artifacts were removed. After synchronizing the timestamps, stable multimodal time-series data of ECG HRV, blood oxygen saturation, perfusion index, respiratory rate, activity intensity, and sleep quality were obtained. The time-domain feature extraction unit performed sliding calculations with a 5-minute window to obtain the mean, standard deviation, and other statistical measures and linear trends of these data. The power spectral density and spectral entropy frequency domain features were further extracted by Fast Fourier Transform. All features were normalized and concatenated into a comprehensive feature vector. Gaussian mixture model was used to fit these vectors. Through expectation-maximization algorithm iteration, the baseline probability distribution of Mr. Zhang's static physiological characteristics in multidimensional space was determined. Hidden Markov Model was introduced. Using the comprehensive feature vector as the observation sequence, the hidden state was learned through Baum-Welch algorithm to obtain the transition law between deep rest and recovery after mild activity. The physiological baseline establishment unit integrates the static distribution described by Gaussian mixture model with the dynamic temporal law captured by Hidden Markov Model to establish Mr. Zhang's personal initial dynamic physiological baseline. This baseline is not a single fixed value, but a composite model that can characterize the dynamic range and evolution law of physiological patterns in the stable state of his early rehabilitation. After establishing the baseline, the system enters the continuous monitoring phase. One afternoon, Mr. Zhang rested after doing light housework. The deviation time series evolution trajectory module started operating in real time. The real-time data acquisition unit forms a sliding window every 5 minutes. It performs the same preprocessing and feature extraction on the real-time collected physiological data as when the baseline was established, and obtains the physiological feature vector of the current window. The Euclidean distance is calculated between this vector and the feature vectors of each baseline in the initial dynamic physiological baseline to find the best matching baseline time. The mean vector and covariance matrix of the Gaussian mixture model parameters corresponding to this time are obtained. The Mahalanobis distance between the current feature vector and the Gaussian distribution of this matching baseline is calculated. This distance value is the deviation index of the actual-initial baseline data state at this moment. The system runs continuously. The short-term fluctuation extraction unit uses an exponentially weighted moving average to analyze the changes in the deviation index in the most recent hour, smoothing random fluctuations to observe short-term trends. It analyzes 24-hour data through an autoregressive integral moving average model to identify whether there is a long-term increase or periodic anomaly. It continuously calculates the cumulative deviation. The system finds that from a certain moment, the deviation index of multiple consecutive time windows has been continuously higher than the historical average level, causing the cumulative deviation value to rise monotonically. When the cumulative value exceeds the preset decision threshold b, the system determines that the moment is a sudden change point where the physiological state deviates adversely. It generates a clear deviation time series evolution trajectory graph on the operation interface. The horizontal axis of the graph is time, the vertical axis is the deviation index, and the short-term trend line, long-term trend line, and the identified sudden change point are marked. It intuitively shows the dynamic deviation process of Mr. Zhang's current cardiopulmonary function relative to his initial baseline. When the deviation time-series evolution trajectory shows a continuous abnormality, the graded early warning module is activated. The best matching pattern acquisition unit uses the dynamic time warping algorithm to compare the generated deviation trajectory with the time-series templates of each pattern in the patient's clinical pathophysiological pattern library. The calculation finds that the current trajectory has the shortest warping distance with the heart failure aggravation pattern, characterized by a nighttime RR increase of more than 25% accompanied by a trend of decreasing blood oxygen saturation. The constructed multi-level Bayesian network begins inference. The network uses various pathophysiological patterns as root nodes, cardiovascular, respiratory and other subsystems as intermediate nodes, and real-time HRV, blood oxygen saturation and RR as leaf nodes. Mr. Zhang's current real-time observation data is used as evidence input to the leaf nodes. Through network propagation calculation, the posterior probability of the heart failure aggravation pattern is as high as 0.82. Since the probability exceeded the preset probability threshold of 0.7, the system determined that the match was valid and classified it according to the preset rules. Because the matching mode was heart failure aggravation and the system detected that his blood oxygen saturation dropped by 5.2% in the past hour, it met the conditions for a level one warning. The system immediately triggered a level one warning automatically, and a red emergency alarm popped up on the hospital's medical management platform. At the same time, the alarm, along with a snapshot of Mr. Zhang's real-time vital signs data and suggestions, was immediately sent to the patient to assess whether he needed to adjust his diuretics or seek emergency treatment. Personalized intervention suggestions were pushed to the mobile terminals of his attending physician and the on-duty nurse. Mr. Zhang also received a warning message on his wristband APP, reminding him to rest and wait for medical staff to contact him. This completed a complete closed loop from intelligent identification of abnormal trajectory to triggering a level one warning and pushing intervention suggestions. To verify the effectiveness and practicality of the system and method described in this invention, we conducted a six-month clinical trial, enrolling 60 post-cardiac surgery patients, including those who underwent coronary artery bypass grafting or heart valve replacement. The patients were randomly divided into an experimental group (using this system) and a control group (using the traditional telephone follow-up + outpatient re-examination model). The differences between the two groups in terms of timely risk warning, readmission rate, and patient satisfaction were compared and analyzed. The experimental group achieved early abnormal trajectory identification through this system, with an average warning time 3.2 to 1.1 days earlier than the control group. Especially in the risk mode of heart failure aggravation and arrhythmia, the system triggered the warning on average 48-72 hours before the appearance of clinical symptoms, providing a valuable time window for clinical intervention. The system triggered 312 warning signals, of which 286 were confirmed as valid warnings after clinical review, with an overall warning accuracy rate of 91.7%. False alarms mainly originated from motion artifacts or temporary equipment detachment. After algorithm optimization, the false alarm rate decreased from 12.4% in the initial stage to 8.3%. The unplanned readmission rate within 30 days was 6.7% in the experimental group and 20.0% in the control group. The emergency room visit rate was reduced by 45% in the experimental group, indicating that the system can effectively reduce the occurrence of acute events caused by delayed identification. Six weeks post-surgery, the experimental group showed an average improvement of 18.5% in the 6-minute walk test distance, compared to 9.2% in the control group. The experimental group also showed better self-reported quality of life scores using the SF-36 scale compared to the control group. The system automatically generates reports on patients' physiological deviation trends and provides tiered early warning suggestions, saving the medical team an average of about 1.5 hours per person per day for data processing and preliminary assessment, and improving the screening efficiency of high-risk patients by about 40%. Patients in the experimental group scored higher than those in the control group on ease of use, real-time feedback, and sense of security. 94.3% of the patients indicated that they were willing to continue using the system in their subsequent rehabilitation. The experimental data above show that the intelligent home follow-up assessment system for cardiopulmonary function in post-cardiac surgery patients proposed in this invention not only achieves multimodal data fusion analysis and personalized dynamic modeling in terms of technology, but also improves the sensitivity and specificity of early risk identification in clinical practice, reduces the readmission rate, improves the quality of patient rehabilitation and the efficiency of medical and nursing work, and has good value for promotion and application.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A home-based intelligent follow-up assessment system for cardiopulmonary function in post-cardiac surgery patients, characterized in that: include: Initial dynamic physiological baseline module, deviation time-series evolution trajectory module, graded early warning module; Among them, the deviation time-series evolution trajectory module is electrically connected to the initial dynamic physiological baseline module, and the graded early warning module is electrically connected to the deviation time-series evolution trajectory module; The initial dynamic physiological baseline module is used to establish a personalized initial dynamic physiological baseline for patients based on their postoperative stable period multimodal physiological time series data through Gaussian mixture model and hidden Markov model. The deviation time-series evolution trajectory module is used to acquire the patient's physiological feature vector in real time, and obtain the deviation index of the actual-initial baseline data state by matching the initial dynamic physiological baseline and calculating the Mahalanobis distance, thereby generating the deviation time-series evolution trajectory. The graded early warning module is used to dynamically time-normalize and match the deviation time evolution trajectory with a preset clinical pathophysiological pattern library, calculate the posterior probability through a multi-level Bayesian network, and trigger graded early warning based on the posterior probability.

2. The home-based intelligent follow-up assessment system for cardiopulmonary function in post-cardiac surgery patients according to claim 1, characterized in that, The initial dynamic physiological baseline module specifically includes: The patient physiological data acquisition unit is based on a medical-grade smart bracelet worn by the patient. The bracelet contains a PPG photoelectric heart rate sensor, a three-axis accelerometer, and a body temperature sensor. The device is calibrated and data is collected for more than 30 minutes at a fixed time every day for 3-7 days after the patient's postoperative discharge in a resting state. Multimodal physiological time-series data of the patient in a stable state are obtained, including the patient's ECG HRV data, blood oxygen saturation and PI data, respiratory RR data, activity intensity and steps, and sleep quality data. The ECG signal is denoised and motion artifacts are removed and unified to the same timestamp. The temporal feature extraction unit calculates the mean, standard deviation, extreme values, and interquartile range of the multimodal physiological time series data under the patient's stable state using a 5-minute sliding time window. It then obtains the linear regression slope and autocorrelation coefficient of the multimodal physiological time series data under the patient's stable state and extracts the temporal features of the multimodal physiological time series data under the patient's stable state. The frequency domain feature extraction unit uses Fast Fourier Transform to convert the time domain features of the patient's multimodal physiological time series data under steady-state conditions to the frequency domain, calculates its power spectral density, main frequency component, frequency band energy ratio and spectral entropy, and extracts the frequency domain features of the patient's multimodal physiological time series data under steady-state conditions.

3. The home-based intelligent follow-up assessment system for cardiopulmonary function in post-cardiac surgery patients according to claim 2, characterized in that, The initial dynamic physiological baseline module also includes: The initial physiological pattern unit integrates the temporal and frequency domain features of the multimodal physiological time series data of the patient in a stable state, performs normalization processing, and splices them to obtain the comprehensive feature vector of the multimodal physiological time series data of the patient in a stable state, and extracts the initial physiological pattern of the patient in the early stage of rehabilitation. The mean vector and covariance matrix establishment unit takes the comprehensive feature vector of multimodal physiological time series data under the patient's stable state as input, and uses the Bayesian information criterion to automatically select the optimal component of the comprehensive feature vector of multimodal physiological time series data under the patient's stable state, and randomly initializes the mean vector and covariance matrix of K optimal components of the comprehensive feature vector of multimodal physiological time series data under the patient's stable state. The static physiological feature baseline distribution unit is iteratively optimized using the expectation-maximization algorithm to calculate the log-likelihood function of the comprehensive feature vector of the multimodal physiological time series data under the patient's steady state. Its mean vector and covariance matrix are continuously updated until the log-likelihood function converges, thereby obtaining the Gaussian mixture function model parameters of the patient's multimodal physiological time series data. The probability density distribution of the patient's multimodal physiological time series data in the multidimensional space under the patient's steady state is determined as the patient's static physiological feature baseline distribution.

4. The home-based intelligent follow-up assessment system for cardiopulmonary function in post-cardiac surgery patients according to claim 3, characterized in that, The initial dynamic physiological baseline module also includes: The temporal evolution law unit introduces a hidden Markov model. The comprehensive feature vector of multimodal physiological temporal data under the steady state of the patient is used as the observation sequence. The number of hidden states of the model is set, and the parameters of the hidden Markov model are randomly initialized, including the initial state probability distribution, the state transition probability matrix and the probability density distribution of multimodal physiological temporal data under the steady state of the patient in the multidimensional space. The Baum-Welch algorithm is used to estimate the maximum likelihood of the model parameters. The model is iteratively optimized until convergence, and a hidden Markov model of multimodal physiological temporal data under the steady state of the patient is established to capture the temporal evolution law of its physiological state. The physiological baseline establishment unit establishes a personalized initial dynamic physiological baseline for each patient based on the patient's static physiological characteristic baseline distribution and the temporal evolution of physiological state.

5. The home-based intelligent follow-up assessment system for cardiopulmonary function in post-cardiac surgery patients according to claim 1, characterized in that, The deviation time series evolution trajectory module specifically includes: The real-time data acquisition unit sets a fixed sliding time window of 5 minutes, monitors the patient's physiological data in real time within the time window, performs normalization processing, and extracts the patient's physiological feature vector within the time window. The matching baseline calculation unit compares the patient's physiological feature vector within the time window with the patient's personalized initial dynamic physiological baseline. It calculates the Euclidean distance between the patient's physiological feature vector within the current time window and each baseline feature vector in the patient's personalized initial dynamic physiological baseline. The baseline with the shortest Euclidean distance is selected as the matching baseline. The mean vector and covariance matrix of the patient's multimodal physiological time series data under the steady state on the baseline with the shortest Euclidean distance are calculated to obtain its baseline Gaussian distribution parameters. The deviation index calculation unit calculates the Mahalanobis distance between the patient's physiological feature vector and the matching baseline within the current time window, which serves as the deviation index of the patient's actual physiological state from the initial baseline data.

6. The home-based intelligent follow-up assessment system for cardiopulmonary function in post-cardiac surgery patients according to claim 5, characterized in that, The deviation time series evolution trajectory module also includes: The short-term fluctuation extraction unit uses time as the horizontal axis and the deviation index of the patient's actual physiological state from the initial baseline data as the vertical axis to generate a continuous time series of deviation index. Using the exponential weighted moving average method, it calculates the deviation trend of the patient's actual physiological state from the initial baseline data within 1 hour and captures its short-term fluctuations. The long-term trend identification unit uses the autoregressive integral moving average method to calculate the deviation trend of the patient's actual physiological state from the initial baseline data within 24 hours, and identifies its long-term trend and periodic changes.

7. The home-based intelligent follow-up assessment system for cardiopulmonary function in post-cardiac surgery patients according to claim 6, characterized in that, The deviation time series evolution trajectory module also includes: The deviation time-series evolution trajectory unit calculates the average deviation of the patient's actual physiological state from the initial baseline data within a time window, using it as the target value. It calculates the difference between the deviation index of the patient's actual physiological state from the initial baseline data and the target value at each moment, and performs cumulative calculation to obtain the cumulative deviation value of the patient's actual physiological state from the initial baseline data. A decision threshold b is set. If the cumulative deviation value of the patient's actual physiological state from the initial baseline data exceeds the decision threshold b, an alarm is triggered. The point at which the cumulative deviation value of the patient's actual physiological state from the initial baseline data begins to rise monotonically to exceed the decision threshold is determined as the mutation point where the physiological state has deviated adversely. The time-series evolution trajectory of the patient's current cardiopulmonary function relative to its initial physiological baseline is generated.

8. The home-based intelligent follow-up assessment system for cardiopulmonary function in post-cardiac surgery patients according to claim 7, characterized in that, The tiered early warning module specifically includes: The mode definition unit, based on clinical guidelines and expert knowledge, is defined as follows: if a patient's ECG HRV data decreases by more than 30% in standard deviation and is accompanied by an increase in heart rate, the patient is defined as being in the arrhythmia risk mode; if nocturnal respiratory RR data increases by more than 25% and is accompanied by a trend of decreasing blood oxygen saturation, the patient is defined as being in the heart failure exacerbation mode; if the resting heart rate continues to increase and exceeds 20% of the patient's personalized initial dynamic physiological baseline within the time window and is accompanied by a slight increase in body temperature, the patient is defined as being in the infection risk mode; if the recovery time after activity is prolonged and the time for the heart rate to return to the patient's personalized initial dynamic physiological baseline within the time window exceeds 10 minutes, the patient is defined as being in the activity intolerance mode. The patient clinical pathophysiology model database establishment unit integrates arrhythmia risk model, heart failure aggravation model, infection risk model and activity intolerance model to establish a patient clinical pathophysiology model database. The optimal matching pattern acquisition unit compares and analyzes each pattern in the patient's clinical pathophysiological pattern library based on the temporal evolution trajectory of the patient's current cardiopulmonary function deviation. Using the Sakoe-Chiba bandwidth-constrained search path, it calculates the minimum regular distance between the patient's current cardiopulmonary function deviation temporal evolution trajectory and any pattern in the patient's clinical pathophysiological pattern library, and selects the pattern with the shortest distance as the optimal matching pattern. The model building unit uses any pattern in the patient clinical pathophysiology pattern library as the root node, the patient physiological subsystem as the intermediate node, and the specific observed variables in the pattern as the leaf nodes to build a multi-level Bayesian network model. The posterior probability acquisition unit uses the current real-time observation data of the patient as evidence, inputs it into the leaf nodes of the multi-level Bayesian network model, performs probabilistic inference, and calculates the posterior probability of each pathophysiological pattern of the root node based on the current real-time observation data of the patient. The graded early warning unit sets a probability threshold greater than 0.7 as the pattern matching benchmark. If the posterior probability exceeds the probability threshold, it determines that the current patient's cardiopulmonary function status matches the matching pattern, automatically identifies abnormal evolution trajectories that conform to specific clinical pathophysiological patterns, and triggers graded early warnings associated with the abnormal patterns.

9. The home-based intelligent follow-up assessment system for cardiopulmonary function in post-cardiac surgery patients according to claim 8, characterized in that, The tiered early warning module also includes: If the processing unit matches a patient in a heart failure aggravation mode or a high-risk arrhythmia mode, and the blood oxygen saturation drops by more than 5% within 1 hour, it is determined as a Level 1 warning. The system automatically pushes an emergency alert to the mobile terminals of the attending physician and on-duty nurse, along with a real-time snapshot of the patient's data and suggestions. If the matching patient is in an infection risk mode or an activity intolerance mode, and the posterior probability of each pathophysiological mode at the root node based on the current real-time observation data of the patient is greater than 0.85, it is determined as a Level 2 warning. The system generates a prominent prompt on the doctor's workstation and suggests arranging a telephone follow-up on the same day or scheduling an outpatient follow-up appointment for the next day. If any mode in the patient's clinical pathophysiological mode library is matched, but the posterior probability of each pathophysiological mode at the root node based on the current real-time observation data of the patient is between 0.7 and 0.85, it is determined as a Level 3 warning. The system marks the patient in the patient management list and generates a summary report the next day for the doctor's morning meeting to review.

10. A method for intelligent home-based follow-up assessment of cardiopulmonary function in post-cardiac surgery patients, characterized in that... The system for performing the intelligent home-based follow-up assessment of cardiopulmonary function in post-cardiac surgery patients as described in claims 1-9 includes: S1 is used to establish a personalized initial dynamic physiological baseline for patients based on multimodal physiological time-series data of patients during the postoperative stable period through Gaussian mixture model and hidden Markov model. S2 is used to acquire the patient's physiological feature vector in real time, and obtain the deviation index of the actual-initial baseline data by matching the initial dynamic physiological baseline and calculating the Mahalanobis distance, thereby generating the deviation time series evolution trajectory. S3 is used to dynamically time-normalize and match the deviation time-series evolution trajectory with a preset clinical pathophysiological model library, calculate the posterior probability through a multi-level Bayesian network, and trigger a graded early warning based on the posterior probability.

Images