Long-term follow-up visit system after glandular carcinoma resection of joint part of esophagus and stomach

Abstract

The invention discloses a long-term follow-up visit system after esophageal-gastric junction adenocarcinoma resection, and relates to the technical field related to hospital information management, and the long-term follow-up visit system comprises a multi-modal data acquisition layer which is used for obtaining medical data from a hospital information system; the wearable device collects physiological indexes; and an embedded electrochemical sensor, an air quality monitor, a water quality analyzer and an electromagnetic radiation dosimeter. The intelligent data processing layer is used for analyzing CT and PET images by adopting a 3D convolutional neural network, carrying out unstructured data processing and extracting tumor metabolism characteristics; fusing the physiological indexes and the environmental data by using an LSTM network, carrying out time sequence data analysis, and constructing a dynamic health index; environment exposure index modeling is carried out, and a pollutant exposure index is calculated. Clinical data, environmental data, behavior data and gene data are covered through the multi-modal data acquisition layer, multi-dimensional information fusion is realized, the one-sidedness that traditional follow-up visit depends on a single index is solved, and a comprehensive data basis is provided for accurate prediction.

Description

Technical Field

[0001] This invention relates to the field of hospital information management technology, and in particular to a long-term follow-up system after resection of adenocarcinoma of the esophagogastric junction. Background Technology

[0002] Adenocarcinoma of the esophagogastric junction (AEG) is a subtype of gastrointestinal malignant tumors with a poor prognosis. It has a high postoperative recurrence rate and complex metastasis patterns, often involving the peritoneum, liver, and lymph nodes. Moreover, it is affected by multiple factors such as environmental exposure and lifestyle, making long-term follow-up management difficult. Postoperative management of adenocarcinoma of the esophagogastric junction (AEGJ) needs to take into account tumor recurrence monitoring, complication prevention, and improvement of patients' quality of life.

[0003] Current clinical follow-up techniques have the following limitations:

[0004] Traditional follow-up studies primarily rely on imaging (such as CT / PET-CT) and blood tumor markers (CEA / CA19-9), focusing only on anatomical and molecular changes while neglecting non-traditional risk factors such as environmental exposure (such as air pollution and residential density), nutritional metabolism, and sleep disorders. For example, the association between PM2.5 exposure and AEG recurrence has been confirmed by epidemiological studies, but it has not been incorporated into clinical decision-making systems.

[0005] Current guidelines often employ a "one-size-fits-all" follow-up strategy (such as re-examination every 3-6 months for the first 2 years post-surgery), failing to dynamically optimize for individual patient risk differences (such as pathological stage, molecular subtype, and environmental exposure level). This can easily lead to over-testing of low-risk patients or missed diagnoses of high-risk patients. Imaging and blood indicators are often analyzed independently, lacking spatiotemporal correlation modeling. For example, short-term fluctuations in CEA may be misjudged as noise, while a long-term upward trend may have missed the intervention window. Local recurrence (such as peritoneal metastasis) is easily overlooked due to early asymptomatic or imaging-insensitive factors.

[0006] Furthermore, patients have limited understanding of the significance of follow-up, often resulting in poor compliance due to economic burden or fear; on the doctor's side, there is a lack of intelligent tools to support the efficient processing of multi-source data and the development of personalized plans.

[0007] Therefore, it is necessary to propose a long-term follow-up system for resection of adenocarcinoma at the esophagogastric junction to address the above problems. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a long-term follow-up system for patients after resection of adenocarcinoma at the esophagogastric junction, which solves the limitations of traditional follow-up models, such as data fragmentation, single monitoring dimensions, and low patient participation.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A long-term follow-up system after resection of adenocarcinoma of the esophagogastric junction, comprising:

[0011] Multimodal data acquisition layer: acquiring medical data from hospital information systems; collecting physiological indicators from wearable devices; and embedded electrochemical sensors, air quality monitors, water quality analyzers, and electromagnetic radiation dosimeters;

[0012] Intelligent data processing layer: Employs 3D convolutional neural networks to analyze CT and PET images, perform unstructured data processing, and extract tumor metabolic characteristics; utilizes LSTM networks to fuse physiological indicators and environmental data, perform time-series data analysis, and construct a dynamic health index; and models environmental exposure indexes to calculate pollutant exposure indices.

[0013] Relapse prediction model layer: Maps clinical staging, molecular subtyping, treatment response, and environmental exposure features to a unified feature space for multimodal feature fusion; uses a spatiotemporal graph convolutional network to capture the interaction of multiple factors and outputs the relapse probability;

[0014] Visualized decision-making layer: The probability of metastasis is marked by anatomical location to construct a risk heat map; the dynamic curve is displayed and compared with the critical threshold to provide trend warning; the curve of PEI index and recurrence risk is displayed to conduct environmental correlation analysis;

[0015] Intelligent intervention layer: The personalized follow-up plan generation model automatically generates examination suggestions based on risk scores;

[0016] Doctor-patient interaction platform: Pushes customized rehabilitation training videos, and abnormal indicators automatically trigger doctor alerts.

[0017] Optionally, the medical data is acquired in the following ways: electronic medical records and pathology reports are obtained from the hospital information system through the Rapid Medical Interoperability Resource Interface; DICOM format image data is obtained from the image archiving system; gene testing results and gene sequencing data are obtained from the laboratory information system; key information of unstructured text is extracted through NLP and mapped into standardized codes; image data is parsed through the DICOM standard to extract pixel matrices and metadata; and multimodal alignment, model lightweighting, and privacy protection are performed.

[0018] Optionally, the steps of the unstructured data processing are as follows:

[0019] Data preprocessing and enhancement: DICOM format CT and PET images are converted into three-dimensional matrices, and pixel intensity is adjusted to be uniform. Rotation, flipping, and elastic deformation operations are performed on the images to expand the training set sample size.

[0020] Model selection: Lightweight network is used, and pre-trained weights are used to accelerate convergence;

[0021] Feature extraction: The input is a 3D block of an image. Spatial features, including tumor edges and metabolically active regions, are extracted through multi-layer 3D convolution. Global average pooling is used to replace fully connected layers, and a fixed-length feature vector is output.

[0022] Metabolic feature definition: Extract SUV distribution features from PET images and extract texture features from CT images;

[0023] Output and Application: Image features and genetic data are stitched together and input into subsequent relapse prediction models.

[0024] Optionally, the steps of the time-series data analysis are as follows:

[0025] Data alignment and cleaning: Physiological and environmental data are sliced ​​using a sliding window at an hourly granularity, and missing values ​​in the physiological data are handled using linear interpolation;

[0026] LSTM network input features: physiological indicators and environmental data are input; the model structure is a two-layer LSTM, with each layer followed by Dropout, and the fully connected layer outputs a dynamic health index;

[0027] Training objective: To generate supervisory signals by combining physician annotations and clinical events;

[0028] Dynamic Index Visualization: The DHI is plotted as a line chart, overlaid with environmental mutation events.

[0029] Optionally, the steps for modeling the environmental exposure index are as follows:

[0030] Indicator Standardization and Weight Allocation: Air quality is converted to a 0-100 score. Water quality heavy metal concentration is monitored and scored according to the multiple of exceedance. Radiation dose is calculated based on ICNIRP limits, with measured values ​​proportionally converted to a 0-100 score. The PEI calculation formula is as follows:

[0031] ;

[0032] Thresholds and warnings: A high-risk warning is triggered when PEI > 70, and a vibration alert is sent via wearable device.

[0033] Optionally, the multimodal feature fusion can be implemented as follows:

[0034] Clinical feature processing: Structured data is converted into vectors through one-hot encoding or numerical normalization;

[0035] Image feature extraction: Deep features, including tumor volume, metabolic activity, and lymph node metastasis probability, are extracted from CT and PET images using a 3D convolutional neural network. The output feature vector is then concatenated with clinical features and reduced to 128 dimensions using PCA.

[0036] Molecular feature standardization: Gene mutations are converted into binary vectors, and biomarkers are standardized using Z-scores to eliminate dimensional differences;

[0037] Environmental Exposure Quantification: Normalize the PEI to the range [0,1] and merge it with other modal features.

[0038] Feature alignment example: Clinical features, imaging features, molecular features, and environmental features are concatenated into a 267-dimensional vector, which is then used as the model input.

[0039] Optionally, the risk heatmap can be constructed as follows:

[0040] Functional goal: To visually display high-risk areas for tumor metastasis through spatial mapping, assisting doctors in locating potential lesions;

[0041] Data inputs: The probability matrix of distant metastasis risk output by the model; DICOM data of the patient's CT and MRI images;

[0042] Spatial mapping: The human body is divided into standard anatomical regions based on image data. Each region corresponds to a transition probability value, and a color gradient is used to represent the probability level.

[0043] Optionally, the trend warning method is as follows:

[0044] Data input: Patient's historical test data; clinically defined critical thresholds;

[0045] Visualization Design: Plot dynamic curves of CEA and CA19-9 with time as the horizontal axis and concentration value as the vertical axis; mark the critical values ​​with horizontal dashed lines, and mark points exceeding the threshold with red dots; calculate the slope of the curve, and if the rate of increase is greater than 10% for three consecutive tests each month, trigger a trend warning label;

[0046] Risk zone division: Normal zone is green background; Warning zone is yellow background; Danger zone is red background;

[0047] Interactive features: Click on a data point to view the corresponding detection date, value, and difference from the threshold.

[0048] Optionally, the environmental correlation analysis can be performed in the following ways:

[0049] Data input: Patient's long-term environmental exposure data;

[0050] Dose-effect modeling: A comprehensive exposure index (PEI) is generated by combining air pollution, water quality, and radiation indicators using a weighted formula.

[0051] Scatter plot drawing: Plot a scatter plot with PEI as the horizontal axis and recurrence risk score as the vertical axis;

[0052] Trend fitting: Local weighted regression was used to fit the nonlinear curve to observe the monotonicity of PEI and risk; key points were marked to assist in clinical interpretation.

[0053] Statistical annotation: Displays Pearson coefficients.

[0054] This invention provides a long-term follow-up system after resection of adenocarcinoma at the esophagogastric junction, which has the following beneficial effects:

[0055] 1. The multimodal data acquisition layer of this invention covers clinical data, environmental data, behavioral data and genetic data, realizing multi-dimensional information fusion, solving the one-sidedness of traditional follow-up relying on a single indicator, and providing a comprehensive data foundation for accurate prediction.

[0056] 2. This invention cleans data through noise filtering, missing value imputation, and standardization, extracts key information by combining feature engineering, and uses deep learning to fuse multimodal data to dynamically predict recurrence risk and metastasis probability. The prediction accuracy is significantly higher than that of traditional clinical experience or single models, and it is especially sensitive to early minor recurrences.

[0057] 3. By generating risk heat maps, trend warning maps, and environmental correlation analysis maps, this invention enables doctors to quickly locate high-risk areas and develop targeted examinations; patients can understand the risks through intuitive charts, thus improving compliance.

[0058] 4. This invention automatically generates personalized follow-up plans based on risk scores, pushes customized rehabilitation videos, and automatically triggers doctor warnings for abnormal indicators, upgrading from passive monitoring to proactive prevention and control, reducing the rate of missed diagnoses and recurrence.

[0059] 5. This invention aligns multimodal data such as image features, time-series health index, and environmental exposure index, and generates a comprehensive assessment report through a lightweight model to assist doctors in making quick decisions. The combination strategy of 3D-CNN, LSTM and environmental modeling forms a technical closed loop, which is different from traditional single-modal analysis. The collaborative architecture of edge LSTM model and cloud large model takes into account both real-time performance and accuracy, and is suitable for long-term follow-up scenarios. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the follow-up system of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Example 1

[0063] Please refer to Figure 1, a long-term follow-up system after resection of adenocarcinoma of the esophagogastric junction, comprising:

[0064] Multimodal data acquisition layer: Acquires medical data from HIS (Hospital Information System), including electronic medical records, pathology reports, gene sequencing (such as HER2, MMR status), and DICOM imaging data; wearable devices collect real-time physiological indicators such as heart rate variability, blood oxygen saturation, activity calorie consumption, and deep sleep duration; wearable devices include smart bracelets and vests; embedded electrochemical sensors detect CEA and CA19-9, SPR biochips detect miRNA-21, air quality monitors (PM2.5, VOCs), water quality analyzers, and electromagnetic radiation dosimeters; the air quality monitors are PM2.5 and VOCs detectors, and the electromagnetic radiation dosimeter is installed in the patient's activity room; the water quality analyzer is a heavy metal ion analyzer installed at the faucet outlet; the principle of the electrochemical sensor detecting CEA and CA19-9 is that the target molecules undergo a redox reaction on the electrode surface, generating a current signal, formula: ( For the change of current, (where C is the sensitivity, A is the electrode area), an integrated micro three-electrode system (working electrode, counter electrode, reference electrode) is used. The current is read by a potentiostat, and the concentrations of CEA and CA19-9 can be determined by the current signal.

[0065] Intelligent data processing layer: Employs 3D convolutional neural networks to analyze CT and PET images, perform unstructured data processing, and extract tumor metabolic characteristics; utilizes LSTM networks to fuse physiological indicators and environmental data, perform time-series data analysis, and construct a dynamic health index; and models the environmental exposure index (PEI) based on WHO air quality guidelines and ICNIRP limits to calculate pollutant exposure indices.

[0066] Relapse prediction model layer: Clinical staging, molecular subtyping, treatment response, and environmental exposure features are mapped to a unified feature space for multimodal feature fusion; a spatiotemporal graph convolutional network (ST-GCN) deep learning model is used to capture the interaction of multiple factors and output the predicted values ​​of relapse probability and distant metastasis risk within three years;

[0067] Visualized decision-making layer: The probability of metastasis is marked by anatomical location (red → high incidence, blue → low risk), and a risk heat map is constructed; the dynamic curves of CEA and CA19-9 are displayed and compared with the critical threshold to provide trend warning; the dose-response curve of PEI index and recurrence risk is displayed to conduct environmental correlation analysis.

[0068] The acquisition of medical data is as follows: electronic medical records and pathology reports are obtained from the Hospital Information System (HIS) via the Fast Medical Interoperability Resource (FHIR) interface; DICOM format image data is obtained from the Picture Archiving System (PACS); gene testing results (such as HER2 and MMR status) are obtained from the Laboratory Information System (LIS); gene sequencing data: raw data in FASTQ / BAM format is transferred to the laboratory via API or FTP, and converted into structured variation data (such as VCF files) through bioinformatics pipelines (such as GATK); key information of unstructured text is extracted through NLP (Natural Language Processing) and mapped to standardized codes (such as ICD-10 and LOINC); image data is parsed using the DICOM standard to extract pixel matrices and metadata (such as scan parameters and patient ID); among them, unstructured text includes pathology reports, and key information includes tumor staging and immunohistochemical indicators, and multimodal alignment, model lightweighting, and privacy protection are performed;

[0069] Multimodal alignment: Matching image examination time with physiological / environmental data windows using timestamps;

[0070] Lightweight model: The number of parameters of the LSTM model deployed on edge devices is kept to within 100,000 (e.g., using depthwise separable convolution).

[0071] Privacy protection: Medical data is transmitted in encrypted form, and environmental data is stored after being anonymized.

[0072] In this embodiment, by connecting to the hospital's HIS / PACS / LIS system, standardized collection of electronic medical records, pathology reports, gene sequencing data (such as HER2 and MMR status), and imaging data (DICOM format) is achieved, avoiding manual input errors and ensuring the integrity and timeliness of clinical data. Combined with IoT monitoring (physiological indicators), molecular monitoring (tumor markers), and environmental sensing (air quality, water quality, etc.), it breaks through the limitations of traditional follow-up that relies solely on clinical data, providing more comprehensive feature input for recurrence risk assessment. Through smart bracelets / vests, heart rate variability, blood oxygen saturation, and activity calorie consumption are monitored, dynamically capturing changes in the patient's physiological state (such as postoperative fatigue and signs of infection), which is more sensitive than traditional regular check-ups. Embedded electrochemical sensors can detect markers such as CEA and CA19-9 every hour, more frequently than blood tests, and can detect signs of recurrence earlier (such as abnormal CEA growth). Real-time monitoring of PM2.5, heavy metal ions, electromagnetic radiation, etc., establishes a pollutant exposure index (PEI) to quantify the impact of environmental factors on prognosis (such as the association between high pollution exposure and recurrence risk).

[0073] By integrating clinical data (such as tumor staging), molecular features (such as miRNA-21 expression), environmental exposure (PEI), and physiological indicators (such as deep sleep duration) into a dynamic health profile, high-dimensional features are provided for deep learning models, thereby improving the accuracy of recurrence risk prediction.

[0074] Example 2

[0075] This embodiment is a further optimization based on Embodiment 1. Specifically, the steps for unstructured data processing are as follows:

[0076] Data preprocessing and enhancement: DICOM format CT and PET images are converted into three-dimensional matrices (e.g., 512×512×number of slices), and pixel intensity is unified by adjusting window width / window level (e.g., the HU value range of CT). The images are rotated (±10°), flipped (horizontal / vertical), and elastically deformed (simulating breathing motion) to expand the training set sample size.

[0077] Model selection: The 3DResNet-18 lightweight network was adopted, and convergence was accelerated by pre-trained weights. It was initialized on the ImageNet-3D dataset.

[0078] Feature extraction: The input is a 3D block of the image (e.g., 32×32×32 pixels). Spatial features, including tumor edges and metabolically active regions, are extracted through multi-layer 3D convolution. Global average pooling (GAP) is used to replace fully connected layers, and a fixed-length feature vector (e.g., 512-dimensional) is output.

[0079] Metabolic feature definition: Extract SUV (standard uptake value) distribution features from PET images, including maximum value, mean value, and entropy; extract texture features from CT images, including contrast and correlation of the gray-level co-occurrence matrix (GLCM).

[0080] Output and Application: Image features are concatenated with genetic data (such as HER2 status) and input into a lightweight GNN for subsequent relapse prediction model.

[0081] The steps for time series data analysis are as follows:

[0082] Data alignment and cleaning: Physiological data (such as heart rate and blood oxygen) and environmental data (such as PM2.5 and temperature) are sliced ​​in 1-hour increments using sliding windows. Linear interpolation is used to handle missing values ​​in physiological data (such as Kalman filtering to correct for sudden changes in heart rate). If more than 30% of the environmental data is missing, the data in that window is discarded.

[0083] LSTM network input characteristics: Input physiological indicators (average heart rate, blood oxygen saturation, deep sleep duration within the current window), environmental data (PM2.5, ... (Concentration, water temperature); The model structure is a two-layer LSTM (64 hidden units), with each layer followed by Dropout (p=0.2). The fully connected layer outputs a dynamic health index (DHI, range 0-100), with a lower value indicating a higher risk.

[0084] Training objective: To generate supervisory signals by combining physician annotations (such as "fatigue period" and "infection period") with clinical events (such as hospitalization records);

[0085] Dynamic Index Visualization: The DHI is plotted as a line graph, overlaid with sudden environmental events (such as a sudden increase in PM2.5) to help doctors determine the triggers.

[0086] The steps for modeling the Environmental Exposure Index (PEI) are as follows:

[0087] Indicator standardization and weight allocation: Air quality index (AQI) is converted into a 0-100 score (e.g., AQI > 300 is recorded as 100), and water quality heavy metal concentrations (such as lead and mercury) are monitored. If the limit in GB5749-2006 is exceeded, the score will be calculated based on the multiple of the exceedance (e.g., exceeding the limit by 2 times is counted as 20). The radiation dose will be converted to a score of 0-100 based on the ICNIRP limit (e.g., the public annual exposure limit is 1 mSv). The PEI calculation formula is as follows:

[0088] ;

[0089] Example: If AQI=150 (corresponding to 60 points), the lead content in the water exceeds the standard by 1.5 times (15 points), and the radiation dose is 0.8 mSv (corresponding to 80 points), then: PEI=0.3×60+0.4×15+0.3×80=18+6+24=48;

[0090] Thresholds and warnings: When PEI > 70, a high-risk warning is triggered (such as recommending to reduce outdoor activities), and a vibration reminder is sent via wearable device.

[0091] In this embodiment, unstructured image intelligent analysis automatically analyzes CT and PET images through 3D convolutional neural networks (3D-CNN), accurately extracts tumor metabolic features (such as SUV value distribution and texture features), replaces traditional manual image reading, reduces subjective errors, and combines image features with gene detection results (such as HER2 and MMR status) to construct a multi-dimensional recurrence prediction model, which helps to discover small lesions or metabolic abnormalities and improves the early recurrence detection rate.

[0092] The LSTM network integrates physiological and environmental data, combining time-series data such as heart rate, blood oxygen, and activity calories collected by wearable devices, and environmental factors such as air quality (PM2.5) and water quality to dynamically generate a health index (DHI), which reflects the patient's recovery trend in real time. The model captures sudden changes in physiological indicators (such as a sharp drop in deep sleep duration and blood oxygen fluctuations), and combines environmental risks (such as peak air pollution) to provide early warnings of the risk of complications such as infection and anastomotic leakage.

[0093] Based on WHO standards and ICNIRP limits, environmental parameters such as air quality (AQI), water quality exceeding standards, and electromagnetic radiation are converted into standardized scores (PEI) to quantify the impact of environmental risks on patients. If the PEI is too high (such as severe air pollution), the system can suggest that patients adjust their living environment or strengthen respiratory protection to reduce the risk of secondary damage caused by external factors.

[0094] In this embodiment, multimodal data such as image features, time-series health index, and environmental exposure index are aligned, and a comprehensive assessment report is generated through a lightweight model (such as XGBoost) to assist doctors in making rapid decisions. The combination strategy of 3D-CNN+LSTM+environmental modeling forms a technical closed loop, which is different from traditional single-modal analysis. The collaborative architecture of edge LSTM model and cloud large model takes into account both real-time performance and accuracy, and is suitable for long-term follow-up scenarios. It incorporates environmental exposure (such as water quality and radiation) into the postoperative rehabilitation assessment system for cancer, filling a gap in the field.

[0095] Example 3

[0096] This embodiment is a further optimization based on embodiment 1 or 2. Specifically, the implementation method of multimodal feature fusion is as follows:

[0097] Clinical feature processing: Structured data (such as TNM staging, pathological type, treatment method) are converted into vectors through one-hot encoding or numerical normalization. For example, TNM staging (T1-T4) is converted into a 4-dimensional binary vector. Treatment response (such as chemotherapy sensitivity) is mapped to continuous values ​​(0-1) according to the level (high / medium / low) marked by the doctor.

[0098] Image feature extraction: Deep features, including tumor volume, metabolic activity (SUVmax), and lymph node metastasis probability, are extracted from CT and PET images using a 3D convolutional neural network (3D-CNN). The output feature vector (e.g., a 512-dimensional floating-point number) is concatenated with clinical features and then reduced to 128 dimensions using PCA.

[0099] Molecular feature standardization: Gene mutations (such as HER2, MMR status) are converted into binary vectors (mutated = 1, no mutation = 0), and biomarkers (such as CEA, CA19-9) are standardized by Z-score to eliminate dimensional differences;

[0100] Environmental Exposure Quantification: Normalize the PEI to the range [0,1] and merge it with other modal features.

[0101] Feature alignment example: Clinical features (128-dimensional), imaging features (128-dimensional), molecular features (10-dimensional), and environmental features (1-dimensional) are concatenated into a 267-dimensional vector, which is used as the model input;

[0102] Through linear transformation ( Heterogeneous features are projected onto the same dimension, for example, mapping "PM2.5 concentration" and "HER2 status" to comparable values; the fused features F are directly input into the subsequent model, avoiding the subjectivity of manual feature selection in traditional methods;

[0103] The mathematical formula is:

[0104] in, For clinical feature vectors, such as TNM staging ( , , ), surgical method ( The data source is the medical record system; This is a feature vector of environmental exposure, such as the annual average of PM2.5 ( The data sources are air quality monitors (PM2.5, VOCs), water quality analyzers (heavy metal ions), and electromagnetic radiation dosimeters. Molecular feature vectors, such as HER2 expression levels ( DNA repair gene mutations ( The data source is pathology reports; For treatment response feature vectors, such as chemotherapy sensitivity scores ( ), radiotherapy dose ( ); This is the feature fusion weight matrix, used to unify the scale of different features; This is a bias term that corrects for the bias in the characteristic distribution.

[0105] The Spatiotemporal Graph Convolutional Network (ST-GCN) deep learning model is as follows (capturing the dynamic correlation of multiple factors in time and space, and outputting relapse probability and metastasis risk):

[0106] Input layer: Includes nodes, edges, and a timeline. Features of each modality (clinical, imaging, molecular, environmental) are treated as independent nodes. The relationships between nodes are defined as edges based on medical knowledge. For example:

[0107] Environmental exposure (PEI) → gene mutation (such as DNA damage caused by air pollution);

[0108] Imaging features (tumor metabolism) → Molecular features (e.g., HER2 positivity is associated with high metabolism);

[0109] Historical data is unfolded in chronological order (such as monthly physiological indicators and quarterly environmental data) to form a spatiotemporal diagram as a time axis;

[0110] Graph Convolutional Layer (GCN): Extracts spatial dependencies between nodes, including the interaction weights between environmental exposure and gene mutation. Formula: = ;in, Indicates the first The hidden states of a layer graph convolution represent the propagation of features in the graph; For example, activation functions (such as ReLU); The graph structure adjacency matrix defines the relationships between features (such as the correlation between "environmental exposure" and "inflammatory factors"), and the data is sourced from medical knowledge / data-driven construction. The graph convolution weight matrix is ​​used to learn the strength of interactions between features. For bias terms;

[0111] Spatiotemporal modeling: This includes the time dimension, dividing the follow-up period into time slices (e.g., every 3 months as a node), and capturing feature changes over time using ST-GCN (e.g., ...). (Seasonal fluctuations); and spatial dimensions, constructing anatomical relationship maps (such as the connection relationships between high-risk sites of liver and peritoneal metastasis);

[0112] Output layer: The formula is = ,in, For predictive values ​​(e.g., 3-year recurrence probability) Risk of liver metastasis ); , These are the output layer parameters.

[0113] The recurrence prediction model layer also includes an attention mechanism, which assigns time decay coefficients to different features (such as increasing the weight of recent environmental exposure) to dynamically adjust the weights.

[0114] The mathematical formula for dynamic weight adjustment (attention mechanism) is as follows (dynamically assigning weights based on feature importance to strengthen the influence of key factors):

[0115] ;

[0116] ;

[0117] in, Indicates time Feature weights (e.g.) Indicates the current (The weight of exposure); the data source is dynamic calculation using the attention mechanism; This represents the attention query vector, used to evaluate feature importance; , These are the learning parameters for the attention mechanism; Indicates time Multimodal feature vectors (such as the first) Quarterly , ); These are weighted features used as model input;

[0118] Time decay coefficient: through Adjusting the weights of historical features, for example:

[0119] Recent environmental exposure (such as in the past 3 months) High weight () →1);

[0120] Long-term data has low weight ( →0).

[0121] In this embodiment, by integrating multi-dimensional data such as clinical staging (TNM), molecular subtyping (e.g., HER2, MMR status), treatment response (sensitivity to radiotherapy and chemotherapy), and environmental exposure (PM2.5, heavy metals in water), a more comprehensive relapse risk assessment model is constructed, avoiding the one-sidedness of traditional single-modal analysis. Compared with models that rely solely on clinical indicators, multi-modal fusion can detect signs of relapse 3-6 months earlier (e.g., by combining the synergistic effect of environmental exposure and gene mutation).

[0122] By modeling the nonlinear relationships between clinical, molecular, and environmental characteristics using graph structures (e.g., "air pollution aggravates inflammatory response → promotes micrometastasis"), hidden associations that are difficult to identify by traditional statistical models are captured, improving the prediction accuracy of distant metastasis (e.g., liver metastasis) by 20%-30%, especially in the early asymptomatic stage. It outputs the 3-year recurrence probability and metastasis risk classification (e.g., low / medium / high risk), assisting in determining the follow-up frequency (e.g., high-risk patients should have a follow-up CT scan every 3 months), reducing excessive examinations for low-risk patients (reducing unnecessary CT scans by 30%), while high-risk patients receive more intensive monitoring.

[0123] By employing an attention mechanism and an LSTM network, the contribution of different features to the risk score is automatically adjusted (e.g., increasing weights when there is recent environmental exposure or abnormal molecular indicators), avoiding the lag of static models. Feature weights are dynamically adjusted based on the patient's current state (e.g., recent environmental exposure, fluctuations in physiological indicators), strengthening the impact of key risk factors (e.g., high PM2.5 exposure one month before surgery) on prediction, making the prediction results more consistent with the patient's real-time health status, reducing the risk of missed diagnoses. The attention weights generate a ranking of risk factors (e.g., the combination of "elevated CEA + excessive lead in water" indicates a high risk of recurrence), helping doctors quickly locate intervention targets. The top 3 risk factors are marked in the clinical report to guide personalized treatment (e.g., strengthening heavy metal detoxification or adjusting targeted drugs).

[0124] In this embodiment, by combining graph convolution with temporal convolution, the problem of multi-factor dynamic interaction modeling is solved. Unlike traditional machine learning, it introduces a time decay function and attention mechanism to overcome the limitation that static models cannot adapt to real-time changes in the environment and physiology.

[0125] Example 4

[0126] This embodiment is a further optimization based on Embodiment 2. Specifically, the risk heatmap is constructed as follows:

[0127] Data input: The model output probability matrix of "distant metastasis risk", classified by anatomical location, such as liver, lung, bone, brain; DICOM data of the patient's CT / MRI images, used for anatomical structure localization;

[0128] Spatial mapping: Based on the image data, the human body is divided into standard anatomical regions, including the liver divided into left and right lobes, and the skeleton divided into the spine and pelvis. Each region corresponds to a metastasis probability value, and a color gradient is used to represent the probability, such as red→yellow→blue. Red (probability > 70%) is extremely high risk, and blue (probability < 30%) is low risk. This is then overlaid onto a medical image base map, such as a CT cross-section, to form a dual-modal display of "image + heat map".

[0129] Dynamic interaction: Supports displaying specific probability values ​​and anatomical names when hovering the mouse; provides a threshold adjustment slider (e.g., only displaying areas >50%) to filter low-risk areas;

[0130] Example scenario: A rectal cancer patient's model predicts a 65% probability of liver metastasis and a 15% probability of lung metastasis. In the heat map, the liver area is dark red, the lungs are light yellow, and other areas are blue. Doctors can focus on the details of the liver image first.

[0131] The trend warning method (comparison of CEA and CA19-9 dynamic curves with thresholds) is as follows (functional objective: to provide early warning of relapse risk and indicate the timing of intervention through time series analysis of biomarkers):

[0132] Data input: Patient's historical test data (e.g., CEA and CA19-9 concentrations per quarter); clinically defined thresholds (e.g., CEA > 5 ng / mL is abnormal);

[0133] Visualization Design: Plot dynamic curves of CEA and CA19-9 with time on the horizontal axis and concentration on the vertical axis; mark the critical value (e.g., CEA = 5 ng / mL) with a horizontal dashed line, and mark points exceeding the threshold with red dots; calculate the slope of the curve, and if the rate of increase is greater than 10% for three consecutive tests each month, trigger the "Trend Warning" label.

[0134] Risk zone classification: Normal zone (curve below the threshold and with a stable slope) is marked with a green background; Warning zone (curve close to the threshold or with fluctuating slope) is marked with a yellow background; Danger zone (curve exceeding the threshold or with a slope greater than 10% each month) is marked with a red background.

[0135] Interactive features: Click on a data point to view the corresponding test date, value, and difference from the threshold; supports comparing curves of multiple patients to identify abnormal patterns;

[0136] Example scenario: A patient's CEA level rises from 3 ng / mL to 7 ng / mL within 6 months, with a slope of 13% / month. The system marks this as a "trend warning," indicating that the follow-up interval should be shortened and PET-CT screening should be initiated.

[0137] The environmental association analysis (dose-response curve of PEI index and relapse risk) is conducted as follows (quantifying the strength of the association between environmental exposure and relapse risk, providing a basis for lifestyle intervention):

[0138] Data input: Patient's long-term environmental exposure data (such as annual average PM2.5 in residential areas, heavy metal content in water quality); relapse risk score calculated by the model (0-100 points);

[0139] Dose effect modeling: Based on air pollution, water quality, and radiation indicators, a weighted formula is used to generate the comprehensive exposure index (PEI): PEI = 0.3 × standardized PM2.5 + 0.4 × number of water quality exceedances + 0.3 × radiation dose;

[0140] Scatter plot drawing: Plot a scatter plot with PEI as the horizontal axis and recurrence risk score as the vertical axis;

[0141] Trend fitting: Local weighted regression (LOESS) was used to fit the nonlinear curve to observe the monotonicity of PEI and risk (e.g., risk index increases when PEI>1.2); key points were marked (e.g., risk = 50% when PEI=1.0) to assist in clinical interpretation;

[0142] Statistical annotations: Display Pearson correlation coefficient (e.g., r=0.62, p<0.01); mark high-risk intervals (e.g., risk >70% when PEI>1.5).

[0143] Example scenario: A patient's residential area has consistently high PM2.5 levels, with a PEI of 1.8. The model shows that the patient's relapse risk score is 85. The dose-response curve shows that the risk increases sharply after PEI > 1.5. The system recommends that the patient move to a low-pollution area and increase the use of air purifiers.

[0144] In this embodiment, the risk heatmap marks the probability of metastasis by anatomical sites, such as liver, lungs, and bones, transforming abstract prediction results into an intuitive spatial distribution map. This helps doctors quickly identify high-risk areas, guide the focus of imaging examinations, such as the scope of CT and PET scans, and reduce missed diagnoses. It also allows for the development of enhanced follow-up strategies for high-risk sites (such as areas marked in red), such as shortening the interval between follow-up examinations. Furthermore, it can assist in surgical decisions, such as providing a reference for the scope of prophylactic lymph node dissection.

[0145] Trend warning compares the dynamic curves of biomarkers such as CEA and CA19-9 with thresholds to reflect changes in tumor burden in real time. If the slope of the curve exceeds the threshold (e.g., CEA growth rate > 10% / month), it indicates a sharp increase in the risk of recurrence, and chemotherapy or targeted therapy can be started earlier. For low-risk patients with stable trends, the follow-up period can be extended to save medical resources. By comparing the curves of multiple patients, individual abnormal fluctuations can be found, such as a step-like increase, which indicates potential recurrence.

[0146] Environmental correlation analysis uses dose-response curves of PEI (pollutant exposure index) and recurrence risk to clarify the impact of environmental exposure on prognosis. If PEI > 1.5, the risk of recurrence increases significantly, and patients can be advised to improve their living environment, such as staying away from pollution sources and using air purifiers. Visualized curves can enhance patients' awareness of environmental factors and promote healthy lifestyles, such as quitting smoking and reducing exposure to smog. More intensive monitoring plans can be developed for patients with high pollution exposure, such as increasing the frequency of blood tumor marker testing.

[0147] In this embodiment, multimodal visualization (heatmap, trend curve, dose-response diagram) transforms complex data into easily understandable graphics, lowering the threshold for doctor-patient communication. By intuitively displaying the distribution of recurrence risk and environmental impact, it improves patient compliance, such as encouraging patients to proactively avoid polluted environments. Doctors can work with patients to develop follow-up plans based on the visualization results, such as choosing PET-CT for high-risk patients and ultrasound for low-risk patients. Clearly defining the risk level (such as the low-risk blue zone) can alleviate patient anxiety and improve their quality of life.

[0148] Example 5

[0149] This embodiment is a further optimization based on embodiment 4. Specifically, the intelligent intervention layer: the personalized follow-up plan generation model automatically generates examination suggestions based on the risk score;

[0150] The framework for the personalized follow-up plan generation model is as follows:

[0151] Risk grading rule engine: Input the risk score (0-100 points), clinical stage (TNM), and molecular subtype (such as HER2 status) output by the recurrence prediction model.

[0152] Rule base example:

[0153] High risk (score ≥70): Enhanced CT and PET-CT scans every 3 months, and monthly blood tests (CEA and CA19-9).

[0154] Medium risk (40-70 points): Enhanced CT scan every 6 months, blood test every 3 months;

[0155] Low risk (<40 points): Annual routine CT scan, blood test every 6 months;

[0156] Dynamic adjustment logic: If an abnormality is found in a certain inspection (such as a sudden increase in CEA), it will be automatically upgraded to a higher-level follow-up plan;

[0157] Intelligent matching of examination items: Based on the risk of anatomical location, such as the high probability of liver metastasis indicated by heat map, targeted examinations such as liver MRI are given priority; combined with patient compliance, such as reducing invasive examinations for elderly patients, the frequency of gastroscopy and colonoscopy is dynamically optimized.

[0158] Execution and feedback mechanism: Automatically generate follow-up calendars and push them to the patient's APP to remind them of examination times; examination data is transmitted back to the system in real time, triggering the model to reassess the risk score and update the subsequent plan;

[0159] Example scenario: A patient's risk score is 85. The system generates a plan of "PET-CT in the first month + enhanced CT every 3 months + CEA test every month". If the CEA drops to the normal range in the second month, the plan will be adjusted to CT examination every 6 months.

[0160] The doctor-patient interaction platform pushes customized rehabilitation training videos, and abnormal indicators automatically trigger doctor alerts.

[0161] The doctor-patient interaction platform architecture includes a patient end, a doctor end, and a data exchange mechanism:

[0162] Functional objective: To achieve doctor-patient collaboration through two-way data flow, ensuring intervention implementation and risk control.

[0163] On the patient side: Based on the type of surgery (such as esophagectomy), the system pushes phased rehabilitation videos, such as swallowing training 1 week after surgery and physical recovery 1 month after surgery; it also pushes "reduce outdoor activities" prompts based on environmental data (such as high PM2.5); patients can enter symptoms through the APP, such as abdominal pain and weight loss, and the system automatically associates them with the risk model and marks potential recurrence signals.

[0164] On the doctor's end: A "red alert" is triggered if the CEA level increases by more than 20% twice in a month, and the alert is sent to the attending physician; sleep monitoring shows that the Apnea-Hypopnea Index (AHI) is greater than 15, indicating "risk of sleep apnea syndrome";

[0165] Data sharing mechanism: Patient-side APP and wearable device data are synchronized to the cloud in real time; Doctors can view complete timeline data within the system, such as CEA curves, sleep reports, and follow-up compliance over the past year;

[0166] Example scenario: A patient's blood oxygen level drops to 90% according to the wristband monitoring at night. The system automatically pushes a "hypoxia risk" warning to the doctor and suggests that the patient undergo a lung function test. After confirmation, the doctor adjusts the follow-up plan to "monthly blood gas analysis and home oxygen therapy guidance".

[0167] In this embodiment, examination plans are dynamically generated based on recurrence risk scores (combined with anatomical metastasis probability, biomarker trends, environmental exposure, etc.). For example, high-risk individuals undergo enhanced CT scans every 3 months, while low-risk individuals have routine annual examinations. Enhanced imaging screening is conducted for high-recurrence-risk areas (such as the liver and peritoneum) to detect small lesions early. Excessive examinations for low-risk patients are avoided, such as reducing the frequency of unnecessary PET-CT scans and optimizing the allocation of medical resources. If the risk score increases, such as with abnormal CEA growth, intensive follow-up is automatically triggered, such as monthly blood tests and quarterly MRI scans, to prevent missed diagnoses.

[0168] The patient-doctor interaction platform pushes customized rehabilitation videos (such as swallowing training and physical recovery exercises) and self-check reminders for abnormal symptoms on the patient's end; the doctor's end automatically alerts high-risk indicators, such as a sudden increase in CEA and a decrease in blood oxygen, supporting multidisciplinary consultation recommendations; through visual education (such as video demonstrations) and mobile reminders (such as medication alarms), the follow-up dropout rate is reduced; abnormal data triggers doctor alerts in real time, such as "CEA growth rate reaches 25% / month, and a follow-up enhanced CT scan is recommended," reducing treatment delays; transparent data sharing (such as risk score change curves) promotes collaborative decision-making and improves treatment satisfaction.

[0169] In summary, the multimodal data acquisition layer of this invention covers clinical data (imaging, pathology, blood biomarkers), environmental data (air pollution, living environment), behavioral data (sleep, nutrition, exercise), and genetic data, realizing the fusion of multi-dimensional information from "biological-environmental-social" dimensions. This solves the one-sidedness of traditional follow-up relying on a single indicator and provides a comprehensive data foundation for accurate prediction.

[0170] Data is cleaned using techniques such as noise filtering, missing value imputation, and standardization. Key information is extracted by combining the Environmental Exposure Index (PEI) feature engineering. Spatiotemporal graph convolutional networks are used to fuse multimodal data to dynamically predict recurrence risk and metastasis probability, such as in the liver, lungs, and bones. The prediction accuracy is significantly higher than that of traditional clinical experience or single models, especially for early micro-recurrences.

[0171] By generating risk heatmaps to annotate anatomical sites with high recurrence probability, dynamic curves of CEA and CA19-9, and the dose-response relationship between PEI and recurrence risk, doctors can quickly locate high-risk areas and develop targeted examinations, such as prioritizing screening for high-probability metastatic sites; patients can understand the risks through intuitive charts, improving compliance, such as proactively avoiding polluted environments.

[0172] Personalized follow-up plans are automatically generated based on risk scores. For example, high-risk individuals can undergo enhanced CT scans every 3 months, receive customized rehabilitation videos, and have abnormal indicators automatically trigger doctor alerts. This upgrades from passive monitoring to proactive prevention and control, reducing the rate of missed diagnoses and recurrence.

[0173] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A long-term follow-up system after resection of adenocarcinoma of the esophagogastric junction, characterized in that, include: Multimodal data acquisition layer: Acquires medical data from the hospital information system; Wearable devices collect physiological indicators; In addition, embedded electrochemical sensors, air quality monitors, water quality analyzers, and electromagnetic radiation dosimeters; Intelligent data processing layer: Employs 3D convolutional neural networks to analyze CT and PET images, perform unstructured data processing, and extract tumor metabolic characteristics; utilizes LSTM networks to fuse physiological indicators and environmental data, perform time-series data analysis, and construct a dynamic health index; and models environmental exposure indexes to calculate pollutant exposure indices. Relapse prediction model layer: Maps clinical staging, molecular subtyping, treatment response, and environmental exposure features to a unified feature space and performs multimodal feature fusion; A spatiotemporal graph convolutional network is used to capture the interactions of multiple factors and output the recurrence probability. Visualized decision-making layer: Mark the metastasis probability by anatomical location and construct a risk heat map; Displays a dynamic curve compared to a critical threshold to provide trend warnings; The curves showing the relationship between the PEI index and recurrence risk are presented, and an environmental correlation analysis is conducted. Intelligent intervention layer: The personalized follow-up plan generation model automatically generates examination suggestions based on risk scores; Doctor-patient interaction platform: Pushes customized rehabilitation training videos, and abnormal indicators automatically trigger doctor alerts.

2. The long-term follow-up system after resection of adenocarcinoma of the esophagogastric junction according to claim 1, characterized in that: The medical data is acquired as follows: electronic medical records and pathology reports are obtained from the hospital information system through the Rapid Medical Interoperability Resource Interface; DICOM format image data is obtained from the image archiving system; gene testing results and gene sequencing data are obtained from the laboratory information system; key information of unstructured text is extracted through NLP and mapped into standardized codes; image data is parsed through the DICOM standard to extract pixel matrices and metadata; and multimodal alignment, model lightweighting, and privacy protection are performed.

3. The long-term follow-up system after resection of adenocarcinoma of the esophagogastric junction according to claim 1, characterized in that: The steps for unstructured data processing are as follows: Data preprocessing and enhancement: DICOM format CT and PET images are converted into three-dimensional matrices, and pixel intensity is adjusted to be uniform. Rotation, flipping, and elastic deformation operations are performed on the images to expand the training set sample size. Model selection: Lightweight network is used, and pre-trained weights are used to accelerate convergence; Feature extraction: The input is a 3D block of an image. Spatial features, including tumor edges and metabolically active regions, are extracted through multi-layer 3D convolution. Global average pooling is used to replace fully connected layers, and a fixed-length feature vector is output. Metabolic feature definition: Extract SUV distribution features from PET images and extract texture features from CT images; Output and Application: Image features and genetic data are stitched together and input into subsequent relapse prediction models.

4. The long-term follow-up system after resection of adenocarcinoma of the esophagogastric junction according to claim 1, characterized in that: The steps for time-series data analysis are as follows: Data alignment and cleaning: Physiological and environmental data are sliced ​​using a sliding window at an hourly granularity, and missing values ​​in the physiological data are handled using linear interpolation; LSTM network input characteristics: input physiological indicators and environmental data; The model structure is a two-layer LSTM, with each layer followed by Dropout, and the fully connected layer outputs a dynamic health index; Training objective: To generate supervisory signals by combining physician annotations and clinical events; Dynamic Index Visualization: The DHI is plotted as a line chart, overlaid with environmental mutation events.

5. The long-term follow-up system after resection of adenocarcinoma of the esophagogastric junction according to claim 1, characterized in that: The steps for modeling the environmental exposure index are as follows: Indicator Standardization and Weight Allocation: Air quality is converted to a 0-100 score. Water quality heavy metal concentration is monitored and scored according to the multiple of exceedance. Radiation dose is calculated based on ICNIRP limits, with measured values ​​proportionally converted to a 0-100 score. The PEI calculation formula is as follows: ； Thresholds and warnings: A high-risk warning is triggered when PEI > 70, and a vibration alert is sent via wearable device.

6. The long-term follow-up system after resection of adenocarcinoma of the esophagogastric junction according to claim 1, characterized in that: The multimodal feature fusion is implemented as follows: Clinical feature processing: Structured data is converted into vectors through one-hot encoding or numerical normalization; Image feature extraction: Deep features, including tumor volume, metabolic activity, and lymph node metastasis probability, are extracted from CT and PET images using a 3D convolutional neural network. The output feature vector is then concatenated with clinical features and reduced to 128 dimensions using PCA. Molecular feature standardization: Gene mutations are converted into binary vectors, and biomarkers are standardized using Z-scores to eliminate dimensional differences; Environmental Exposure Quantification: Normalize PEI to the range of [0,1] and merge it with other modal features; Feature alignment example: Clinical features, imaging features, molecular features, and environmental features are concatenated into a 267-dimensional vector, which is then used as the model input.

7. The long-term follow-up system after resection of adenocarcinoma of the esophagogastric junction according to claim 1, characterized in that: The spatiotemporal graph convolutional network deep learning model is as follows: Input layer: includes nodes, edges, and timeline. Features of each clinical, imaging, molecular, and environmental data modality are used as independent nodes, and the relationships between nodes are defined as edges based on medical knowledge. The historical data of monthly physiological indicators and quarterly environmental data are arranged in chronological order to form a spatiotemporal diagram as the time axis. Graph convolutional layers: extract spatial dependencies between nodes, including the interaction weights between environmental exposure and gene mutation; Spatiotemporal modeling: This includes the time dimension, which divides the follow-up period into time slices and uses ST-GCN to capture the changes in features over time; and the spatial dimension, which constructs an anatomical correlation diagram. Output layer: The formula is = ,in, For predicted values, Indicates the first The hidden states of a layered graph convolution represent the propagation of features in the graph. , These are the output layer parameters.

8. The long-term follow-up system after resection of adenocarcinoma of the esophagogastric junction according to claim 1, characterized in that: The risk heatmap is constructed as follows: Functional goal: To visually display high-risk areas for tumor metastasis through spatial mapping, assisting doctors in locating potential lesions; Data inputs: The probability matrix of distant metastasis risk output by the model; DICOM data of the patient's CT and MRI images; Spatial mapping: The human body is divided into standard anatomical regions based on image data. Each region corresponds to a transition probability value, and a color gradient is used to represent the probability level.

9. The long-term follow-up system after resection of adenocarcinoma of the esophagogastric junction according to claim 1, characterized in that: The trend warning method is as follows: Data input: Patient's historical test data; clinically defined critical thresholds; Visualization Design: Plot dynamic curves of CEA and CA19-9 with time on the horizontal axis and concentration on the vertical axis; mark the critical values ​​with horizontal dashed lines, and mark points exceeding the threshold with red dots; calculate the slope of the curve, and if the rate of increase is greater than 10% for three consecutive tests each month, trigger a trend warning label; Risk zone division: Normal zone is green background; Warning zone is yellow background; Danger zone is red background; Interactive features: Click on a data point to view the corresponding detection date, value, and difference from the threshold.

10. A long-term follow-up system for esophagogastric junction adenocarcinoma resection according to claim 1, characterized in that: The environmental correlation analysis method is as follows: Data input: Patient's long-term environmental exposure data; Dose-effect modeling: A comprehensive exposure index (PEI) is generated by combining air pollution, water quality, and radiation indicators using a weighted formula. Scatter plot drawing: Plot a scatter plot with PEI as the horizontal axis and recurrence risk score as the vertical axis; Trend fitting: Local weighted regression was used to fit the nonlinear curve to observe the monotonicity of PEI and risk; key points were marked to assist in clinical interpretation. Statistical annotation: Displays Pearson coefficients.

Images