Diabetic nephropathy risk screening method and system optimized by intelligent algorithm

Abstract

The present application relates to a diabetes nephropathy risk screening method and system optimized by intelligent algorithm, relates to the field of intelligent medical treatment, obtains longitudinal time sequence data of a diabetes target user, carries out stage evolution mode labeling on structured physiological index sequences among them, identifies the progress time and the rollback time of each stage based on the time sequence change of the kidney function related index, aligns the image features extracted from the unstructured image data with the labeled structured index on the time axis, and constructs a multi-modal feature sequence; based on the multi-modal feature sequence, a risk state variable adjusted bidirectionally with time is constructed, and a continuous risk trajectory curve is generated; according to the inflection point analysis of the continuous risk trajectory curve, a multi-scale screening result is output based on the inflection point time interval. The technical problem of insufficient utilization of dynamic evolution information in long-term follow-up data, resulting in insufficient accuracy and comprehensiveness of risk assessment, is solved, and the effect of accurate, comprehensive and dynamic early warning of diabetes nephropathy risk screening is achieved.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare technology, and in particular to a method and system for screening the risk of diabetic nephropathy using intelligent algorithms. Background Technology

[0002] Diabetic nephropathy is one of the most common microvascular complications of type 2 diabetes. While the early microalbuminuria stage is somewhat reversible, it becomes irreversible once it progresses to the stage of massive proteinuria. Currently, clinical practice mainly relies on the urinary microalbumin-to-creatinine ratio (UACR) and estimated glomerular filtration rate (eGFR) for regular screening and staging management of diabetic patients. Staging thresholds (e.g., UACR = 30, 300 mg / g) are based on population statistical standards. Some studies have attempted to predict the risk of diabetic nephropathy using traditional machine learning or basic deep learning models. These models typically use a patient's historical structured indicators as input and output the probability of future kidney disease. In addition, there are also methods that utilize fundus images to extract microvascular lesion features for non-invasive screening of diabetic nephropathy.

[0003] However, the existing solutions mentioned above still have shortcomings in the following aspects: they ignore early reversibility and therefore cannot output reference information about disease regression; fixed staging thresholds do not take into account individual differences, such as sensitivity changes caused by factors such as age and disease course; the asynchronous acquisition time points of structured indicators and multimodal data such as fundus images lack an effective alignment and fusion mechanism; and the risk output is mostly static probability values, lacking analysis of the dynamic trend of risk changes, making it difficult to provide early warning before disease progression occurs. Summary of the Invention

[0004] This invention addresses the technical problem in existing diabetic nephropathy risk screening methods that do not fully utilize the dynamic evolution information in long-term patient follow-up data, resulting in insufficient accuracy and comprehensiveness of risk assessment. It provides a diabetic nephropathy risk screening method and system optimized by intelligent algorithms to solve this problem.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a method for screening the risk of diabetic nephropathy using intelligent algorithms, comprising: acquiring longitudinal time-series data of a target diabetic user, wherein the longitudinal time-series data includes structured physiological indicator sequences and unstructured image data collected at non-uniform time intervals; annotating the structured physiological indicator sequences with stage evolution patterns, identifying the progression and regression times between each stage based on the temporal changes of renal function-related indicators, and extracting microvascular lesion-related features from the unstructured image data; aligning the extracted image features with the annotated structured indicators on the time axis to construct a multimodal feature sequence; constructing a risk state variable that adjusts bidirectionally over time based on the multimodal feature sequence, wherein the risk state variable dynamically accumulates or decays risk values ​​according to the changing trend of indicator features at the current time relative to the previous time, generating a continuous risk trajectory curve; performing inflection point analysis based on the continuous risk trajectory curve, and outputting multi-scale screening results based on the inflection point time interval determined by the analysis.

[0006] Secondly, this invention provides a diabetic nephropathy risk screening system optimized by intelligent algorithms, comprising: a time-series data acquisition module for acquiring longitudinal time-series data of target diabetic users, wherein the longitudinal time-series data includes structured physiological indicator sequences and unstructured image data acquired at non-uniform time intervals; a feature sequence construction module for annotating the structured physiological indicator sequences with stage evolution patterns, identifying the progression and regression times between each stage based on the temporal changes of renal function-related indicators, extracting microvascular lesion-related features from the unstructured image data, aligning the extracted image features with the annotated structured indicators on the time axis, and constructing a multimodal feature sequence; a risk trajectory analysis module for constructing a risk state variable that adjusts bidirectionally over time based on the multimodal feature sequence, wherein the risk state variable dynamically accumulates or decays risk values ​​according to the changing trend of indicator features at the current time relative to the previous time, generating a continuous risk trajectory curve; and a multi-scale screening module for performing inflection point analysis based on the continuous risk trajectory curve, and outputting multi-scale screening results based on the inflection point time interval determined by the analysis.

[0007] The beneficial effects of this invention include: by acquiring non-uniformly sampled longitudinal multimodal data, the algorithm can fully utilize the dynamic changes in patients during long-term follow-up, overcoming the shortcomings of insufficient information in traditional cross-sectional data, and laying a data foundation for subsequent accurate analysis. Stage evolution annotation of structured indicators simultaneously identifies progression and regression points, incorporating the reversibility of early kidney injury into the quantitative system for the first time. This bidirectional annotation mechanism not only focuses on the disease progression path but also captures the possibility of improvement, thus solving the problem that existing algorithms can only predict progression unidirectionally and cannot reflect regression. It allows screening results to provide both probabilities of deterioration and improvement, improving the comprehensiveness of prediction. Simultaneously, the algorithm extracts microvascular lesion features from fundus images and aligns them with structured indicators on the time axis to construct a multimodal feature sequence. Through cross-modal complementarity, information about fundus lesions deteriorating earlier than kidney function indicators is effectively fused, solving the problems of difficult alignment of multi-source heterogeneous data and fragmented eye-kidney information, significantly enhancing the algorithm's early warning capability. The algorithm constructs a risk state variable that can be adjusted bidirectionally over time based on multimodal feature sequences. It dynamically accumulates or decays risk values ​​according to the trend of the current indicator relative to the previous time, generating a continuous risk trajectory curve. Because the curve completely preserves the temporal dynamics information such as the rate and acceleration of risk change, the output is upgraded from static probability to dynamic trajectory, solving the problems of lack of continuity and traceability in risk expression and making the screening results more intuitive. Finally, the algorithm performs inflection point analysis on the risk trajectory curve, identifying inflection points of abnormal risk acceleration by calculating the first and second-order rates of change, and providing graded warnings based on acceleration stability. This allows for the issuance of multi-scale screening results before the indicators reach the clinical diagnostic threshold, achieving true early intervention. From data fusion, bidirectional annotation, dynamic trajectory generation to inflection point graded warnings, this algorithm effectively solves the technical problem of insufficient utilization of the dynamic evolution information of long-term follow-up data in existing technologies, significantly improving the accuracy, comprehensiveness, and clinically practical screening effect of diabetic nephropathy risk assessment. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the diabetic nephropathy risk screening method optimized by intelligent algorithms provided by the present invention.

[0009] Figure 2 This is a schematic diagram of the structure of the diabetic nephropathy risk screening system optimized by intelligent algorithms provided by the present invention.

[0010] Figure labeling: Time series data acquisition module 1, feature sequence construction module 2, risk trajectory analysis module 3, multi-scale screening module 4. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0013] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0014] Example 1: like Figure 1 As shown, this embodiment of the invention provides a method for screening the risk of diabetic nephropathy using intelligent algorithms, including: Step S10: Obtain longitudinal time-series data of target users with diabetes, wherein the longitudinal time-series data includes structured physiological index sequences and unstructured image data collected at non-uniform time intervals.

[0015] Specifically, the raw data used for subsequent analysis and modeling is obtained, including data from sources such as hospital information systems, electronic health records, diabetes co-care management systems, or research team databases. After anonymization or user authorization, the data is collected from users diagnosed with type 2 diabetes. Longitudinal time-series data refers to multiple observation records collected from the same target user at at least three different time points, preferably with a continuous follow-up period of no less than 3 years and an interval of no more than 12 months between adjacent follow-ups. Each time point record is called a follow-up point. Due to the uncertainty of patient visit behavior, the time interval between adjacent follow-up points is usually non-uniform, i.e., a non-uniform time interval is used. For example, a patient may have follow-up records at 1 month, 4 months, 9 months, 18 months, and 30 months after diagnosis, with time intervals of 3 months, 5 months, 9 months, and 12 months, respectively. This invention does not require forced uniformization of the time intervals but retains the original interval information as part of subsequent modeling.

[0016] Structured physiological index sequences refer to quantitative indicators stored in numerical or categorical form that can be directly processed by algorithms. These include indicators closely related to diabetic nephropathy, such as core indicators of renal function: the urinary microalbumin to creatinine ratio (UACR, unit: mg / g); and estimated glomerular filtration rate (eGFR, unit: mL / min / 1.73m). 2 Serum creatinine (Cr, μmol / L). Glucose control indicators: Glycated hemoglobin (HbA1c, %); Fasting plasma glucose (FPG, mmol / L); 2-hour postprandial glucose (2hPG, mmol / L). Blood pressure and metabolic indicators: Systolic blood pressure (SBP, mmHg); Diastolic blood pressure (DBP, mmHg); Total cholesterol (TC), Triglycerides (TG), High-density lipoprotein (HDL-C), Low-density lipoprotein (LDL-C). Other auxiliary indicators: Height, weight, Body mass index (BMI), waist circumference, hip circumference, duration of diabetes, medication records, etc.

[0017] Unstructured image data includes fundus images of the target user. These images are acquired using a standard fundus camera, following standard operating procedures to capture color fundus photographs centered on the macula and with a 45° field of view. At least one image is acquired for each eye, typically a representative image of the right eye or the eye with more severe disease. Fundus images are stored in DICOM, JPEG, or PNG format, with a pixel resolution of at least 1024×1024, and sufficient clarity to identify diabetic retinopathy-related features such as microaneurysms, hemorrhages, hard exudates, and soft exudates. In certain embodiments, kidney ultrasound or CT images may also be included, but fundus images are a minimum requirement.

[0018] To meet the modeling requirements for metabolic memory effects and long-term evolution patterns in subsequent steps, the total follow-up period for the longitudinal time-series data is preferably no less than 3 years, preferably 5 years or more, and most preferably 10 years. This is because the progression of diabetic nephropathy from the normal phase to the microalbuminuria phase typically takes several years, and short-term data cannot reflect long-term trends. The maximum permissible interval between two adjacent follow-ups should not exceed 12 months. If an interval exceeds 12 months, the follow-up is considered insufficient, and the data may be discarded or require special processing, such as being marked as a long-interval record. Preferably, there is one complete test record per year containing both UACR and eGFR.

[0019] The collected longitudinal time-series data are preferably stored in the database in the form of structured tables. Each row represents a record of a follow-up point, including the user's unique identifier, timestamp, and the values ​​and missing bits of all structured indicators. Fundus images are associated with follow-up points in the form of file paths.

[0020] Step S20: The structured physiological index sequence is labeled with a stage evolution pattern. Based on the temporal changes of renal function-related indicators, the progression and regression times between each stage are identified. Microvascular lesion-related features are extracted from the unstructured image data. The extracted image features are aligned with the labeled structured indicators on the time axis to construct a multimodal feature sequence.

[0021] Specifically, the acquired raw longitudinal time-series data is converted into structured, multimodal feature sequences rich in clinical evolutionary semantics, including two dimensions: one is the stage evolution labeling of structured physiological indicator sequences. Since the pathological progression of diabetic nephropathy follows a clear staging pattern: normal stage – microalbuminuria stage – macroalbuminuria stage, and the early stages are reversible, simply using absolute values ​​of indicators cannot reflect the dynamic transition between stages. Therefore, based on the temporal trajectory of the core renal function indicators—the urinary microalbumin-to-creatinine ratio (UACR) and estimated glomerular filtration rate (eGFR)—each follow-up point is labeled with a stage classification, including: normal stage, microalbuminuria stage, or macroalbuminuria stage. Based on this labeling, progression and regression points are further identified. 2. Microvascular Lesion Feature Extraction and Alignment from Unstructured Image Data: Since diabetic nephropathy and diabetic retinopathy share a common microvascular pathological basis, features such as microaneurysms, hemorrhages, and exudations in fundus images can indirectly reflect the health status of renal microvessels. Furthermore, fundus lesions often precede renal disease. Therefore, extracting microvascular lesion-related features from unstructured image data such as fundus images is crucial for assisting early risk prediction. When extracting microvascular lesion-related features from unstructured image data such as fundus images, a fundus image encoder is pre-trained using self-supervised contrastive learning. The encoder is trained to distinguish fundus images from different stages of diabetic retinopathy. The embedded features output by the encoder are cross-modal aligned with contemporaneous structured indicators. When fundus images show microvascular abnormalities that precede the deterioration of renal function indicators, the degree of abnormality is used as a feedforward gain for the risk state variable, causing the risk value to rise earlier. The aligned image features are concatenated with the structured indicator vectors and stage annotation results to form a multimodal feature vector for follow-up points. All multimodal feature vectors for follow-up points are arranged chronologically to form a multimodal feature sequence.

[0022] In a preferred embodiment, the fundus image encoder is a deep learning model. The input is a fundus image, and the output is a fixed-length numerical vector, typically 128-dimensional, 256-dimensional, or 512-dimensional. This vector is called the embedding feature or representation vector. Images with similar fundus lesion features will have output vectors that are close to each other in the numerical space, while images with different lesion features will have output vectors that are far apart. The fundus image encoder employs a convolutional neural network architecture or a visual Transformer architecture. Taking ViT-L as an example, it contains 24 Transformer modules, each consisting of a multi-head self-attention layer and a feedforward network, which can globally model lesion features such as vascular morphology, microaneurysms, and hemorrhages in fundus images. Pre-training uses self-supervised contrastive learning. The training process involves randomly selecting an image from the fundus image dataset and performing two different data augmentation transformations, such as random cropping and scaling, rotation, flipping, and brightness / contrast perturbation, to generate two views as positive sample pairs. The encoder is forced to make the output vectors of these two views as close as possible in the latent space, i.e., maximizing similarity. Simultaneously, other images are selected from the dataset as negative samples, forcing the encoder to keep the vector of the current image far away from the vector of the negative samples, i.e., minimizing similarity. The encoder parameters are optimized by a contrastive loss function, enabling the encoder to learn representations sensitive to lesion features without manual annotation. After pre-training, supervised fine-tuning can be further performed on a publicly available diabetic retinopathy grading dataset to enhance the encoder's ability to distinguish different stages of retinopathy. After the encoder is trained, inference is performed on all available fundus images, and the embedded feature vector corresponding to each image is extracted. Since the acquisition time of fundus images and the acquisition time of structured indicators are usually not synchronized, it is necessary to align the image features and structured indicators on the time axis. A time proximity alignment strategy is adopted: for each follow-up point, a set of structured indicators and stage annotation results are provided. The fundus image acquisition record closest to t is found on the time axis. If the time interval is less than a preset threshold, such as 180 days, the embedded features of the fundus image are associated with the structured indicators of the current follow-up point. If the time interval exceeds the threshold, the corresponding follow-up point is marked as having no effective image features. After alignment, the aligned image embedding features, structured index vectors, stage annotation results, and follow-up time intervals are concatenated to form the multimodal feature vectors of the corresponding follow-up points. The multimodal feature vectors of all follow-up points are arranged in chronological order to form a multimodal feature sequence.

[0023] Furthermore, the structured physiological indicator sequence is labeled with a stage evolution pattern. Based on the temporal changes of renal function-related indicators, the progress and regression times between each stage are identified. This includes: sliding cumulative expansion of the renal function-related indicator sequence with a preset time window, merging follow-up points continuously within the same state range into a cumulative state block to establish a state evolution sequence. The cumulative window width of the sliding cumulative expansion adaptively increases with the continuous occurrence of the same state until the state changes or reaches a preset maximum cumulative duration; feature matching mapping is performed on the state evolution sequence according to preset stage evolution segmentation change characteristics to determine the stage evolution pattern; based on the temporal sequence of the stage evolution pattern in the state evolution sequence, sequential progress analysis and reverse regression analysis are performed to identify the progress and regression times.

[0024] Furthermore, the preset stage evolution segmentation change characteristics include: initial offset characteristics, which are the first state block in the state evolution sequence that deviates from the normal state and whose duration is less than a first threshold; critical crossing characteristics, which are the boundary points between adjacent state blocks of different risk levels and the duration of both the preceding and following state blocks is greater than a second threshold; and stability confirmation characteristics, which are the duration of the state block exceeding a third threshold and no regression event detected within at least one follow-up period after the state block ends.

[0025] Because patient follow-up time points are uneven and subject to short-term random fluctuations, this invention does not directly use the original follow-up points. Instead, it employs a sliding cumulative expansion method to merge follow-up points that are consecutively within the same state range into a cumulative state block, thereby obtaining a stable state evolution sequence. Specifically, step one uses the first follow-up point as the starting point of the window and sets an initial window width, such as covering the first two follow-up points. The window is then expanded forward in time. Whenever a follow-up point with the same state as an existing point appears at the end of the window, the window end time is automatically extended to that follow-up point. If a follow-up point with a different state is encountered, the expansion stops. Step two merges all consecutive follow-up points with the same state within the current window into a cumulative state block and records the state type: normal / micro / massive, start time, end time, and duration. Step three moves the window starting point to the follow-up point where the state changes and repeats steps one to two until all follow-up points have been processed. During the expansion process, the window width adaptively increases with the continuous occurrence of consecutive identical states until the state changes or the preset maximum cumulative duration is reached. If the duration of consecutive identical states exceeds the maximum cumulative duration, it is forcibly truncated and split into multiple state blocks to preserve time segmentation information. Through processing, the original non-uniform discrete follow-up point sequence is transformed into a state evolution sequence composed of cumulative state blocks arranged in chronological order.

[0026] After obtaining the state evolution sequence, three preset stage evolution segmentation change features are used for matching and mapping to determine the stage evolution pattern. The first feature is the initiation offset feature: the first state block in the state evolution sequence that deviates from the normal state, i.e., the first trace or large block, and the duration of this state block is less than the first threshold, such as 6 months. The starting time of the state block is the initiation offset time, used to mark the early warning point of the appearance of abnormal symptoms of the disease. The second feature is the critical crossing feature: the boundary point between two adjacent state blocks of different risk levels, and the duration of both the preceding and following state blocks is greater than the second threshold, such as 3 months. This boundary point is the critical crossing time, used to identify stable and substantial stage transitions, such as normal-trace or trace-large. The third feature is the stability confirmation feature: the duration of a state block exceeds the third threshold, such as 12 months, and no regression event is detected within at least one follow-up period after the end of the state block, i.e., no transition to a low-risk stage has occurred. The starting point of the state block is the stability confirmation time, used to mark that the patient has stably entered a new stage and that it is irreversible.

[0027] Based on the temporal position of the aforementioned stage evolution patterns in the state evolution sequence, sequential progress analysis and reverse regression analysis were performed. Sequential progress analysis traversed adjacent state blocks in forward time. When the state level increased, such as from normal to trace or trace to large, the critical transition moment was preferentially taken as the progress moment. If there was no critical transition, the starting offset moment or the starting point of the next block was taken. Reverse regression analysis traversed in reverse time. When the state level decreased, such as from large to trace or trace to normal, and conditions such as the duration of the next block and no rapid re-progression after regression were met, the regression moment was identified. Through bidirectional analysis, all progress and regression moments for each patient during long-term follow-up were finally obtained, serving as key monitoring signals for subsequent risk modeling.

[0028] For example, patient data: UACR time series is normal 12 months - trace 6 months - normal 12 months - trace 14 months - large 20 months, with the first threshold set to 6 months, the second threshold to 3 months, and the third threshold to 12 months. After processing, the state evolution sequence is obtained: normal block (12) - trace block (5) - normal block (12) - trace block (14) - large block (20). Initial offset feature: the first trace block lasts for 5 months < 6 months - marked as initial offset, the initial offset time = trace block start point. Critical crossing feature: the second trace block (14) - large block (20), the duration of both blocks is > 3 months - marked as critical crossing, the time = trace block end / large block start boundary. Stability confirmation feature: the large block lasts for 20 months > 12 months, and there is no regression afterward, then it is marked as stable confirmation, the time = large block start point.

[0029] Sequential progression analysis yields: Progression time = Initial offset time (early warning) and critical crossing time (substantial progression). Reverse regression analysis yields: Scanning from back to front reveals a micro-block (14) - normal block (12), which is actually a transition from normal to micro, not a regression. However, it is noted that the first micro-block (6) - normal block (12) pair is a regression in the forward direction, i.e., from micro to normal, and the subsequent block is a normal block that lasts for 12 months. There is no further progression during the follow-up period after the regression. In fact, there is a subsequent micro-block (14), but it only appears 12 months after the end of the normal block (12), which exceeds the set follow-up period of 6 months. Therefore, it can be considered to meet the conditions, and the regression time is the starting point of the normal block. The regression time can be identified through reverse analysis. This achieves the effect of automatically and accurately marking the progression time and regression time for each patient, providing key temporal supervision signals for subsequent risk state variable modeling.

[0030] Furthermore, the adaptive growth rule of the sliding cumulative expansion is as follows: whenever a follow-up point with the same state as the current window appears at the end of the current window, the window end time is automatically extended to the current follow-up point, and the expansion operation is not included in the statistics of the preset maximum cumulative duration. The first follow-up point when the state changes is used as the starting point of the next state block.

[0031] Specifically, in constructing the cumulative state block, this invention employs an adaptive sliding window to traverse the time sequence of renal function-related indicators. The window's end time is not a fixed step size, but rather dynamically determined based on whether the state of the newly added follow-up point matches the existing states within the current window. First, the patient's renal function-related indicators are arranged chronologically, forming a series of time-stamped follow-up points. Each follow-up point is determined to be in the normal period, microalbuminuria period, or macroalbuminuria period based on its UACR or eGFR value. The algorithm starts from the first follow-up point, establishing an empty cumulative state block, using the state of the first follow-up point as the block's state, and the time of that follow-up point as the block's start time. Then, each subsequent follow-up point is examined sequentially in forward time. For the currently examined follow-up point, the following judgment is performed: State consistency extension judgment: If the state of the currently examined follow-up point is the same as the state of all points already in the current cumulative state block, then the follow-up point is included in the current block, and the end time of the current block is automatically extended to the time of the follow-up point. After extension, the next follow-up point is examined. The extension operation itself is not counted in any independent count of extension counts or extension steps, but the actual time consumed by the extension, i.e., the time interval from the block start point to the current follow-up point, is still included in the actual duration of the block. To limit the indefinite extension of a single state block, a maximum cumulative duration is preset as a forced truncation threshold, such as 24 months. The forced truncation threshold is judged only after each extension is completed and before the next follow-up point arrives. If the time interval from the block start point to the latest included follow-up point has reached or exceeded the maximum cumulative duration, then even if the state of the next follow-up point is the same, extension will not continue, and the current block will be forcibly closed. The maximum cumulative duration limit prevents a single state block from becoming too long and losing its internal time distribution information.

[0032] Block sealing and new block initiation during state changes: If, during the observation process, the state of the current follow-up point differs from the state of the current cumulative state block, the expansion of the current block is immediately stopped. The current block is sealed, and its end time is the time of the previous follow-up point, i.e., the time of the last point with the same state as the block. This sealed block becomes the cumulative state block in the state evolution sequence. The first follow-up point where the state changed—the point that caused the state difference—is taken as the starting point of the next new cumulative state block. The state of the new block is taken from the state of that follow-up point, and its start time is the time of that point. Then, starting from this new point, the above expansion and judgment steps are repeated.

[0033] Special handling for forced truncation: When a situation arises where the states are identical but the time interval has exceeded the maximum cumulative duration, the system performs forced truncation, closing the current block and setting the end time to the block start time plus the maximum cumulative duration. The current follow-up point under observation—the point that triggered the truncation—is then used as the starting point for a new block, with the new block's state identical to that point, beginning a new round of accumulation. This prevents a long-term stable state, such as several consecutive years of normal operation, from being compressed into an excessively large block, thus losing time segmentation information. Through forced truncation, the long-term stable state is split into multiple consecutive blocks with durations not exceeding the maximum cumulative duration.

[0034] Handling of individual follow-up points: In real-world data, there may be isolated follow-up points with different states before and after, such as an isolated micro-point appearing after a normal period, followed by a return to the normal period. In this case, the algorithm treats this isolated point as a separate cumulative state block, with its start and end times both being the time of the follow-up point, and its duration recorded as 0. State blocks with a duration of 0 are still meaningful in subsequent feature matching because they signify a brief state deviation.

[0035] All follow-up points were processed one by one. When the last follow-up point was included in a block or used as the starting point of a new block, the block was closed and saved. Finally, the entire follow-up sequence was converted into a set of cumulative state blocks arranged in chronological order. Each block recorded the state type, start time, end time, and duration. These blocks constituted the state evolution sequence, which served as the input for subsequent three-feature matching and bidirectional analysis.

[0036] This adaptive growth rule divides the entire time series data into a series of continuous cumulative state blocks, which can flexibly handle non-uniformly sampled clinical follow-up data. It can preserve the continuity of long-term stable states and accurately capture the boundaries of state transitions. At the same time, the forced truncation avoids excessive length of individual state blocks, providing a clear and complete state evolution sequence for subsequent stage evolution analysis.

[0037] Furthermore, identifying the progression and regression times includes: calculating the transition confidence scores for multiple indicators related to renal function, where different indicators have different confidence weights in different transition types; for progression from the normal phase to the microalbuminuria phase, the microalbumin-to-creatinine ratio has a higher weight than the estimated glomerular filtration rate (GFR); for progression from the microalbuminuria phase to the macroalbuminuria phase, the estimated GFR has a higher weight than the microalbumin-to-creatinine ratio; for regression events, the microalbumin-to-creatinine ratio has the highest weight, and it is required that the estimated GFR has not shown an irreversible decline; the final transition time is determined by the weighted confidence scores of each indicator exceeding a fusion threshold.

[0038] Specifically, in identifying progression and regression points, to overcome the noise or lag that may occur with a single indicator, such as relying solely on UACR, this invention introduces multiple renal function-related indicators and assigns differentiated confidence weights to different indicators in different transition types. A weighted fusion approach is used to comprehensively determine the transition point. At least two core renal function indicators are selected: the urinary microalbumin-to-creatinine ratio (UACR) and the estimated glomerular filtration rate (eGFR). UACR reflects damage to the glomerular filtration membrane barrier function and is sensitive to early renal injury; eGFR reflects glomerular filtration function and is sensitive to mid-to-late-stage renal function decline. The dynamic changes of these two indicators have different clinical significance at different stages of diabetic nephropathy. For each candidate transition point, such as candidate time points identified by changes in state before and after cumulative state block boundaries or follow-up points, the transition confidence scores for UACR and eGFR are calculated separately. The transition confidence score is a value between 0 and 1, indicating the degree to which the indicator supports the occurrence of a transition, i.e., progression or regression.

[0039] For progression from the normal phase to the microalbuminuria phase, the UACR conversion confidence level is based on the ratio of the current UACR value to the UACR value at the previous follow-up period, as well as the absolute change. If the UACR increases from below 30 mg / g to above 30 mg / g with an increase exceeding 20%, the confidence level is set at 0.9; if it only slightly exceeds the threshold (e.g., 30-35 mg / g) with a small increase, the confidence level is set at 0.5; if there is no significant increase or decrease, the confidence level is 0. For progression from the microalbuminuria phase to the macroalbuminuria phase, the UACR confidence level is calculated similarly, based on a 300 mg / g threshold, but with a reduced weighting.

[0040] For the progression from the trace phase to the large phase, the confidence level of eGFR is based on the rate of decline and the absolute value. If the eGFR decreases by more than 30% from baseline and the current value is below 60 mL / min / 1.73 m, the confidence level is determined. 2 If the decline is slow or the absolute value is still above 60, the confidence level is set to 0.8; if the decline is slow or the absolute value is still above 60, the confidence level is lower. For rollback events, the confidence level of eGFR must be stable or improved, and there must be no irreversible decline, that is, eGFR must continue to decline for more than 3 months without any upward trend, otherwise the rollback determination will be rejected.

[0041] The importance of each indicator varies across different types of events, therefore different weights are assigned: Progression from the normal phase to the microalbuminuria phase: UACR is weighted at 0.7, and eGFR at 0.3. This is because early kidney damage primarily manifests as microalbuminuria, at which point eGFR often has not yet shown a significant decline. Progression from the microalbuminuria phase to the macroalbuminuria phase: UACR is weighted at 0.4, and eGFR at 0.6. In the middle and late stages, the decline in glomerular filtration rate becomes more critical, and once eGFR irreversibly decreases, the prognosis is worse. Regression events, from the macroalbuminuria phase to the microalbuminuria phase or from the microalbuminuria phase to the normal phase: UACR is weighted at 0.8, and eGFR at 0.2. An additional constraint is added: it must be confirmed that eGFR has not shown an irreversible decline, i.e., the recent trend of eGFR change is stable or increasing, and there has been no sustained decline for more than 3 months; otherwise, even if UACR improves, it is not considered an effective regression. This reflects the clinical consensus that once irreversible damage to kidney function occurs, improvement in proteinuria alone does not represent a true recovery.

[0042] For each candidate transition point, calculate the overall transition confidence score: Overall Confidence Score = (UACR Confidence Score × UACR Weight) + (eGFR Confidence Score × eGFR Weight), with a preset fusion threshold, such as 0.65. When the overall confidence score reaches or exceeds this threshold, the candidate point is determined as a valid transition moment, progress moment, or regression moment. If the overall confidence score is below the threshold, it is not marked as a transition. The fusion threshold can be optimized and determined based on the validation set data through methods such as grid search to ensure the best batching effect. Preferably, the fusion threshold is set between 0.6 and 0.7.

[0043] For example, in the early progression candidate point, UACR increased by 52% from 25 to 38 mg / g (confidence level 0.9), and eGFR decreased by 2% from 95 to 93 (confidence level 0.2). The overall score is 0.9 × 0.7 + 0.2 × 0.3 = 0.69 ≥ 0.65, indicating effective progression. In the late progression candidate point, UACR increased by 0.9% from 80 to 350 mg / g (confidence level 0.9, weight 0.4), and eGFR decreased by 35% from 70 to 45 (confidence level 0.9, weight 0.6). The overall score is 0.9 × 0.4 + 0.9 × 0.6 = 0.9, indicating effective progression. In the regression candidate point, UACR decreased by 0.9% from 60 to 25 mg / g (confidence level 0.9, weight 0.8), and eGFR remained stable from 80 to 82 (confidence level 0.9, weight 0.2), with no irreversible decrease. The overall score is 0.9, indicating effective regression. Candidate point for regression but eGFR continues to decline: If UACR improves, but eGFR declines from 75 to 55 for three consecutive months without recovery, regression is directly rejected and not marked as regression point.

[0044] Furthermore, the step of merging follow-up points that are continuously within the same state range into a cumulative state block is adopted using a boundary probability distribution method. Specifically, for each follow-up point, the probability distribution of belonging to the normal period, microalbuminuria period, and macroalbuminuria period is calculated. When the probability of adjacent follow-up points belonging to the same state exceeds a preset threshold, they are merged into a probability block. The boundary of the state block is defined as the fuzzy region of the probability distribution transformation, where the width of the fuzzy region serves as a weighting factor for the evolutionary features of the matching stage.

[0045] Specifically, in the process of merging follow-up points continuously within the same state range into a cumulative state block using sliding cumulative expansion, a boundary probability distribution method can be used as a feasible implementation to more accurately determine the state range to which each follow-up point belongs. This method calculates the probability distribution for each follow-up point belonging to the normal period, microalbuminuria period, or macroalbuminuria period based on the measured values ​​of renal function indicators and the coefficient of variation. When the probability of adjacent follow-up points belonging to the same state exceeds a preset threshold, these points are merged into a single probability block, thus forming a cumulative state block in the state evolution sequence. The boundary of the state block is defined as the fuzzy region of the probability distribution transformation, and the width of the fuzzy region serves as a weighting factor for the evolutionary features in the subsequent matching stage, used to quantify the abruptness of the transformation. In practice, it is assumed that the UACR measurement value follows a normal distribution with the mean of the detected value and the standard deviation of the coefficient of variation, which is typically 5%-10%. For each follow-up point, the probability of UACR falling into the following three intervals is calculated: normal period interval <30mg / g, microalbuminuria period interval 30-300mg / g, and macroalbuminuria period interval ≥300mg / g. Simultaneously, the auxiliary constraint of eGFR is considered: if eGFR <30mL / min / 1.73m 2 If the eGFR is between 30 and 60, the probability of the massive albuminuria phase is forcibly set to 1, and the probability of the other two phases is 0. If the eGFR is between 30 and 60, the probability of the trace phase is appropriately reduced. The final state probability at each follow-up point can be the normalized result of the three factors. For example, if the UACR is 28 mg / g and the coefficient of variation is 5%, the probability of the normal phase is about 0.85, the probability of the trace phase is about 0.15, and the probability of the massive phase is about 0. If the UACR is 32 mg / g, the probability of the normal phase is about 0.45, the probability of the trace phase is about 0.55, and the probability of the massive phase is about 0.

[0046] The probability block merging rule is as follows: A probability threshold is set, and follow-up points are processed sequentially in forward time. If the probability of a certain state of the current follow-up point exceeds the probability threshold, the point is temporarily assigned to that state. When the probability of multiple consecutive follow-up points belonging to the same state exceeds the probability threshold, these points are merged into a probability block. The starting point of the block is the first point whose probability exceeds the threshold, and the ending point is the last point whose probability exceeds the threshold. The duration is the time interval between the two points. If the highest probability state of a follow-up point is different from the points before and after it, or if the highest probability is lower than the probability threshold, it is not merged into any block but is treated as a separate boundary point or processed together with adjacent ambiguous regions. Through the merging rule of this invention, each probability block represents a high-confidence continuous state interval. For example, if the normal period probabilities of three consecutive follow-ups are 0.9, 0.85, and 0.92, respectively, all exceeding 0.7, they are merged into a normal probability block.

[0047] Furthermore, when the probability distribution of adjacent follow-up points changes, such as gradually transitioning from a probability-dominant period to a probability-dominant period, the region where the probabilities of the two are close is called the fuzzy region. Specifically, the transition start point is defined as the first follow-up point where the probability of the original state is below 0.6 and the probability of the new state is above 0.4. The transition end point is defined as the last follow-up point where the probability of the new state exceeds 0.7 and the probability of the original state is below 0.3, or the first point to stably enter the new state. The width of the fuzzy region = the time of the transition end point - the time of the transition start point. If the transition occurs very rapidly, such as a direct jump between two follow-ups, the width of the fuzzy region is approximately the time interval between the two follow-ups. If the transition is gradual, such as a gradual change over multiple follow-ups, the width is larger. The width of the fuzzy region is recorded and used as a weighting factor in subsequent evolutionary feature matching, such as critical crossing features. A smaller width indicates a more abrupt transition, corresponding to a higher significance of risk change, and the confidence score of that transition point is increased during feature matching; a larger width indicates a more gradual transition, possibly suggesting a gradual process, and the risk weight is correspondingly reduced. In practical applications, the width can be normalized and multiplied by a preset coefficient to adjust the overall confidence level in the aforementioned multi-indicator weighted fusion, or it can be directly used as an additional condition for critical transition features, such as requiring the width to be less than a certain threshold to be marked as an abrupt transition. The boundary probability distribution method ensures the robustness of state block merging, avoids boundary jitter caused by hard thresholds, and the width of the fuzzy region provides additional temporal dynamics information for the stage evolution features, making it possible to quantify the abruptness of the transition later, thereby improving the accuracy and interpretability of the overall prediction model.

[0048] Step S30: Based on the multimodal feature sequence, construct a risk state variable that adjusts bidirectionally over time. The risk state variable dynamically accumulates or decays risk values ​​according to the changing trend of the indicator characteristics at the current moment relative to the previous moment, generating a continuous risk trajectory curve.

[0049] Furthermore, generating a continuous risk trajectory curve includes: extracting short-term change features, medium-term trend features, and long-term memory features from the multimodal feature sequence; using exponential decay weighting to simulate the metabolic memory effect for the long-term memory features; initializing risk state values ​​and updating them bidirectionally point-by-point based on time sequence; where, when an indicator deteriorates, the risk value is positively accumulated according to the rate of deterioration, with faster accumulation occurring the faster the deterioration; when an indicator improves or external intervention occurs, the risk value is negatively decayed at a limited rate, with the decay rate decreasing as the current risk value increases, and the intervention effect decaying exponentially over time; weighting and fusing the updated risk state value with the short-term change features, medium-term trend features, and long-term memory features to obtain the fused risk trajectory points; dynamically adjusting the short-term weights based on the consistency between the short-term change direction and the long-term trend; and smoothing the fused risk trajectory points to generate a continuous risk trajectory curve.

[0050] Specifically, a risk state variable that adjusts bidirectionally over time is constructed based on multimodal feature sequences, generating a continuous risk trajectory curve that reflects dynamic risk changes. This transforms dispersed multimodal time-series information into a continuous, bidirectionally evolving risk trajectory curve, achieving an upgrade from static probability to dynamic process risk expression. This provides intuitive and quantifiable technical support for early warning and individualized intervention in diabetic nephropathy. For each follow-up point, three different time-scale features are extracted from the multimodal feature sequences: short-term change features of instantaneous fluctuations, medium-term trend features of recent evolution direction, and long-term memory features of metabolic memory effects. Among them, short-term change features reflect the speed and magnitude of indicator changes between two adjacent follow-ups. That is, for each core indicator, such as the urinary microalbumin to creatinine ratio (UACR), estimated glomerular filtration rate (eGFR), systolic blood pressure, and glycated hemoglobin, the change relative to the previous follow-up is calculated, and then divided by the number of days between the two follow-ups to obtain the daily change rate. For example, if UACR increases from 25 mg / g to 40 mg / g over a 90-day interval, the daily change rate is approximately 0.167 mg / g / day. The daily change rates of all indicators are aggregated into a vector as a short-term change characteristic. If an indicator is missing, the value of that dimension is set to 0 and an additional missing flag is recorded.

[0051] The intermediate-term trend characteristics are calculated using the current time point as the endpoint, taking the indicator values ​​from the past 6 months or at least 3 follow-up visits, and using a linear fitting method to calculate the trend. The trend is represented by the slope; a positive slope indicates that the indicator is rising recently, and a negative slope indicates that it is falling. The absolute value reflects the rate of increase or decrease. For example, if the UACR values ​​for the past 6 months are 28, 32, and 38 mg / g, and the slope of the fitted value is positive, it indicates that microalbuminuria is gradually increasing. If the follow-up period is less than 6 months, all available data are used.

[0052] Long-term memory is characterized by the fact that blood glucose, blood pressure, and other indicators from several years ago can still affect the current risk in diabetic nephropathy patients; this is known as metabolic memory. This invention uses an exponentially decaying weighted average to simulate this effect. Historical values ​​further back in time have lower weights, while more recent historical values ​​have higher weights. For each indicator, the weighted average of all historical values ​​prior to the current time point is calculated. The weight decreases exponentially with the retrospective time, with a half-life set at 180 days. This means that an indicator from six months ago has approximately 50% of the current weight, an indicator from one year ago has approximately 25%, and so on. Missing historical values ​​are not included in the calculation, and the denominator is adjusted accordingly.

[0053] The two-way update mechanism for risk status variables includes positive accumulation and negative decay. Positive accumulation involves initializing the risk status value to 0, then processing each follow-up point sequentially over time, updating the risk status value after each point is processed. If a core indicator in the short-term change characteristics is deteriorating, the cumulative amount of this deterioration is calculated. The magnitude of the cumulative amount is proportional to the rate of deterioration; the faster the deterioration, the larger the cumulative amount. Specific rules: First, find the largest daily rate of change among all deteriorating indicators. The base cumulative amount = deterioration rate × base coefficient, set to 0.1, indicating that for every 1 mg / g / day increase in the deterioration rate, the risk increases by 0.1. If the rate of deterioration exceeds a preset rapid deterioration threshold, such as a daily increase in UACR exceeding 0.5 mg / g / day, an additional fixed acceleration value of 0.2 is added to reflect the high risk of rapid deterioration. Furthermore, if both the medium-term trend and long-term memory indicate an upward trend (i.e., both slopes are positive), the cumulative amount is multiplied by 1.2 to strengthen the risk signal from trend consistency. The final positive cumulative amount = base cumulative amount + possible acceleration value, multiplied by the trend consistency coefficient. Negative decay specifically involves calculating the decay amount if short-term changes show improvement in indicators, or if there is an external intervention event such as medication adjustment, dietary control, or increased exercise. A maximum decay rate is set, such as a maximum daily decay of 0.05 risk values. The actual decay amount is calculated as: maximum decay rate × (1 - current risk value squared) × intervention effect factor. If there is no intervention and only natural improvement, the intervention effect factor is set to 0.8. If there is intervention and patient compliance is good, the factor can be determined based on follow-up visit records or medication dispensing records, and is set to 1.2. Furthermore, the decay amount cannot exceed 20% of the current risk value to prevent excessive decay at once. The intervention effect decays over time. If the intervention relied upon for this update occurred a long time ago, its effect needs to be discounted. Specifically, each intervention generates a decay pulse, and the pulse intensity decreases exponentially with the number of days since the intervention, such as a half-life of 35 days. Multiple interventions can be accumulated.

[0054] Risk status is updated based on positive accumulation and negative decay. The new risk status value = old risk status value + positive accumulation - negative decay - residual intervention decay pulse. The new value is then limited to between 0 and 1, where 0 represents no risk and 1 represents extremely high risk. Thus, the risk status variable can move in both directions, rising when the indicator deteriorates and falling when it improves or is intervened. The rate of rise and fall is constrained by clinical characteristics.

[0055] Relying solely on the updated risk status value may be too sensitive. Therefore, it is necessary to integrate medium-term trends and long-term memory features to obtain the final risk trajectory point. The integration formula uses a weighted sum, with the initial weights set as follows: short-term 20%, medium-term trend 30%, and long-term memory 50%. At the same time, to prevent interference from short-term fluctuations, a dynamic adjustment mechanism is added. The direction of short-term changes (whether most indicators are rising) is compared with the direction of the long-term trend (the overall trend of long-term memory feature values). If both are in the same direction (rising or falling), the short-term weight remains unchanged at 20%. If the two are in opposite directions, the short-term weight is reduced to 10%, and the saved 10% weight is distributed to the medium and long term proportionally to weaken short-term noise that contradicts the long-term trend and improve the stability of the trajectory.

[0056] Finally, the obtained series of discrete risk trajectory points are smoothed. Each point corresponds to one follow-up time. Due to the uneven follow-up intervals, in order to obtain a smooth trajectory curve that is continuous in time and can support risk queries at any time point, interpolation and smoothing are required to obtain a continuous risk trajectory curve.

[0057] Step S40: Perform inflection point analysis based on the continuous risk trajectory curve, and output multi-scale screening results based on the inflection point time interval determined by the analysis.

[0058] Furthermore, inflection point analysis is performed based on the continuous risk trajectory curve, and multi-scale screening results are output based on the inflection point time interval determined by the analysis. This includes: calculating the first-order and second-order rates of change of the continuous risk trajectory curve at each follow-up point, and statistically analyzing the sign change frequency of the second-order rate of change; classifying the inflection points into three levels—stable acceleration type, fluctuating acceleration type, and oscillating observation type—based on the absolute value of the second-order rate of change and the sign change frequency, corresponding to Level 1, Level 2, and Level 3 warnings, respectively; the output content includes: the predicted probability of progression to the massive albuminuria stage, the predicted probability of regression to the normal stage, the identified inflection point level and inflection point time interval, and a visual graph of the risk trajectory curve.

[0059] Specifically, based on the continuous risk trajectory curve, by calculating the rate of change, acceleration of change, and stability of acceleration of risk, different types of risk inflection points are identified. Different levels of screening feedback information are output according to the characteristics of the inflection points to detect turning points where risk accelerates, providing more refined decision support for clinical intervention. The continuous risk trajectory curve is a smooth curve with time on the horizontal axis and risk status value on the vertical axis. Each point on the curve represents the patient's kidney disease risk level at a certain moment, with values ​​between 0 and 1, where 0 represents no risk and 1 represents extremely high risk. The first-order rate of change is the slope of the tangent line at a point on the continuous risk trajectory curve, representing the rate of change of risk. A positive first-order rate of change indicates that the risk is increasing, and a negative value indicates that the risk is decreasing; the larger the absolute value, the faster the change. The rate of change of the first-order rate of change represents the risk acceleration, i.e., the second-order rate of change. A positive second-order rate of change indicates that the rate of risk increase is accelerating, and a negative value indicates that the rate of risk increase is slowing down or the rate of risk decrease is accelerating. To assess the stability of risk acceleration, a certain number of historical points (excluding the current point) are traced back at each follow-up point. The signs of the second-order rates of change at these points are recorded, and the number of times these signs change is counted as the sign change frequency. For example, within a given time window, if the second-order rate of change changes from positive to negative or from negative to positive in the order of positive, positive, negative, then the sign change occurs once (from positive to negative); if it is positive, negative, positive, then two changes occur. The sign change frequency is the number of changes within the specified window. The length of the given window can be correlated with the median follow-up time interval. For example, if patients are followed up on average every 90 days, the traceback window can be taken as approximately 270 days from the three most recent follow-ups. An adaptive window can better adapt to the differences in follow-up frequency among different patients. The sign change frequency reflects the stability of risk acceleration. A low frequency indicates a stable acceleration direction and a single trend in risk change, while a high frequency indicates frequent sign changes in acceleration and an oscillating risk trajectory.

[0060] Based on the absolute value of the second-order rate of change and the frequency of sign change, inflection points are divided into three levels, each corresponding to different warning levels and clinical recommendations. These include: Level 1, a stable accelerating inflection point, where the absolute value of the second-order rate of change exceeds a preset acceleration threshold and the frequency of sign change does not exceed once. This indicates that the risk is accelerating at a stable rate, and the condition is in a rapid deterioration phase. In this case, it is recommended that doctors immediately review the current treatment plan, consider strengthening intervention measures, such as adjusting antihypertensive drugs, hypoglycemic drugs, or adding nephroprotective drugs, and shorten the follow-up interval.

[0061] The second level is the inflection point of accelerated fluctuation. When the absolute value of the second-order rate of change exceeds the acceleration threshold, but the frequency of sign change reaches two or more times, it indicates that the risk is accelerating and accompanied by violent fluctuations, suggesting that the condition is very unstable and may be affected by multiple factors such as poor adherence to intermittent medication, infection, and dietary fluctuations. The urgency of this inflection point is higher than that of the first level. It is recommended to urgently assess the patient's condition, investigate possible influencing factors in the near future, and consider hospitalization or enhanced outpatient management.

[0062] The third level is the oscillating observation inflection point. When the absolute value of the second-order rate of change is below the acceleration threshold, but the sign change frequency reaches three or more times, it indicates that although the risk has not risen rapidly, the direction of risk change is frequently switching, reflecting that the system is in an oscillating state, which is often a precursor to a more serious inflection point. At this time, it is recommended to increase the monitoring frequency, such as monthly review, to closely track the trend, and at the same time assess the patient's lifestyle and psychological state to prevent a rapid deterioration in the future.

[0063] The acceleration threshold can be determined based on the second-order rate of change distribution of the training dataset, such as taking the 75th percentile of the sample, or using a grid search to best match the warning with actual clinical deterioration events. Finally, a multi-scale screening result report is output, including: a predicted probability of progression to the massive albuminuria stage based on the current risk value and recent trends; an assessment of the probability of recovery for patients currently in the microalbuminuria stage; a predicted probability of regression to the normal stage; the identified inflection point level and corresponding inflection point time interval, such as a Level 1 warning occurring between May and August 2023; and a visual graph of the risk trajectory curve. The curve uses different colors or markers to indicate the positions of each inflection point, and can include annotations of key events such as medication adjustments and blood pressure fluctuations to help clinicians intuitively understand the risk evolution process. This invention's inflection point grading warning mechanism transforms a simple risk curve into multi-level alarm signals with clear clinical implications, enhancing the model's interpretability and practical value, and achieving an effective connection from risk assessment to clinical intervention.

[0064] Example 2: like Figure 2 As shown, based on the same inventive concept as the diabetic nephropathy risk screening method optimized by intelligent algorithms provided in Embodiment 1, this embodiment of the invention also provides a diabetic nephropathy risk screening system optimized by intelligent algorithms, the system comprising: The time-series data acquisition module 1 is used to acquire longitudinal time-series data of target users with diabetes, including structured physiological index sequences and unstructured image data acquired at non-uniform time intervals.

[0065] The feature sequence construction module 2 is used to annotate the stage evolution pattern of the structured physiological index sequence, identify the progress and regression times between each stage based on the temporal changes of renal function-related indicators, extract microvascular lesion-related features from the unstructured image data, align the extracted image features with the annotated structured indicators on the time axis, and construct a multimodal feature sequence.

[0066] Risk trajectory analysis module 3 is used to construct a risk state variable that adjusts bidirectionally over time based on the multimodal feature sequence. The risk state variable dynamically accumulates or decays risk values ​​according to the changing trend of the indicator characteristics at the current moment relative to the previous moment, generating a continuous risk trajectory curve.

[0067] The multi-scale screening module 4 is used to perform inflection point analysis based on the continuous risk trajectory curve and output multi-scale screening results based on the inflection point time interval determined by the analysis.

[0068] Furthermore, the feature sequence construction module includes: The state evolution module is used to perform sliding cumulative expansion on the sequence of renal function-related indicators within a preset time window. Follow-up points that are continuously in the same state range are merged into a cumulative state block to establish a state evolution sequence. The cumulative window width of the sliding cumulative expansion adaptively increases with the continuous occurrence of the same state until the state changes or the preset maximum cumulative duration is reached.

[0069] The stage evolution module is used to perform feature matching mapping on the state evolution sequence according to the preset stage evolution segmentation change characteristics to determine the stage evolution mode.

[0070] The sequential progress analysis module is used to perform sequential progress analysis and reverse rollback analysis based on the temporal sequence of the state evolution pattern of the stage evolution mode, and to identify the progress time and rollback time.

[0071] Furthermore, the pre-defined stage evolutionary segmentation and change characteristics include: The initial offset feature is the first state block in the state evolution sequence that deviates from the normal state, and its duration is less than a first threshold.

[0072] The critical crossing feature is the boundary point between adjacent state blocks of different risk levels, where the duration of both the preceding and following state blocks is greater than a second threshold.

[0073] The stable confirmation feature is that the duration of the state block exceeds a third threshold, and no rollback event is detected within at least one follow-up period after the state block ends.

[0074] Furthermore, the adaptive growth rule of the sliding cumulative expansion is as follows: whenever a follow-up point with the same state as the current window appears at the end of the current window, the window end time is automatically extended to the current follow-up point, and the expansion operation is not included in the statistics of the preset maximum cumulative duration. The first follow-up point when the state changes is used as the starting point of the next state block.

[0075] Furthermore, the sequential progress analysis module is used to perform the following steps: The transition confidence scores of multiple indicators related to kidney function were calculated separately, with different confidence scores having different weights for different indicators in different transition types.

[0076] For the progression from the normal phase to the microalbuminuria phase, the weight of the urinary microalbumin to creatinine ratio is higher than that for estimating the glomerular filtration rate.

[0077] For the progression from microalbuminuria to macroalbuminuria, the weight of the estimated glomerular filtration rate is higher than that of the microalbumin-to-creatinine ratio.

[0078] For regression events, the ratio of urinary microalbumin to creatinine has the highest weight, and it is required that the estimated glomerular filtration rate has not shown an irreversible decline.

[0079] The final transition time is determined by the weighted confidence scores of each indicator exceeding the fusion threshold.

[0080] Furthermore, the step of merging follow-up points that are continuously in the same state range into a cumulative state block is adopted using the boundary probability distribution method. Specifically, the probability distribution of each follow-up point belonging to the normal period, microalbuminuria period, and macroalbuminuria period is calculated. When the probability of adjacent follow-up points belonging to the same state exceeds a preset threshold, they are merged into a probability block.

[0081] The boundary of the state block is defined as the fuzzy region of the probability distribution transformation, where the width of the fuzzy region serves as a weighting factor for the evolutionary features of the matching stage.

[0082] Furthermore, the risk trajectory analysis module is used to perform the following steps: Short-term change features, medium-term trend features, and long-term memory features are extracted from the multimodal feature sequences. The long-term memory features are simulated by exponential decay weighting to simulate the metabolic memory effect.

[0083] Initialize the risk status value and update it bidirectionally point by point based on time sequence. When the indicator deteriorates, the risk value is positively accumulated according to the deterioration rate, and the faster the deterioration, the faster the accumulation. When the indicator improves or there is external intervention, the risk value is negatively decayed at a limited rate, and the decay rate decreases as the current risk value increases. The intervention effect decays exponentially over time.

[0084] The updated risk status value is weighted and fused with short-term change characteristics, medium-term trend characteristics, and long-term memory characteristics to obtain the fused risk trajectory point. The short-term weight is dynamically adjusted according to the consistency between the short-term change direction and the long-term trend.

[0085] The merged risk trajectory points are smoothed to generate a continuous risk trajectory curve.

[0086] Furthermore, the multi-scale screening module is used to perform the following steps: Calculate the first-order and second-order rates of change of the continuous risk trajectory curve at each follow-up point, and statistically analyze the sign change frequency of the second-order rate of change.

[0087] Based on the absolute value of the second-order rate of change and the frequency of sign change, inflection points are classified into three levels: stable acceleration type, fluctuating acceleration type, and oscillating observation type, corresponding to Level 1, Level 2, and Level 3 warnings, respectively. The output includes: the predicted probability of progressing to the massive albuminuria stage, the predicted probability of regressing to the normal stage, the identified inflection point level and inflection point time interval, and a visual graph of the risk trajectory curve.

[0088] Through the foregoing detailed description of the diabetic nephropathy risk screening method optimized by intelligent algorithms, those skilled in the art can clearly understand the diabetic nephropathy risk screening system optimized by intelligent algorithms in this embodiment. As the system disclosed in the embodiment corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to the method section description.

[0089] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for screening the risk of diabetic nephropathy using intelligent algorithms, characterized in that, include: Acquire longitudinal time-series data of target users with diabetes, including structured physiological index sequences and unstructured image data acquired at non-uniform time intervals. The structured physiological index sequence is labeled with a stage evolution pattern. Based on the temporal changes of renal function-related indicators, the progression and regression times between each stage are identified. Microvascular lesion-related features are extracted from the unstructured image data. The extracted image features are aligned with the labeled structured indicators on the time axis to construct a multimodal feature sequence. Based on the multimodal feature sequence, a risk state variable that adjusts bidirectionally over time is constructed. The risk state variable dynamically accumulates or decays risk values ​​according to the changing trend of the indicator characteristics at the current moment relative to the previous moment, generating a continuous risk trajectory curve. Based on the continuous risk trajectory curve, inflection point analysis is performed, and multi-scale screening results are output based on the inflection point time interval determined by the analysis.

2. The method for screening the risk of diabetic nephropathy using intelligent algorithms as described in claim 1, characterized in that, The structured physiological indicator sequence is labeled with a stage evolution pattern, and the progression and regression points between each stage are identified based on the temporal changes of renal function-related indicators, including: The sequence of renal function-related indicators is extended by sliding accumulation within a preset time window. Follow-up points that are continuously in the same state range are merged into a cumulative state block to establish a state evolution sequence. The width of the cumulative window of sliding accumulation extension adaptively increases with the continuous continuous same state until the state changes or the preset maximum accumulation time is reached. According to the preset stage evolution segmentation change characteristics, the state evolution sequence is subjected to feature matching mapping to determine the stage evolution mode; Based on the temporal sequence of the state evolution pattern of the stage evolution model, sequential progress analysis and reverse rollback analysis are performed to identify the progress time and rollback time.

3. The method for screening the risk of diabetic nephropathy using intelligent algorithms as described in claim 2, characterized in that, The pre-defined stage evolutionary segmentation and change characteristics include: The initial offset feature is the first state block in the state evolution sequence that deviates from the normal state, and its duration is less than a first threshold. The critical crossing feature is the boundary point between adjacent state blocks of different risk levels, and the duration of both the preceding and following state blocks is greater than a second threshold. The stable confirmation feature is that the duration of the state block exceeds a third threshold, and no rollback event is detected within at least one follow-up period after the state block ends.

4. The method for screening the risk of diabetic nephropathy using intelligent algorithms as described in claim 2, characterized in that, The adaptive growth rule of the sliding cumulative expansion is as follows: whenever a follow-up point with the same state as the current window appears at the end of the current window, the window end time is automatically extended to the current follow-up point, and the expansion operation is not included in the statistics of the preset maximum cumulative duration. The first follow-up point when the state changes is used as the starting point of the next state block.

5. The method for screening the risk of diabetic nephropathy using intelligent algorithms as described in claim 2, characterized in that, Identifying the progress time and rollback time includes: The transition confidence scores of multiple indicators related to renal function were calculated separately, with different confidence scores having different weights for different indicators in different transition types; For the progression from the normal phase to the microalbuminuria phase, the weight of the urinary microalbumin to creatinine ratio is higher than that of the estimated glomerular filtration rate. For the progression from microalbuminuria to macroalbuminuria, the weight of the estimated glomerular filtration rate is higher than that of the urinary microalbumin to creatinine ratio. For regression events, the ratio of urinary microalbumin to creatinine has the highest weight, and it is required that the estimated glomerular filtration rate has not shown an irreversible decline. The final transition time is determined by the weighted confidence scores of each indicator exceeding the fusion threshold.

6. The method for screening the risk of diabetic nephropathy using intelligent algorithms as described in claim 2, characterized in that, The method of merging follow-up points that are continuously in the same state range into a cumulative state block adopts the boundary probability distribution method. Specifically, the probability distribution of each follow-up point belonging to the normal period, microalbuminuria period, and macroalbuminuria period is calculated. When the probability of adjacent follow-up points belonging to the same state exceeds a preset threshold, they are merged into a probability block. The boundary of the state block is defined as the fuzzy region of the probability distribution transformation, where the width of the fuzzy region serves as a weighting factor for the evolutionary features of the matching stage.

7. The method for screening the risk of diabetic nephropathy using intelligent algorithms as described in claim 1, characterized in that, Based on the multimodal feature sequence, a risk state variable that adjusts bidirectionally over time is constructed. This risk state variable dynamically accumulates or decays risk values ​​according to the changing trend of the indicator characteristics at the current moment relative to the previous moment, generating a continuous risk trajectory curve, including: Short-term change features, medium-term trend features, and long-term memory features are extracted from multimodal feature sequences. The long-term memory features are simulated by exponential decay weighting to mimic the metabolic memory effect. Initialize the risk status value and update it bidirectionally point by point based on time sequence. When the indicator deteriorates, the risk value is positively accumulated according to the deterioration rate, and the faster the deterioration, the faster the accumulation. When the indicator improves or there is external intervention, the risk value is negatively decayed at a limited rate, and the decay rate decreases as the current risk value increases. The intervention effect decays exponentially over time. The updated risk status value is weighted and fused with short-term change characteristics, medium-term trend characteristics and long-term memory characteristics to obtain the fused risk trajectory point. The short-term weight is dynamically adjusted according to the consistency between the short-term change direction and the long-term trend. The merged risk trajectory points are smoothed to generate a continuous risk trajectory curve.

8. The method for screening the risk of diabetic nephropathy using intelligent algorithms as described in claim 1, characterized in that, Based on the continuous risk trajectory curve, inflection point analysis is performed, and multi-scale screening results are output based on the inflection point time interval determined by the analysis, including: Calculate the first-order and second-order rates of change of the continuous risk trajectory curve at each follow-up point, and statistically analyze the sign change frequency of the second-order rate of change. Based on the absolute value of the second-order rate of change and the frequency of sign change, inflection points are classified into three levels: stable acceleration type, fluctuating acceleration type, and oscillating observation type, corresponding to level one, level two, and level three early warning, respectively. The output includes: the predicted probability of progression to the massive albuminuria stage, the predicted probability of regression to the normal stage, the identified inflection point level and inflection point time interval, and a visualization of the risk trajectory curve.

9. A diabetic nephropathy risk screening system optimized using intelligent algorithms, characterized in that, For implementing the method for screening the risk of diabetic nephropathy using intelligent algorithms as described in any one of claims 1-8, the system comprises: The time-series data acquisition module is used to acquire longitudinal time-series data of target users with diabetes. The longitudinal time-series data includes structured physiological index sequences and unstructured image data acquired at non-uniform time intervals. The feature sequence construction module is used to annotate the stage evolution pattern of the structured physiological index sequence, identify the progress and regression times between each stage based on the temporal changes of renal function-related indicators, extract microvascular lesion-related features from the unstructured image data, align the extracted image features with the annotated structured indicators on the time axis, and construct a multimodal feature sequence. The risk trajectory analysis module is used to construct a risk state variable that adjusts bidirectionally over time based on the multimodal feature sequence. The risk state variable dynamically accumulates or decays risk values ​​according to the changing trend of the indicator characteristics at the current moment relative to the previous moment, generating a continuous risk trajectory curve. The multi-scale screening module is used to perform inflection point analysis based on the continuous risk trajectory curve and output multi-scale screening results based on the inflection point time interval determined by the analysis.

Images