Machine learning based method and system for remote volume management of heart failure patients

Abstract

The embodiment of the present disclosure provides a kind of machine learning-based heart failure patient remote volume management method and system, comprising: collecting multi-source data, accurately identifying and nutrition analysis to diet image by deep learning model, combine the time trend analysis of physiological monitoring data and the professional evaluation of clinical test data, construct multilevel prediction model, realize the accurate prediction of volume overload risk;Automatic generation of graded warning information according to risk level, and matching individualized intervention scheme is pushed to mobile application terminal.The embodiment of the present disclosure can effectively improve the self-management ability of heart failure patients, realize early intervention of volume problem.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, specifically to a remote volume management method and system for heart failure patients based on machine learning, used to achieve remote monitoring, assessment and intervention management of the volume status in heart failure patients. Background Technology

[0002] Heart failure is a common end-stage manifestation of cardiovascular disease. It is caused by impaired cardiac pumping function, leading to volume imbalances in the body, often resulting in symptoms such as fluid retention and dyspnea. With the development of medical technology, remote management of heart failure patients has become an important research direction, aiming to monitor and intervene in patients' daily fluid volume status in real time through information technology.

[0003] Traditional management of heart failure patients primarily relies on in-hospital follow-up and patient self-monitoring, such as regular weight measurement and recording of diet and medication administration. With the development of IoT technology, some remote monitoring systems have been applied clinically, such as monitoring patients' heart rate, blood pressure, and other physiological indicators through wearable devices, or measuring and uploading weight data using dedicated equipment. Currently, more advanced remote management systems for heart failure patients typically use single or limited data sources for volume assessment, such as relying mainly on weight changes or biochemical indicators like NT-proBNP to indirectly assess the patient's fluid status. While these systems can achieve basic remote monitoring functions, their data acquisition methods are relatively limited, lacking the ability to accurately quantify and analyze factors that directly affect volume balance, such as daily dietary intake, water content, and sodium content.

[0004] The main problems with existing technologies include: on the one hand, the lack of comprehensive analysis capabilities for multimodal data, making it impossible to effectively integrate patients' dietary images, physiological indicators, and clinical data; on the other hand, the accuracy and personalization of prediction models are insufficient, making it impossible to provide accurate warnings of capacity overload risk based on individual patient differences, thus limiting the timeliness and pertinence of intervention measures. Summary of the Invention

[0005] The purpose of this disclosure is to address the problems existing in the prior art by providing a remote volume management method and system for heart failure patients based on machine learning. This disclosure achieves accurate assessment and early warning of the volume status of heart failure patients through multimodal data acquisition and analysis, and provides personalized intervention plans, thereby improving the efficiency and effectiveness of remote management of heart failure patients.

[0006] To achieve the above objectives, the present disclosure provides the following technical solutions: In a first aspect, embodiments of this disclosure provide a remote volume management method for heart failure patients based on machine learning, including: Multimodal data of patients with heart failure are acquired, including dietary image data, physiological monitoring data and clinical test data. The multimodal data is preprocessed to obtain a preprocessed multimodal dataset. Multi-source data analysis is performed on the preprocessed multimodal dataset, including image recognition and nutritional analysis of dietary image data, time series analysis of physiological monitoring data, and indicator evaluation of clinical laboratory data. Features are extracted and feature engineering is performed to obtain a feature-engineered dataset. A multi-level prediction model is constructed based on the feature engineering dataset. The multi-level prediction model includes a first-stage feature-focused model group and a second-stage meta-learning integration model. The first-stage feature-focused model group outputs preliminary prediction results, which are then input into the second-stage meta-learning integration model for integrated processing. The model outputs a capacity overload risk value and obtains the final risk prediction result. Based on the final risk prediction results, risk is classified, classified early warning information is generated and pushed to the mobile application terminal to obtain the early warning push results. Based on the early warning push results, a personalized intervention plan is matched and pushed to the mobile application terminal, providing intervention execution confirmation and effect tracking functions to obtain intervention management results.

[0007] As one implementation method, the preprocessing of the multimodal data to obtain a preprocessed multimodal dataset includes: Based on the multimodal data, a data cleaning algorithm is applied to detect and process missing values, outliers, and duplicate records in the data. A threshold range is set for the physiological monitoring data, and physiological monitoring data that exceeds the threshold range is marked or corrected. The dietary image data is subjected to quality screening to remove blurry, dark, or incomplete dietary images. The clinical test data is subjected to integrity verification to obtain the cleaned multimodal data. The cleaned multimodal data is format-standardized. The physiological monitoring data and the clinical test data are standardized with unified timestamp format and the units and precision of numerical data are standardized. The dietary image data is normalized in size and converted in color space, and naming and storage rules are established to obtain format-standardized multimodal data. Key features, including weight change rate, cumulative sodium intake, and medication adherence indicators, are extracted from the standardized multimodal data. Data from different sources and types are then time-aligned and feature-fused to obtain the preprocessed multimodal dataset.

[0008] As one implementation method, image recognition and nutritional analysis are performed on the dietary image data in the preprocessed multimodal dataset, including: The dietary image data in the preprocessed multimodal dataset is subjected to image preprocessing, including brightness adjustment, contrast enhancement, cropping and scaling, to obtain the preprocessed dietary images. A labeled food image dataset is obtained, which contains various types of food image data and corresponding water content and sodium content parameters. A convolutional neural network model is constructed and trained using the transfer learning method to obtain a trained food recognition model. The preprocessed food image is input into the trained food recognition model, and the model infers and identifies the food category in the image to obtain the food recognition result. Based on the food identification results and the standard tableware reference in the image, the food volume is estimated by analyzing the depth information of the image, and the food weight is calculated by combining the food density parameters to obtain the food weight estimation result. Based on the estimated food weight, the food database is queried to obtain the water content and sodium content per unit weight of the corresponding food, and the total water content and total sodium content are calculated to obtain the dietary volume analysis results.

[0009] As one implementation method, time-series analysis is performed on the physiological monitoring data in the preprocessed multimodal dataset, including: Physiological monitoring data, including weight data, blood pressure data, heart rate data, and blood oxygen saturation data, are extracted from the preprocessed multimodal dataset. The physiological monitoring data is then arranged in time series to obtain physiological monitoring time series data. A sliding time window is set for the weight data in the physiological monitoring time series data, and the weight change rate within the sliding time window is calculated. When the weight change rate exceeds a preset threshold, it is identified as an abnormal weight gain pattern, and the weight change trend characteristics are obtained. At the same time, the start and end times of the sliding time window are recorded as time window characteristics. Trend analysis is performed on the blood pressure, heart rate and blood oxygen saturation data in the physiological monitoring time series data. The moving average method is used to smooth the data series, the exponential smoothing method is used to predict the data trend, the deviation between the actual value and the predicted value is calculated, and when the deviation exceeds the preset fluctuation threshold, it is identified as an abnormal fluctuation pattern, thus obtaining the physiological indicator change trend characteristics. By comprehensively analyzing the weight change trend characteristics, the physiological indicator change trend characteristics, and the time window characteristics, a multi-dimensional physiological change feature vector is constructed to obtain the physiological change trend results.

[0010] As one implementation method, the clinical laboratory data in the preprocessed multimodal dataset are evaluated using metrics, including: Clinical test data, including N-terminal pro-brain natriuretic peptide (NT-proBNP) data, serum sodium concentration data, serum potassium concentration data, renal function index data, and hemoglobin data, are extracted from the preprocessed multimodal dataset to obtain a set of clinical test indicators. Based on the NT-proBNP data in the aforementioned clinical laboratory indicator set, and in accordance with the heart failure severity grading criteria, the patient's cardiac function status was assessed, and cardiac function assessment results were obtained. Based on the serum sodium concentration data, serum potassium concentration data, and renal function index data in the aforementioned clinical test index set, the electrolyte balance and renal compensatory capacity are assessed to obtain the metabolic status assessment results. Based on the hemoglobin data in the aforementioned clinical laboratory indicator set, the degree of anemia and its impact on cardiac function were assessed, and the anemia impact assessment results were obtained. The clinical status assessment results are obtained by combining the cardiac function assessment results, the metabolic status assessment results, and the anemia impact assessment results.

[0011] As one implementation method, the step of extracting features and performing feature engineering processing to obtain a feature-engineered dataset includes: Dietary intake features are extracted from the dietary volume analysis results; weight change trend features, physiological indicator change trend features, and weight monitoring time window features are extracted from the physiological change trend results; and clinical indicator features and symptom features are extracted from the clinical status assessment results to obtain the original key feature set. Based on the original key feature set, a variance filter is applied to remove features with variance less than a preset threshold. The recursive feature elimination method combined with the random forest algorithm is used to evaluate the importance of features. The correlation coefficient matrix between features is calculated, and highly correlated features with correlation coefficients greater than a preset threshold are identified and removed to obtain the filtered feature set. Logarithmic transformation is applied to the NT-proBNP index in the filtered feature set to address the skewed distribution. A time-series interactive feature is created based on the weight change trend feature and the dietary intake feature. The time-series interactive feature includes the product of the weight change rate and the cumulative value of sodium intake. A dietary intake time window feature is constructed based on the total water content in the dietary intake feature. The dietary intake time window feature includes the trend of water intake changes over the past 3 days. The feature engineering dataset is then obtained.

[0012] As one implementation method, constructing a multi-level prediction model based on the feature engineering dataset includes: Based on the feature engineering dataset, the dataset is divided into time-series weight data, dietary feature data, clinical indicator data, and symptom data according to feature type. A time-series weight model is constructed using a long short-term memory network for the time-series weight data, a dietary feature model is constructed using a random forest algorithm for the dietary feature data, a clinical indicator model is constructed using a gradient boosting decision tree algorithm for the clinical indicator data, and a symptom model is constructed using a logistic regression algorithm for the symptom data, forming the first-stage feature-focused model group. Each feature-focused model in the first-stage feature-focused model group outputs a preliminary risk score and feature importance weight to obtain the preliminary prediction result. The preliminary risk score and feature importance weights in the preliminary prediction results are used as meta-features. Combined with the original key features in the feature engineering dataset, a gradient boosting decision tree algorithm is used to construct a meta-learning ensemble model as the second-stage meta-learning ensemble model. The outputs of each feature-focused model in the first-stage feature-focused model group are weighted and fused using Stacking ensemble technology. The weights of each feature-focused model in the first-stage feature-focused model group in the ensemble are dynamically adjusted, and the capacity overload risk value is output to obtain the final risk prediction result.

[0013] As one implementation method, the step of classifying risks based on the final risk prediction result, generating graded early warning information, and pushing it to the mobile application terminal to obtain the early warning push result includes: Based on the volume overload risk value in the final risk prediction result, and combined with the preset risk grading threshold, the volume overload risk value is divided into four levels: normal, mild risk, moderate risk, and high risk, to obtain the patient's volume status grading result. Based on the patient capacity status classification results, a prompt-level warning message is generated for mild risk level, a warning-level warning message is generated for moderate risk level, and an emergency-level warning message is generated for high risk level. The classified warning messages are sent to the patient's and medical staff's mobile application terminals in real time through a message push service. For the high risk level, an SMS or telephone reminder is added to obtain the warning push result.

[0014] As one implementation method, the personalized intervention plan is matched based on the early warning push result and pushed to the mobile application terminal, providing intervention execution confirmation and effect tracking functions to obtain intervention management results, including: Based on the risk level and the patient's specific risk status in the early warning push results, including sodium intake exceeding the preset daily sodium intake limit, water intake exceeding the preset daily water intake limit, or weight gain rate exceeding the preset weight change threshold, a corresponding intervention plan is matched from the preset intervention rule base. The intervention plan includes dietary adjustment suggestions to limit sodium intake, medication adjustment prompts to adjust diuretic dosage, and lifestyle guidance to reduce water intake, resulting in a matched personalized intervention plan. The personalized intervention plan is pushed to the patient's end of the mobile application terminal, and the intervention execution confirmation function is provided on the medical staff end. The medical staff can review and adjust the personalized intervention plan. The effect tracking function records the patient's intervention execution and capacity status changes to obtain the intervention management results.

[0015] Secondly, embodiments of this disclosure also provide a machine learning-based remote capacity management system for heart failure patients, comprising: A memory, wherein a computer program is stored; A processor for loading the computer program to implement the machine learning-based remote volume management method for heart failure patients as described in the first aspect.

[0016] Compared with the prior art, the embodiments disclosed herein have the following beneficial effects: This disclosure utilizes multimodal data fusion technology to comprehensively analyze dietary image data, physiological monitoring data, and clinical laboratory data, overcoming the limitations of traditional single data sources. Secondly, visual technology is used to analyze dietary images, enabling precise quantification of food water and sodium content. Simultaneously, a multi-level machine learning model is employed to improve the accuracy of volume overload risk prediction. A tiered early warning and personalized intervention mechanism is designed to achieve closed-loop management from monitoring, assessment, early warning to intervention, effectively improving the accuracy and efficiency of remote volume management for heart failure patients. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the remote volume management method for heart failure patients in this embodiment. Figure 2 This is a schematic diagram of the main page of the remote capacity management system for heart failure patients in this embodiment; Figure 3 This is a schematic diagram of a list of common high-sodium foods used in the remote volume management system for heart failure patients in this embodiment. Figure 4 This is a schematic diagram of the remote capacity management system for heart failure patients in this embodiment. Detailed Implementation

[0019] The preferred embodiments of this disclosure will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit this disclosure.

[0020] Example 1: Please see Figure 1 , Figure 1 This is a flowchart illustrating the remote volume management method for heart failure patients provided in this embodiment. This remote volume management method for heart failure patients can be executed by a computer system.

[0021] like Figure 1 As shown in this embodiment, the remote volume management method for heart failure patients based on machine learning includes the following steps: Step S1: Obtain multimodal data of patients with heart failure, including dietary image data, physiological monitoring data and clinical test data. Preprocess the multimodal data to obtain a preprocessed multimodal dataset.

[0022] Please see Figure 2 , Figure 2 The diagram shows the main page of the remote volume management system for heart failure patients in this embodiment. In this step, the patient can operate from this page to collect dietary image data, such as meal photos, through a mobile application terminal; obtain the patient's physiological monitoring data, including weight data, blood pressure data, heart rate data, and blood oxygen saturation data; at the same time, the patient's clinical test data, such as NT-proBNP data and serum sodium concentration data, can be obtained from the hospital information system.

[0023] Step S2: Perform multi-source data analysis on the preprocessed multimodal dataset, including image recognition and nutritional analysis of dietary image data, time series analysis of physiological monitoring data, and indicator evaluation of clinical test data. Extract features and perform feature engineering processing to obtain a feature-engineered dataset.

[0024] In step S2, the dietary image data obtained in the above steps will be preprocessed and analyzed using deep learning to identify food types and calculate water and sodium content; time-series analysis will be performed on physiological monitoring data to identify abnormal weight gain patterns and abnormal fluctuations in physiological indicators; and indicator evaluation will be performed on clinical laboratory data to assess the patient's cardiac function and metabolic status.

[0025] Step S3: Construct a multi-level prediction model based on the feature engineering dataset. The multi-level prediction model includes a first-stage feature-focused model group and a second-stage meta-learning integration model. The first-stage feature-focused model group outputs preliminary prediction results, which are then input into the second-stage meta-learning integration model for integrated processing. The model outputs a capacity overload risk value to obtain the final risk prediction result.

[0026] Understandably, in this step, the first-stage feature-focused model group includes a time-series weight model, a dietary feature model, a clinical indicator model, and a symptom model, with each model focusing on a specific type of feature; the second-stage meta-learning integration model uses ensemble learning techniques to weight and fuse the outputs of each focus model to obtain the final risk prediction result.

[0027] Step S4: Based on the final risk prediction result, perform risk classification, generate classified early warning information and push it to the mobile application terminal to obtain the early warning push result.

[0028] Step S4 categorizes the patient's volume status into four levels: normal, low risk, moderate risk, and high risk, based on the volume overload risk value. At the same time, it generates corresponding warning information for different risk levels and sends it to the mobile application terminal via push notification service.

[0029] Step S5: Based on the early warning push result, match a personalized intervention plan and push it to the mobile application terminal, providing intervention execution confirmation and effect tracking functions to obtain intervention management results.

[0030] In this step, appropriate intervention plans will be matched from a pre-set intervention rule base based on the patient's risk level and specific risk status. These plans include dietary adjustment suggestions to limit sodium intake, medication adjustment prompts to adjust diuretic dosage, and lifestyle guidance to reduce water intake. These plans will be pushed to the patient via a mobile application terminal, while providing an intervention execution confirmation function on the medical staff's terminal.

[0031] The steps S1-S5 described above will be further described below.

[0032] In this embodiment, step S1, which involves preprocessing the multimodal data to obtain a preprocessed multimodal dataset, includes: Step S11: Based on the multimodal data, apply a data cleaning algorithm to detect and process missing values, outliers, and duplicate records in the data; set a threshold range for the physiological monitoring data; mark or correct physiological monitoring data that exceeds the threshold range; perform quality screening on the dietary image data to remove blurry, dark, or incomplete dietary images; and perform integrity verification on the clinical test data to obtain cleaned multimodal data.

[0033] Understandably, the data cleaning process begins with a comprehensive scan and analysis of the collected multimodal data. For physiological monitoring data (such as weight, blood pressure, heart rate, and blood oxygen saturation), threshold ranges based on clinical guidelines and individual patient baselines were established. For example, the weight change rate threshold was set to no more than 1.5% of the patient's baseline weight per day; the systolic blood pressure range was set at 80-180 mmHg; the diastolic blood pressure range at 50-110 mmHg; the heart rate range at 40-120 beats / minute; and the blood oxygen saturation range at 90%-100%.

[0034] When data points exceeding these thresholds are detected, these outliers are automatically marked, and different processing strategies are adopted according to the degree of anomaly. For slightly deviated data (such as exceeding the threshold by less than 10%), a local regression smoothing algorithm is used for correction; for severely deviated data (such as exceeding the threshold by more than 30%), it is marked as a possible erroneous input, and a confirmation request is sent to the user.

[0035] For food image data, this embodiment employs a multi-layered quality screening mechanism. First, image quality is assessed by calculating the image's brightness histogram and sharpness index (based on image gradient variance). Specifically, the average brightness value of the image is calculated, and images with an average brightness below 30 (within the range of 0-255) are marked as too dark; simultaneously, the Brenner gradient function value of the image is calculated, and images below a preset threshold of 1000 are marked as blurry. Second, the system detects whether the image contains complete information about food and reference objects (such as standard tableware). Edge detection algorithms are used to identify the outlines of major objects in the image to determine the completeness of the image content. Finally, the system automatically removes food images that do not meet quality requirements and sends a prompt to the user to retake the image.

[0036] In practical applications, approximately 15-20% of user-uploaded images are judged to be of substandard quality and require re-capture.

[0037] For clinical laboratory data, this embodiment has developed a dedicated integrity verification process. First, it checks the existence and validity period of key test items (such as NT-proBNP, serum sodium, serum potassium, creatinine, and hemoglobin) to ensure the data is within the validity period for medical decisions (usually test results from the last 7 days). Second, it checks the consistency of units in the test data, ensuring that all NT-proBNP values ​​are expressed in pg / mL and all electrolyte concentrations in mmol / L. Finally, it applies logical rules based on clinical experience to check the reasonableness of the data; for example, NT-proBNP values ​​should not be negative, and serum sodium concentrations should generally not be lower than 120 mmol / L or higher than 160 mmol / L.

[0038] It should be noted that when clinical data that does not meet the integrity requirements is found, the system will mark the relevant records and send a data verification request to medical staff.

[0039] Step S12: Standardize the format of the cleaned multimodal data, unify the timestamp format of the physiological monitoring data and the clinical test data, standardize the units and precision of the numerical data, normalize the size and convert the color space of the diet image data, and establish naming and storage rules to obtain standardized multimodal data.

[0040] In the format standardization stage of step S12, the cleaned multimodal data will undergo comprehensive format standardization processing to ensure that all data have a consistent representation in subsequent analysis.

[0041] Specifically, for time information, the ISO 8601 standard (YYYY-MM-DDThh:mm:ss±hh:mm format) is uniformly adopted, converting the timestamps of all data sources to this standard format and unifying them to UTC+8 time zone (China Standard Time), thus effectively solving the problem of inconsistent time formats recorded by different devices and systems. For numerical data, this embodiment defines a standard unit mapping table; for example, weight is uniformly expressed in kilograms (kg), blood pressure in millimeters of mercury (mmHg), and NT-proBNP in pg / mL. At the same time, this embodiment specifies the precision standards for various types of data, such as retaining weight to one decimal place, blood pressure as an integer, and setting different precisions for laboratory test indicators according to their clinical significance (e.g., retaining serum sodium concentration to one decimal place).

[0042] In this embodiment, a strict standardization process was implemented for the dietary image data. First, all images were normalized to a uniform size of 224×224 pixels (the standard input size for most pre-trained CNN models). Bicubic interpolation was used for image scaling to maintain key image details while ensuring consistent size. Second, all images were converted from their original color spaces (e.g., RGB, sRGB, Adobe RGB) to the standard RGB color space, and a color correction matrix was applied to ensure consistent color reproduction across images taken by different devices. Finally, this embodiment established unified image naming and storage rules, using the format patientID_YYYYMMDD_mealtype_imagenumber (e.g., P10086_20230315_lunch_01) for naming and a hierarchical directory structure for storage, facilitating efficient retrieval and management.

[0043] Based on the above standardization process, this embodiment significantly improves the accuracy and stability of subsequent image recognition algorithms.

[0044] Step S13: Extract key features from the standardized multimodal data, including weight change rate, cumulative sodium intake and medication adherence index, and perform time alignment and feature fusion on data from different sources and types to obtain the preprocessed multimodal dataset.

[0045] In the key feature extraction stage of step S13, this embodiment extracts a feature set that is highly correlated with the volume status of heart failure patients from standardized multimodal data.

[0046] Weight change rate is a core indicator for assessing fluid retention. This embodiment calculates weight change rates for different time windows, including 24-hour, 3-day, and 7-day rates. The weight change rate is calculated using a relative percentage change method: (current weight - baseline weight) / baseline weight × 100%. The baseline weight is determined by the patient's dry weight (stable weight without significant edema) or the lowest weight over the past 30 days. This embodiment also calculates a weight variability index, namely the standard deviation of weight measurements over seven consecutive days, to assess the stability of the patient's weight.

[0047] Those skilled in the art will recognize that cumulative sodium intake is a key dietary factor affecting the volume status of heart failure patients. This embodiment calculates a patient's daily sodium intake using dietary image recognition and nutritional analysis, generating cumulative indicators for multiple time windows, including 3-day cumulative sodium intake and 7-day average daily sodium intake. Furthermore, the sodium intake excess rate, i.e., the percentage of days exceeding the prescribed sodium intake limit out of the total monitoring days, is calculated; this indicator reflects the patient's long-term dietary adherence.

[0048] Medication adherence is typically a crucial dimension for assessing treatment effectiveness. Diuretic adherence rate, calculated by examining patient medication records, is the ratio of actual dosage to prescribed dosage. Simultaneously, the regularity of medication timing is analyzed to calculate the medication timing deviation index, which is the standard deviation of the difference between actual and prescribed medication times; a smaller deviation indicates more consistent medication timing. This embodiment also calculates a standardized diuretic intensity index based on the type and dosage of diuretic to quantify the intensity differences among different diuretic regimens.

[0049] It should be noted that data time alignment is a crucial step in feature fusion. This embodiment employs an event-based time alignment strategy, using the patient's main monitoring events (such as weight measurement, dietary records, and medication records) as anchors to organize data from different sources into a unified time frame. This embodiment uses an adaptive time window mechanism, setting appropriate time granularity for alignment based on the collection frequency and variation characteristics of different data types. For example, weight data is typically collected daily and aligned on a daily basis; while blood pressure and heart rate may be measured multiple times, and are summarized using intraday averages or time-period averages.

[0050] In summary, this embodiment employs a multi-level fusion strategy in the feature fusion stage. First, features are combined within the same data type (e.g., combining raw weight data and the rate of change into a weight feature set). Then, correlated features are constructed between different data types (e.g., the time correlation between weight change and sodium intake). This embodiment also simultaneously employs a feature integrity compensation mechanism; when some data sources are missing, relevant features from other available data sources are estimated and supplemented to ensure the integrity of the feature set. Finally, a multimodal dataset containing multiple feature dimensions is generated, providing comprehensive input information for subsequent risk prediction models.

[0051] The following will provide further explanation of step S2.

[0052] In this embodiment, step S2 involves performing image recognition and nutritional analysis on the dietary image data in the preprocessed multimodal dataset, including: Step S211: Perform image preprocessing on the dietary image data in the preprocessed multimodal dataset, including brightness adjustment, contrast enhancement, cropping, and scaling, to obtain the preprocessed dietary image.

[0053] Understandably, image preprocessing in step S211 is a crucial step in improving the accuracy of food identification. In this step, the image undergoes adaptive brightness adjustment using an automatic brightness correction algorithm based on histogram analysis. This algorithm calculates the image's brightness histogram, analyzes its distribution characteristics, and optimizes images where brightness distribution is overly concentrated in low-value areas (dark areas) or high-value areas (bright areas) by applying logarithmic transformation or gamma correction.

[0054] Specifically, the mean and standard deviation of the image's brightness are calculated. When the mean is below a preset threshold (typically 35% of the image's brightness range), a brightness enhancement function is applied; when the standard deviation is below the preset threshold, a contrast stretching function is applied. This adaptive adjustment can effectively handle food images taken under different lighting conditions, enhancing details in dark areas to a recognizable level.

[0055] Specifically, contrast enhancement is a key technique for improving image texture details. This embodiment employs a local adaptive contrast enhancement algorithm. This algorithm segments the image into multiple overlapping small regions, optimizes contrast independently within each region, and then fuses the results from each region using bilinear interpolation, avoiding boundary artifacts that may occur with traditional global contrast enhancement. This local adaptive method is particularly suitable for processing food images with uneven lighting, simultaneously preserving detail information in both bright and shadow areas, thus improving the recognizability of food texture features.

[0056] In the image cropping stage, this embodiment employs a content-aware intelligent cropping algorithm. This algorithm first uses an object detection network to locate food regions and tableware references in the image. Then, based on the detection results, it automatically calculates the optimal cropping boundary, preserving key content areas while removing irrelevant background. For complex images containing multiple food items, the system uses a hierarchical clustering method to group and prioritize the detected food regions, ensuring that the cropped image contains complete main food items and necessary references.

[0057] It should be noted that image scaling is a necessary step in standardizing image size. This embodiment employs an intelligent scaling strategy that preserves aspect ratio. First, a scaling factor is calculated based on the target size (224×224 pixels). Then, a bicubic interpolation algorithm is applied for high-quality scaling. This algorithm performs cubic polynomial interpolation between pixels in the original image, maximizing the preservation of texture details and edge information during image scaling. For images whose aspect ratio does not match the target size, the system first performs proportional scaling by aligning the long side, then adjusts the image to a square using a padding operation (using the average color value of the image edge pixels), and finally scales it to the target size. This processing method effectively avoids image distortion caused by direct stretching and maintains the true shape characteristics of the food.

[0058] Step S212: Obtain the labeled food image dataset, which contains various types of food image data and corresponding water content and sodium content parameters. Convolutional neural network model is constructed and trained through transfer learning method to obtain the trained food recognition model.

[0059] In step S212, the construction of the food image dataset is the foundation of the food recognition model. Understandably, the system collects and organizes a large-scale image dataset containing commonly seen Chinese and Western foods. These foods cover various types, including staple foods (such as rice, noodles, bread, etc.), meats (such as pork, beef, chicken, etc.), vegetables, fruits, soups, and compound dishes. Images of each food type were collected from different angles, under different lighting conditions, with different cooking methods and different plating styles, averaging 200-300 high-quality images per food category. All images were precisely labeled by a team of professional nutritionists, with annotation information including food category, main ingredients, estimated volume, and weight.

[0060] Of particular importance is that each food item is accompanied by detailed nutritional data, especially the water content (expressed in grams per 100 grams of food) and sodium content (expressed in milligrams per 100 grams of food). This nutritional data comes from food composition tables and measurements taken by professional food composition analysis laboratories.

[0061] In this step, transfer learning is a key method for building a high-performance food recognition model. This embodiment selects a ResNet-50 deep convolutional neural network pre-trained on the ImageNet dataset as the base model.

[0062] The ResNet-50 model has four residual blocks and approximately 25 million parameters, enabling it to effectively extract hierarchical features from images. Specifically, this embodiment employs a two-stage transfer learning strategy: first, the weights of the first three residual blocks of the ResNet-50 are frozen, and only the last residual block and the newly added fully connected classification layer are trained, allowing the model to initially adapt to the food recognition task; then, all layers are unfrozen, and end-to-end fine-tuning is performed using a small learning rate (initial learning rate set to 0.0001), allowing the entire network to better adapt to the specific food recognition task.

[0063] To improve the model's ability to recognize food images, this step employs various data augmentation techniques, including random horizontal flipping (probability 0.5), random rotation (±15 degrees), random brightness and contrast adjustment (within ±10%), random scaling, and cropping. These techniques effectively expand the diversity of training samples and improve the model's generalization ability. Considering the characteristics of the food recognition task, this embodiment also adopts food-specific data augmentation strategies, such as simulating changes in food appearance under different lighting conditions, and adding tableware and background variations.

[0064] During model training, this embodiment uses the cross-entropy loss function to evaluate classification performance, employs the Adam optimizer for parameter updates, sets the initial learning rate to 0.001, and applies a learning rate decay strategy (reducing it to 0.1 times the original rate every 10 epochs). The training process is set to 100 epochs with a batch size of 32, and implements an early stopping mechanism (training stops if the validation set accuracy does not improve after 10 consecutive epochs). Furthermore, to address the issue of imbalanced sample sizes across different categories in the dataset, this embodiment uses a class-weighted loss function, assigning higher weights to food categories with fewer samples. Based on this, the trained food recognition model achieves higher accuracy on the independent test set.

[0065] Step S213: Input the preprocessed food image into the trained food recognition model, and use the model to infer and identify the food category in the image to obtain the food recognition result.

[0066] In step S213, the food identification process is a crucial step in automatically analyzing the food images uploaded by patients using a trained deep learning model. Understandably, after a patient takes and uploads a food image via a mobile application, the pre-processed image is sent to a cloud server for identification. First, feature extraction is performed on the input image. Through cascaded convolutional layers of the ResNet-50 model, a 2048-dimensional feature vector is generated. These feature vectors capture multi-level visual features such as texture, shape, and color in the image, providing a foundation for subsequent classification.

[0067] Specifically, for complex food images containing multiple foods, this embodiment employs a two-stage recognition strategy. In the first stage, a Faster R-CNN object detection network is used to perform region analysis on the image, identifying and locating each food item and generating multiple candidate regions and their bounding box coordinates. In the second stage, each candidate region is individually cropped and fed into the food recognition model to obtain the food classification results for each region.

[0068] It is important to note that assessing the reliability of the identification results is a crucial step in ensuring system stability. This involves not only outputting the predicted category for each food item but also calculating the corresponding confidence score. For example, if the highest confidence score for a food item falls below a preset threshold (usually set to 0.7), a manual confirmation process is triggered, presenting the patient with multiple possible food category options and requesting confirmation or manual selection.

[0069] In addition, this embodiment applies confidence calibration technology, which adjusts the probability distribution of the original network output through temperature scaling, so that the confidence score more accurately reflects the actual accuracy and avoids misidentification caused by model overconfidence.

[0070] For the identified food categories, this embodiment also performs a contextual consistency check. For example, when the identification results contain food combinations that would not normally occur together (such as Western fast food and Chinese dim sum), the confirmation threshold for suspicious identification results is increased. This embodiment also applies a temporal consistency check, comparing the current identification results with the patient's historical dietary records to identify possible abnormal changes. These multiple verification mechanisms effectively reduce the impact of identification errors and improve the reliability of the entire analysis process.

[0071] Step S214: Based on the food identification results and the standard tableware reference in the image, the food volume is estimated by analyzing the depth information of the image, and the food weight is calculated by combining the food density parameters to obtain the food weight estimation result.

[0072] In step S214, food weight estimation is a crucial step from image to nutritional analysis. This step employs a reference-object-based size inference method, requiring the user to include standard tableware (such as a standard plate with a diameter of 22cm or a standard bowl with a capacity of 250mL) as a reference object when capturing the food image. In this embodiment, a pre-trained object detection network identifies the type of reference tableware in the image and retrieves its standard size parameters from the system database. Using the known actual size of the reference object, the conversion ratio (pixels / cm) between image pixels and actual physical size can be calculated, providing a benchmark for subsequent food size measurement.

[0073] As those skilled in the art will recognize, food volume estimation is a technical challenge because two-dimensional images lack direct depth information. This embodiment employs a shape-specific volume estimation strategy, selecting an appropriate volume calculation model based on the typical geometry of different foods.

[0074] For example, for approximately cylindrical foods (such as rice balls), the volume of a cylinder is calculated by measuring its diameter and height in the image and applying the cylinder volume formula; for irregularly shaped foods, a segmentation-based volume estimation method is applied, dividing the food region into multiple simple geometric shapes, calculating the volume of each separately, and then summing the results.

[0075] To address the issue of food occlusion, this implementation also employs a shape prior-based completion algorithm, capable of estimating the volume of food obscured by tableware or other food portions. Food density is a key factor in the volume-to-weight conversion. This embodiment constructs a detailed food density database, containing typical density values ​​for various foods under different cooking methods. For example, the density of cooked white rice is approximately 0.7 g / cm³. 3 The content of stir-fried leafy green vegetables is approximately 0.3g / cm³. 3 The braised pork belly contains approximately 0.9g / cm³. 3Based on the identified food category and cooking method, the corresponding density parameters are retrieved from the database. The estimated volume is then multiplied by the density to obtain the food weight. For compound foods (such as stir-fries), a component ratio model is used to estimate the proportion of each component based on a typical recipe, and the total weight is calculated by applying the corresponding density parameters.

[0076] This embodiment also includes a multi-angle calibration mechanism, encouraging users to photograph the same food from different angles to improve the accuracy of volume estimation through triangulation. For food types with high uncertainty (such as highly irregularly shaped foods), users will be prompted to input reference information (such as a small bowl, half the size of a palm, etc.) to assist in weight estimation.

[0077] Step S215: Based on the food weight estimation results, query the food database to obtain the water content per unit weight and sodium content per unit weight of the corresponding food, calculate the total water content and total sodium content, and obtain the dietary volume analysis results.

[0078] In step S215, dietary volume analysis is a crucial step in converting food identification and weight estimation results into nutritional indicators required for volume management. This embodiment constructs a comprehensive food nutrient database, containing detailed nutritional data for over 2000 common foods. This database is based on authoritative nutritional reference materials and has been reviewed and supplemented by a team of professional nutritionists. For each food, the database not only includes standard nutrient contents (such as protein, fat, and carbohydrates), but also specifically emphasizes parameters directly related to volume management in patients with heart failure, including water content (g / 100g food) and sodium content (mg / 100g food).

[0079] Specifically, the nutritional parameters of the corresponding food are retrieved from the database based on the food identification results. For new foods or variations not directly included in the database, this embodiment uses a nearest neighbor estimation method based on similarity to find the known food with the closest nutritional composition as a reference. The influence of cooking methods on the nutritional composition of food is also considered; for example, raw and cooked spinach of the same weight have significant differences in water and sodium content. Furthermore, regional differences in cooking styles are taken into account, such as the difference in salt usage between northern and southern cuisines. Users can specify the regional characteristics of their dietary habits in the system settings, and the system will adjust the estimation parameters accordingly.

[0080] Specifically, the total water content and total sodium content are calculated using a weighted summation method. For each food item identified in a meal, the system multiplies its estimated weight by the water content and sodium content parameters per unit weight, and then sums them to obtain the total water content and total sodium content of that meal. Simultaneously, the system records and summarizes the patient's total daily intake, calculating the daily average and cumulative values ​​over multiple days. To facilitate patient understanding, the calculation results are converted into an intuitive representation. For example, sodium intake is converted to the usual salt intake (approximately 40% of salt is sodium), showing that approximately 5.2 grams of salt have been consumed today, representing 87% of the prescribed limit (6 grams / day).

[0081] This embodiment also includes a dietary volume trend analysis function. By calculating the average intake and trends over 7 and 14 consecutive days, it generates visual charts to show changes in the patient's eating habits. For intake exceeding the prescribed limits, specific dietary adjustment suggestions are generated, such as reducing the use of soy sauce and pickled foods, and trying to use herbs and lemon juice for seasoning to help patients improve their eating habits. These analysis results and suggestions are integrated into the patient's volume management report, providing an important basis for subsequent risk prediction and intervention decisions.

[0082] In this embodiment, step S2, which involves performing time-series analysis on the physiological monitoring data in the preprocessed multimodal dataset, includes: Step S221: Extract physiological monitoring data from the preprocessed multimodal dataset. The physiological monitoring data includes weight data, blood pressure data, heart rate data, and blood oxygen saturation data. Arrange the physiological monitoring data in time series to obtain physiological monitoring time series data.

[0083] In step S221, the time-series processing of physiological monitoring data is fundamental to assessing changes in volume status in patients with heart failure. First, four key physiological monitoring data categories are extracted from the preprocessed multimodal dataset: weight data, blood pressure data, heart rate data, and blood oxygen saturation data. These data come from diverse sources, including manual input by patients, automatic uploads from smart health devices, and test records provided by medical institutions. The data are then time-stamped and standardized, converting all data points into a standard time-series format, where each data point contains a precise timestamp and a corresponding measurement value.

[0084] Understandably, the inconsistency in data collection frequency is the primary challenge for time-series data processing, especially for the aforementioned physiological monitoring data. For example, weight data is typically measured once a day (ideally after urination in the morning and before eating), while blood pressure may be measured 2-3 times a day, and heart rate and blood oxygen data may be continuously monitored through wearable devices, generating a high-frequency data stream.

[0085] Based on this, this embodiment adopts a multi-frequency data coordination mechanism, which applies a time window aggregation method to high-frequency data (such as heart rate) to calculate the statistical characteristics (mean, maximum, minimum, standard deviation, etc.) of each time period (such as hourly, different time periods of day) so that it is aligned with low-frequency data in terms of time granularity.

[0086] Meanwhile, for cases where data collection times are irregular, this embodiment applies time-series normalization technology. First, standard collection time points or time periods are defined for each type of data; for example, the standard collection time for weight data is 6:00-9:00 AM daily. Then, the actual data collection patterns of patients are analyzed to identify their personal habitual time points; for example, some patients may habitually measure their weight in the evening. Based on these time patterns, irregular data points are mapped to the closest standard time point, and time offset markers are added to facilitate consideration of the impact of time factors in subsequent analysis.

[0087] Specifically, data integrity is another key issue in time series analysis. For time points with missing data, this embodiment employs a context-aware interpolation strategy. For short-term missing data (such as a single day's missed measurement), linear interpolation or spline interpolation methods are used to estimate the missing value; for long-term missing data (such as multiple consecutive days without data), these intervals are marked as insufficient data to avoid unreliable estimations affecting subsequent analysis. Furthermore, this embodiment also includes an anomaly detection mechanism to identify abnormal data points caused by possible measurement errors or equipment malfunctions, preventing these outliers from interfering with the time series analysis results.

[0088] After time series alignment, a structured time series is constructed for each type of physiological data, specifically including a complete time index, measurement values, and data quality markers. This standardized time series data provides a reliable foundation for subsequent trend analysis and pattern recognition, enabling precise capture of the dynamic changes in the patient's physiological state.

[0089] Step S222: Set a sliding time window for the weight data in the physiological monitoring time series data, calculate the weight change rate within the sliding time window, and identify the abnormal weight growth pattern when the weight change rate exceeds a preset threshold to obtain the weight change trend characteristics. At the same time, record the start and end times of the sliding time window as the time window characteristics.

[0090] In step S222, weight change analysis is a crucial part of volume monitoring in patients with heart failure, as weight gain is often a direct manifestation of fluid retention. This step uses a sliding time window technique to dynamically analyze weight data, setting multiple time scales, including a 3-day window (short-term changes), a 7-day window (medium-term changes), and a 14-day window (long-term changes). The sliding time window is updated daily, calculating the weight change characteristics within the window to form a continuous weight monitoring sequence.

[0091] Among these, the rate of weight change is a key indicator for assessing the speed of fluid retention. This step employs multiple calculation methods to capture different types of weight change patterns. The basic rate of change is calculated as (current weight - initial weight of the window) / initial weight of the window × 100%, representing the cumulative change over the entire window period. This implementation also calculates the daily average rate of change, which is the total weight change within the window divided by the number of days, reflecting the average speed of change. For weight data with large fluctuations, a linear regression method is applied to calculate the trend slope, reducing the impact of random fluctuations and more accurately reflecting the overall trend.

[0092] Typically, the identification of abnormal weight gain is based on threshold rules established through clinical practice and research evidence. For example, a weight gain of more than 1.5 kg within 3 days, or an increase exceeding 2% of the patient's baseline weight, is defined as short-term acute weight gain; a weight gain of more than 2.5 kg within 7 days, or an increase exceeding 4% of the baseline weight, is defined as medium-term significant weight gain. This embodiment also defines a set of composite rules, such as a daily weight gain of more than 0.5 kg for 3 consecutive days as a warning sign of progressive fluid retention. When a patient's weight changes meet these rules, the corresponding time period is marked as a period of abnormal weight gain, and the degree to which the threshold is exceeded is calculated as a severity indicator.

[0093] It is important to note that time window characteristics are a crucial medium for linking weight changes with other physiological indicators and behavioral factors. The start and end times of each detected abnormal growth period were recorded, and a 3-7 day time window was traced back as a potential trigger analysis period. These time windows were then correlated with dietary records, medication use, and symptom changes during the same period to identify potential triggering factors, such as a significant increase in sodium intake 3 days prior to abnormal growth or failure to take diuretics on time 2 days prior to abnormal growth. These temporal correlation characteristics help understand the mechanisms of volume changes in individual patients, providing a basis for personalized interventions.

[0094] This embodiment also calculates weight stability indicators, such as the standard deviation and coefficient of variation of weight over 7 days, to assess the degree of weight fluctuation in patients. This is because even if average weight does not increase significantly, increased weight fluctuation may be an early sign of heart failure instability. This embodiment integrates these weight change trend characteristics, abnormal pattern recognition results, and time window characteristics into a weight analysis report, providing important input for subsequent risk prediction.

[0095] Step S223: Perform trend analysis on the blood pressure data, heart rate data and blood oxygen saturation data in the physiological monitoring time series data, use the moving average method to smooth the data series, use the exponential smoothing method to predict the data trend, calculate the deviation between the actual value and the predicted value, and identify the abnormal fluctuation pattern when the deviation exceeds the preset fluctuation threshold to obtain the physiological indicator change trend characteristics.

[0096] In step S223, time-series trend analysis of physiological indicators is an important supplementary method for detecting changes in volume status. Systematic trend processing and anomaly identification are performed on three key physiological indicators: blood pressure, heart rate, and blood oxygen saturation. First, the raw data is smoothed using a moving average method to reduce the impact of random fluctuations and highlight the underlying trends. For blood pressure data, a 5-point central moving average is used; for heart rate and blood oxygen saturation data, which may contain more noise, a wider 7-point moving average window is used. Understandably, this smoothing process effectively filters out short-term fluctuations and measurement errors, improving the reliability of trend analysis.

[0097] Specifically, trend prediction employs exponential smoothing, a method that assigns higher weights to recent data points and decreasing weights to older data points, making it suitable for capturing gradual trends in physiological indicators. This step utilizes an adaptive parameter exponential smoothing algorithm, automatically adjusting the smoothing parameter α (typically within the range of 0.2-0.4) to suit the data characteristics of different patients. Predictive models are constructed for systolic blood pressure, diastolic blood pressure, heart rate, and blood oxygen saturation, predicting trends over the next 24-72 hours based on historical data, and calculating a 95% confidence interval as the expected fluctuation range.

[0098] In this step, abnormal fluctuation identification is a key mechanism for detecting sudden changes in physiological state. This embodiment calculates the deviation between the actual observed value and the model's predicted value and compares this deviation with preset fluctuation thresholds. Different threshold standards are set for different physiological indicators; for example, the fluctuation threshold for systolic blood pressure is ±15 mmHg, diastolic blood pressure is ±10 mmHg, heart rate is ±15 beats / minute, and blood oxygen saturation is ±3%. When the deviation of two or more consecutive measurements exceeds these thresholds, an abnormal fluctuation alarm is triggered. Furthermore, this embodiment also designs composite rules to identify specific abnormal patterns. For example, a sustained increase in systolic blood pressure accompanied by an increase in heart rate may indicate sympathetic nerve activation, an early signal of worsening heart failure.

[0099] This step focuses not only on changes in individual indicators but also on analyzing the synergistic patterns of changes among multiple indicators. For example, it calculates the correlation between blood pressure and heart rate. Under normal circumstances, the two are often negatively correlated (heart rate decreases when blood pressure rises, and vice versa). However, in cases of worsening heart failure, an abnormal pattern of simultaneous increases in both may occur. Simultaneously, it analyzes changes in the difference between daytime and nighttime blood pressure. Insufficient nighttime blood pressure reduction (non-dipping blood pressure) may indicate volume overload.

[0100] In summary, this embodiment generates comprehensive physiological indicator change trend characteristics through these multi-dimensional trend analyses and pattern recognitions, providing multi-faceted physiological evidence for capacity status assessment.

[0101] Step S224: The weight change trend characteristics, the physiological indicator change trend characteristics, and the time window characteristics are comprehensively analyzed to construct a multi-dimensional physiological change feature vector and obtain the physiological change trend results.

[0102] In step S224, the comprehensive analysis of physiological changes is a crucial step in integrating multi-dimensional monitoring data to form a unified physiological state assessment. This step employs a feature fusion method to systematically integrate the weight change trend characteristics, blood pressure and heart rate change trend characteristics, and various time window characteristics obtained from the previous analysis, constructing a multi-dimensional physiological change feature vector. This vector not only contains the absolute change values ​​of each indicator but also includes deeper information such as the rate of change, fluctuation patterns, and temporal correlations.

[0103] Specifically, the feature vector construction adopts a hierarchical structure design. The first layer is the basic feature layer, which contains the raw statistical features of various physiological indicators, such as the 3-day and 7-day weight change rates, mean blood pressure, and heart rate variability. The second layer is the trend feature layer, which contains the slope of change, periodic fluctuation characteristics, and abrupt change point characteristics of each indicator. The third layer is the correlation feature layer, which contains the time lag relationship between weight change and blood pressure change, and the co-change pattern of blood pressure and heart rate. This hierarchical design ensures that the feature vector contains both intuitive surface changes and deep physiological correlation information.

[0104] This step also includes the design of a physiological change pattern library, which includes typical physiological manifestations of volume changes in heart failure. For example, the volume overload progression pattern is characterized by a combination of gradual weight gain, fluctuating blood pressure, increased nocturnal heart rate, and a slight decrease in blood oxygen saturation; the angiotensin-converting enzyme inhibitor (ACEI) adverse reaction pattern is characterized by a combination of significant blood pressure decrease but no significant weight change. This embodiment identifies the most likely type of physiological state change by matching the patient's actual physiological changes with these preset patterns based on similarity.

[0105] Understandably, time-series analysis is a crucial tool for discovering causal relationships. By performing time-series correlation analysis on changes in different physiological indicators and calculating cross-correlation functions, the time lag relationships between these indicators can be identified. For example, a typical time-series pattern can be observed where increased sodium intake leads to weight gain 2-3 days later, followed by an increase in systolic blood pressure 3-5 days later. This temporal relationship helps in understanding the physiological response mechanisms and progression rates of individual patients. By encoding these temporal correlation features into time-dependent networks, the dynamic influence paths between various physiological indicators can be better captured.

[0106] Based on the above steps, the comprehensive analysis results of this embodiment are presented in the form of a physiological change trend report, which includes several key components: assessment of volume status changes (e.g., stable, mild retention, significant retention, or rapid deterioration), assessment of the rate of change (e.g., slow, moderate, or rapid progression), estimation of warning time (e.g., expected symptom worsening within 48-72 hours), and analysis of possible triggers (e.g., related to increased sodium intake, related to decreased medication adherence). This comprehensive physiological change trend result provides a solid data foundation for subsequent risk prediction and intervention decisions, accurately capturing early signals of volume changes and enabling prospective management.

[0107] In this embodiment, step S2 involves evaluating the clinical laboratory data in the preprocessed multimodal dataset, including: Step S231: Extract clinical test data from the preprocessed multimodal dataset. The clinical test data includes N-terminal pro-brain natriuretic peptide (NT-proBNP) data, serum sodium concentration data, serum potassium concentration data, renal function index data, and hemoglobin data to obtain a set of clinical test indicators.

[0108] In step S231, clinical laboratory data is a crucial component in assessing the internal physiological state of patients with heart failure, providing vital information that cannot be directly obtained through external monitoring. This step extracts multiple core clinical laboratory indicators from the preprocessed multimodal dataset to construct a comprehensive clinical assessment foundation.

[0109] Understandably, clinical laboratory data primarily originates from patients' routine outpatient examinations, inpatient tests, and follow-up examinations at primary healthcare institutions. This step employs a multi-source clinical data integration mechanism, enabling the acquisition of patient laboratory data through various means, including hospital information system interfaces, electronic health record export functions, and manual data entry by healthcare professionals on mobile applications.

[0110] Specifically, during the data extraction process, the timeliness of the test data is assessed. Different test indicators have different timeliness values: NT-proBNP data is typically tested every 1-3 months, with priority given to results from the most recent 30 days; results from 30-90 days ago are given decreasing weight; and results older than 90 days are only used as historical references. Electrolyte and renal function data are typically tested every 1-2 months, with a 45-day validity period. Hemoglobin data is typically tested every 3 months, with a 90-day validity period. For data exceeding its validity period, its timeliness will be clearly marked in the assessment report, and medical staff will be reminded to consider scheduling new tests.

[0111] Standardization of laboratory data is a crucial step in constructing clinical indicator sets. This embodiment establishes a detailed mapping table for laboratory tests, effectively handling differences in nomenclature and units among different medical institutions and laboratories. For example, NT-proBNP may be expressed in pg / mL or ng / L, which will be uniformly converted to pg / mL; renal function assessments may use creatinine (Cr) or estimated glomerular filtration rate (eGFR), and standardized eGFR values ​​will be calculated based on available data. For cases where units and reference ranges are inconsistent, conversion rules are applied for standardization to ensure the comparability of test results from different sources.

[0112] This step also includes a targeted verification mechanism for test results. By comparing the results with historical patient data and clinical rationale rules, potential outliers or data entry errors are identified. For example, if a certain indicator deviates significantly from the patient's historical level (such as a sudden 90% drop in NT-proBNP value) and there is no corresponding clinical intervention record, the result is marked as requiring verification, and medical staff are prompted to confirm. This multi-layered data extraction and verification mechanism ensures the completeness, accuracy, and timeliness of the clinical test indicator set, providing a reliable foundation for subsequent clinical status assessment.

[0113] Step S232: Based on the NT-proBNP data in the clinical test index set, and in accordance with the heart failure severity grading criteria, assess the patient's cardiac function status and obtain the cardiac function assessment results.

[0114] In step S232, NT-proBNP assessment is one of the standards for assessing cardiac function in patients with heart failure. Based on the latest guidelines for the diagnosis and treatment of heart failure, this embodiment constructs a multi-level NT-proBNP interpretation framework.

[0115] First, diagnostic thresholds are set based on patient age stratification: for patients under 75 years old, NT-proBNP > 450 pg / mL is considered a possible indication of heart failure; for patients 75 years and older, considering the higher baseline levels in the elderly, the threshold is adjusted to NT-proBNP > 900 pg / mL. For the exclusion criteria of acute heart failure, the system uses a higher threshold of NT-proBNP < 300 pg / mL; when a patient's value is below this threshold, the system will mark it as having a low probability of acute heart failure.

[0116] Typically, the severity of cardiac function is assessed using a multi-level stratified standard. Specifically, NT-proBNP values ​​are divided into four severity levels: mild (300-1000 pg / mL), moderate (1000-3000 pg / mL), severe (3000-10000 pg / mL), and very severe (>10000 pg / mL). This grading considers not only absolute values ​​but also the patient's baseline levels and trends. The NT-proBNP change ratio is defined as the ratio of the patient's current NT-proBNP value to their historical baseline value (average level during stable periods). A ratio >1.5 is marked as a significant increase; a ratio >3 is marked as a severe increase, suggesting a possible acute deterioration in cardiac function.

[0117] This embodiment also considers confounding factors affecting NT-proBNP levels. Specifically, for patients with concomitant renal insufficiency, a correction formula is applied to adjust NT-proBNP values, as decreased renal clearance can lead to elevated NT-proBNP levels without necessarily reflecting worsening cardiac function. Typically, the correction method is based on the patient's eGFR value; when eGFR < 60 mL / min / 1.73 m... 2 In some cases, the explanatory threshold for NT-proBNP is adjusted upwards based on the degree of renal function. Similarly, for obese patients with a body mass index (BMI) > 30 kg / m², the explanatory threshold is adjusted downwards to account for the potential effect of adipose tissue reducing NT-proBNP levels.

[0118] It should be noted that dynamic assessment of cardiac function is a key function of this embodiment. By comparing multiple consecutive NT-proBNP measurements, the system calculates the rate of change and the pattern of change. Studies have shown that a decrease in NT-proBNP value of ≥30% or a decrease to <1000 pg / mL generally indicates effective treatment; while a sustained increase or a lack of significant decrease indicates poor treatment efficacy.

[0119] In summary, this embodiment generates a cardiac function trend assessment based on these change patterns, such as continuous improvement, stable cardiac function, gradual deterioration, or potential acute deterioration of cardiac function. Finally, by combining absolute level assessments and dynamic change assessments, a comprehensive cardiac function status report is generated, providing a reference for clinical decision-making.

[0120] Step S233: Based on the serum sodium concentration data, serum potassium concentration data, and renal function index data in the clinical test index set, assess the electrolyte balance and renal compensatory capacity to obtain the metabolic status assessment results.

[0121] In step S233, electrolyte balance and renal function assessment are important parts of the metabolic status assessment for patients with heart failure. In this step, a comprehensive assessment of serum sodium concentration is first performed using a three-level classification standard, including: hyponatremia (<135 mmol / L), normal range (135-145 mmol / L), and hypernatremia (>145 mmol / L).

[0122] Specifically, hyponatremia is further subdivided into mild (130-134 mmol / L), moderate (125-129 mmol / L), and severe (<125 mmol / L). Those skilled in the art will recognize that hyponatremia is a common complication of heart failure and an independent predictor of poor prognosis. By analyzing the possible mechanisms of hyponatremia, including dilutional hyponatremia due to water retention and depletional hyponatremia due to diuretic overdose, and by combining patient weight change trends, diuretic use, and clinical symptoms, the most likely type of hyponatremia is inferred, and targeted intervention recommendations are generated.

[0123] Typically, serum potassium levels are assessed using a three-tiered classification: hypokalemia (<3.5 mmol / L), normal range (3.5-5.5 mmol / L), and hyperkalemia (>5.5 mmol / L). This embodiment pays particular attention to dynamic changes in serum potassium, as commonly used medications in heart failure patients, such as diuretics, ACEIs / ARBs, and aldosterone antagonists, can affect potassium balance. This embodiment identifies potential drug-related electrolyte disturbances by correlating changes in serum potassium with medication records. Specifically, for severe abnormalities such as serum potassium <3.0 mmol / L or >6.0 mmol / L, a high-priority alert is generated, indicating the possibility of urgent intervention.

[0124] It should be noted that renal function assessment is crucial for determining the compensatory capacity of patients with heart failure. This embodiment prioritizes eGFR as the standard for renal function assessment, calculated using the CKD-EPI formula, taking into account the patient's age, sex, race, and serum creatinine level. Renal function classification follows the Kidney Disease International Guidelines (KDIGO) criteria: G1 stage (eGFR ≥ 90 mL / min / 1.73 mL / min). 2 G2 phase (60-89 mL / min / 1.73 m) 2 ), G3a stage (45-59mL / min / 1.73m 2 G3b phase (30-44 mL / min / 1.73 m) 2 ), G4 stage (15-29mL / min / 1.73m 2 ) and G5 phase (<15mL / min / 1.73m) 2In addition, the rate of change in renal function was calculated. Acute renal deterioration was marked when eGFR decreased by >25% or serum creatinine increased by >0.3 mg / dL within 14 days. This may be a sign of worsening heart failure or diuretic overdose.

[0125] Among these steps, assessing renal compensatory capacity is a crucial step in this embodiment. First, the concept of renal diuretic reserve is defined and assessed by analyzing the patient's response pattern to diuretics. Specifically, the diuretic efficiency index is calculated, which is the increase in urine volume and weight loss induced by each milligram of furosemide (or an equivalent dose of another diuretic). A significant decrease in this index over time suggests a reduction in renal diuretic reserve, potentially requiring adjustment of the diuretic strategy.

[0126] In summary, by comprehensively assessing electrolyte status and renal function, a metabolic status report is generated, including the current status classification (e.g., good electrolyte balance, adequate renal function compensation, or electrolyte disturbance, inadequate renal function compensation) and risk warnings (e.g., risk of diuretic resistance or increased risk of worsening cardiorenal syndrome).

[0127] Step S234: Based on the hemoglobin data in the set of clinical test indicators, assess the degree of anemia and its impact on cardiac function to obtain the anemia impact assessment results.

[0128] In step S234, anemia assessment is a crucial component of comprehensive management for patients with heart failure, as anemia can increase the burden on the heart and worsen heart failure symptoms. This embodiment adopts the World Health Organization (WHO) anemia criteria, setting thresholds based on gender differences: hemoglobin <130 g / L for men and <120 g / L for women is defined as anemia. The system further classifies anemia into mild (110-129 g / L for men, 110-119 g / L for women), moderate (80-109 g / L), and severe (<80 g / L). For elderly patients (>65 years old), considering physiological changes, the system appropriately adjusts the thresholds (<120 g / L for men, <110 g / L for women) to reduce false positives.

[0129] Specifically, anemia type analysis is a key function of this embodiment. By integrating data on erythrocyte indices (such as MCV, MCH, and MCHC) and iron metabolism indicators (such as ferritin and transferrin saturation), the most likely type of anemia is inferred. Common types include iron deficiency anemia (common in patients with heart failure, especially when using antiplatelet or anticoagulant drugs), chronic disease anemia (chronic inflammatory states associated with heart failure), and renal anemia (in patients with concurrent renal insufficiency). This embodiment uses a decision tree algorithm for anemia type analysis, providing a basis for targeted intervention.

[0130] It is worth noting that assessing the impact of anemia on cardiac function is a crucial step in this embodiment. This embodiment defines an anemia burden index to quantify the additional burden of anemia on the heart. This index comprehensively considers the degree of anemia (the extent of hemoglobin reduction), duration, and rate of change.

[0131] Those skilled in the art will know that for every 10 g / L decrease in hemoglobin, cardiac output needs to increase by approximately 10% to maintain tissue oxygenation, thus increasing the burden on the heart. Based on this physiological principle, and combined with the patient's baseline cardiac function status (such as ejection fraction and NT-proBNP levels), the percentage of additional cardiac load that anemia may cause is calculated. When the anemia load index > 20%, it is marked as anemia significantly affecting cardiac function; when the index > 40%, it is marked as anemia severely affecting cardiac function.

[0132] This embodiment also focuses on the correlation between anemia and heart failure symptoms. Specifically, by comparing the time relationship between patients' self-reported symptom data (such as dyspnea score and fatigue score) and hemoglobin levels, it is possible to identify whether the worsening of symptoms is caused by the aggravation of anemia or by the deterioration of cardiac function itself. This helps medical staff to develop more precise intervention strategies.

[0133] Based on the above steps, an anemia impact assessment report is finally generated. The report includes the current anemia status, possible anemia types, the degree of impact on cardiac function, and treatment priority recommendations (such as recommending to correct anemia first or recommending to manage cardiac function and anemia simultaneously).

[0134] Step S235: Combine the cardiac function assessment results, the metabolic status assessment results, and the anemia impact assessment results to obtain the clinical status assessment results.

[0135] In step S235, the comprehensive clinical status assessment is a crucial step in integrating the results of each individual assessment into a comprehensive clinical profile. This embodiment employs a multidimensional clinical status modeling method to create a holistic representation of the patient's current state. First, a clinical status scoring matrix is ​​constructed, quantifying cardiac function, metabolic status, and anemia status into standardized scores of 0-100 points. Specifically, the cardiac function score is primarily based on NT-proBNP levels and trends, the metabolic status score integrates electrolyte balance and renal function, and the anemia score is based on hemoglobin levels and their inferred impact on cardiac function. This standardized scoring allows for comparison and synthesis of clinical status across different dimensions.

[0136] Specifically, clinical status classification is a crucial output of the assessment results. This embodiment defines several typical clinical status types: stable compensated (stable indicators with no obvious abnormalities), early decompensated (cardiac function indicators begin to deteriorate, but metabolic function is still acceptable), metabolic imbalance (mainly manifested as electrolyte disturbances and / or decreased renal function), anemia-dominant (significant anemia with a significant impact on cardiac function), multi-system decompensated (cardiac function, metabolic status, and anemia all deteriorate together), and acute exacerbation (cardiac function indicators deteriorate rapidly, often accompanied by acute metabolic disorders). This embodiment uses a pattern recognition algorithm to match the patient's multidimensional scores with these preset types to identify the most suitable clinical status type.

[0137] This embodiment also incorporates an assessment of clinical status stability. By calculating the standard deviation and coefficient of variation of clinical status scores over the past 30 days, patient status stability is categorized into several levels, including highly stable (coefficient of variation <10%), basically stable (10-20%), fluctuating (20-40%), and highly unstable (>40%). It should be noted that stability assessment helps predict the risk of acute events in patients in the short term and has significant reference value for the development of clinical management strategies.

[0138] In this step, clinical risk stratification is the final step in the comprehensive assessment. Based on clinical status scores and stability assessments, combined with factors such as patient age, comorbidities, and history of acute events, this embodiment calculates the patient's event risk in different time frames.

[0139] Specifically, this embodiment defines several key risk indicators: 30-day readmission risk, 90-day cardiovascular event risk, and 180-day mortality risk. These risk assessments are based on predictive models validated by large-scale heart failure cohort studies, and individualized risk calculations are performed by integrating current clinical status data. Ultimately, a comprehensive clinical status assessment report is generated, including current status classification, stability rating, short-term risk prediction, and intervention priority recommendations, providing systematic clinical decision support for healthcare professionals.

[0140] In this embodiment, step S2, which involves extracting features and performing feature engineering to obtain a feature-engineered dataset, includes: Step S241: Extract dietary intake features from the dietary volume analysis results, extract weight change trend features, physiological indicator change trend features, and weight monitoring time window features from the physiological change trend results, and extract clinical indicator features and symptom features from the clinical status assessment results to obtain the original key feature set.

[0141] In step S241, feature extraction is a fundamental step in building an efficient prediction model. This step extracts important features related to the risk of capacity overload from the results of multi-source data analysis, forming a comprehensive original feature set. First, the dietary intake features extracted from the dietary capacity analysis results include the following key dimensions: sodium intake features, water intake features, dietary regularity features, and nutritional balance features.

[0142] Specifically, sodium intake characteristics include average daily sodium intake (mg / day), 3-day cumulative sodium intake (mg), 7-day cumulative sodium intake (mg), sodium intake exceedance rate (percentage of days exceeding the standard out of the total monitoring days), and sodium intake fluctuation coefficient (standard deviation of daily intake divided by mean). Water intake characteristics include average daily water intake (mL / day), 3-day cumulative water intake (mL), water intake compliance rate, and drinking water regularity index (assessing the degree of regularity by calculating the consistency of daily drinking time and quantity).

[0143] In step S241, the weight change trend features extracted from the physiological change trend results are important indicators for volume monitoring. Specifically, this includes extracting weight change indicators at multiple time scales: 24-hour weight change rate (%), 3-day weight change rate (%), 7-day weight change rate (%), weight gain rate (kg / day), weight fluctuation index (standard deviation / mean), and weight trend slope (calculated through linear regression). Furthermore, this embodiment also defines characteristic indicators for several abnormal weight patterns: acute growth index (quantifying the degree of increase >1.5 kg within 3 days), gradual growth index (quantifying the degree of steady increase within 7 days), and morning-evening weight difference (weight difference measured within the same day, reflecting intraday fluid retention).

[0144] The physiological indicator trend characteristics in this embodiment include dynamic information on blood pressure, heart rate, and blood oxygen saturation. Key extracted indicators include: mean systolic blood pressure, systolic blood pressure trend (slope), systolic blood pressure variability index, mean diastolic blood pressure, pulse pressure difference (systolic blood pressure - diastolic blood pressure), mean heart rate, heart rate trend, intra-day heart rate variability (highest value - lowest value), correlation coefficient between heart rate and body weight (reflecting the autonomic nervous system's response to volume changes), and blood oxygen saturation decline index (frequency and degree of decline below 95%).

[0145] It should be noted that the characteristics of the weight monitoring time window are an important dimension for analyzing the triggering factors of weight change. This embodiment defines and extracts a series of time-related features, including: the onset time of abnormal weight gain, the duration of abnormal weight gain, the acceleration of weight change (change in the rate of change), the peak time of sodium intake before weight gain, and the temporal relationship between weight gain and changes in medication adherence. These time-series correlation features help to understand the dynamic process and causal mechanism of volume change.

[0146] In this step, clinical indicators extracted from the clinical status assessment results include NT-proBNP values ​​(original and logarithmically transformed), NT-proBNP change rate, electrolyte balance index (a composite index of the degree of abnormality in serum sodium and potassium), renal function index (standardized score based on eGFR), renal function trend, and anemia index (based on the difference between hemoglobin and sex-specific normal values). Symptom characteristics include dyspnea score (0-10 points), exercise tolerance score (based on the 6-minute walk test or self-assessment of daily activities), edema score (0-4 points, based on the degree and extent of pitting), frequency of paroxysmal nocturnal dyspnea, and occurrence of orthopnea.

[0147] Drug-related characteristics are another important dimension of this embodiment, including diuretic type, standardized dose (converted to equivalent furosemide dose), medication adherence index (actual dose / prescribed dose × 100%), diuretic effect index (weight loss / diuretic dose), and ACEI / ARB / ARNI usage.

[0148] In summary, by integrating the above multi-dimensional features, this embodiment constructs an original key feature set containing multiple candidate features, providing a comprehensive data foundation for subsequent feature selection and engineering.

[0149] Step S242: Based on the original key feature set, apply a variance filter to remove features with variance less than a preset threshold, use a recursive feature elimination method combined with a random forest algorithm to evaluate feature importance, calculate the correlation coefficient matrix between features, identify and remove highly correlated features with correlation coefficients greater than a preset threshold, and obtain the filtered feature set.

[0150] In step S242, feature selection is a crucial step in filtering out the most predictive features from the original feature set. This embodiment implements a multi-stage feature selection process to ensure the efficiency and interpretability of the final model.

[0151] First, a variance filter is applied to calculate the variance of each feature on the training dataset, with a variance threshold set to 0.01 (the optimal value determined through cross-validation). Typically, features with very small variance show little variation, contain limited information, and contribute little to prediction. In typical applications, the variance filter removes approximately 10% of features on average, such as almost unchanging baseline features or features with incomplete documentation.

[0152] Next, after variance filtering, a recursive feature elimination method based on random forests (RF-RFE) is applied to evaluate feature importance. Understandably, RF-RFE, as an iterative feature selection method, first trains a random forest model using all features (number of trees = 500, maximum depth = 10), calculating the importance score for each feature (based on its contribution to model accuracy). Then, features are sorted in descending order of importance, and in each iteration, the lowest 10% of features are removed. The model is retrained and its performance is evaluated until performance begins to decline significantly or a preset minimum number of features is reached (usually 30% of the original number of features). Typically, the subset of features with the best model performance is selected as the candidate feature set.

[0153] Understandably, the advantage of RF-RFE lies in its ability to capture the nonlinear contributions of features and the interactions between features, which is particularly important for complex capacity prediction tasks.

[0154] In this embodiment, feature correlation analysis is a crucial step in avoiding feature redundancy. This embodiment calculates the Pearson correlation coefficients between all feature pairs and constructs a correlation coefficient matrix. Feature pairs with absolute correlation coefficient values ​​greater than a preset threshold (usually set to 0.85) are identified as highly correlated features. To determine which feature to retain, the maximum correlation minimum redundancy (mRMR) principle is adopted: the correlation between each feature and the target variable (capacity overload risk) is calculated, while its average correlation with other features is evaluated; features with high correlation to the target but low redundancy with other features are prioritized for retention.

[0155] For example, when the 3-day weight change rate and the 7-day weight change rate are highly correlated, the decision to retain one or the one that contributes more to the predictive ability will be based on the strength of their correlation with the risk of capacity overload.

[0156] This embodiment also employs a domain-knowledge-based feature preservation mechanism. Certain clinically recognized important features, such as NT-proBNP values ​​and acute weight gain index, are retained in the final feature set even if they are not optimal features in statistical analysis. This embodiment maintains a list of important features to ensure that these clinically significant features are not erroneously removed during automated screening. Therefore, through this feature selection strategy combining statistical methods and professional knowledge, a final feature set containing multiple (25-30) high-value features is obtained, significantly reducing model complexity while maintaining predictive power.

[0157] Step S243: Apply logarithmic transformation to the NT-proBNP index in the filtered feature set to process the skewed distribution. Create a time-series interactive feature based on the weight change trend feature and the dietary intake feature. The time-series interactive feature includes the product of the weight change rate and the cumulative value of sodium intake. Construct a dietary intake time window feature based on the total water content in the dietary intake feature. The dietary intake time window feature includes the trend of water intake changes over the past 3 days. The feature engineering dataset is obtained.

[0158] In step S243, feature engineering is a key technology to improve the prediction capability of the original features. In this embodiment, new features with more expressive power are created through mathematical transformation and feature combination.

[0159] First, an appropriate mathematical transformation is applied to the characteristic of a significantly skewed distribution. NT-proBNP values ​​exhibit a typical right-skewed distribution, with a range spanning several orders of magnitude (from hundreds to tens of thousands of pg / mL), and relatively few patients exhibiting high values. Applying a natural logarithmic transformation (ln(NT-proBNP)) to the NT-proBNP values ​​makes its distribution closer to a normal distribution.

[0160] Based on the above transformations, the clinical significance of NT-proBNP value increments across different concentration ranges becomes more consistent (e.g., an increase from 1000 to 2000 has similar importance on a logarithmic scale as an increase from 5000 to 6000); simultaneously, it can mitigate the excessive influence of extreme values ​​on the model, improving the model's generalization ability in the general population. Similar logarithmic or square root transformations can also be applied to other severely skewed features (such as certain drug dosage characteristics).

[0161] Secondly, the temporal interaction feature is a key step in this embodiment, capable of capturing the dynamic interactions between different physiological factors. Typically, the product of the weight change rate and the cumulative sodium intake is a typical temporal interaction feature, quantifying the potential impact of sodium intake on weight change. Specifically, the calculation method involves multiplying the 3-day weight change rate (%) by the cumulative sodium intake (mg) over the previous 3-5 days to form a composite index. The biological basis for this feature is that increased sodium intake may lead to water retention and weight gain, but this effect exhibits individual differences and a time lag.

[0162] This embodiment effectively captures this nonlinear interaction by constructing product features, thereby improving the sensitivity of prediction. Similar time-series interaction features include the ratio of weight change rate to diuretic adherence (reflecting the effectiveness of diuretic therapy in weight control) and the interaction features of blood pressure changes and sodium intake (quantifying the degree of influence of sodium intake on blood pressure).

[0163] Furthermore, defining dietary intake time windows is a crucial step in capturing temporal patterns of dietary behavior. Based on total water content within dietary intake characteristics, this example constructs features for multiple time windows, including the trend of water intake changes over the past three days. This feature is obtained by calculating the slope of a linear regression of water intake over three consecutive days; positive values ​​indicate an upward trend in intake, while negative values ​​indicate a downward trend.

[0164] This embodiment also calculated the water intake fluctuation index (the standard deviation of daily intake over 3 days divided by the mean) and the sodium-to-water ratio (the ratio of sodium intake to water intake, reflecting the salt concentration of the diet). These time window characteristics can reflect the dynamic changes in patients' short-term eating behavior and provide more predictive information than single-point intake.

[0165] In this step, this embodiment also employs feature transformation based on clinical knowledge, converting original features into new features with clear clinical interpretation. For example, a volume overload index is defined, which standardizes the difference between a patient's current weight and their dry weight (the ideal weight determined by the physician or the lowest stable weight within the last 30 days), more intuitively reflecting the degree of volume overload. Similarly, the diuretic resistance index quantifies the degree of reduced responsiveness of patients to diuretic therapy by calculating the ratio of standardized diuretic dosage to weight loss effect. These features not only improve model performance but also enhance the clinical interpretability of the model output.

[0166] In summary, this embodiment, through these complex feature engineering techniques, ultimately forms a feature engineering dataset containing multiple high-quality features, significantly improving the expressive power and predictive value of the original features, and laying a solid foundation for building a high-performance risk prediction model.

[0167] In this embodiment, step S3, constructing a multi-level prediction model based on the feature engineering dataset, includes: Step S31: Based on the feature engineering dataset, the dataset is divided into time-series weight data, dietary feature data, clinical indicator data, and symptom data according to feature type. A time-series weight model is constructed using a long short-term memory network for the time-series weight data. A dietary feature model is constructed using a random forest algorithm for the dietary feature data. A clinical indicator model is constructed using a gradient boosting decision tree algorithm for the clinical indicator data. A symptom model is constructed using a logistic regression algorithm for the symptom data, forming the first-stage feature-focused model group. Each feature-focused model in the first-stage feature-focused model group outputs a preliminary risk score and feature importance weight to obtain the preliminary prediction result.

[0168] In step S31, the first-stage feature-focused model group design forms the foundational layer of the system's risk prediction architecture. It involves selecting the most suitable algorithm to construct specialized models for different types of features. The system first divides the feature engineering dataset into several main feature subsets based on data characteristics and clinical significance: time-series weight data (including time-series weight characteristics and their variation patterns), dietary feature data (including sodium and water intake-related features), clinical indicator data (including laboratory test and physiological monitoring data), and symptom data (including patient self-reported symptoms and signs). This feature grouping not only considers the similarity of data types but also reflects the logic of classifying and processing information from different sources in clinical thinking.

[0169] Specifically, the time-series weight model is constructed using a Long Short-Term Memory (LSTM) network, specifically designed to handle time-series weight data. LSTM is a special type of recurrent neural network with the ability to capture long-term dependencies, making it suitable for processing data with significant temporal characteristics, such as weight. The LSTM model constructed in this embodiment comprises a two-layer LSTM structure, with 64 neurons in each layer, followed by a fully connected layer to output risk prediction values. The model input consists of daily weight measurements over 14 consecutive days, along with their first-order differences (daily variation) and second-order differences (acceleration of change), forming a feature sequence. Furthermore, to address the issue of missing measurements, this embodiment also employs an interpolation mechanism based on historical data and a missing value mask vector to ensure the model can effectively utilize incomplete time-series data.

[0170] Understandably, LSTM models can identify complex time patterns, such as gradual growth, periodic fluctuations, or sudden changes, which play an important role in early warning of capacity overload.

[0171] The dietary feature model is constructed using the random forest algorithm to analyze the impact of dietary behavior on volume status. Random forests improve model stability and accuracy by constructing multiple decision trees and averaging their predictions. This embodiment configures a random forest containing 200 decision trees, with a maximum depth limit of 8 for each tree to balance model complexity and generalization ability.

[0172] Understandably, random forests can be used to process dietary feature data, specifically: first, there are complex nonlinear relationships and interactions among dietary features; second, dietary features may contain noise and missing values, and random forests are robust to this type of data; and third, random forests can automatically assess feature importance, helping to identify the most critical dietary factors. The model input includes various sodium and water intake features, dietary regularity indicators, and time-series interaction features, and the output is a diet-related volume risk score.

[0173] The clinical indicator model is constructed using the Gradient Boosting Decision Tree (GBDT) algorithm to process laboratory test and physiological monitoring data. This embodiment selects XGBoost as the specific implementation method, whose efficient GBDT framework offers excellent predictive performance and scalability.

[0174] Typically, XGBoost model configurations include a learning rate of 0.1, a maximum tree depth of 6, a subsampling rate of 0.8, and L1 / L2 regularization terms to prevent overfitting. Clinical indicator data usually includes multiple physiological and biochemical indicators, among which there are complex associations and threshold effects (i.e., the risk increases significantly after certain indicators exceed a specific threshold). The GBDT algorithm can automatically learn these complex nonlinear relationships and threshold effects without requiring manual specification of specific thresholds or relationship forms. Model inputs include clinical indicators such as NT-proBNP, electrolytes, renal function, and hemoglobin, as well as their trend characteristics, and the output is a risk score based on clinical indicators.

[0175] Specifically, the symptom model was constructed using a logistic regression algorithm to analyze the relationship between symptom data and capacity risk. A logistic regression model with L2 regularization (ridge regression) was implemented, and the regularization coefficient was determined to be 0.01 through cross-validation.

[0176] Typically, symptom data primarily comes from patient self-reports and clinical assessments, including subjective scores such as the degree of dyspnea, edema, and exercise tolerance. These symptom characteristics usually exhibit strong linear separability and clear clinical interpretability. Logistic regression maintains the model's simplicity and interpretability while providing sufficient predictive power. Furthermore, the probability values ​​output by logistic regression can be directly interpreted as risk probabilities, facilitating clinical understanding and application. Model inputs include various symptom scores and physical signs, and the output is a symptom-based risk score.

[0177] In summary, this embodiment employs independent training and evaluation processes for each feature-focused model. Training utilizes 5-fold cross-validation to ensure model performance stability. Each model not only outputs a preliminary risk score within the 0-1 range but also calculates feature importance weights to quantify the contribution of each feature to the prediction. The feature importance evaluation method varies depending on the model type: LSTM models are evaluated using a permutation importance method (randomly shuffling feature values ​​and observing performance degradation); Random Forest and GBDT are evaluated using their built-in feature importance calculation mechanisms; and Logistic Regression is evaluated using the absolute value of the coefficients. These preliminary risk scores and feature importance weights constitute the prediction results of the first stage, serving as input to the meta-learning integration model in the second stage.

[0178] Step S32: Using the preliminary risk score and feature importance weight in the preliminary prediction result as meta-features, and combining them with the original key features in the feature engineering dataset, a gradient boosting decision tree algorithm is used to construct a meta-learning fusion model as the second-stage meta-learning fusion model. The outputs of each feature-focused model in the first-stage feature-focused model group are weighted and fused using Stacking fusion technology. The weights of each feature-focused model in the first-stage feature-focused model group in the fusion are dynamically adjusted, and the capacity overload risk value is output to obtain the final risk prediction result.

[0179] In step S32, the second-stage meta-learning fusion model is a high-level layer of the system's prediction architecture. It intelligently fuses the outputs of the various focus models from the first stage using meta-learning methods to form the final risk prediction result. Meta-learning is a method of learning how to learn; by constructing a meta-feature set, the fusion model can understand and rationally utilize the predictive capabilities of each focus model.

[0180] The meta-feature set includes the following key information: the first category is the preliminary risk score output by each focused model, which directly reflects the prediction results of different feature subsets; the second category is the feature importance weight, which reveals the degree of dependence of each model on different features; and the third category is the high-value features in the original key features, which ensures that the meta-model can directly access some core predictive factors and enhance decision-making flexibility.

[0181] In this embodiment, the meta-learning ensemble model is constructed using the Gradient Boosting Decision Tree (GBDT) algorithm, with LightGBM as the implementation framework. LightGBM, as an efficient variant of GBDT, can handle a large number of features and complex relationships. Typically, the configured LightGBM model includes 500 trees, a maximum depth of 5, and a learning rate of 0.05. It employs a leaf-wise growth strategy to improve training efficiency, while feature subsampling and L2 regularization are applied to prevent overfitting. As a meta-model, GBDT can automatically learn complex nonlinear fusion rules, identify the prediction strengths and weaknesses of each focused model in different scenarios, and achieve more intelligent ensemble decision-making.

[0182] It's important to note that Stacking ensemble technology is a crucial mechanism for meta-learning ensembles. Traditional ensemble methods, such as simple averaging or majority voting, struggle to handle variations in model performance across different data distributions. Stacking trains a meta-model to learn how to optimally combine the predictions of the base models. This embodiment employs an improved two-stage cross-validation Stacking framework: First, the training data is divided into K folds (typically K=5). A first-stage focused model is trained on K-1 folds, generating predictions on the remaining 1 fold. This process is repeated K times to obtain unbiased predictions for all training data. Then, these predictions are used as meta-features to train the second-stage ensemble model. This method effectively avoids information leakage, ensuring that the meta-model learns the true performance characteristics of the model.

[0183] In this step, dynamic weight adjustment is a key part of this embodiment, enabling the model to adapt to the differences in data quality among different patients. This embodiment defines data quality assessment metrics, including scores for the completeness, consistency, and timeliness of various types of data.

[0184] For example, when there is a significant amount of missing patient weight measurement data, the reliability of the time-series weight model decreases, and its weight in the ensemble is automatically reduced; conversely, when patients have recent NT-proBNP test results, the weight of the clinical indicator model is increased accordingly. This dynamic adjustment is achieved by adding data quality indicators to the meta-features, enabling the meta-model to automatically learn the optimal weight allocation strategy based on data characteristics.

[0185] Meanwhile, this embodiment also employs an uncertainty estimation mechanism. When the predictive consistency of all basic models is low or the predictive confidence is insufficient, the integrated model will increase the uncertainty range of the prediction results, reminding medical staff to use it cautiously in conjunction with clinical judgment.

[0186] In summary, the final risk prediction result output in this embodiment is represented by a risk score of 0-100, intuitively reflecting the degree of capacity overload risk, while providing a 95% confidence interval to estimate the uncertainty range. Simultaneously, this embodiment also generates a risk interpretation report, including analysis of key risk factors (identifying features that contribute most to risk prediction), feature threshold analysis (identifying indicators that exceed safe limits), and model reliability assessment (based on data quality and model consistency). It is understood that this multi-level prediction architecture not only improves prediction accuracy but also enhances the interpretability and reliability of the prediction results, supporting healthcare professionals in making more scientific clinical decisions.

[0187] In this embodiment, step S4, which involves classifying risks based on the final risk prediction result, generating graded early warning information, and pushing it to the mobile application terminal to obtain the early warning push result, includes: Step S41: Based on the volume overload risk value in the final risk prediction result, and combined with the preset risk grading threshold, the volume overload risk value is divided into four levels: normal, mild risk, moderate risk, and high risk, to obtain the patient's volume status grading result.

[0188] In step S41, risk stratification is a crucial step in translating continuous risk predictions into clinically actionable decisions. Specifically, a stratified threshold strategy is employed to classify the capacity overload risk value within the range of 0-100 into four clinically significant risk levels.

[0189] The grading thresholds were set based on two important principles: first, the correlation with clinical prognosis, meaning each risk level should correspond to a different degree of adverse clinical event risk; and second, the rationality of intervention timing, meaning the grading should guide the appropriate timing and intensity of clinical intervention. This embodiment, through analysis of large cohort data of heart failure patients (including follow-up data from over 2000 patients) and in conjunction with the consensus of medical staff, determined the final grading thresholds: 0-30 points for normal (low risk), 30-50 points for mild risk, 50-70 points for moderate risk, and 70-100 points for high risk.

[0190] It should be noted that the risk stratification process in this step considers not only individual risk values ​​but also a comprehensive assessment of risk change trends. In this embodiment, the rate and acceleration of change in the patient's risk value over the past 7 days are defined as the risk dynamic coefficient.

[0191] For example, when a patient's risk score is in the borderline zone (e.g., 48-52 points, close to the boundary between mild and moderate risk) and the risk dynamic coefficient is positive (indicating an upward trend in risk), the patient will be classified into a higher risk level to achieve proactive risk management. Conversely, if the risk score is in the borderline zone but the risk dynamic coefficient is negative and lasts for more than 3 days, the patient will be kept at the current risk level to avoid intervention instability caused by frequent fluctuations in risk level.

[0192] In this step, this embodiment also employs an individualized risk stratification adjustment mechanism. For patients with abundant historical data, their risk value distribution characteristics are analyzed to identify individual-specific risk patterns. For example, for some patients with high baseline risk (long-term high risk values ​​but stable clinical status), the stratification threshold is appropriately adjusted based on their individual risk distribution characteristics to avoid alarm fatigue caused by consistently being at a high risk level. This individualized adjustment is achieved by calculating the patient's individual risk Z-score (the standardized deviation from their historical average), making the risk stratification more consistent with the patient's individual characteristics.

[0193] Meanwhile, the risk assessment also considers the uncertainty of model predictions, calculating a 95% confidence interval for each predicted risk value to reflect the reliability of the prediction. When the confidence interval is wide (indicating high prediction uncertainty) and spans multiple risk levels, an uncertainty warning is displayed in the risk grading results to remind medical staff to make careful decisions based on clinical judgment. In summary, this transparent uncertainty handling mechanism enhances the reliability and clinical applicability of the risk grading system, avoiding the potential risks of mechanical decision-making.

[0194] Step S42: Based on the patient capacity status classification results, generate a prompt-level warning message for mild risk level, a warning-level warning message for moderate risk level, and an emergency-level warning message for high risk level. Send the classified warning messages to the patient's and medical staff's mobile application terminals in real time through message push service. For the high risk level, add SMS or telephone reminders to obtain the warning push results.

[0195] In step S42, generating early warning information is a crucial step in transforming risk classification results into specific action guidelines. This embodiment designs a three-level early warning mechanism based on different risk levels, including alert-level, warning-level, and emergency-level early warnings. Each level of early warning has different information content, urgency, and intervention recommendations.

[0196] The alert level is for patients with mild risk (risk score 30-50). The alert information is mild, highly advisory, and emphasizes preventative management. For example, typical alert-level alert messages include: "Your volume status has changed slightly; please closely monitor your weight and follow a low-sodium diet plan," or "You have recently gained a little weight; it is recommended to reduce fluid intake and ensure you take your medication on time." This type of information is usually accompanied by educational content, such as low-sodium diet tips or guidance on proper weight monitoring, to help patients improve their self-management abilities.

[0197] Warning-level alerts are for patients at moderate risk (risk score 50-70). These alerts are clear, instructive, and emphasize timely adjustments. For example, typical warning-level alert messages include: "Your volume status indicates moderate risk; your weight gain is accelerating; please strictly control your salt and water intake; consider adjusting your diuretic dosage." or "An increased risk of volume overload has been detected; please immediately check for worsening edema and contact your doctor for medication adjustment advice." These messages usually include specific action steps and goals, such as limiting daily fluid intake to no more than 1000 mL or contacting your doctor today to discuss diuretic adjustments, guiding patients to take clear action.

[0198] Emergency alerts are for high-risk patients (risk score 70-100). The alert messages are urgent, highly directive, and emphasize immediate intervention. For example, typical emergency alert messages include: "Your volume status indicates a high risk of acute heart failure worsening; please contact your doctor immediately and prepare for medical attention," or "Severe risk of volume overload detected; please stop activity immediately, rest in a semi-recumbent position, and contact emergency services." These messages usually include clear medical instructions and emergency contact information to guide patients to obtain medical help quickly. For emergency alerts, a notification is also sent to the patient's designated emergency contact (such as a family member) to ensure the patient receives timely assistance.

[0199] Typically, the aforementioned early warning messages employ a multi-channel, multi-level strategy to ensure timely delivery and sufficient attention. Specifically, alert-level warnings are sent via in-app push notifications and display a reminder icon on the app's homepage; warning-level warnings, in addition to in-app push notifications, also send system notifications and display a pop-up reminder when the app launches; emergency-level warnings activate a comprehensive notification mechanism, including in-app push notifications, system notifications, and SMS reminders. For extremely high-risk situations (risk score > 85), an automated voice outbound calling system will also dial the patient's pre-set emergency contact number to ensure immediate information delivery. Simultaneously, all warning information is sent to the healthcare application, and high-risk warnings trigger priority handling prompts for healthcare personnel.

[0200] The early warning push service in this embodiment is implemented based on a real-time message queue architecture and uses a publish-subscribe pattern for message distribution. The system uses Apache Kafka as the message middleware to ensure high throughput and low latency in message processing.

[0201] Meanwhile, to address network instability, the system employs a multi-level message persistence and retry mechanism: when a patient's application is detected to be offline, the alert information is stored in the server-side message queue, with message priority and expiration time set; when the patient reconnects, unexpired messages are pushed in priority order. For emergency alerts, a message delivery confirmation mechanism is also implemented. If no confirmation feedback is received from the patient within a preset time (usually 30 minutes), the notification method is automatically escalated, such as sending an additional SMS or initiating a phone call reminder, ensuring that high-risk situations are handled promptly.

[0202] In this embodiment, step S5, which involves matching a personalized intervention plan based on the early warning push result and pushing it to the mobile application terminal, provides intervention execution confirmation and effect tracking functions to obtain intervention management results, includes: Step S51: Based on the risk level and the patient's specific risk status in the early warning push result, the patient's specific risk status includes sodium intake exceeding the preset daily sodium intake limit, water intake exceeding the preset daily water intake limit, or weight gain rate exceeding the preset weight change threshold. Match the corresponding intervention plan from the preset intervention rule base. The intervention plan includes dietary adjustment suggestions to limit sodium intake, medication adjustment prompts to adjust diuretic dosage, and lifestyle guidance to reduce water intake, to obtain a matched personalized intervention plan.

[0203] In step S51, matching personalized intervention plans is a crucial step in translating risk assessment into concrete treatment actions. First, a multi-dimensional analysis of the patient's specific risk profile is conducted to identify the main factors contributing to the increased risk of volume overload.

[0204] Specifically, sodium intake assessment uses a tiered standard: for patients with heart failure, the standard limit is set at 2000 mg of sodium per day (approximately 5 g of salt), with mild excess at 2000-3000 mg / day, moderate excess at 3000-4000 mg / day, and severe excess at >4000 mg / day. This embodiment also calculates the patient's sodium intake trend. When the average sodium intake over three consecutive days exceeds the limit and shows an upward trend, it is marked as persistent sodium intake excess, which is generally a better predictor of volume problems than a single day's excess.

[0205] Fluid intake assessment also employs individualized standards: based on the patient's weight and clinical status, a daily fluid intake limit is calculated, generally ranging from 1200-1800 mL / day. Assessment of excessive fluid intake considers not only the total daily intake but also intake patterns. For example, identifying concentrated evening drinking patterns (consuming more than 50% of the total daily fluid intake after 6 PM), which may increase nighttime urination burden and the risk of fluid retention during sleep. Furthermore, the relationship between fluid intake and activity and temperature is considered; in hot weather or after patients report strenuous activity, the fluid intake limit assessment standards are appropriately adjusted.

[0206] Weight change assessment is a direct indicator of the rate of fluid retention. This embodiment sets multi-level warning thresholds for weight change, specifically including: an increase of >0.9 kg or 1.2% of the baseline weight within 24 hours is marked as rapid growth; an increase of >1.5 kg or 2% of the baseline weight within 3 days is marked as significant growth; and slow growth (>0.2 kg per day) lasting more than 7 days is marked as progressive fluid retention.

[0207] It should be noted that this embodiment not only assesses the absolute growth amount, but also focuses on the growth pattern, especially the accelerated growth pattern (the growth rate itself is increasing), which usually indicates a higher risk of capacity runaway.

[0208] Typically, the intervention rule base serves as the knowledge foundation for matching personalized intervention plans to the system, employing a multi-layered structured design. The top layer is the intervention category library, encompassing four main categories: dietary intervention, pharmacological intervention, lifestyle intervention, and medical service intervention. The middle layer is the specific intervention strategy library, with 10-15 basic intervention strategies under each category. The bottom layer is the intervention implementation details library, containing the specific implementation methods, frequency, intensity, and monitoring indicators for each strategy.

[0209] This embodiment employs a rule-based matching engine to retrieve the most suitable intervention combination from the intervention rule base based on the patient's risk level and specific risk status. The matching process follows a problem-oriented principle: first, the main risk factors (such as persistent excessive sodium intake) are identified, and then the main intervention strategy (such as a phased strict salt restriction plan) and auxiliary intervention strategies (such as guidance on the use of alternative seasonings) targeting these factors are matched.

[0210] Specifically, dietary intervention programs form the basis of volume management. For patients with excessive sodium intake, a tiered salt restriction plan is provided: a mild salt restriction plan is recommended for mild excess, suggesting reducing the intake of processed foods and reducing the amount of salt used in cooking; a moderate salt restriction plan is recommended for moderate excess, providing a detailed list of high-sodium foods and suggestions for alternatives; and a strict salt restriction plan is recommended for severe excess, providing a complete daily dietary plan, including specific food choices and cooking methods.

[0211] For example, Figure 3 A list of some common high-sodium foods and their sodium content standards are provided, which patients can use to effectively control their sodium intake.

[0212] The system also identifies the main sources of individual sodium intake (such as pickled foods, processed meats, and sauces) based on the patient's dietary image analysis results, and provides targeted alternative suggestions, such as using lemon juice and herbs instead of soy sauce for seasoning or choosing fresh meat instead of smoked meat and sausages.

[0213] Medication intervention suggestions are an important tool for regulating volume status. This example generates medication adjustment recommendations based on the patient's risk profile, current medication regimen, and historical medication response. For patients experiencing rapid weight gain, a temporary increase in diuretic dosage may be recommended, with specific suggestions including the magnitude of the increase and precautions. For patients with recurrent fluid retention, optimization of the diuretic regimen may be recommended, such as considering adding a thiazide diuretic as adjunctive therapy under a physician's guidance, or adjusting the diuretic administration time, with the main dose taken in the morning and the adjunctive dose taken in the afternoon. It is particularly important to emphasize that medication interventions must be performed under the guidance of a physician, and all medication adjustment recommendations include clear medical consultation tips.

[0214] In summary, lifestyle intervention guidance complements dietary and pharmacological interventions, comprehensively improving patients' fluid management capabilities. For patients with excessive fluid intake, a refined drinking plan is provided, such as recommending the use of a measuring cup to evenly distribute fluid intake throughout the day: 600mL in the morning, 400mL at noon, and 200mL in the evening. For patients with insufficient activity tolerance, graded activity guidance is provided, such as suggesting 2-3 sessions of 5-10 minutes of light walking daily, followed by appropriate rest. This implementation also focuses on psychological support elements, providing stress management techniques and psychological adjustment suggestions, such as trying 10 minutes of breathing relaxation exercises to alleviate anxiety. These psychological interventions help improve patients' treatment adherence and self-efficacy.

[0215] Step S52: Push the personalized intervention plan to the patient's end of the mobile application terminal, provide an intervention execution confirmation function on the medical staff end, receive the review and adjustment of the personalized intervention plan by medical staff, record the patient's intervention execution status and capacity status changes through the effect tracking function, and obtain the intervention management results.

[0216] Step S52, where the intervention plan is delivered, is a crucial step in effectively communicating personalized recommendations to patients. This step employs a multi-layered information presentation architecture to ensure that complex intervention plans are presented to patients in a clear and actionable manner.

[0217] Specifically, the intervention plan push notifications adopt a clear hierarchy design principle: the homepage displays key intervention points, presenting the most critical action recommendations with eye-catching icons and concise text, such as strictly limiting daily salt intake to less than 3 grams or adding 20 mg of furosemide as prescribed for two consecutive days; the details page provides the complete intervention plan, a task list arranged chronologically, and detailed implementation guidelines. This embodiment employs a context-aware push mechanism, selecting the optimal push timing based on the patient's daily behavior patterns, such as sending dietary recommendations before the time the patient usually prepares meals, or sending medication reminders before medication administration time.

[0218] It should be noted that the intervention confirmation function on the healthcare provider's end is a crucial guarantee for ensuring the safety and professionalism of the intervention. In this embodiment, the generated personalized intervention plan is first sent to the healthcare provider for review, especially suggestions involving medication adjustments, which must be confirmed by the doctor before being pushed to the patient.

[0219] Typically, the healthcare interface displays the patient's risk assessment report, intervention recommendations, and historical intervention effects. Doctors can directly confirm the system's recommendations, make adjustments, or add professional annotations. This step incorporates a tiered review mechanism: routine intervention recommendations for patients with mild risk (such as dietary and lifestyle adjustments) can be reviewed and confirmed by nurses; intervention recommendations for patients with moderate to high risk, or recommendations involving medication adjustments, must be reviewed and confirmed by doctors.

[0220] Meanwhile, in order to improve work efficiency, a default approval mechanism has been implemented for routine interventions with low risk and no drug adjustment. That is, if medical staff do not take clear intervention within a set time (usually 4 hours), intervention suggestions will be automatically pushed and marked as the system default suggestion, reminding patients that they should still consult medical staff when conditions permit.

[0221] Intervention implementation confirmation is a crucial step in closed-loop management. This embodiment implements multiple intervention implementation confirmation mechanisms on the patient side: for dietary interventions, patients can upload pictures of their meals to record dietary adherence; for medication interventions, patients can confirm medication use by scanning the pillbox or manually recording the information; for lifestyle interventions, it supports automatic synchronization of activity data or manual recording.

[0222] Meanwhile, this embodiment sets execution status markers (completed, partially completed, incomplete) and execution difficulty feedback (a 5-level scale from very easy to very difficult) for each intervention task, helping medical staff understand the actual situation of intervention execution and patients' subjective experience. Intervention execution data is automatically summarized into an intervention compliance report, calculating the execution rate and difficulty score of various intervention measures, providing a basis for subsequent intervention plan adjustments.

[0223] Effectiveness tracking is a key function in evaluating the effectiveness of an intervention. This embodiment employs a comprehensive intervention effectiveness monitoring mechanism, tracking changes in multi-dimensional volume status indicators. Short-term effectiveness indicators include trends in weight change (whether there is weight loss or a slowdown in the rate of weight gain within 3-5 days after the intervention), the degree of symptom improvement (such as changes in dyspnea scores), and the resolution of edema (through patient self-assessment or image comparison); medium-term effectiveness indicators include trends in volume risk scores (changes in risk values ​​over 7-14 consecutive days), the time to return to normal volume status, and the duration of stability; long-term effectiveness indicators include trends in intervention adherence, changes in the frequency of volume fluctuations, and reductions in healthcare service utilization (such as a decrease in emergency room visits and hospitalizations). Through these multi-timescale effectiveness indicators, the effectiveness of the intervention strategy is comprehensively evaluated.

[0224] Intervention adjustment is a key step in optimizing intervention strategies based on effect tracking results. This embodiment designs an adaptive intervention adjustment mechanism: when poor intervention effects are detected (such as continued weight gain or no symptom improvement), the intervention intensity is automatically increased or a change in intervention strategy is recommended; when patients report difficulty in intervention implementation or low compliance, implementation obstacle analysis and simplified intervention recommendations are provided; when the intervention effect is good and stable, a gradual transition to maintenance and relapse prevention strategies is achieved. This feedback-based dynamic adjustment ensures that the intervention program always maintains optimal effectiveness and acceptability, achieving truly personalized capacity management.

[0225] The mobile application terminal of this embodiment will now be described in detail.

[0226] In this embodiment, the mobile application terminal is deployed in a system architecture that includes a terminal layer, a cloud layer, and a technical support layer. Specifically: The terminal layer includes a patient-side application and a medical care-side application developed based on the Android system. The patient-side application integrates functions such as diet recording, weight recording, medication recording, and symptom self-assessment. The medical care-side application integrates functions such as clinical data entry and intervention execution confirmation.

[0227] The cloud layer is deployed on a cloud server, configured with image recognition and machine learning services, and designed with an SQL database structure including a patient basic information table, a clinical data table, a monitoring data table, a food database table, and an abnormal rules table.

[0228] The technical support layer interfaces with computer vision APIs, encapsulates machine learning algorithm libraries, integrates cloud computing service interfaces, and realizes data interaction between the terminal layer and the cloud layer through a secure communication mechanism based on the HTTPS protocol.

[0229] First, the terminal layer is the front-end interface where the system directly interacts with the user. It is developed based on the Android operating system (compatible with Android 8.0 and above) and uses the Material Design language to ensure a consistent and intuitive user experience. The patient-side application adopts a modular architecture, divided into four main functional modules: core monitoring, health management, education support, and communication collaboration. The interface design has been optimized for the characteristics of heart failure patients (mostly middle-aged and elderly). This embodiment uses a high-contrast color scheme, adjustable text size, and a simplified navigation structure to improve the application's accessibility and ease of use.

[0230] Secondly, the diet recording function is a crucial module of the patient-side application. This embodiment supports multiple recording methods to adapt to different scenarios. Image recording mode is the primary method; users simply click the "Record Diet" button on the main interface to use their phone's camera to take a photo of the food. Real-time shooting guidance is provided (e.g., ensuring the plate is completely in the frame or using standard tableware as a reference) to improve image quality. For situations where taking photos is not possible, a quick recording mode is provided, allowing users to select food types and estimated portions from a commonly used food library. This embodiment also supports voice recording, allowing users to verbally describe their meals, which are then converted into structured dietary data using voice recognition technology. All diet records are automatically categorized by time into breakfast, lunch, dinner, and snacks (see [link to relevant documentation]). Figure 3 This creates a dietary calendar view, making it easy to review and analyze dietary patterns.

[0231] The weight recording function supports both manual entry and automatic synchronization with smart devices. In manual entry mode, a numeric keypad and shortcut buttons are provided, allowing users to input their daily weight. The system automatically displays the change compared to the previous day, using color coding to indicate the direction of change (red for increase, green for decrease, and blue for stability). The smart device synchronization mode supports various brands of Bluetooth scales. Users simply stand on the scale to complete the measurement, and the data is automatically transmitted to the mobile application and recorded via Bluetooth. This embodiment also supports a weight measurement reminder function, automatically adjusting the reminder frequency based on the patient's risk status (once daily under normal conditions, potentially increasing to twice daily when risk is elevated). Weight data is displayed visually in trend chart format, highlighting short-term (3-day) and medium-term (7-day) trends to help patients promptly detect abnormal weight fluctuations.

[0232] The medication record function helps patients manage complex medication regimens. The system supports drug barcode scanning for data entry; users simply scan the barcode on the drug packaging, and the system automatically identifies the drug name and specifications, simplifying the data entry process. Users can also set the time, dosage, and frequency of administration for each medication, and the system generates a personalized medication plan and provides scheduled reminders.

[0233] Medication confirmation typically employs multiple methods, including one-click confirmation, photo recording (taking photos of the pills and packaging as evidence of medication use), and NFC tag confirmation (quick confirmation by bringing a mobile phone close to an NFC tag pre-attached to the pillbox). The system also includes a drug interaction check function; when a potential drug interaction risk is detected, it generates an alert and recommends consulting a doctor. Medication adherence reports are usually generated automatically, displaying the accuracy, on-time performance, and missed doses for various medications, helping healthcare professionals assess treatment adherence.

[0234] The symptom self-assessment function provides an intuitive symptom recording interface, covering common heart failure symptoms such as shortness of breath, fatigue, edema, and nocturnal cough. The system employs multiple assessment scales, including a modified Borg scale for assessing shortness of breath (0-10 points), the New York Heart Association (NYHA) classification for assessing exercise tolerance (grades I-IV), and the Visual Analogue Scale (VAS) for assessing fatigue (0-10 points). Edema assessment uses an image-assisted method; the system displays reference images of pitting edema of varying degrees, and users select the image that best reflects their own condition for scoring. The system also provides a symptom diary function, allowing users to record any unusual sensations and possible triggers, creating a symptom record that combines structured and free text. All symptom data is stored in time-series format, and the system automatically analyzes symptom trends and correlates them with other monitoring data (such as weight changes) to generate a comprehensive symptom report.

[0235] The healthcare application employs a role-based access control system, supporting doctors, nurses, and health managers, each with distinct functional permissions. Key functions include patient management, clinical data entry, intervention management, and statistical analysis. The patient management function supports grouping, risk ranking, and rapid filtering, facilitating efficient processing of large volumes of patient data by healthcare professionals. The clinical data entry interface has been professionally optimized, supporting the rapid entry of laboratory test results, physical examination findings, and clinical assessment results.

[0236] Typically, the system provides data validity verification, automatically checking whether input values ​​are within a reasonable range and alerting users to potential input errors. The intervention execution confirmation interface displays the intervention suggestions generated by the system and the patient's execution feedback. Medical staff can review and modify the suggestions and add professional annotations. All intervention decisions and execution results are compiled into an intervention log, facilitating subsequent treatment evaluation and accountability.

[0237] Secondly, the cloud layer is the system's critical computing and storage center. This embodiment employs a distributed microservice architecture deployed on Alibaba Cloud Elastic Compute Service (ECS) to ensure high availability and scalability. The cloud server configuration uses a multi-instance cluster deployment, with the main compute nodes using ecs.g6e.4xlarge instances and the GPU nodes dedicated to machine learning inference using ecs.gn6i.8xlarge instances. This embodiment adopts a containerized deployment strategy, using Kubernetes to manage the container cluster, enabling automatic scaling and fault self-healing of the service, ensuring stable operation 24 / 7.

[0238] The image recognition service is a dedicated module for processing food images uploaded by patients, deployed independently using a microservice architecture. The core recognition engine is based on the TensorFlow Serving framework, loading a pre-trained food recognition CNN model to provide high-performance inference services. Typically, the image recognition process includes the following main steps: a preprocessing service handles image standardization; a detection service locates food items and reference objects in the image; a classification service identifies food types; and a quantization service estimates food portion size and nutritional composition. This embodiment employs model version management and A / B testing mechanisms, supporting online model updates without service interruption, ensuring continuous optimization of the recognition algorithm.

[0239] To improve service efficiency, the system typically employs request batching technology, which merges multiple image recognition requests received within a short period of time for processing. This fully leverages the parallel computing capabilities of the GPU, maintaining low latency response even under high load conditions (average processing time <200ms).

[0240] The machine learning service, a crucial engine for risk prediction and intervention decisions, is also deployed using a microservice architecture. The prediction service loads pre-trained multi-level prediction models, receives feature data, and outputs risk prediction results. The feature engineering service handles data preprocessing and feature transformation, converting raw data into feature vectors required by the model. The model management service manages model version control, performance monitoring, and regular retraining. This embodiment supports multiple model formats, including PMML (Predictive Model Markup Language) and ONNX (Open Neural Network Exchange), facilitating the deployment of models developed using different algorithm frameworks. Furthermore, the machine learning service employs a model interpretability module, capable of generating feature importance analysis and decision path visualizations to help healthcare professionals understand the basis for prediction results.

[0241] The database design employs a hybrid architecture combining relational and non-relational databases. Critical business data is constructed using a relational database built with MySQL 8.0, containing several main tables: a patient basic information table storing static information such as demographic data, contact information, and emergency contacts; a clinical data table using a time-series structure storing test results, diagnostic records, and clinical assessment data; a monitoring data table storing monitoring records from patient self-tests and device uploads, including weight, blood pressure, heart rate, and symptom scores; a food database table storing food nutritional information, including parameters such as calories, macronutrients, water content, and sodium content; and an anomaly rule table storing early warning and intervention rule definitions, including threshold parameters, risk level assessment criteria, and intervention strategy mappings. Large volumes of unstructured data (such as dietary images) are stored in Alibaba Cloud Object Storage Service (OSS), with the database only storing object references.

[0242] It's worth noting that the database design emphasizes time-dimensional management, employing a robust temporal data processing mechanism. All health-related data utilizes a dual timestamp strategy: transaction time (data recording time) and validity time (actual occurrence time), supporting complex time-based queries and analysis. The database also implements a version control mechanism, preserving data modification history for easy auditing and error recovery. For frequently accessed monitoring data, a multi-level caching strategy is implemented: a memory cache stores the most recent 7 days of frequently accessed data; a Redis cache stores the most recent 30 days of summary data; and raw data is stored in the main database. This multi-layered caching design significantly improves system response speed, maintaining good performance even under heavy concurrent user access.

[0243] Finally, the technical support layer serves as an intermediate service layer connecting the terminal layer and the cloud layer, providing unified API interfaces, secure communication mechanisms, and technical service integration functions. It interfaces with various computer vision APIs, forming a matrix of visual analysis capabilities. The core visual service is provided by a self-developed food recognition model, supplemented by the general image recognition API from Alibaba Cloud's Visual Intelligence Open Platform. When the self-developed model's recognition confidence is insufficient, the general API is called for auxiliary judgment. This embodiment also integrates a dedicated medical image analysis API for processing edema photos uploaded by patients, assessing the degree of edema by analyzing skin texture and depression features. The technical support layer implements an API call management mechanism, including request rate limiting, failure retries, and result caching, ensuring the stability and cost-effectiveness of external service calls.

[0244] It's important to note that the encapsulation of machine learning algorithm libraries is a crucial function of the technical support layer, providing a unified intelligent analysis interface for upper-layer applications. This embodiment encapsulates several open-source machine learning frameworks, including scikit-learn (for traditional machine learning algorithms), XGBoost and LightGBM (for gradient boosting decision tree algorithms), and TensorFlow and PyTorch (for deep learning algorithms). The algorithm library encapsulation adopts a microservice design pattern, with each algorithm type deployed as an independent service, providing services through RESTful APIs and the gRPC protocol. Simultaneously, the system employs algorithm performance monitoring and automatic degradation mechanisms: when an abnormal response time or increased error rate is detected in a certain algorithm service, it automatically switches to a backup or simplified algorithm to ensure service continuity. Furthermore, the algorithm library also supports online learning, enabling incremental updates of model parameters using newly collected data, allowing the prediction model to continuously optimize as data accumulates.

[0245] The cloud computing service interface integration enables unified management and invocation of various cloud services. It primarily integrates the following cloud services: storage services (Alibaba Cloud OSS object storage, for unstructured data management), messaging services (Alibaba Cloud Messaging Service MNS, for internal component communication and push notifications), computing services (Function Compute FC, for elastic processing of compute-intensive tasks), and security services (Access Control RAM and Key Management KMS, for authentication and data encryption).

[0246] Understandably, the technical support layer implements a service abstraction layer, which encapsulates the specific interfaces of various cloud services into a unified internal API, reducing the dependence of upper-layer applications on specific cloud services and improving the portability and maintainability of the system.

[0247] It should be noted that secure communication mechanisms are a key technology for ensuring secure data transmission. This embodiment employs a multi-layered secure communication architecture: transport layer security is based on the HTTPS protocol (TLS 1.3), using a 2048-bit RSA certificate for server authentication and encryption; application layer security uses JWT (JSON Web Token) for user authentication and authorization, with a token validity period of 24 hours and support for a refresh mechanism; data layer security uses the AES-256 algorithm for end-to-end encryption of sensitive data, with the key generated from user credentials via a key derivation function (KDF). This embodiment also includes a complete security audit mechanism, logging all critical operations, including data access, modification, and transmission events, facilitating security event tracing and compliance verification.

[0248] It's important to note that the data interaction between the terminal and the cloud employs an efficient synchronization strategy. Considering the potential instability of network connections on mobile devices, the system utilizes a bidirectional incremental synchronization mechanism: the terminal maintains a local database, recording unsynchronized data changes and timestamps; upon reconnection, the terminal sends the change records to the cloud while simultaneously retrieving the latest changes from the cloud. For potential synchronization conflicts (the same data being modified by multiple parties while offline), a rule-based conflict resolution strategy is implemented, prioritizing the most recent timestamp or changes made by the healthcare provider. To optimize bandwidth usage, the system applies data compression and differential update techniques, transmitting only the changed data portions, significantly reducing data transmission volume (saving over 70% bandwidth in typical scenarios). The technical support layer also implements an offline working mode, ensuring that patients can still record health data when the network is unavailable, automatically synchronizing it to the cloud once the network is restored.

[0249] Example 2: like Figure 4 As shown, Figure 4 A schematic diagram of the structure of a remote capacity management system for heart failure patients provided in an embodiment is shown. The system (hereinafter referred to as a computer system) includes at least a processor 101, a communication interface 102, and a memory 103.

[0250] The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the computer system, capable of parsing various instructions and processing various data within the computer system. The communication interface 102 may optionally include standard wired or wireless interfaces (such as Wi-Fi or mobile communication interfaces), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system used to store programs and data. It is understood that the memory 103 here can include the computer system's built-in memory, or it can include extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system; this invention does not limit this storage space.

[0251] In one embodiment, the processor 101 executes the remote volume management method for heart failure patients provided in Embodiment 1 by running a computer program in the memory 103.

[0252] The above description is only a preferred embodiment of this disclosure and is not intended to limit this disclosure. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the protection scope of this disclosure.

Claims

1. A remote volume management method for heart failure patients based on machine learning, characterized in that, include: Multimodal data of patients with heart failure are acquired, including dietary image data, physiological monitoring data and clinical test data. The multimodal data is preprocessed to obtain a preprocessed multimodal dataset. Multi-source data analysis is performed on the preprocessed multimodal dataset, including image recognition and nutritional analysis of dietary image data, time series analysis of physiological monitoring data, and indicator evaluation of clinical laboratory data. Features are extracted and feature engineering is performed to obtain a feature-engineered dataset. A multi-level prediction model is constructed based on the feature engineering dataset. The multi-level prediction model includes a first-stage feature-focused model group and a second-stage meta-learning integration model. The first-stage feature-focused model group outputs preliminary prediction results, which are then input into the second-stage meta-learning integration model for integrated processing. The model outputs a capacity overload risk value and obtains the final risk prediction result. Based on the final risk prediction results, risk is classified, classified early warning information is generated and pushed to the mobile application terminal to obtain the early warning push results. Based on the early warning push results, a personalized intervention plan is matched and pushed to the mobile application terminal, providing intervention execution confirmation and effect tracking functions to obtain intervention management results.

2. The method according to claim 1, characterized in that, The preprocessing of the multimodal data to obtain a preprocessed multimodal dataset includes: Based on the multimodal data, a data cleaning algorithm is applied to detect and process missing values, outliers, and duplicate records in the data. A threshold range is set for the physiological monitoring data, and physiological monitoring data that exceeds the threshold range is marked or corrected. The dietary image data is subjected to quality screening to remove blurry, dark, or incomplete dietary images. The clinical test data is subjected to integrity verification to obtain the cleaned multimodal data. The cleaned multimodal data is format-standardized. The physiological monitoring data and the clinical test data are standardized with unified timestamp format and the units and precision of numerical data are standardized. The dietary image data is normalized in size and converted in color space, and naming and storage rules are established to obtain format-standardized multimodal data. Key features, including weight change rate, cumulative sodium intake, and medication adherence indicators, are extracted from the standardized multimodal data. Data from different sources and types are then time-aligned and feature-fused to obtain the preprocessed multimodal dataset.

3. The method according to claim 2, characterized in that, Image recognition and nutritional analysis are performed on the dietary image data in the preprocessed multimodal dataset, including: The dietary image data in the preprocessed multimodal dataset is subjected to image preprocessing, including brightness adjustment, contrast enhancement, cropping and scaling, to obtain the preprocessed dietary images. A labeled food image dataset is obtained, which contains various types of food image data and corresponding water content and sodium content parameters. A convolutional neural network model is constructed and trained using the transfer learning method to obtain a trained food recognition model. The preprocessed food image is input into the trained food recognition model, and the model infers and identifies the food category in the image to obtain the food recognition result. Based on the food identification results and the standard tableware reference in the image, the food volume is estimated by analyzing the depth information of the image, and the food weight is calculated by combining the food density parameters to obtain the food weight estimation result. Based on the estimated food weight, the food database is queried to obtain the water content and sodium content per unit weight of the corresponding food, and the total water content and total sodium content are calculated to obtain the dietary volume analysis results.

4. The method according to claim 3, characterized in that, Time-series analysis is performed on the physiological monitoring data in the preprocessed multimodal dataset, including: Physiological monitoring data, including weight data, blood pressure data, heart rate data, and blood oxygen saturation data, are extracted from the preprocessed multimodal dataset. The physiological monitoring data is then arranged in time series to obtain physiological monitoring time series data. A sliding time window is set for the weight data in the physiological monitoring time series data, and the weight change rate within the sliding time window is calculated. When the weight change rate exceeds a preset threshold, it is identified as an abnormal weight gain pattern, and the weight change trend characteristics are obtained. At the same time, the start and end times of the sliding time window are recorded as time window characteristics. Trend analysis is performed on the blood pressure, heart rate and blood oxygen saturation data in the physiological monitoring time series data. The moving average method is used to smooth the data series, the exponential smoothing method is used to predict the data trend, the deviation between the actual value and the predicted value is calculated, and when the deviation exceeds the preset fluctuation threshold, it is identified as an abnormal fluctuation pattern, thus obtaining the physiological indicator change trend characteristics. By comprehensively analyzing the weight change trend characteristics, the physiological indicator change trend characteristics, and the time window characteristics, a multi-dimensional physiological change feature vector is constructed to obtain the physiological change trend results.

5. The method according to claim 4, characterized in that, The clinical laboratory data in the preprocessed multimodal dataset are evaluated using metrics, including: Clinical test data, including N-terminal pro-brain natriuretic peptide (NT-proBNP) data, serum sodium concentration data, serum potassium concentration data, renal function index data, and hemoglobin data, are extracted from the preprocessed multimodal dataset to obtain a set of clinical test indicators. Based on the NT-proBNP data in the aforementioned clinical laboratory indicator set, and in accordance with the heart failure severity grading criteria, the patient's cardiac function status was assessed, and cardiac function assessment results were obtained. Based on the serum sodium concentration data, serum potassium concentration data, and renal function index data in the aforementioned clinical test index set, the electrolyte balance and renal compensatory capacity are assessed to obtain the metabolic status assessment results. Based on the hemoglobin data in the aforementioned clinical laboratory indicator set, the degree of anemia and its impact on cardiac function were assessed, and the anemia impact assessment results were obtained. The clinical status assessment results are obtained by combining the cardiac function assessment results, the metabolic status assessment results, and the anemia impact assessment results.

6. The method according to claim 5, characterized in that, The extraction of features and the subsequent feature engineering processing yield a feature-engineered dataset, including: Dietary intake features are extracted from the dietary volume analysis results; weight change trend features, physiological indicator change trend features, and weight monitoring time window features are extracted from the physiological change trend results; and clinical indicator features and symptom features are extracted from the clinical status assessment results to obtain the original key feature set. Based on the original key feature set, a variance filter is applied to remove features with variance less than a preset threshold. The recursive feature elimination method combined with the random forest algorithm is used to evaluate the importance of features. The correlation coefficient matrix between features is calculated, and highly correlated features with correlation coefficients greater than a preset threshold are identified and removed to obtain the filtered feature set. Logarithmic transformation is applied to the NT-proBNP index in the filtered feature set to address the skewed distribution. A time-series interactive feature is created based on the weight change trend feature and the dietary intake feature. The time-series interactive feature includes the product of the weight change rate and the cumulative value of sodium intake. A dietary intake time window feature is constructed based on the total water content in the dietary intake feature. The dietary intake time window feature includes the trend of water intake changes over the past 3 days. The feature engineering dataset is then obtained.

7. The method according to claim 1, characterized in that, The construction of a multi-level prediction model based on the feature engineering dataset includes: Based on the feature engineering dataset, the dataset is divided into time-series weight data, dietary feature data, clinical indicator data, and symptom data according to feature type. A time-series weight model is constructed using a long short-term memory network for the time-series weight data, a dietary feature model is constructed using a random forest algorithm for the dietary feature data, a clinical indicator model is constructed using a gradient boosting decision tree algorithm for the clinical indicator data, and a symptom model is constructed using a logistic regression algorithm for the symptom data, forming the first-stage feature-focused model group. Each feature-focused model in the first-stage feature-focused model group outputs a preliminary risk score and feature importance weight to obtain the preliminary prediction result. The preliminary risk score and feature importance weights in the preliminary prediction results are used as meta-features. Combined with the original key features in the feature engineering dataset, a gradient boosting decision tree algorithm is used to construct a meta-learning ensemble model as the second-stage meta-learning ensemble model. The outputs of each feature-focused model in the first-stage feature-focused model group are weighted and fused using Stacking ensemble technology. The weights of each feature-focused model in the first-stage feature-focused model group in the ensemble are dynamically adjusted, and the capacity overload risk value is output to obtain the final risk prediction result.

8. The method according to claim 1, characterized in that, The step of classifying risks based on the final risk prediction results, generating graded early warning information, and pushing it to the mobile application terminal to obtain early warning push results includes: Based on the volume overload risk value in the final risk prediction result, and combined with the preset risk grading threshold, the volume overload risk value is divided into four levels: normal, mild risk, moderate risk, and high risk, to obtain the patient's volume status grading result. Based on the patient capacity status classification results, a prompt-level warning message is generated for mild risk level, a warning-level warning message is generated for moderate risk level, and an emergency-level warning message is generated for high risk level. The classified warning messages are sent to the patient's and medical staff's mobile application terminals in real time through a message push service. For the high risk level, an SMS or telephone reminder is added to obtain the warning push result.

9. The method according to claim 1, characterized in that, The process involves matching personalized intervention plans based on the early warning push results and pushing them to the mobile application terminal, providing intervention execution confirmation and effect tracking functions, and obtaining intervention management results, including: Based on the risk level and the patient's specific risk status in the early warning push results, including sodium intake exceeding the preset daily sodium intake limit, water intake exceeding the preset daily water intake limit, or weight gain rate exceeding the preset weight change threshold, a corresponding intervention plan is matched from the preset intervention rule base. The intervention plan includes dietary adjustment suggestions to limit sodium intake, medication adjustment prompts to adjust diuretic dosage, and lifestyle guidance to reduce water intake, resulting in a matched personalized intervention plan. The personalized intervention plan is pushed to the patient's end of the mobile application terminal, and the intervention execution confirmation function is provided on the medical staff end. The medical staff can review and adjust the personalized intervention plan. The effect tracking function records the patient's intervention execution and capacity status changes to obtain the intervention management results.

10. A remote capacity management system for heart failure patients based on machine learning, characterized in that, include: A memory, wherein a computer program is stored; A processor for loading the computer program to implement the machine learning-based remote volume management method for heart failure patients as described in any one of claims 1-9.

Images