A model for assessing risk of diabetic nephropathy

Abstract

The application relates to the technical field of medical data processing, and discloses a diabetic nephropathy risk assessment model. The model comprises collecting time-series medical data of a diabetic patient, including blood test values, body measurement data and diagnosis and treatment records. For each time point, feature elements are extracted and a temporary risk assessment value is calculated. A historical medical database is connected, similar patient categories are matched, the temporary risk assessment value is corrected by using the risk history of the similar patient categories, and a unified risk assessment value is obtained. The difference between the temporary risk assessment value and the unified risk assessment value is taken as an accuracy index. In combination with the accuracy index and the time-series change of the features, the future nephropathy risk grade of the patient is predicted. The application calibrates individual assessment by using group historical data, generates an internal accuracy index, and improves the accuracy of risk assessment and the reliability of clinical decision-making.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, specifically to a risk assessment model for diabetic nephropathy. Background Technology

[0002] Existing methods for assessing the risk of diabetic nephropathy largely rely on static analysis of a patient's blood test values ​​and basic anthropometric indicators over a single instance or period. These methods are typically based on fixed clinical thresholds or simple risk scoring cards, determining risk levels by comparing a patient's current data with general standards. Their assessment logic is linear, failing to fully utilize the dynamic, time-series data characteristics of individual patients, and failing to calibrate individual data within the context of clinical outcomes in a broader population of similar patients.

[0003] Conventional technical solutions for processing time-series medical data often focus solely on the trends of biomarkers over time. They lack a mechanism to correlate and compare provisional risk conclusions calculated based on the latest individual data with group experience from large-scale historical databases. This makes the assessment results susceptible to short-term individual fluctuations or measurement errors, resulting in insufficient robustness and personalization of the assessment.

[0004] Existing risk assessment systems typically output a single risk score or rating, but rarely provide a built-in, quantitative indicator of the assessment's reliability or accuracy. When interpreting risk reports, clinicians often struggle to intuitively determine the extent to which the assessment results reflect the patient's true condition or the degree of uncertainty. This lack of information regarding the accuracy of the assessment process itself limits the practical value of risk prediction results in developing prospective, differentiated follow-up strategies. Summary of the Invention

[0005] The purpose of this invention is to provide a risk assessment model for diabetic nephropathy to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a risk assessment model for diabetic nephropathy, the model comprising:

[0007] Collect a time-series medical data set of diabetic patients, which includes blood test values, anthropometric data, and medical records.

[0008] For each time point, feature elements are isolated from the medical dataset, and a provisional risk estimate for the feature elements is calculated using predefined rules;

[0009] Connect to the historical medical data storage system, identify matching patient categories, and use the risk history of patient categories to correct the provisional risk estimate, thereby generating a unified risk estimate;

[0010] Calculate the magnitude of the difference between the provisional risk estimate and the uniform risk estimate;

[0011] The magnitude of the difference is recorded as an indicator of the accuracy of the assessment;

[0012] Using accuracy indicators and temporal changes in characteristic elements, the risk level of diabetic nephropathy in patients at future time points can be determined.

[0013] Preferably, the method for generating the unified risk valuation includes:

[0014] Highly frequent feature elements were selected from historical medical data storage systems as the main feature elements.

[0015] For each major characteristic element, analyze its numerical distribution characteristics and temporal correlation, and in combination with the statistical relationship with the kidney disease outcome, assign the influence weight of the major characteristic element;

[0016] For non-primary feature elements, examine their co-occurrence patterns and temporal stability with primary feature elements, and assign influence weights to non-primary feature elements.

[0017] The impact weights of all characteristic elements are aggregated to form a unified risk assessment for each point in time.

[0018] Preferably, the influence weights of the assigned key feature elements include:

[0019] For each key feature element and each historical time period, extract the numerical sequence of the key feature element within the historical time period;

[0020] Calculate the average level, range of variation, and trend slope of the numerical sequence;

[0021] By inputting the average level, range of variation, and trend slope into the weighting mechanism, the initial weights of the main feature elements in the historical time period are obtained.

[0022] The initial weights of the main feature elements for different historical time periods are weighted and fused over time to generate the final influence weight.

[0023] Preferably, the weighting of the influence of non-primary feature elements includes:

[0024] For each non-primary feature and for each historical time period in which it occurs, calculate the correlation strength between the non-primary feature and each primary feature;

[0025] The strongest correlation was selected as the primary correlation value.

[0026] Assess the degree of numerical oscillation of the non-primary feature elements within the historical time period;

[0027] By combining the main correlation values ​​with the degree of numerical oscillation and performing scaling processing, the basic weights of the non-main feature elements in the historical time period are obtained.

[0028] The baseline weights are adjusted based on the clinical weights of non-primary feature elements to obtain the influence weights of non-primary feature elements.

[0029] Preferably, the formation of a uniform risk assessment at each time point includes:

[0030] The numerical range affecting the weight is divided into several continuous intervals, with the number of intervals matching the number of risk levels.

[0031] Each interval corresponds to a risk level, and the interval with the highest impact weight corresponds to the highest risk level.

[0032] For each feature element at each time point, the standard risk level of the feature element at that time point is determined based on the interval into which its influence weight falls.

[0033] The standard risk level, which integrates all characteristic elements, is converted into a unified risk valuation.

[0034] Preferably, the method for recording the accuracy index includes:

[0035] For each feature element, obtain the difference between the temporary risk valuation and the uniform risk valuation of the feature element at multiple time points;

[0036] Take the absolute value of all the differences and sum them. Then standardize the negative value of the sum to obtain the accuracy index of the feature element.

[0037] Preferably, determining the patient's risk level of diabetic nephropathy at future time points includes:

[0038] When the accuracy index is higher than the preset accuracy threshold, the future risk level is calculated based on the temporary risk assessment and historical risk assessment sequence of the feature elements at the current time.

[0039] When the accuracy index is lower than or equal to the preset accuracy threshold, the future risk level is calculated based on the unified risk assessment and temporary risk assessment of the feature elements at the current time.

[0040] Preferably, when the accuracy index is higher than a preset accuracy threshold, calculating the future risk level based on the temporary risk assessment and historical risk assessment sequence of the feature elements at the current time includes:

[0041] The historical risk estimates of the aforementioned feature elements are arranged in chronological order to form a risk estimate sequence. The risk estimate sequence is then fitted with a trend to obtain the slope of the fitted line.

[0042] The direction of risk change is determined by the sign of the slope. The current risk assessment is added to the adjustment amount to obtain the future risk level.

[0043] Preferably, when the accuracy index is lower than or equal to a preset accuracy threshold, calculating the future risk level based on the unified risk assessment and provisional risk assessment of the feature elements at the current time point includes:

[0044] At the current point in time, if the unified risk assessment is greater than the provisional risk assessment, then the unified risk assessment will be used as the future risk level.

[0045] If the uniform risk valuation is less than or equal to the provisional risk valuation, then the provisional risk valuation will be used as the future risk level.

[0046] Preferably, the present invention also includes a diabetic nephropathy risk assessment model, the model including a storage device, a processing device, and a computer program stored in the storage device and running on the processing device.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] By connecting to historical medical data storage systems, matching patient categories are identified, and the risk history of these categories is used to correct temporary risk estimates calculated based on current individual data, resulting in a unified risk assessment. This technical solution leverages collective intelligence to calibrate individual instantaneous risk assessments, effectively mitigating assessment biases caused by short-term fluctuations in individual data or measurement noise. The collective risk history contains long-term outcome information from patients with similar clinical pathways, interpreting individual risk within the context of collective experience. This enhances the contextual and clinical relevance of risk assessments, making the outputs more relevant and practically instructive.

[0049] By calculating the difference between the provisional risk estimate and the unified risk estimate corrected for population data, this difference is recorded as an accuracy indicator for the assessment. This scheme provides an inherent quality control signal for each risk assessment. When individual data and population experience are highly consistent, the difference is small, indicating high confidence in the assessment; when significant deviations occur, the difference is large, suggesting the need to be alert to data anomalies or unusual changes in the patient's condition. This dynamically generated accuracy indicator, together with the temporal changes in characteristic elements, is used to determine future risks, quantifying the uncertainty of prediction and incorporating it into the decision-making process. This ensures that the final risk level judgment is no longer a mechanical calculation but incorporates considerations of the reliability of the assessment process, enhancing the prudence of risk warnings and the depth of decision support. Attached Figure Description

[0050] Figure 1 This is a schematic diagram illustrating the working principle of the diabetic nephropathy risk assessment model described in this invention.

[0051] Figure 2 A flowchart for generating standardized risk valuations;

[0052] Figure 3 A flowchart for assigning weights to key feature elements;

[0053] Figure 4 A comparison chart showing the preset accuracy thresholds for each feature element and their actual accuracy.

[0054] Figure 5 A comparison chart of predicted future risk levels. Detailed Implementation

[0055] 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.

[0056] Please see Figure 1 This invention provides a risk assessment model for diabetic nephropathy. The model includes: collecting a chronologically ordered set of medical data from diabetic patients, comprising blood test values, anthropometric data, and medical records. For each time point, feature elements, such as specific indicators like blood glucose, blood pressure, and creatinine, are extracted from the medical data set, and provisional risk estimates for these feature elements are calculated using predefined rules, which can be simple risk scoring tables based on clinical guidelines. A historical medical data storage system is connected, containing a large amount of anonymized historical data from diabetic patients. By matching the feature element patterns of current patients, patient categories with similar characteristics are identified in the system, and the provisional risk estimates are corrected using the risk historical data of these patient categories, resulting in a more robust and unified risk estimate. The difference between the provisional risk estimate and the unified risk estimate is calculated, reflecting the degree of deviation between the assessment based on current single-point data and the assessment corrected based on group historical data. The calculated difference is recorded as the accuracy index of this assessment, quantifying the reliability of the current assessment result. By combining this accuracy index with the changing patterns of feature elements over time, the risk level of a patient developing diabetic nephropathy at a future point in time can be determined, thereby achieving early warning and dynamic monitoring of the risk.

[0057] Example 1: See Figure 2 In practice, the historical medical data storage system contains a large amount of anonymized longitudinal medical data on diabetic patients, covering blood test values, anthropometric data, and medical records. The selection process is based on the statistical frequency of feature elements in the historical dataset, identifying those with a frequency higher than a preset threshold as primary features. High-frequency features typically include glycated hemoglobin, systolic blood pressure, diastolic blood pressure, the ratio of urinary albumin to creatinine, serum creatinine, and estimated glomerular filtration rate. These indicators have significant clinical significance and prevalence in the development of diabetic nephropathy. In practice, for each indicator identified as a primary feature, its numerical distribution characteristics need to be analyzed in depth. This analysis includes calculating the central tendency indicators of the primary feature across all patient historical data, such as the arithmetic mean and median, as well as dispersion indicators, such as the standard deviation and interquartile range. Temporal correlation analysis focuses on the patterns exhibited by the values ​​of the primary feature over time, such as whether there is a continuous upward or downward trend, or whether there are periodic fluctuations. The statistical relationship analysis with kidney disease outcomes uses statistical methods such as logistic regression or Cox proportional hazards models to quantify the association strength between different numerical levels of key characteristic factors and the eventual occurrence of diabetic nephropathy. Based on the above analysis results, an initial influence weight is assigned to each key characteristic factor, reflecting the predictive importance of that factor in isolation.

[0058] In practical implementation, for non-primary feature elements—those that appear less frequently than a preset threshold in historical datasets, such as certain serum electrolyte levels, transient abnormalities in liver function indicators, or records of specific types of drug use—different weighting strategies are required. The core is to examine the association patterns between non-primary and primary feature elements. Co-occurrence pattern analysis aims to determine whether the values ​​of primary feature elements tend to exhibit a certain regularity when a non-primary feature element shows a specific value; for example, when non-primary feature element A has a high value, primary feature element B also tends to have a high value simultaneously. This association strength can be quantified using methods such as chi-square tests, correlation coefficients, or association rule mining. Temporal stability analysis assesses the fluctuation of the non-primary feature element's value over time; a non-primary feature element with drastically oscillating values ​​is generally less reliable than a non-primary feature element with relatively stable values. Combining the results of co-occurrence pattern and temporal stability analyses, an initial influence weight is assigned to the non-primary feature element, which is usually lower than the weight of the primary feature element.

[0059] In practice, aggregating the influence weights of all feature elements to form a unified risk assessment at each time point is a structured process. The influence weights of all identified feature elements, including primary and non-primary features, are collected. The numerical range of these influence weights is pre-divided into multiple consecutive, non-overlapping intervals. The number of intervals strictly corresponds to the number of risk levels to be presented. For example, if five risk levels are set (low, low-medium, medium, high-medium, high), the weight range is also divided into five intervals. Each interval is explicitly mapped to a risk level; the interval with the highest influence weight corresponds to the highest risk level, and the interval with the lowest influence weight corresponds to the lowest risk level. For each feature element at each specific time point, based on its calculated influence weight value, it is determined which pre-defined interval the weight falls into, thus assigning a standard risk level to that feature element at that time point. In essence, this process discretizes continuous weight values ​​into clinically meaningful grading labels.

[0060] In practice, integrating the standard risk levels of all feature elements to generate a final unified risk estimate involves a decision-making or aggregation step. After obtaining the standard risk level of each feature element at a specific time point, the system needs to combine these multiple discrete risk level information into a comprehensive risk assessment value. One implementation is to perform a weighted average of the standard risk levels of all feature elements, with weights pre-set based on the influence weight of the feature element itself or its clinical importance. Another implementation is to take the highest level among all the standard risk levels of all feature elements as the unified risk estimate at that time point; this is a more conservative and safer strategy designed to avoid underestimating the risk. Optionally, a voting mechanism or a rule-based reasoning system can also be used for integration. The unified risk estimate, as a comprehensive scalar output or category label, represents the overall assessment of the patient's risk of diabetic nephropathy at that time point, after correction based on historical population data.

[0061] In some embodiments, updates to the historical medical data storage system dynamically influence the selection of key features. As new diabetes patient data is continuously incorporated into the historical medical data storage system, the statistical results of feature frequency occurrence will change. Previously infrequent, non-key features may become key features due to the accumulation of sufficient case data. Conversely, the clinical significance of certain features may be reassessed as medical advancements progress. The list of key features is not a static inventory, but a dynamic set that needs to be recalculated and updated periodically or based on triggering conditions. This dynamism ensures that the risk assessment model can adapt to the evolution of medical knowledge and changes in population characteristics.

[0062] In some embodiments, a non-equal-width interval partitioning method can be used when dividing the numerical intervals of influencing weights. Interval partitioning does not require that the numerical span of each interval be equal; it can be based on the actual distribution of influencing weights in historical data, using methods such as quantiles. For example, each interval can contain approximately the same number of historical sample feature points. This partitioning method is more adaptable to uneven weight distribution, resulting in better homogeneity of cases within each risk level. It can be understood that equal-frequency partitioning helps address the risk level imbalance problem caused by skewed data distribution.

[0063] Optionally, for the conversion of standard risk levels, fuzzy logic or probabilistic models can be introduced to handle uncertainty. When the influence weight of a feature element is very close to the boundary between two adjacent intervals, rigidly classifying it to a certain risk level may not be accurate enough. In this case, the membership degree or probability of the feature element belonging to each possible risk level can be calculated, and then these probability distributions can be comprehensively considered when integrating all feature element information, thereby obtaining a smoother and more robust unified risk estimate. This method reduces the risk of abrupt changes in assessment results caused by rigid division of interval boundaries. Unstructured text information in medical records can be transformed into features that can be used for co-occurrence pattern analysis using natural language processing techniques. For example, key phrases such as "poor blood pressure control" or "microalbuminuria" can be extracted from doctors' medical notes, and these text features can be quantified. These quantified text features can be used as non-primary feature elements to participate in the co-occurrence pattern analysis between them and the primary feature elements in structured data, thereby making more comprehensive use of medical information.

[0064] Example 2: See Figure 3In practical implementation, allocating the influence weights of key characteristic elements requires systematic processing for each key characteristic element and each selected historical time period. A key characteristic element might be glycated hemoglobin (HbA1c), and a historical time period might be the last thirty-six months. All values ​​of the key characteristic elements within the historical time period are extracted to form a HbA1c numerical sequence. The average level of the HbA1c numerical sequence is calculated using the arithmetic mean. The range of variation of the HbA1c numerical sequence is calculated using the standard deviation. The trend slope of the HbA1c numerical sequence is calculated, expressed as the slope of the regression line through linear regression analysis. Since the average level, range of variation, and trend slope have different dimensions and scales, they must be standardized to convert them into dimensionless scales for comprehensive calculations. Standardization uses the Z-score method: for each statistical characteristic, the average of all key characteristic elements in the historical data on that statistical characteristic is subtracted, and then divided by the corresponding standard deviation. After standardization, the average level, range of variation, and trend slope become dimensionless values. These dimensionless statistical feature values ​​are input into a pre-defined weighting mechanism, which can be a multiple linear combination model. This model outputs the preliminary weights of the main feature elements over a historical time period. These preliminary weights are dimensionless numerical values. The model form of the weighting mechanism can be expressed as:

[0065]

[0066] in: Key characteristic elements in historical time periods Preliminary weights, Key characteristic elements in historical time periods The average level after standardization of the Inner Canon. Key characteristic elements in historical time periods The range of variation after standardization of the Inner Canon, Key characteristic elements in historical time periods The trend slope after standardization of the Inner Canon, , , These are the model coefficients predetermined in the weight allocation mechanism.

[0067] In practice, assigning influence weights to non-primary characteristic elements involves different calculation processes. A non-primary characteristic element is, for example, blood potassium concentration. For a non-primary characteristic element and each historical time period in which it occurs, the correlation strength between the non-primary characteristic element and each primary characteristic element needs to be calculated. The correlation strength is quantified by calculating the Pearson correlation coefficient, a dimensionless value. From the series of correlation strength values ​​calculated between the non-primary characteristic element and all primary characteristic elements, the largest correlation strength value is selected as the primary correlation value of the non-primary characteristic element in the historical time period. The degree of numerical oscillation of the non-primary characteristic element within the historical time period is assessed. This oscillation is evaluated by calculating the coefficient of variation (COP) of the non-primary characteristic element's numerical sequence, also a dimensionless value. The dimensionless primary correlation value is combined with the dimensionless numerical oscillation, for example, by scaling the primary correlation value by dividing it by a dimensionless adjustment factor positively correlated with the numerical oscillation, to obtain the base weight of the non-primary characteristic element in the historical time period. This base weight is also a dimensionless value. The baseline weights are adjusted based on the clinical weights of non-primary feature elements. The clinical weights are dimensionless correction factors pre-set based on medical literature or expert consensus. The adjusted result is the influence weight of the non-primary feature elements, which is also a dimensionless value.

[0068] Example 3: In specific implementation, the method for recording accuracy indicators begins with the acquisition and processing of data for a single feature element at multiple time points. The feature element is, for example, systolic blood pressure. The system acquires provisional risk estimates for the feature element at a series of consecutive time points. These provisional risk estimates are dimensionless risk scores calculated based on predefined rules. Simultaneously, the system acquires unified risk estimates for the feature element at the exact same series of time points, after correction by a historical medical data storage system. These unified risk estimates are also dimensionless risk scores. The difference between the provisional and unified risk estimates at each time point is calculated using arithmetic subtraction. Since both provisional and unified risk estimates are dimensionless, the difference is also dimensionless. After acquiring the differences between the provisional and unified risk estimates for the feature element at multiple time points, the absolute values ​​of all these differences are taken. Absolute value operations do not change the dimension, and the result remains dimensionless. The absolute values ​​of the differences calculated at all time points are summed to obtain a cumulative difference sum reflecting the overall deviation of the feature element's assessment results across all selected time points. This cumulative difference sum is also a dimensionless value. The purpose of negativeing ​​the cumulative differences is to ensure that the final index is inversely proportional to the degree of difference; negativeing ​​does not change the units of measurement. The negativeed results are then standardized, mapping them to the interval between 0 and 1. The standardized value is the accuracy index of the feature element. The standardization process uses the following formula:

[0069]

[0070] in: An accuracy index representing a dimensionless quantity. Represents the dimensionless cumulative difference sum. This represents a dimensionless constant used to adjust the offset, typically taken as the maximum value of the cumulative difference sum from historical data. This represents a dimensionless constant used to control the scaling range, typically taken as the difference between the maximum and minimum values ​​of the cumulative sum of differences in historical data. By adjusting... and The values ​​ensure the accuracy of the indicators. It falls within the expected interval [0,1]. Since all variables and constants in the formula are dimensionless, the accuracy index... It is also dimensionless. The higher the value of the accuracy index, the better the consistency between the provisional risk estimate and the uniform risk estimate of the feature elements at historical points in time, and the higher the reliability of the assessment.

[0071] Example 4: In specific implementation, determining a patient's future risk level of diabetic nephropathy relies on comparing an accuracy index with a preset accuracy threshold. The preset accuracy threshold is a predefined numerical threshold used to classify the confidence level of the assessment results. Its value is set based on historical backtesting data or clinical expert consensus; for example, a preset accuracy threshold value of 0.75 might be set. The system-calculated feature element accuracy index is compared with this preset accuracy threshold value. When the accuracy index is higher than the preset accuracy threshold, the current feature element-based assessment is considered to have high reliability. The system will employ a prediction strategy, which primarily relies on the provisional risk estimate of the feature element at the current time point and the trend information presented by the historical risk estimate sequence of that feature element to extrapolate the future risk level. When the accuracy index is lower than or equal to the preset accuracy threshold, the current assessment is considered to have significant uncertainty. The system will then employ a more cautious prediction strategy, which comprehensively considers the unified risk estimate and provisional risk estimate of the feature element at the current time point.

[0072] In practice, the preset accuracy threshold can be dynamically adjusted based on the characteristics of different feature elements. Different feature elements may have varying baseline reliability due to differences in measurement stability and their correlation with kidney disease outcomes. Therefore, different preset accuracy thresholds can be set for different types of feature elements. For example, for feature elements like glycated hemoglobin, which are relatively stable in measurement and clearly correlated with kidney disease development, a relatively high preset accuracy threshold, such as 0.80, can be set, requiring higher consistency to be considered reliable. For indicators with larger fluctuations or weaker correlations, a slightly lower preset accuracy threshold, such as 0.70, can be set. This differentiated threshold setting allows the selection of prediction strategies to better adapt to the inherent characteristics of different feature elements. See Table 1, which illustrates a possible differentiated preset accuracy threshold setting.

[0073] Table 1: Preset accuracy thresholds for different feature elements

[0074]

[0075] In practical implementation, the selection logic for the prediction strategy can be implemented through conditional statements. The system process first calculates or obtains the accuracy index of the current feature element, and then compares it with a preset accuracy threshold. The comparison operation can be implemented through a simple conditional branch structure. If the accuracy index is greater than the preset accuracy threshold, the program flow enters the high-confidence prediction branch, calling the future risk level calculation module based on temporary risk valuation and historical risk valuation sequences. If the accuracy index is less than or equal to the preset accuracy threshold, the program flow enters the low-confidence prediction branch, calling the future risk level calculation module based on a comparison of unified risk valuation and temporary risk valuation. This branching logic ensures that the risk assessment method can adaptively adjust the prediction basis according to the assessment quality.

[0076] In some embodiments, the preset accuracy threshold can be designed as a time-adaptive variable. The preset accuracy threshold is not a fixed constant, but rather periodically adjusted based on the overall accuracy distribution of all patient assessment results over a historical period. For example, the system can calculate the median of the accuracy indicators of all feature elements over the past quarter each quarter and set the preset accuracy threshold as a percentage of this median, such as 90%. This adaptive adjustment mechanism allows the threshold value to reflect the overall level of current data quality and model performance, maintaining the timeliness of the assessment criteria. For prediction strategies where the accuracy indicator exceeds the preset accuracy threshold, uncertainty interval estimation can be introduced. In addition to calculating a definite future risk level point estimate, the prediction interval for the future risk level can be calculated based on the volatility of the historical risk estimate sequence. For example, after obtaining the slope of the trend fit, the standard error of the predicted value can be further calculated, thus providing a range into which the future risk level may fall, such as a 95% confidence interval. Providing interval estimation helps clinicians better understand the uncertainty of the prediction results and make more comprehensive decisions.

[0077] Optionally, when the accuracy metric is below or equal to a preset accuracy threshold, the calculation of future risk levels can employ a weighted average instead of a simple binary choice. Instead of directly selecting a uniform risk valuation or a provisional risk valuation, a weighted combination of the uniform and provisional risk valuations is used, with the weights related to the accuracy metric. The lower the accuracy metric, the higher the weight of the uniform risk valuation, because low accuracy implies poor reliability of provisional risk assessments, requiring greater reliance on historical group information for correction. The weighting formula can be expressed as:

[0078]

[0079] in: This represents the calculated level of future risk. Represents a temporary risk assessment. Represents a unified risk valuation. Represents the weighting coefficient, weighting coefficient It is an accuracy indicator functions, such as (when hour, (To set a preset accuracy threshold), but the weighting coefficients need to be limited. The value range is [0,1]. This method provides a smooth transition between provisional and uniform assessments in low-confidence scenarios.

[0080] Optionally, the selection of a prediction strategy can be based on a comprehensive judgment of the accuracy indices of multiple feature elements. When predicting the overall risk of a patient, the assessment results of multiple feature elements may be considered comprehensively. The system can calculate the average or minimum value of the accuracy indices of all relevant feature elements and compare this comprehensive accuracy index with a preset accuracy threshold to determine which overall prediction strategy to adopt. For example, if the minimum value of the accuracy indices of all feature elements is higher than the threshold, a high-confidence prediction strategy is adopted; otherwise, a low-confidence prediction strategy is adopted. This judgment based on the weakest link or the overall level can improve the robustness of the overall prediction.

[0081] See Figure 4 This diagram serves as the core framework for selecting future risk level prediction strategies based on accuracy indicators. It focuses on the preset accuracy thresholds, actual accuracy, and confidence level determination for six characteristic elements, representing a key technological presentation of quantifying reliability assessment to select prediction strategies. This diagram supports an innovative design that incorporates assessment uncertainty into the decision-making process. It allows high-confidence elements to employ a prediction strategy combining temporary risk estimates and historical trends, while low-confidence elements utilize a prudent strategy comparing unified and temporary risk estimates. This upgrades risk assessment from a single output to dynamic decision-making with a reliability dimension, providing a complete logical chain for diabetic nephropathy risk warning—quantifying credibility, selecting strategies, and outputting risk levels. It reflects the technical solution's consideration of assessment robustness and enhances the practicality of risk prediction in clinical scenarios. It is a concrete expression of the innovation of integrating reliability considerations into the assessment process to improve decision-making depth, thereby enhancing the scientific rigor and clinical guidance value of risk assessment.

[0082] Example 5: In specific implementation, when the accuracy index exceeds the preset accuracy threshold, the calculation of future risk levels relies on trend analysis of historical risk valuation sequences. Historical risk valuations of feature elements are arranged chronologically to form a risk valuation sequence, which includes risk scores for feature elements at multiple historical time points. Trend fitting is performed on the risk valuation sequence. Trend fitting can be achieved using least squares linear regression to obtain an optimally fitted straight line. The slope of this fitted line quantifies the average rate and direction of risk change over time. The sign of the slope determines the direction of risk change: a positive slope indicates an upward trend in the risk valuation sequence, with the risk increasing; a negative slope indicates a downward trend in the risk valuation sequence, with the risk decreasing. The temporary risk valuation of the feature element at the current time point is combined with the risk change direction determined by the slope. Specifically, an adjustment is added to the temporary risk valuation at the current time point. This adjustment is determined by the slope of the fitted line and the predicted time span, thereby calculating the predicted risk level for the future target time point. The calculation process can be expressed as:

[0083]

[0084] in: This represents the risk level at a predicted future point in time. This represents a temporary risk assessment of representative characteristic elements at the current point in time. This represents the slope of the fitted line obtained after linearly fitting the historical risk valuation series. This represents the time span from the current point in time to the future target point in time. The slope of the fitted line. This reflects the average rate of change of risk level over time, and is compared with the forecast time span. Multiply by this to obtain the total expected change in risk level over the entire forecast period. Then multiply this change by the current risk level. Adding them together gives us the point estimate of the future risk level.

[0085] In practice, when the accuracy index is lower than or equal to a preset accuracy threshold, the strategy for calculating the future risk level shifts to a direct comparison of two key valuations at the current time point. At the current time point, the system acquires a unified risk valuation and a provisional risk valuation for the characteristic elements. The unified risk valuation is then compared with the provisional risk valuation. If the unified risk valuation is greater than the provisional risk valuation, it indicates that the risk assessment result, corrected based on historical group data, is higher than the assessment result obtained solely based on current data. Out of prudence for risk management, the unified risk valuation is directly used as the predicted future risk level. If the unified risk valuation is less than or equal to the provisional risk valuation, the provisional risk valuation is used as the future risk level. This strategy tends to use the higher of the two valuations when the reliability of the assessment result is low, reflecting the consideration of a "safety margin" in medical risk assessment.

[0086] In some embodiments, trend fitting can employ nonlinear fitting methods instead of linear fitting. When the changing pattern of the historical risk valuation series is clearly not linear, such as exhibiting exponential or logarithmic growth trends, multinomial regression, exponential function fitting, or other nonlinear models can be used for trend fitting. The functional expression obtained from nonlinear fitting is used to calculate future time points. The function value at that point, which is related to the current risk level. The difference or relative change can be used as a quantitative input for the direction of risk change, and then the future risk level can be determined through appropriate transformation rules. Nonlinear fitting can more accurately capture complex risk evolution patterns.

[0087] In some embodiments, the determination of the direction of risk change can be combined with statistical significance testing. The slope is obtained by linearly fitting the historical risk estimation series. Subsequently, instead of directly using the slope value, a significance test is first performed on the slope, such as calculating the p-value. If the p-value of the slope is greater than the preset significance level (e.g., 0.05), the observed trend is considered to be likely caused by random fluctuations and is not statistically significant. In this case, the direction of risk change can be regarded as "stable" or "no significant change," and the adjustment amount... Set to zero, future risk level Equal to the current risk level This method avoids over-reliance on insignificant trends for prediction.

[0088] Optionally, when the accuracy metric is below or equal to a preset accuracy threshold, the calculation of future risk levels can consider recent changes in risk valuation. Even under low confidence conditions, if the provisional or uniform risk valuations at several recent time points show a clear and consistent unidirectional change, such as a continuous rise, a small adjustment reflecting the recent trend can be added to the selected valuation. The adjustment can be determined based on the simple difference average of the valuations at several recent time points. This fine-tuning allows the forecast results to remain prudent while also responding to recent dynamics.

[0089] See Figure 5 This chart is a comprehensive representation of the prediction strategy and future risk level calculation method based on accuracy indicators. It focuses on five time points: now, 3 months later, 6 months later, 9 months later, and 12 months later, presenting the difference in risk level between high-confidence and low-confidence predictions. This chart upgrades risk prediction from a single output to a dual-track output of confidence and risk level dimensions. Clinically, this chart allows for a simultaneous understanding of the severity and reliability of the risk assessment, providing a direct basis for differentiated follow-up strategies. It is a concrete expression of incorporating assessment uncertainty into the decision-making process to enhance the depth of risk warning, strengthening the clinical practicality and decision support of diabetic nephropathy risk assessment.

[0090] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0091] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A diabetic nephropathy risk assessment model, characterized by, The model comprises a storage device, a processing device, and a computer program stored in the storage device and running on the processing device, wherein the processing device, when executing the computer program, implements a method for assessing the risk of diabetic nephropathy, the method comprising: collecting a time-ordered set of medical data of a diabetic patient, the set of medical data comprising blood test values, anthropometric data, and medical record documents; for each time point, separating feature elements from the set of medical data and calculating a provisional risk estimate of the feature elements using predefined rules; connecting to a historical medical data storage system, finding a matching patient category, and correcting the provisional risk estimate using the risk history of the patient category to produce a unified risk estimate; calculating the difference between the provisional risk estimate and the unified risk estimate; recording the difference as an accuracy indicator of the assessment; using the accuracy indicator and the time-ordered changes in the feature elements, determining the risk level of diabetic nephropathy of the patient at a future time point; the determination of the risk level of diabetic nephropathy of the patient at a future time point comprises: when the accuracy indicator is higher than a preset accuracy threshold, calculating the future risk level based on the provisional risk estimate and a sequence of historical risk estimates of the feature elements at the current time point; when the accuracy indicator is lower than or equal to the preset accuracy threshold, calculating the future risk level based on the unified risk estimate and the provisional risk estimate of the feature elements at the current time point; the calculation of the future risk level based on the provisional risk estimate and the sequence of historical risk estimates of the feature elements at the current time point when the accuracy indicator is higher than the preset accuracy threshold comprises: arranging the historical risk estimates of the feature elements in time order to form a sequence of risk estimates, and performing trend fitting on the sequence of risk estimates to obtain the slope of the fitting line; determining the direction of risk change according to the sign of the slope, adding the current risk estimate to an adjustment amount to obtain the future risk level, wherein the adjustment amount is determined by the slope of the fitting line and the predicted time span; the calculation of the future risk level based on the unified risk estimate and the provisional risk estimate of the feature elements at the current time point when the accuracy indicator is lower than or equal to the preset accuracy threshold comprises: at the current time point, if the unified risk estimate is greater than the provisional risk estimate, taking the unified risk estimate as the future risk level; if the unified risk estimate is less than or equal to the provisional risk estimate, taking the provisional risk estimate as the future risk level.

2. The diabetic nephropathy risk assessment model of claim 1, wherein, the method for producing the unified risk estimate comprises: selecting feature elements with high frequency of occurrence from the historical medical data storage system as primary feature elements; for each primary feature element, analyzing its numerical distribution characteristics and time correlation, and combining with the statistical relationship with the outcome of nephropathy to assign an impact weight to the primary feature element; for non-primary feature elements, investigating their co-occurrence patterns and time sequence stability with the primary feature elements, and assigning an impact weight to the non-primary feature elements; summarizing the impact weights of all feature elements to form a unified risk estimate at each time point.

3. The diabetic nephropathy risk assessment model of claim 2, wherein, the assignment of the impact weight of the primary feature element comprises: For each primary feature element and each historical time period, a numerical sequence of the primary feature element in the historical time period is extracted; The average level, variation range and trend slope of the numerical sequence are calculated; The average level, variation range and trend slope are input into a weight allocation mechanism to obtain a preliminary weight of the primary feature element in the historical time period; The preliminary weights of the primary feature element in different historical time periods are time-weighted fused to generate a final impact weight.

4. The diabetic nephropathy risk assessment model of claim 2, wherein, The allocation of the impact weight of the non-primary feature element includes: For each non-primary feature element and each historical time period in which it appears, the association strength of the non-primary feature element with each primary feature element is calculated; The maximum association strength is selected as the primary association value; The numerical oscillation degree of the non-primary feature element in the historical time period is evaluated; The primary association value and the numerical oscillation degree are combined and scaled to obtain a basic weight of the non-primary feature element in the historical time period; The basic weight is adjusted according to the clinical weight of the non-primary feature element to obtain the impact weight of the non-primary feature element.

5. The diabetic nephropathy risk assessment model of claim 2, wherein, The formation of the unified risk estimate at each time point includes: The numerical range of the impact weight is divided into several consecutive intervals, and the number of intervals is consistent with the number of risk levels; Each interval corresponds to a risk level, and the interval with the highest impact weight corresponds to the highest risk level; For each feature element at each time point, the standard risk level of the feature element at the time point is determined according to the interval in which the impact weight falls; The standard risk levels of all feature elements are integrated and converted into a unified risk estimate.

6. The diabetic nephropathy risk assessment model of claim 1, wherein, The recording method of the accuracy index includes: For each feature element, the difference value between the temporary risk estimate and the unified risk estimate of the feature element at multiple time points is obtained; The absolute values of all difference values are summed, and the negative values of the sum are standardized to obtain the accuracy index of the feature element.

Images