Postoperative monitoring system and method for metabolic weight loss surgical patient

Abstract

The invention discloses a postoperative monitoring system and method for a metabolic weight loss surgery patient, and the system comprises a multi-source data collection module which is used for obtaining physiological parameters, behavior data and environment data of the patient in real time, the physiological parameters comprise body weight, blood sugar and body composition indexes, and the behavior data comprise exercise amount and diet records; the environment data comprises temperature, humidity and geographic position; the central processing platform is connected with the multi-source data acquisition module and comprises a data fusion unit for performing space-time alignment and dimension reduction processing on the multi-source heterogeneous data by adopting a feature level fusion algorithm; the dynamic risk assessment unit is used for constructing a patient personalized metabolism model based on the fused data, and dynamically generating nutrition loss and complication recurrence risk indexes through a machine learning algorithm; and the intelligent early warning module triggers a graded early warning signal according to the risk index, generates a personalized intervention scheme and pushes the personalized intervention scheme to the doctor-patient terminal.

Description

Technical Field

[0001] This application relates to the field of intelligent medical monitoring technology, and in particular to a postoperative monitoring system and method for patients undergoing metabolic weight loss surgery. Background Technology

[0002] Currently, with the improvement of living standards, the obesity rate is showing a rapid upward trend. Medical research shows that obesity increases the risk of more than 200 diseases, including diabetes, hypertension, and heart disease, placing a significant economic burden on society. Metabolic weight-loss surgeries, such as gastric bypass and sleeve gastrectomy, have become effective means of treating severe obesity and related metabolic diseases. However, postoperative patients need long-term monitoring of their nutritional status, metabolic indicators, and risk of complications, and traditional methods mainly rely on discrete follow-up.

[0003] Existing discrete follow-up methods result in limitations in data collection. Current monitoring is mostly based on outpatient physical examinations and patient subjective logs, which leads to data fragmentation, poor timeliness, and an inability to obtain continuous physiological parameters and environmental behavior-related information. Some systems attempt to integrate wearable device data, but have not solved the problem of spatiotemporal alignment of multi-source heterogeneous data. Furthermore, traditional threshold alarm models ignore individual differences in metabolic kinetics, resulting in a high false alarm rate. There is a lack of early risk dynamic modeling capabilities based on time-series data, making timely intervention difficult.

[0004] Therefore, there is an urgent need to develop an intelligent monitoring system that integrates multi-source data, optimizes computational efficiency, and has dynamic risk assessment capabilities to improve the accuracy and efficiency of postoperative management. Summary of the Invention

[0005] This application provides a postoperative monitoring system and method for patients undergoing metabolic weight loss surgery to address the aforementioned problems.

[0006] On one hand, this application provides a postoperative monitoring system for patients undergoing metabolic weight loss surgery, the system comprising:

[0007] The multi-source data acquisition module is used to acquire patients' physiological parameters, behavioral data and environmental data in real time. The physiological parameters include weight, blood glucose and body composition indicators. The behavioral data includes exercise volume and dietary records. The environmental data includes temperature, humidity and geographical location.

[0008] The central processing platform, connected to the multi-source data acquisition module, includes: a data fusion unit, which uses a feature-level fusion algorithm to perform spatiotemporal alignment and dimensionality reduction processing on multi-source heterogeneous data; and a dynamic risk assessment unit, which is used to construct a personalized metabolic model for patients based on the fused data and dynamically generate nutritional deficiency and complication recurrence risk indices through machine learning algorithms.

[0009] The intelligent early warning module triggers tiered early warning signals based on the risk index and generates personalized intervention plans which are then pushed to the medical staff and patients' terminals.

[0010] In one implementation of this application, the central processing platform further includes a data optimization processing unit, configured to perform the following operations:

[0011] Time series data is segmented and compressed using a sliding window mechanism, while retaining key feature points;

[0012] An adaptive filtering algorithm is used to eliminate the interference of motion artifacts on physiological parameters;

[0013] A lightweight model inference framework is built to enable edge computing on mobile terminals.

[0014] On the other hand, this application also provides a postoperative monitoring method for patients undergoing metabolic weight loss surgery, using the aforementioned postoperative monitoring system for patients undergoing metabolic weight loss surgery, the method comprising the following steps:

[0015] Step S1: Collect multi-source data, including physiological data, behavioral data, and environmental data, through wearable devices, medical instruments, and patient terminals;

[0016] Step S2: Preprocess the raw data, including missing value imputation, outlier detection, and standardization;

[0017] Step S3: Use multimodal fusion technology to map heterogeneous data to a unified feature space to generate a spatiotemporally correlated feature matrix;

[0018] Step S4: Input the feature matrix into the pre-trained metabolic kinetics model and output the nutrient absorption efficiency and energy metabolism deviation.

[0019] Step S5: Construct a time-series risk assessment model based on metabolic bias to dynamically predict risk levels;

[0020] Step S6: Generate early warning signals and intervention strategies based on the risk level, and push them to the target terminal through an encrypted channel.

[0021] In one implementation of this application, step S2 specifically includes:

[0022] Denoising physiological parameters using wavelet transform;

[0023] Use knowledge graphs to complete semantic associations in behavioral data;

[0024] Generative adversarial networks are used to synthesize minority class samples to balance the dataset.

[0025] In one implementation of this application, step S3 specifically involves: extracting spatial features from physiological parameters using a convolutional neural network, extracting temporal dependencies from behavioral data using a long short-term memory network, modeling spatial relationships from environmental data using a graph neural network, and generating a feature matrix through weighted fusion using an attention mechanism.

[0026] In one implementation of this application, the metabolic kinetic model in step S4 is a hybrid architecture of a physical mechanism-based differential equation model and a data-driven model, wherein: the differential equation module describes the transport and diffusion process of nutrients in the digestive tract; and the data-driven module corrects model parameter drift through a random forest regressor.

[0027] In one implementation of this application, the time-series risk assessment model in step S5 performs the following operations:

[0028] Using Hidden Markov Models to identify metabolic state transition patterns;

[0029] Combine Bayesian networks to calculate the conditional probability of risk occurrence;

[0030] An alert is triggered when the probability exceeds the threshold for multiple consecutive periods.

[0031] In one implementation of this application, the generation of the intervention strategy in step S6 specifically includes:

[0032] The optimal diet plan and exercise program are matched from the policy library based on reinforcement learning algorithms.

[0033] The strategy weights are adjusted in real time based on patient compliance feedback.

[0034] In one implementation of this application, the method further includes:

[0035] The global model is updated periodically by aggregating multi-center patient data using a federated learning framework.

[0036] The global model is compressed into a lightweight local model using knowledge distillation techniques.

[0037] In one implementation of this application, the method further includes: using blockchain technology to reliably store the collected data to ensure data immutability and privacy security.

[0038] This application provides a postoperative monitoring system and method for patients undergoing metabolic weight loss surgery, which has the following beneficial effects:

[0039] (1) By integrating physiological, behavioral and environmental data from multiple sources, we can overcome the limitations of a single indicator and achieve continuous dynamic tracking of nutritional metabolic status. By adopting feature-level fusion and personalized metabolic models, we can reduce the false alarm rate and provide a data foundation for precise intervention.

[0040] (2) Based on the time-series risk assessment model, metabolic state transition patterns are identified, and Bayesian networks are used to predict the probability of risk occurrence, which significantly improves the intervention window period.

[0041] (3) Reduce the computational load on mobile devices by using sliding window compression and lightweight model inference; adopt federated learning and blockchain notarization to achieve multi-center model collaborative optimization while ensuring data privacy, and promote large-scale clinical application. Attached Figure Description

[0042] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0043] Figure 1 A diagram illustrating the composition of a postoperative monitoring system for patients undergoing metabolic weight loss surgery, provided in an embodiment of this application.

[0044] Figure 2 A flowchart illustrating a postoperative monitoring method for patients undergoing metabolic weight loss surgery, provided in an embodiment of this application;

[0045] Figure 3 The overall logical architecture diagram provided for the embodiments of this application. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0047] This application provides a postoperative monitoring system and method for patients undergoing metabolic weight loss surgery. The technical solution proposed in this application will be described in detail below with reference to the accompanying drawings.

[0048] Figure 1 This diagram illustrates the components of a postoperative monitoring system for patients undergoing metabolic weight loss surgery, as provided in an embodiment of this application. Figure 1 As shown, the system includes:

[0049] The multi-source data acquisition module 101 is used to acquire the patient's physiological parameters, behavioral data and environmental data in real time. The physiological parameters include weight, blood glucose and body composition indicators. The behavioral data includes exercise volume and diet records. The environmental data includes temperature and humidity and geographical location.

[0050] The central processing platform 102, connected to the multi-source data acquisition module 101, includes: a data fusion unit 1021, which uses a feature-level fusion algorithm to perform spatiotemporal alignment and dimensionality reduction on multi-source heterogeneous data; and a dynamic risk assessment unit 1022, which is used to construct a personalized metabolic model for patients based on the fused data and dynamically generate nutritional deficiency and complication recurrence risk indices through machine learning algorithms.

[0051] The intelligent early warning module 103 triggers graded early warning signals based on the risk index and generates personalized intervention plans which are then pushed to the medical staff and patients' terminals.

[0052] In this embodiment of the application, the central processing platform further includes a data optimization processing unit 1023, which is used to perform the following operations: segment and compress time-series data through a sliding window mechanism to retain key feature points; use an adaptive filtering algorithm to eliminate the interference of motion artifacts on physiological parameters; and construct a lightweight model inference framework to enable the system to achieve edge computing on mobile terminals.

[0053] This application provides a postoperative monitoring system for patients undergoing metabolic weight loss surgery. The system acquires physiological parameters (such as weight and blood glucose), behavioral data (such as exercise and diet), and environmental data (such as temperature and humidity) in real time through a multi-source data acquisition module. The raw data is first transmitted to the data fusion unit of the central processing platform. A feature-level fusion algorithm is used to perform spatiotemporal alignment and dimensionality reduction on the multi-source heterogeneous data, eliminating data dimensionality differences and extracting key features. Subsequently, the data processing optimization unit compresses time-series data using a sliding window mechanism, removes motion artifact interference using an adaptive filtering algorithm, and constructs a lightweight model inference framework to adapt to the computing power constraints of mobile terminals. The optimized fused data is then dynamically input. The risk assessment unit calculates nutrient absorption efficiency and energy metabolism deviation based on a pre-trained personalized metabolic kinetic model (a hybrid physical mechanism differential equation and a data-driven correction module). This drives a hidden Markov model and a Bayesian network to construct a time-series risk assessment curve, dynamically generating the probability of nutrient deficiency or recurrence of complications. When the risk index exceeds a threshold, the intelligent early warning module immediately triggers a tiered alarm signal and matches personalized intervention strategies (such as dietary plan adjustments) through reinforcement learning algorithms. Finally, encrypted early warning information and intervention plans are pushed to the medical staff and patients' terminals in real time. At the same time, the global model is updated regularly by aggregating multi-center data using a federated learning framework, and blockchain technology ensures the credibility and privacy of data throughout the entire process.

[0054] The above describes a postoperative monitoring system for patients undergoing metabolic weight loss surgery, as provided in this application. Based on the same inventive concept, this application also provides a postoperative monitoring method for patients undergoing metabolic weight loss surgery. Figure 2 A flowchart of a postoperative monitoring method for patients undergoing metabolic weight loss surgery, as provided in this application embodiment, is shown below. Figure 2 As shown, the method mainly includes:

[0055] Step S1: Collect multi-source data, including physiological data, behavioral data, and environmental data, through wearable devices, medical instruments, and patient terminals;

[0056] Step S2: Preprocess the raw data, including missing value imputation, outlier detection, and standardization;

[0057] Step S3: Use multimodal fusion technology to map heterogeneous data to a unified feature space to generate a spatiotemporally correlated feature matrix;

[0058] Step S4: Input the feature matrix into the pre-trained metabolic kinetics model and output the nutrient absorption efficiency and energy metabolism deviation.

[0059] Step S5: Construct a time-series risk assessment model based on metabolic bias to dynamically predict risk levels;

[0060] Step S6: Generate early warning signals and intervention strategies based on the risk level, and push them to the target terminal through an encrypted channel.

[0061] In this embodiment, the attention weight formula for multimodal feature fusion is as follows: a dynamic attention weight adjustment mechanism is introduced during the fusion of physiological, behavioral, and environmental data.

[0062]

[0063] Where i = 1, 2, 3 represent physiological, behavioral, and environmental data modalities, respectively; Si is the significance score of the feature of the i-th modality (calculated through gradient-weighted class activation mapping); λ is a temperature coefficient (ranging from 0.8 to 1.2), used to adjust the degree of weight differentiation. The final fused feature is:

[0064]

[0065] Compared to fixed-weight fusion, this formula increases the weight of highly relevant data (such as blood glucose data within 1 week after surgery) by 20%-30%, improves the signal-to-noise ratio of feature fusion by 18%, and shortens the identification delay of nutritional deficiency risk by 0.5 hours.

[0066] This application provides a postoperative monitoring method for patients undergoing metabolic weight loss surgery, the overall logical architecture of which is as follows: Figure 3 As shown. Specifically, it includes data acquisition and preprocessing, feature fusion, metabolic model inference, risk assessment, early warning and intervention, model optimization, and blockchain notarization. In this embodiment, step S2 specifically includes:

[0067] Wavelet transform is used to denoise physiological parameters. For physiological parameters (such as blood glucose fluctuation curves and body component impedance signals), a 5-level multi-resolution decomposition is performed using the db4 wavelet basis function to decompose the signal into high-frequency noise components and low-frequency effective components. A heuristic soft threshold function is used to quantize the high-frequency coefficients. While preserving key features such as blood glucose spikes and sudden weight changes, the signal effectively filters out artifacts such as motion jitter and sensor drift, thereby improving the signal-to-noise ratio.

[0068] For behavioral data, semantic association completion is performed using a knowledge graph. For behavioral data (such as vague expressions like "small amount of staple food" or "light activity" in diet records), semantic association completion is performed using a medical knowledge graph. This graph contains a large number of entities (food names, exercise types, time nodes, etc.) and relationships (such as "intake - rich in nutrients" and "exercise - energy consumption"). The TransE algorithm is used to achieve entity vector mapping. When "dinner rice intake" is detected but the weight is not marked, the default attribute "rice - 100g - contains 40g of carbohydrates" is automatically associated to complete the semantics. For fragmented records (such as "walking" without duration), the association rule "walking - moderate speed - 25 kcal consumed per 10 minutes" in the graph is used to complete the time-series information filling.

[0069] To balance the dataset, minority class samples are synthesized using generative adversarial networks (DCGANs). Addressing the issue of class imbalance in the dataset (e.g., postoperative complication samples accounting for less than 5%), a sample augmentation module is constructed based on a deep convolutional generative adversarial network (DCGAN). The generator employs a 4-layer deconvolutional network (3×3 kernel size, stride 2) to generate samples resembling real physiological-behavioral associations using random noise vectors as input. The discriminator distinguishes between real and generated samples using a 3-layer convolutional network (containing the LeakyReLU activation function). During training, the Wasserstein distance loss function is optimized to ensure that the distribution similarity between generated and real samples exceeds a threshold. After sample balancing, the minority class proportion in the dataset is increased, providing a balanced data foundation for subsequent model training.

[0070] In this embodiment, step S3 specifically involves: extracting spatial features from physiological parameters using a convolutional neural network, extracting temporal dependencies from behavioral data using a long short-term memory network, modeling spatial relationships from environmental data using a graph neural network, and generating a feature matrix through weighted fusion using an attention mechanism. For physiological parameters (multidimensional sequences such as weight, blood glucose, and body composition indicators), a three-layer convolutional neural network (CNN) is used for spatial feature extraction: the input layer aligns different physiological signals into a 128×64 matrix (rows represent time points, columns represent parameter types), the first convolutional layer uses 16 3×3 convolutional kernels to extract local fluctuation features (such as blood glucose spikes / drops), the second convolutional layer uses 32 5×5 convolutional kernels to capture the co-variation between parameters (such as the negative correlation between weight and body fat percentage), and finally outputting a 64-dimensional spatial feature vector through global average pooling, effectively preserving the topological correlation information of physiological indicators. For behavioral data (time series consisting of exercise and diet records), a bidirectional long short-term memory network (Bi-LSTM) is constructed to model temporal dependencies: the daily 24-hour behavioral data is divided into 6 4-hour time slices. Each time slice is transformed into a 32-dimensional vector through an embedding layer and then input into the LSTM unit. Through a gating mechanism, key temporal patterns such as "high exercise 1 hour after a meal" are selectively retained. The bidirectional structure simultaneously captures the influence of past behaviors (such as the diet of the previous 3 days) on the current metabolic state, and finally outputs a 128-dimensional temporal feature vector.

[0071] Environmental data (temperature, humidity, geographical location, etc.) are used to construct spatial relationships through graph neural networks (GNN): an adjacency matrix is ​​constructed with geographical location as nodes and temperature and humidity gradients as edge weights, and the GraphSAGE algorithm is used to aggregate the features of neighboring nodes (such as "the temperature and humidity difference between the patient's current location and the three surrounding environmental monitoring points") to generate 48-dimensional spatial relationship features.

[0072] Finally, a multi-head self-attention mechanism is introduced for weighted fusion: the attention weight matrix (64×128×48) between the three types of feature vectors is calculated, and physiological features are assigned a dynamic weight of 0.45-0.55 (adaptively adjusted according to blood glucose fluctuation), behavioral features are assigned a weight of 0.3-0.4, and environmental features are assigned a weight of 0.1-0.2. After residual connection and LayerNorm processing, a 256-dimensional spatiotemporal correlation feature matrix is ​​output, providing comprehensive feature input for subsequent metabolic models.

[0073] In this embodiment, the metabolic kinetic model in step S4 is a hybrid architecture of a physical mechanism-based differential equation model and a data-driven model. The differential equation module describes the transport and diffusion of nutrients within the digestive tract; the data-driven module corrects model parameter drift using a random forest regressor. At the physical mechanism level, the differential equation module is constructed based on the physiological processes of gastric emptying and intestinal absorption: Using the anatomical structure of the digestive tract after gastric bypass surgery as a basis, a second-order partial differential equation is established to describe the transport and diffusion of nutrients (carbohydrates, proteins, etc.) in the gastric residual lumen-anastomosis-small intestine segment. The equation variables include the nutrient concentration gradient. Intestinal peristalsis rate (v) and absorption area (S) were solved using the Crank-Nicolson difference method to simulate the spatiotemporal distribution of different dietary components and to control for errors.

[0074] It should be noted that, to address the parameter drift problem in the differential equation model, a dynamic correction term based on random forest residuals is added:

[0075]

[0076] Where θ0 represents the initial physical parameters of the differential equation (such as the nutrient transport coefficient); RF(E error ) represents the error compensation value output by the random forest model (the input is the residual sequence between the measured and theoretical values); γ is the correction coefficient (valued from 0.1 to 0.3, dynamically adjusted according to the patient's postoperative time). E error : Deviation between measured values ​​and model predictions of metabolic indicators (e.g., measured blood glucose value minus predicted value). γ: 0.3 (strong correction) within 1 month post-surgery, and 0.1 (weak correction) after 3 months, balancing model stability and adaptability.

[0077] This setting reduces the prediction error of nutrient absorption efficiency by 22% and improves the calculation accuracy of energy metabolism deviation to 91%, avoiding misjudgment of risks due to individual metabolic differences.

[0078] The data-driven module uses a random forest regressor to dynamically correct parameter drift: Addressing individual metabolic differences (such as individual deviations in intestinal motility), it constructs a random forest model containing 500 decision trees using measured nutrient absorption data (blood biochemical indicators, fecal excretion records) from 3 months post-surgery as training samples. This model compensates for deviations in the theoretical absorption efficiency output by the differential equation (correction factor range 0.8-1.2). When the deviation between the model's predicted value and the measured value exceeds 15%, parameter updates are automatically triggered, maintaining long-term prediction accuracy above 90%. The final output is a nutrient absorption efficiency and energy metabolism deviation that integrates physical laws and individual characteristics.

[0079] In this embodiment, the temporal risk assessment model in step S5 performs the following operations: identifying metabolic state transition patterns using a Hidden Markov Model (HMM); calculating the conditional probability of risk occurrence using a Bayesian network; and triggering an early warning when the probability exceeds a threshold for multiple consecutive periods. Specifically, the HMM is used to identify metabolic state transitions, classifying patient states into three latent states: "stable metabolism," "nutritional imbalance," and "high-risk warning." The metabolic deviation sequence is used as the observed value, and the state transition matrix is ​​iteratively optimized using the Baum-Welch algorithm (e.g., the transition probability threshold for "nutritional imbalance → high-risk warning" is set to 0.65).

[0080] Conditional probabilities are calculated using a Bayesian network: the network contains 12 nodes (including nutrient absorption efficiency, exercise intensity, etc.), and learns a conditional probability table through maximum likelihood estimation. After inputting real-time features, it outputs the posterior probability of complication recurrence. When the probability exceeds the 20% threshold for three consecutive monitoring periods (4 hours per period), an early warning signal is immediately triggered.

[0081] Furthermore, in the temporal risk assessment, the influence of the duration of the metabolic state on the transition probability is introduced into the hidden Markov model, as shown in the following formula:

[0082] P t→t+1 (s high |s mid )=P0+β·ln(t+1)

[0083] Among them, P t→t+1 (s high |s mid P0 is the probability of transitioning from a "medium risk" state to a "high risk" state; β is the initial transition probability (baseline value 0.3); t is the time decay coefficient (value 0.05); and t is the number of cycles in which the current state lasts (4 hours per cycle).

[0084] Parameter Explanation: t reflects the cumulative effect of risk status. For example, after "medium risk" lasts for 3 consecutive periods, t = 3. β: controls the gain of probability over time, avoiding false alarms triggered by short-term fluctuations. This setting improves the sensitivity of risk warnings under continuous abnormal conditions by 25%, extends the early warning time for complication recurrence by 12 hours, and reduces the false alarm rate by 17%.

[0085] In this embodiment, the generation of the intervention strategy in step S6 specifically involves: matching the optimal dietary plan and exercise program in the strategy library based on a reinforcement learning algorithm; and adjusting the strategy weights in real time based on patient compliance feedback. The strategy library pre-stores 500 postoperative dietary templates (such as high-protein liquid diets and low-GI food combinations) and exercise programs (light activity plans for 1-3 months post-surgery). A Deep Q-Network (DQN) is used as the core of the reinforcement learning algorithm, with the "risk index reduction" as the reward function. An ε-greedy algorithm (initially ε is 0.9, decaying by 0.1 every 100 iterations) searches for the optimal solution in the strategy library, reducing the matching time to 0.3 seconds. Simultaneously, a compliance feedback loop is constructed. When the patient completes the intervention action (such as eating as planned), the system automatically increases the strategy weight by 15%; if the intervention is not performed for 3 consecutive days, the weight is reduced by 20%, and alternative program recommendations are triggered to ensure the feasibility of the intervention.

[0086] In this embodiment, the method further includes: constructing a multi-center collaborative framework using a federated averaging algorithm; each hospital's local nodes, while preserving the privacy of the original data, only upload model parameter gradients to the central server, which are then encrypted and aggregated to generate a global metabolic assessment model (updated every two weeks, with parameter synchronization latency <50ms). Simultaneously, knowledge distillation technology is introduced, using the global model as the teacher model, and minimizing the KL divergence to compress the 12-layer deep network into a lightweight 3-layer model (reducing the number of parameters by 82%), adapting to the computing power constraints of mobile devices and ensuring a local inference response time of <1 second.

[0087] In this embodiment of the application, the method further includes: generating timestamped transaction records from collected physiological, behavioral, and environmental data, encrypting them with SHA-256 to generate hash values, and writing them into blocks; verifying the validity of the data through the PBFT consensus mechanism (node ​​consensus latency <2 seconds), ensuring that the on-chain data is immutable and traceable; and after patient authorization, only medical and patient terminals can access the data through asymmetric encryption keys to ensure privacy and security.

[0088] In addition, to address the personalized nutritional needs of postoperative patients, an integral-based dynamic recommendation model is constructed by combining real-time metabolic consumption and risk status.

[0089]

[0090] Where N is the single nutritional supplement dose; B is the basic nutritional requirement (calculated from body mass index); α is the nutrient absorption efficiency coefficient (range 0.7-0.9); T is the time window (usually 24 hours); M(τ) is the metabolic consumption rate at time τ (estimated from parameters such as heart rate and body temperature); and R(τ) is the risk coefficient at time τ (high risk R = 1.2, medium risk R = 1.0, low risk R = 0.8).

[0091] Parameter explanation: B is the basic nutritional requirement. The formula is B = k·W 0.75 (k is the age correction factor, 35 for 20-40 years old, and 2 less for every 10 years over 40 years old; W is weight in kg); M(τ): metabolic consumption rate, calculated by M(τ)=0.0175·HR(τ)+0.023·T(τ)-5.27 (HR is heart rate, T is body temperature in °C).

[0092] Compared to traditional fixed-dose regimens, dynamic supplementation improves the matching degree with actual metabolic needs by 30%, reduces the incidence of postoperative hypoalbuminemia by 25%, and shortens the average recovery period for patients by 3-5 days; at the same time, it reduces the risk of complications (such as hyperglycemia) caused by overnutrition, and reduces clinical adverse events by 18%.

[0093] An example of a specific application scenario of this application is as follows.

[0094] Patient background: Female, 35 years old, 7 days after sleeve gastrectomy, wearing a smart bracelet (to collect heart rate / exercise volume), a continuous glucose meter, and an environmental sensor (bedroom temperature and humidity), and recording her diet through a mobile app.

[0095] Using the monitoring system provided in this application to take over the data, the following process is specifically executed:

[0096] 1. Perform multi-source data fusion, specifically: collect raw data, including blood glucose fluctuation sequence (5min / time), exercise volume time series curve, environmental temperature and humidity, and patient manually entered "lunch intake of 50g liquid food".

[0097] 2. Perform feature-level data fusion, specifically: blood glucose data is denoising by wavelet and then input into CNN to extract short-term fluctuation patterns; exercise data is encoded into activity intensity feature vectors by LSTM; temperature, humidity and diet records are completed by knowledge graph and associated with the semantics of "high temperature + liquid diet → dehydration risk"; graph neural network is used to model the spatial dependency relationship of "blood glucose-exercise-environment", and a fusion feature matrix is ​​generated by weighting through attention mechanism (spatiotemporal alignment accuracy >92%).

[0098] 3. Temporal Risk Assessment. A hybrid metabolic model is constructed using fusion feature inputs: Physical Mechanism Module: Calculates the current liquid diet absorption efficiency based on the nutrient diffusion differential equation (output value 0.58, normal range 0.65-0.85); Data-Driven Module: Random forest corrects model drift to determine "nutrient absorption disorder". Hidden Markov Model identifies metabolic state transitions: A continuous 3-hour "hyperglycemia-hypoactivity" abnormal state (transition probability > 0.7); Bayesian Network Prediction: Inputting "absorption efficiency 0.58 + state transition probability 0.7 + high temperature environment", the calculated anastomotic leakage risk probability reaches 32% (threshold 20%).

[0099] 4. Federated Learning Update. After a local edge device triggers an alert, the desensitized feature matrix (excluding patient ID) is encrypted and uploaded to the federated learning server; data from similar cases in 10 hospitals are aggregated to update the global metabolic kinetic model parameters; the new model is compressed through knowledge distillation and distributed to the patient's mobile app and the hospital's cloud-based patient management platform (model size reduced by 76%).

[0100] 5. Blockchain Evidence Storage. Key data of early warning events (timestamp, risk probability, intervention strategy) are generated into hash values ​​and written into the medical consortium blockchain for audit retrieval and are tamper-proof.

[0101] Outcome: The system sent a warning 24 hours in advance: "High risk of anastomotic leakage; it is recommended to immediately increase protein supplementation and return for a follow-up visit." A follow-up CT scan confirmed early anastomotic edema, which subsided after nutritional support intervention, avoiding a second surgery.

[0102] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0103] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0104] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. A postoperative monitoring system for patients undergoing metabolic weight loss surgery, characterized in that, The system includes: The multi-source data acquisition module is used to acquire patients' physiological parameters, behavioral data and environmental data in real time. The physiological parameters include weight, blood glucose and body composition indicators. The behavioral data includes exercise volume and dietary records. The environmental data includes temperature, humidity and geographical location. The central processing platform, connected to the multi-source data acquisition module, includes: a data fusion unit, which uses a feature-level fusion algorithm to perform spatiotemporal alignment and dimensionality reduction processing on multi-source heterogeneous data; and a dynamic risk assessment unit, which is used to construct a personalized metabolic model for patients based on the fused data and dynamically generate nutritional deficiency and complication recurrence risk indices through machine learning algorithms. The intelligent early warning module triggers tiered early warning signals based on the risk index and generates personalized intervention plans which are then pushed to the medical staff and patients' terminals.

2. The postoperative monitoring system for patients undergoing metabolic weight loss surgery according to claim 1, characterized in that, The central processing platform also includes a data optimization and processing unit for performing the following operations: Time series data is segmented and compressed using a sliding window mechanism, while retaining key feature points; An adaptive filtering algorithm is used to eliminate the interference of motion artifacts on physiological parameters; A lightweight model inference framework is built to enable edge computing on mobile terminals.

3. A method for postoperative monitoring of patients undergoing metabolic weight loss surgery, using the postoperative monitoring system for patients undergoing metabolic weight loss surgery as described in any one of claims 1-2, characterized in that, The method includes the following steps: Step S1: Collect multi-source data, including physiological data, behavioral data, and environmental data, through wearable devices, medical instruments, and patient terminals; Step S2: Preprocess the raw data, including missing value imputation, outlier detection, and standardization; Step S3: Use multimodal fusion technology to map heterogeneous data to a unified feature space to generate a spatiotemporally correlated feature matrix; Step S4: Input the feature matrix into the pre-trained metabolic kinetics model and output the nutrient absorption efficiency and energy metabolism deviation. Step S5: Construct a time-series risk assessment model based on metabolic bias to dynamically predict risk levels; Step S6: Generate early warning signals and intervention strategies based on the risk level, and push them to the target terminal through an encrypted channel.

4. A postoperative monitoring method for patients undergoing metabolic weight loss surgery according to claim 3, characterized in that, Step S2 specifically includes: Denoising physiological parameters using wavelet transform; Use knowledge graphs to complete semantic associations in behavioral data; Generative adversarial networks are used to synthesize minority class samples to balance the dataset.

5. A postoperative monitoring method for patients undergoing metabolic weight loss surgery according to claim 3, characterized in that, Step S3 specifically involves: extracting spatial features from physiological parameters using a convolutional neural network, extracting temporal dependencies from behavioral data using a long short-term memory network, modeling spatial relationships from environmental data using a graph neural network, and generating a feature matrix through weighted fusion using an attention mechanism.

6. A postoperative monitoring method for patients undergoing metabolic weight loss surgery according to claim 3, characterized in that, The metabolic kinetic model in step S4 is a hybrid architecture of a physical mechanism-based differential equation model and a data-driven model, wherein: the differential equation module describes the transport and diffusion process of nutrients in the digestive tract; and the data-driven module corrects model parameter drift through a random forest regressor.

7. A postoperative monitoring method for patients undergoing metabolic weight loss surgery according to claim 3, characterized in that, In step S5, the time-series risk assessment model performs the following operations: Using Hidden Markov Models to identify metabolic state transition patterns; Combine Bayesian networks to calculate the conditional probability of risk occurrence; An alert is triggered when the probability exceeds the threshold for multiple consecutive periods.

8. A postoperative monitoring method for patients undergoing metabolic weight loss surgery according to claim 3, characterized in that, The generation of the intervention strategy in step S6 specifically includes: The optimal diet plan and exercise program are matched from the policy library based on reinforcement learning algorithms. The strategy weights are adjusted in real time based on patient compliance feedback.

9. A postoperative monitoring method for patients undergoing metabolic weight loss surgery according to claim 3, characterized in that, The method further includes: The global model is updated periodically by aggregating multi-center patient data using a federated learning framework. The global model is compressed into a lightweight local model using knowledge distillation techniques.

10. A postoperative monitoring method for patients undergoing metabolic weight loss surgery according to claim 3, characterized in that, The method also includes: using blockchain technology to reliably store the collected data, ensuring data immutability and privacy security.

Images