Dual-drive prediction method for diabetic complications

Abstract

The invention discloses a dual-drive prediction method for diabetic complications, and particularly relates to the field of disease prediction. Comprising the following steps: acquiring sample data, wherein a sample comprises a plurality of basic hematology indexes and a plurality of derived immune inflammation characteristics; performing classification data prediction on the samples through the prediction model to obtain a prediction result; sorting all complications according to the confidence coefficient, and taking the complications with the confidence coefficient higher than that of the current complications as correction anchor points; comparing the correlation between the current complication and each correction anchor point with a predicted value of the current complication and each correction anchor point, and adjusting the confidence coefficient of the current complication according to a comparison result to obtain a new confidence coefficient; according to the new confidence coefficient, correcting the corresponding complications in the prediction result; and matching the predicted value of each complication in the corrected prediction result in a knowledge base, and correcting the corresponding complication in the prediction result according to a matching result. Based on the method, the prediction accuracy can be improved.

Description

Technical Field

[0001] This application relates to the field of disease prediction, and more particularly to a dual-drive prediction method for diabetic complications. Background Technology

[0002] The core challenge in diabetes management lies in its series of serious chronic complications, such as cerebral infarction, diabetic nephropathy, retinopathy, neuropathy, and diabetic bladder disease. Therefore, early identification of high-risk patients and individualized intervention are crucial for improving patient prognosis, reducing healthcare costs, and enhancing quality of life. However, achieving effective screening and early warning of diabetic complications still faces numerous challenges.

[0003] Diabetic complications are often interconnected and develop synergistically. Although existing research has made progress in predicting diabetic complications, it has overlooked the correlation between complications—that is, when a patient has one complication, there is also a possibility of developing other complications. Moreover, these studies often focus only on the raw data itself (such as using only blood routine indicators such as glycated hemoglobin HbA1c and fasting blood glucose), failing to delve into the rich information contained in the data. This results in insufficient discriminative power and robustness, leading to low accuracy in predictive results. Summary of the Invention

[0004] The main objective of this application is to provide a dual-drive prediction method for diabetic complications, which aims to address the problem of low accuracy in prediction results of existing methods.

[0005] To achieve the above objectives, this application provides a dual-drive prediction method for diabetic complications, comprising: acquiring sample data, the samples including multiple basic hematological indicators and multiple derived immune inflammatory features; classifying and predicting the data of the samples using a prediction model to obtain prediction results, the prediction results including the predicted value and corresponding confidence level of each complication; sorting all complications according to their confidence levels, and using complications with higher confidence levels than the current complication as correction anchors; comparing the correlation between the current complication and each correction anchor with their predicted values, and adjusting the confidence level of the current complication according to the comparison results to obtain a new confidence level; correcting the corresponding complication in the prediction results when any new confidence level reaches a first preset condition; acquiring a knowledge base containing multiple high-confidence association rules, matching the predicted value of each complication in the corrected prediction results in the knowledge base, and correcting the corresponding complication in the prediction results according to the matching results to obtain the final prediction result.

[0006] Optionally, the correlation between the current complication and each modified anchor point is compared with their predicted values, and the confidence level of the current complication is adjusted according to the comparison results to obtain a new confidence level. This includes: pre-determining the correlation between the anchor point and the complication to be modified; if the correlation between the anchor point and the current complication is consistent with their predicted values, the current confidence level is reduced according to the proportion of the penalty factor; otherwise, the current confidence level is increased according to the proportion of the penalty factor to obtain a new confidence level.

[0007] Optionally, the new confidence level can be determined as follows:

[0008] In the formula, At the current confidence level, The confidence level of the anchor point. Let be the Pearson correlation coefficient between the two, and α be the penalty factor.

[0009] Optionally, the high-confidence association rule is a frequent itemset; the predicted value of each complication in the corrected prediction result is matched in the knowledge base, and the corresponding complication in the prediction result is corrected according to the matching result, including: the complications in the corrected prediction result are sorted in ascending order of confidence, each complication is taken as the target complication, the predicted values ​​of complications other than the target complication are taken as the context, and matched in the frequent itemset; according to the matching result, the support of flipping the prediction and the support of maintaining the original prediction are determined; a nonlinear mapping is performed on the new confidence; the initial violation score of the target complication is determined according to the support of flipping the prediction and the support of maintaining the original prediction, the confidence of all complications after nonlinear mapping, and the first preset relationship; when the initial violation score reaches the second preset condition, the target complication is corrected.

[0010] Optionally, based on the matching results, the support for reversing the prediction and the support for maintaining the original prediction are determined, including: when a successful match is determined and the predicted value of the target complication is consistent with the state corresponding to the frequent itemset, the support of the frequent itemset is used as the support for maintaining the original prediction; when a successful match is determined and the predicted result of the target complication is inconsistent with the state in the frequent itemset, the support of the frequent itemset is used as the support for reversing the prediction.

[0011] Optionally, the first preset relationship is:

[0012] In the formula, The confidence level after nonlinear mapping of the target complication. This is the sum of confidence scores after nonlinear mapping of all complications. The sum of all support levels that maintain the original forecast. The sum of the support for all flip predictions. This is the sum of the number of complications other than the target complication.

[0013] Optionally, the high-confidence association rule is a subset of the frequent itemset; after determining the initial violation score of the target complication, the method further includes: removing the target complication, using the remaining complications to form N combinations, where each combination includes N-1 complications; using the predicted value of the complications in each combination as the context of the target complication, matching it in the subset of the frequent itemset, and determining the support of the flipped prediction and the support of the original prediction for each combination based on the matching results; determining the subset violation score of the target complication corresponding to each combination based on the support of the flipped prediction and the support of the original prediction for each combination, the confidence after nonlinear mapping of all complications, and the second preset relationship; aggregating the initial violation score of the target complication and the subset violation scores of all combinations to obtain the violation probability of the target complication; and correcting the target complication when the violation probability reaches the second preset condition.

[0014] Optionally, the second preset relationship is:

[0015] In the formula, This is the sum of confidence scores after nonlinear mapping of all complications other than the target complication and the ignored complication. This is the sum of the number of complications excluding the target complication and neglected complications. The confidence level after nonlinear mapping of the target complication. The sum of all support levels that maintain the original forecast. This is the sum of the support for all flip predictions.

[0016] Alternatively, the probability of violation can be determined as follows:

[0017]

[0018] In the formula, The initial violation score, It is the product of the violation scores of the subsets corresponding to all combinations. These are the weighting coefficients.

[0019] A dual-drive predictive device for diabetes complications, comprising: The data acquisition module acquires sample data, including multiple basic hematological indicators and multiple derived immune inflammatory features. The initial screening module classifies and predicts data from the samples using a prediction model, obtaining prediction results including the predicted value and corresponding confidence level for each complication. The linear correction module sorts all complications based on their confidence levels, using complications with higher confidence levels as correction anchors. It compares the correlation between the current complication and each correction anchor with their predicted values ​​and adjusts the confidence level of the current complication based on the comparison results to obtain a new confidence level. When any new confidence level reaches a first preset condition, the corresponding complication in the prediction results is corrected. The deep correction module acquires a knowledge base, matches the predicted value of each complication in the corrected prediction results against the knowledge base, and corrects the corresponding complication in the prediction results based on the matching results to obtain the final prediction result.

[0020] Compared with the prior art, the beneficial effects of this application are as follows: The dual-drive prediction method for diabetic complications of this invention performs deep learning-based initial screening by classifying and predicting samples using a prediction model, followed by post-processing using association rules. This dual-drive approach leverages the powerful nonlinear fitting capabilities of deep neural networks while incorporating prior medical knowledge through rule-driven post-processing, effectively preventing overfitting and improving the robustness and generalization performance of the prediction model in real clinical scenarios. By mining routine blood data, it combines basic hematological features with limited discriminative power with deep-level pathophysiological indicators, enhancing the pathophysiological significance and discriminative power of the input features, reducing the black-box nature of the model, and effectively improving the model's ability to represent complex metabolic states, thereby increasing prediction accuracy. Furthermore, it uses only a single type of routine blood data as input, making data collection easy. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a dual-drive prediction method for diabetic complications according to this application. Figure 2 This is a comparison chart of the performance indicators of the dual-drive prediction method for diabetic complications and the initial screening model proposed in this application. Figure 3 This is a curve comparison between the dual-drive prediction method for diabetic complications proposed in this application and the primary screening model PR.

[0022] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] The first embodiment of the present invention provides a dual-drive prediction method for diabetic complications, specifically including the following steps: Step S10: Obtain sample data, which includes multiple basic hematological indicators and multiple derived immune inflammatory features; Among them, the basic hematological indicators can be selected as key basic hematological indicators, and the derived immune inflammatory features are determined based on the basic hematological indicators, which can be SII, SIRI, NLR, PLR, PNR, and MLR, calculated as follows:

[0025]

[0026]

[0027]

[0028]

[0029]

[0030] In the formula, Neutrophil is the neutrophil count, Lymphocyte is the lymphocyte count, Platelet is the platelet count, and Monocyte is the monocyte count.

[0031] In this embodiment, the required input is only routine blood test indicators, without the need for expensive or invasive examinations, making it easy to deploy in primary healthcare institutions.

[0032] Step S20: Classify and predict the samples using the prediction model to obtain prediction results, which include the predicted value and corresponding confidence level for each complication. In this embodiment, the prediction model is obtained through pre-training. In the input layer of the prediction model, the multidimensional feature vector (including multiple basic hematological indicators and multiple derived immune inflammatory features) is first received and fed into two fully connected hidden layers for processing. Each hidden layer is followed by batch normalization and Dropout layers to enhance the model's generalization ability and prevent overfitting. During training, Focal Loss is used as the loss function, the core idea of ​​which is to dynamically adjust two key parameters. and This can effectively address the class imbalance problem, as shown in the following formula:

[0033] In the formula, This represents the model's predicted probability for the true class. For sample data The loss value after classification and prediction To focus on the parameters, the degree to which the weights of simple samples are reduced is controlled. This is a balancing parameter used to balance the importance of positive and negative samples.

[0034] In each batch, the loss function is automatically adjusted based on the ratio of positive to negative samples in the current batch. and The value of Focal Loss increases when a severe imbalance in the sample distribution is detected. The numerical value makes the model pay more attention to samples that are difficult to classify; at the same time, adjust This approach assigns higher weights to classes with fewer samples. In this way, an "optimal decision threshold" adapted to the sample distribution is dynamically generated for each prediction task, significantly improving the model's classification performance on imbalanced data.

[0035] The training dataset was constructed as follows: Complete blood count data were acquired, a missing value threshold of 20% was set, and any feature columns with a missing value rate exceeding this threshold were deleted. Next, for the remaining dataset with a small number of missing values, the average of each column was used to fill in the missing values, completing the initial data cleaning and completion. From the processed dataset, based on the counts of neutrophils, lymphocytes, monocytes, and platelets, SII, SIRI, NLR, PLR, PNR, and MLR were calculated, respectively. Several other key basic hematological indicators were selected, and these basic hematological indicators and derived immune inflammatory features were used as a training sample to construct the training dataset.

[0036] To further ensure the accuracy of the prediction results, after training, a search is performed independently on the validation set for each complication label to find the option that maximizes the F1 score. This score is then used as the optimal classification threshold for each complication label, and finally, the predicted value and confidence level (i.e., predicted probability) for each complication are output. In this embodiment, the optimal classification threshold is calculated as follows:

[0037] In the formula, For the first The optimal decision threshold for each label, This is a function used to evaluate the F1 score of the model. For the first The true value of each label This is a decision rule that defines a prediction result whose output probability is greater than the threshold T as a positive class (1), otherwise as a negative class (0).

[0038] Step S30: Sort the prediction results of the samples from high to low confidence, start from the prediction with the lowest confidence and traverse upwards, and use the complication with a confidence higher than the current complication as the correction anchor point. Step S40: Compare the correlation between the current complication and each modified anchor point with their predicted values, and adjust the confidence level of the current complication based on the comparison results to obtain a new confidence level. Specifically, the Pearson correlation coefficient (correlation) between every two complications is pre-calculated. If the anchor point and the current complication are positively correlated, their predicted values ​​differ (e.g., the predicted value of the current complication is 0, while the predicted value of the current anchor point is 1), indicating a conflict between the prediction and the correlation. In this case, the current confidence level is reduced proportionally to the penalty factor. Conversely, if the anchor point and the current complication are positively correlated, but their predicted values ​​differ, the prediction is consistent with the correlation, and the current confidence level is increased proportionally to the penalty factor, resulting in a new confidence level. The new confidence level is determined as follows:

[0039] In the formula, At the current confidence level, The confidence level of the anchor point. The Pearson correlation coefficient between the two is... This is a penalty factor.

[0040] In this embodiment, by comparing the correlation coefficient matrix of the linear correlation strength between two complications with the prediction results, the prediction results are dynamically rewarded or penalized. The paired, linear complication association logic is explicitly incorporated into the model, making the model output more consistent with the common medical knowledge that complications often occur together in clinical practice, and enhancing the consistency of the prediction logic. Step S50: When the new confidence level reaches the first preset condition, perform linear correction on the corresponding complication in the prediction result; Only predictions that are penalized for “conflict” in the assessment, i.e., predictions of complications with a new confidence level lower than the original confidence level, will be considered for reversal.

[0041] For example, the first preset condition is that the new confidence level is lower than a preset decision threshold. The specific correction method is as follows:

[0042] In the formula, To preset the decision threshold, An indicator of whether the model's current predicted label combination conflicts with frequent itemsets. This is a preset decision threshold.

[0043] Step S60: Obtain the knowledge base, match the predicted value of each complication in the corrected prediction result with the knowledge base, and perform in-depth correction on the corresponding complication in the prediction result based on the matching result to obtain the final prediction result.

[0044] It is worth noting that the knowledge base includes multiple frequent itemsets and corresponding subsets, and the knowledge base is trained on the training dataset using the Apriori algorithm. The analysis revealed that the Apriori algorithm sets a minimum support threshold; only itemsets with support exceeding this threshold are retained. These higher-order frequent itemsets will serve as the core basis for subsequent co-occurrence pattern correction. For example, a frequent itemset can be a 5-frequent itemset, with its subset being a 4-frequent itemset.

[0045] Specifically, in step S61, the complications in the corrected prediction results are sorted in ascending order of confidence (i.e., the predictions with the least certainty in the model are processed first, because they are most likely to be wrong and most in need of correction), and each complication is treated as a target complication. The predicted values ​​of the complications other than the target complication are used as context and matched in the frequent itemset. When a match is found to be successful and the predicted value of the target complication is consistent with the state corresponding to the frequent itemset, the support of the frequent itemset is used as the support to maintain the original prediction. When a match is confirmed and the predicted outcome of the target complication is inconsistent with the state of the frequent itemset, the support of the frequent itemset is used as the support for the flip prediction.

[0046] Step S63: Perform a nonlinear mapping on the new confidence level; Step S64: Based on the support of the flipped prediction and the support of the original prediction, the confidence level after nonlinear mapping of all complications, and the first preset relationship, determine the initial violation score of the target complication; the first preset relationship is:

[0047] In the formula, For the target complication to be evaluated, The confidence level after nonlinear mapping of the target complication. This is the sum of confidence scores after nonlinear mapping of all complications. To maintain the sum of the original forecasts, The sum of the support counts for the flip predictions. This is the sum of the number of complications other than the target complication.

[0048] In step S64, if a match fails, meaning the sum of the original predicted support and the sum of the flipped predicted support are both 0, then the initial violation score is 1. Step S65: When the initial violation score is determined to meet the second preset condition, the target complication is corrected.

[0049] Furthermore, to enhance robustness, this embodiment performs subset matching of frequent itemsets after performing frequent itemset matching, i.e., after obtaining the initial violation score, as follows.

[0050] Remove the target complication and divide the remaining complications into N combinations, where each combination includes N-1 complications; Using the predicted value of the complication in each combination as context, match it in a subset of the frequent itemsets in the knowledge base. When a match is found to be successful and the predicted value of the target complication is consistent with the corresponding state in the frequent itemsets, the support of the frequent itemsets is used as the support to maintain the original prediction. When a match is found to be successful and the predicted result of the target complication is inconsistent with the state in the frequent itemsets, the support of the frequent itemsets is used as the support to flip the prediction. Based on the support of the flipped prediction and the support of the original prediction, the confidence after nonlinear mapping of all complications, and the second pre-defined relationship, the subset violation score of the target complication is determined. The initial violation score and subset violation score of the target complication are aggregated to obtain the violation probability of the target complication; When the probability of violation reaches the second preset condition, the prediction results of the target complications in the prediction results are flipped, and the prediction values ​​of all target complications are output as the final prediction results.

[0051] The second preset relationship is:

[0052] In the formula, This is the sum of confidence scores after nonlinear mapping of all complications other than the target complication and the ignored complication. This is the sum of the number of complications excluding the target complication and neglected complications.

[0053] The probability of violation is determined as follows:

[0054] In the formula, The initial violation score, It is the product of the violation scores of the subsets corresponding to all combinations.

[0055] To make "flipping to at least a few classes" relatively easy and "flipping to the majority class" relatively difficult, as a balancing mechanism to prevent the model from being overly biased towards common predictions, this embodiment calibrates the violation probabilities obtained above. The calibration method is as follows:

[0056] In the formula, These are weighting coefficients, dynamically calculated based on the distribution of the predicted target complications in the dataset. This represents the calibrated probability of violation.

[0057] Specifically, the second precondition is that if the calibrated probability of violation is greater than or equal to the random number, the prediction is flipped (i.e., 0 becomes 1, or 1 becomes 0); otherwise, the original prediction remains unchanged. This probabilistic correction method can more gently integrate higher-order rules into the model, avoiding misjudgments caused by overly rigid rules. The random number mentioned above is a random number between 0 and 1 generated based on the Monte Carlo method principle.

[0058] In this embodiment, the combination of complications for each sample is treated as a "transaction." Algorithms such as Apriori are used to mine frequently occurring comorbidity combinations in the training data, generating high-confidence association rules, i.e., frequent itemsets. All samples are iterated over; if the predicted probability of a sample contradicts a high-confidence rule, the predicted probability that does not conform to the rule is deeply revised based on the rule's confidence level. Through this series of adjustments, the model's final output highly follows clinical logic, making the model not only "accurate in prediction" but also "interpretable." A nonlinear transformation and a probability flipping strategy based on Monte Carlo principles are employed to achieve flexible correction of the prediction results, avoiding misjudgments caused by rigid rules.

[0059] The specific process of step S60 will be explained in detail, taking the prediction results including the predicted values ​​and confidence levels of five complications as an example.

[0060] The predicted results for the sample were as follows: diabetic nephropathy: 1 predicted, 55% confidence level; peripheral neuropathy: 1 predicted, 60% confidence level; cerebral infarction: 0 predicted, 70% confidence level; peripheral vascular disease: 1 predicted, 75% confidence level; diabetic neurogenic bladder: 0 predicted, 80% confidence level.

[0061] The treatment order after sorting will be: diabetic nephropathy (55%) → peripheral neuropathy (60%) → cerebral infarction (70%) → peripheral vascular disease (75%) → diabetic neurogenic bladder (80%).

[0062] The pre-built knowledge base includes 5-frequent itemsets and 4-frequent itemsets, namely the two core functions five() and four(). By calling the two core functions five() and four() respectively, the knowledge base is searched for supporting or opposing evidence for the prediction of the current target complication, and quantified as "corrected probability".

[0063] five() function: Strong correlation engine - full match verification The goal of this function is to find the most direct and strongest evidence. It is very strict, requiring that the other four complications besides the target disease must have a completely matching pattern found in the knowledge base.

[0064] Hypothetical scenario: Calculate the corrected probability for retinal lesions in the queue (index 2, prediction 0). The complete prediction for the current sample is {1, 1, 0, 1, 0}.

[0065] Specific execution steps: Initialization: same = 0 (records the sum of the frequencies of itemsets for which "evidence supports the current prediction"); diff = 0 (records the sum of the frequencies of itemsets for which "evidence opposes the current prediction"); Traverse the knowledge base: Read the "Frequent Itemsets.txt" file line by line. Suppose we read a line: 1 1 1 1 0 60 (meaning the pattern {1,1,1,1,0} appears 60 times).

[0066] Context matching: Examining the context of the frequent itemset, namely the states of diseases 0, 1, 3, and 4 ({1, 1, ?, 1, 0}), and comparing the context of the frequent itemset with the predicted context of the current sample {1, 1, ?, 1, 0}, reveals a perfect match. Therefore, the itemset is considered a valid "strong association evidence".

[0067] After context matching, examining the status of the target complication (retinopathy) in the itemset reveals it to be 1, while the predicted value for the target complication is 0, indicating a discrepancy. Therefore, this evidence opposes the current prediction, and the support of the current frequent itemset is used as the support for flipping the prediction, i.e., diff = diff + 60.

[0068] Continue reading. When you read the itemset 1 1 0 1 0 20, whose target complication status (0) is consistent with the predicted value (0), execute: same = same + 20.

[0069] After traversing all itemsets, assume that the final result is same = 20 and diff = 60.

[0070] Historical data support: same / (same + diff) = 20 / 80 = 0.25, which means that among all relevant historical records, only 25% of people believe the prediction is correct (support), indicating that it is very likely to be wrong.

[0071] To further clarify, the function weights the support score based on the model's confidence level, and finally calculates the corrected probability, i.e., the initial violation score, using the formula result = 1.0 - (support score * confidence score weight). The lower the support score, the higher the result (corrected probability).

[0072] The four() function: Weak association engine - subset aggregation verification This function is a fallback and supplement to five(). When a perfect four-condition match cannot be found, it takes a step back to look for weak evidence of multiple three-condition matches and aggregates them.

[0073] Hypothetical scenario: The five() function fails to provide valid information because it cannot find any matching context for the four diseases, resulting in both same and diff being 0.

[0074] Specific execution steps: Create a subset and loop: Split {0,1,3,4} and perform 4 independent searches in sequence: Round 1: Use {1,3,4} as context to perform subset search.

[0075] Round 2: Use {0,3,4} as context to perform subset search.

[0076] Round 3: Use {0,1,4} as context to perform subset search.

[0077] Round 4: Use {0,1,3} as context to perform subset search.

[0078] Repeat the "match-accumulate" process in each subset: For each iteration of the above loop, perform a "context matching" and "evidence accumulation" process similar to the five() function, except that the number of matching conditions is reduced from four to three. Each iteration calculates the local sameness and difference.

[0079] Calculate and aggregate probabilities: Each iteration calculates a "subset correction rate" based on its own same and diff. The four() function aggregates the influence of all weak evidence (opinions from 4 subsets) to form the final aggregated correction probability by accumulating multiplication and then negating the result: totalRate = 1.0 - totalRate * (subset correction rate).

[0080] The implementation code for step S60 is as follows:

[0081]

[0082] A second embodiment of the present invention provides a dual-drive predictive device for diabetic complications, comprising: The data acquisition module is used to acquire sample data, which includes multiple basic hematological indicators and multiple derived immune and inflammatory features. The initial screening module is used to classify and predict data of samples using a prediction model, and obtain prediction results, including the predicted value and corresponding confidence level for each complication. The linear correction module is used to sort all complications according to confidence level, and take the complication with the highest confidence level as the anchor point; compare the correlation between the anchor point and the remaining complications with the confidence level of the remaining complications, and adjust each confidence level according to the comparison results to obtain a new confidence level; when any new confidence level reaches the first preset condition, the corresponding complication in the prediction result is corrected. The deep correction module is used to acquire knowledge base, match the predicted value of each complication in the corrected prediction results with the knowledge base, and correct the corresponding complication in the prediction results based on the matching results to obtain the final prediction result.

[0083] Example Using a validation sample, predictions were made for (cerebral infarction, diabetic nephropathy, diabetic neurogenic bladder, peripheral neuropathy, and peripheral vascular disease). The prediction results from step S20 (prediction by the initial screening model) and step S60 (prediction by the FICN model) were compared on multiple key performance indicators and the PR (Precision-Recall) curve. The results are as follows: Figure 2-3 As shown.

[0084] like Figure 2 As shown in (a), the FICN model significantly outperformed the initial screening model in most tasks in terms of accuracy. For example, in the "cerebral infarction" task, FICN achieved an accuracy of 0.755, while the initial screening model only achieved 0.678, demonstrating that FICN has stronger learning and expressive capabilities when processing temporal or structured information. Similarly, FICN also showed a significant advantage in the "diabetic neurogenic bladder" classification task. Although the difference was smaller in some tasks such as "diabetic nephropathy" and "peripheral neuropathy," the overall trend still indicates that FICN maintains high accuracy while exhibiting more stable output performance.

[0085] Secondly, from the perspective of accuracy, such as Figure 2 (b) FICN demonstrates higher positive class recognition ability across multiple tasks. In particular, in the "diabetic neurogenic bladder" task, FICN's precision rate is 0.750, far higher than the 0.160 of the initial screening model. This indicates that FICN is more cautious and accurate in determining whether two samples belong to the same category, reducing the chance of mistakenly identifying negative classes as positive classes. This reflects its stronger multi-task coordination and boundary discrimination capabilities when dealing with complex labeling systems.

[0086] The Negative Predictive Value (NPV), which is the probability that a predicted negative class result will actually be negative, is particularly crucial in scenarios where the reliability of negative results is paramount, such as medical diagnosis and anomaly detection. For example... Figure 2 (c) FICN achieved higher NPV values ​​in almost all tasks, indicating that FICN is more reliable in finding negative results than the initial screening model.

[0087] Finally, the Diagnostic Odds Ratio (DOR), as a core indicator for comprehensively evaluating model performance, fully reflects the overall superiority of FICN. DOR combines multiple factors such as sensitivity, specificity, precision, and NPV; a higher DOR indicates a stronger discriminative ability of the model. Figure 2 (d) shows that FICN’s DOR is generally higher than that of the initial screening model across all tasks, indicating that it not only performs well on a single task, but also has an overwhelming advantage in overall performance.

[0088] In summary, the FICN model of this invention significantly improves the ability to understand and model complex data patterns by introducing more advanced architectural design or training strategies. It can not only more accurately identify positive samples but also more effectively eliminate negative class interference, thus comprehensively surpassing the initial screening model in terms of accuracy, stability, and reliability. This improvement is not only reflected in the increase in specific numerical values ​​but, more importantly, in the model's stronger adaptability and robustness when facing complex real-world tasks.

[0089] like Figure 3As shown in the figure, each curve corresponds to an independent detection task. The vertical axis represents precision (the proportion of samples predicted as positive that are actually positive), and the horizontal axis represents recall (the proportion of true positive samples successfully identified by the model out of all positive samples). The average precision (AP) value marked below the curve comprehensively reflects the model's overall discriminative ability under different operational thresholds. Overall, FICN's PR curves are significantly better than the initial screening model in all tasks. It not only has a higher initial precision but also maintains relatively stable performance as recall gradually increases, with a more gradual decline. In contrast, the initial screening model often sacrifices precision rapidly in pursuit of higher recall, indicating that it is more prone to introducing a large number of false positives when identifying positive samples. This difference is quantified in the AP metric: FICN's AP values ​​in all tasks are significantly higher than the baseline method, demonstrating its stronger overall advantage in balancing detection capability and prediction reliability.

[0090] More importantly, FICN demonstrates consistent performance improvements across multiple tasks with varying structures and distributions, indicating its strong generalization ability and robustness. This cross-task stability also proves the effectiveness of data preprocessing and feature engineering in FICN, as well as the invention's targeted optimization of context modeling strategies or class imbalance problems.

[0091] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A dual-drive prediction method for diabetic complications, characterized in that, include: Acquire sample data, which includes multiple basic hematological indicators and multiple derived immune and inflammatory features; The data is classified and predicted by a prediction model to obtain prediction results, which include the predicted value and corresponding confidence level for each complication. All complications are sorted according to the confidence level, and complications with a confidence level higher than the current complication are used as correction anchors; The correlation between the current complication and each modified anchor point is compared with the predicted values ​​of both, and the confidence level of the current complication is adjusted based on the comparison results to obtain a new confidence level. When any new confidence level reaches the first preset condition, the corresponding complication in the prediction result is corrected; A knowledge base containing multiple high-confidence association rules is obtained. The predicted value of each complication in the corrected prediction result is matched in the knowledge base. Based on the matching result, the corresponding complication in the prediction result is corrected to obtain the final prediction result.

2. The dual-drive prediction method for diabetic complications according to claim 1, characterized in that, The process of comparing the correlation between the current complication and each modified anchor point with their predicted values, and adjusting the confidence level of the current complication based on the comparison results to obtain a new confidence level, includes: The correlation between the anchor point and the complication to be modified is determined in advance. If the correlation between the anchor point and the current complication is consistent with the predicted values ​​of both, the current confidence level is reduced according to the proportion of the penalty factor. Otherwise, the current confidence level is increased according to the proportion of the penalty factor to obtain a new confidence level.

3. The dual-drive prediction method for diabetic complications according to claim 2, characterized in that, The new confidence level is determined as follows: In the formula, At the current confidence level, The confidence level of the anchor point. Let be the Pearson correlation coefficient between the two, and α be the penalty factor.

4. The dual-drive prediction method for diabetic complications according to claim 1, characterized in that, The high-confidence association rule is a frequent itemset; the predicted value of each complication in the corrected prediction result is matched in the knowledge base, and the corresponding complication in the prediction result is corrected according to the matching result, including: The complications in the revised prediction results are sorted in ascending order of confidence. Each complication is then taken as the target complication. The predicted values ​​of the complications other than the target complications are used as the context for matching in the frequent itemset. Based on the matching results, the support for flipping the prediction and the support for maintaining the original prediction are determined. Apply a nonlinear mapping to the new confidence level; Based on the support of the flipped prediction and the support of the original prediction, the confidence after nonlinear mapping of all complications, and the first preset relationship, the initial violation score of the target complication is determined. When the initial violation score is determined to meet the second preset condition, the target complication is corrected.

5. The dual-drive prediction method for diabetic complications according to claim 4, characterized in that, The step of determining the support for flipping the prediction and the support for maintaining the original prediction based on the matching results includes: When a match is found to be successful and the predicted value of the target complication is consistent with the state corresponding to the frequent itemset, the support of the frequent itemset is used as the support to maintain the original prediction. When a match is confirmed and the predicted outcome of the target complication is inconsistent with the state of the frequent itemset, the support of the frequent itemset is used as the support for the flip prediction.

6. The dual-drive prediction method for diabetic complications according to claim 4, characterized in that, The first preset relationship is: In the formula, The confidence level after nonlinear mapping of the target complication. This is the sum of confidence scores after nonlinear mapping of all complications. The sum of all support levels that maintain the original forecast. The sum of the support for all flip predictions. This is the sum of the number of complications other than the target complication.

7. The dual-drive prediction method for diabetic complications according to claim 4, characterized in that, The high-confidence association rules are subsets of frequent itemsets; After determining the initial violation score for the target complication, the method further includes: Remove the target complication and use the remaining complications to form N combinations, where each combination includes N-1 complications; The predicted value of the complication in each combination is used as the context of the target complication. Matching is performed in a subset of the frequent itemset. Based on the matching results, the support for flipping the prediction and the support for maintaining the original prediction are determined for each combination. Based on the support of the flip prediction and the support of the original prediction corresponding to each combination, the confidence after nonlinear mapping of all complications, and the second preset relationship, the subset violation score of the target complication corresponding to each combination is determined. The initial violation score of the target complication and the subset violation scores corresponding to all combinations are aggregated to obtain the violation probability of the target complication; When the probability of violation reaches the second preset condition, the target complication is corrected.

8. The dual-drive prediction method for diabetic complications according to claim 6, characterized in that, The second preset relationship is: In the formula, This is the sum of confidence scores after nonlinear mapping of all complications other than the target complication and the ignored complication. This is the sum of the number of complications excluding the target complication and neglected complications. The confidence level after nonlinear mapping of the target complication. The sum of all support levels that maintain the original forecast. This is the sum of the support for all flip predictions.

9. The dual-drive prediction method for diabetic complications according to claim 7, characterized in that, The probability of violation is determined as follows: In the formula, The initial violation score, It is the product of the violation scores of the subsets corresponding to all combinations. These are the weighting coefficients.

10. A dual-drive predictive device for diabetic complications, characterized in that, include: The data acquisition module is used to acquire sample data, which includes multiple basic hematological indicators and multiple derived immune inflammatory features; The initial screening module is used to classify and predict data of samples using a prediction model to obtain prediction results, which include the predicted value and corresponding confidence level for each complication. A linear correction module is used to sort all complications according to the confidence level and use complications with a confidence level higher than the current complication as correction anchors; The correlation between the current complication and each corrected anchor point is compared with the predicted values ​​of the two, and the confidence level of the current complication is adjusted according to the comparison results to obtain a new confidence level; when any new confidence level reaches the first preset condition, the corresponding complication in the prediction results is corrected. The deep correction module is used to acquire a knowledge base containing multiple high-confidence association rules. The predicted value of each complication in the corrected prediction result is matched in the knowledge base. Based on the matching result, the corresponding complication in the prediction result is corrected to obtain the final prediction result.

Images