A method for constructing an early warning model and risk assessment system for type 1 diabetic ketoacidosis in children.
By constructing a multivariate logistic regression model and an interactive web-based risk assessment tool, and utilizing routine clinical laboratory indicators for children with type 1 diabetes, the problem of early identification of DKA was solved. This approach achieves high efficiency, interpretability, and wide applicability in primary healthcare institutions, thereby enhancing the early warning capability and clinical practicality of DKA.
Patent Information
- Application Number
- CN202611577515.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-13
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies are insufficient for the early identification of type 1 diabetic ketoacidosis (DKA) in primary healthcare institutions and emergency settings. Furthermore, existing models lack applicability, interpretability, and deep integration with clinical workflows, failing to achieve the complementary advantages of quantitative prediction and qualitative judgment.
We constructed an early warning system for type 1 diabetes mellitus ketoacidosis in children based on a multivariate logistic regression model. By utilizing routine clinical laboratory indicators such as immediate blood glucose, glycated hemoglobin, and white blood cell count, and combining them with an interactive web-based risk assessment tool, we achieved an organic integration of quantitative prediction and qualitative judgment.
It improves the early identification capability of DKA, enhances the interpretability and clinical acceptability of the model, achieves deep integration of the model with clinical workflow, ensures wide applicability and ease of operation in different medical institutions, lowers the threshold for use, and improves the safety and practicality of risk assessment.
Smart Images

Figure CN123120328A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for constructing an early warning model and a risk assessment system for type 1 diabetic ketoacidosis in children, belonging to the field of medical artificial intelligence and clinical decision-making technology. Background Technology
[0002] Diabetic ketoacidosis (DKA) is a syndrome of hyperglycemia, ketosis, and metabolic acidosis caused by absolute insulin deficiency. It is the most common and dangerous acute complication of type 1 diabetes mellitus (T1DM) in children. Epidemiological data show that approximately 25% to 40% of children with T1DM present with DKA as their first symptom, and there is a 5% to 10% annual recurrence rate of DKA in already diagnosed children. DKA has a rapid onset and progression; if it is not identified and intervened in a timely manner, it can progress to cerebral edema, hypovolemic shock, multiple organ failure, and even death within a short period of time.
[0003] Currently, the diagnosis of DKA mainly relies on arterial blood gas analysis, blood glucose, and blood / urine ketone tests. However, these diagnostic methods have significant limitations: First, arterial blood gas analysis and blood ketone tests are not routinely available in primary healthcare institutions and emergency room initial consultations, and the limited testing conditions often lead to diagnostic delays; second, these standards are post-hoc diagnostic criteria, applicable only after DKA has occurred, and cannot achieve pre-hoc risk prediction and early warning; third, the early clinical manifestations of DKA (such as polydipsia, polyuria, fatigue, nausea, etc.) lack disease specificity and are easily confused with common childhood diseases such as respiratory infections and acute gastroenteritis, leading to misdiagnosis and missed diagnosis.
[0004] In recent years, artificial intelligence technology has made significant progress in the field of medical risk prediction. However, existing DKA prediction technologies suffer from the following prominent problems: First, most existing models are designed for adults with type 2 diabetes (T2DM), while the pathogenesis, metabolic characteristics, and DKA-inducing factors of children with type 1 diabetes (T1DM) differ fundamentally from those of T2DM, resulting in severely limited extrapolation applicability of the models. Second, existing models mostly employ black-box machine learning algorithms (such as deep neural networks and random forests), lacking interpretability in the model decision-making process. Clinicians cannot understand the medical basis of the prediction results, severely restricting the acceptance and trust of the models in clinical practice. Third, existing research is mostly at the algorithm validation stage, failing to develop practical decision support tools that are deeply integrated with clinical workflows, making it difficult to effectively translate prediction results into clinical actions. Fourth, existing models generally neglect the integration and utilization of qualitative information such as clinical causes and symptoms of DKA, failing to achieve the complementary advantages of quantitative prediction and qualitative judgment. Summary of the Invention
[0005] To address the technical problems existing in the prior art, this invention provides a method for constructing an early warning model and a risk assessment system for type 1 diabetes mellitus (T1DM) in children. This early warning technology solution for DKA is specifically designed for children with T1DM, has good interpretability, is deeply integrated into clinical workflows, and takes into account both quantitative prediction and qualitative judgment.
[0006] To achieve the above objectives, the technical solution adopted in this invention is a method for constructing an early warning model for type 1 diabetic ketoacidosis in children, comprising the following steps: S1, Data Acquisition: Extract routine clinical laboratory indicators from the medical information system for children with type 1 diabetes. These indicators include: complete biochemical test indicators, blood gas analysis indicators, glycated hemoglobin, co-infection status, 25-hydroxyvitamin D, liver function test indicators, blood lipid indicators, complete blood count indicators, kidney function indicators, and urine test indicators. S2, Data Preprocessing: Data preprocessing is performed on routine clinical laboratory indicators. Based on the principle of clinical applicability, laboratory indicators that can be quickly obtained in routine outpatient examinations are selected. Missing values are handled using the grouped median imputation method. After grouping by DKA status, the missing values of the corresponding groups are imputed using the median of each group. Samples with missing indicator rates exceeding the preset threshold are identified and deleted. S3, Candidate Variable Screening: Clinical characteristic analysis was performed on body mass index, co-infection status, glycated hemoglobin, total bilirubin, direct bilirubin, indirect bilirubin, total cholesterol, immediate blood glucose, triglycerides, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, blood urea nitrogen, carbon dioxide combining power, pH value, 25-hydroxyvitamin D, C-peptide, white blood cell count, neutrophil percentage, absolute neutrophil count, lymphocyte percentage, absolute lymphocyte count, monocyte percentage, absolute monocyte count, red blood cell count, hematocrit, hemoglobin, mean corpuscular volume, mean corpuscular hemoglobin, mean corpuscular hemoglobin concentration, red blood cell distribution width coefficient of variation, platelet count, C-reactive protein, urinary creatinine, urinary β2-microglobulin, and urinary microalbumin, and Graphpad was used for analysis. Prism 10.1.2 software used univariate logistic regression analysis to screen candidate variables for the indicators, and initially obtained candidate indicators such as glycated hemoglobin, co-infection status, total bilirubin, indirect bilirubin, total cholesterol, triglycerides, urea nitrogen, carbon dioxide combining power, white blood cell count, and 25-hydroxyvitamin D. The results of the univariate logistic regression analysis were obtained using the following formula: ln[P / (1-P)]=β0+β1X, In the formula, P represents the probability of a child developing diabetic ketoacidosis; β0 represents the intercept term; β1 represents the regression coefficient of the independent variable X; and X represents the clinical indicator to be analyzed. S4, Collinearity Diagnosis: Collinearity diagnosis is performed on the above candidate indicators, the variance inflation factor (VIF) is calculated, and collinear variables with VIF ≥ 5 are eliminated; The collinearity analysis results are obtained using the following formula: VIFj = 1 / (1-Rj²), where VIFj is the variance inflation factor of the j-th independent variable, and Rj² is the coefficient of determination obtained by linear regression with the j-th independent variable as the dependent variable and the remaining independent variables as predictors. When VIFj≥5, the independent variable is considered to have multicollinearity and is removed. S5, Predictor identification: Interaction term analysis was performed on the retained candidate indicators. Stepwise regression and multivariate logistic regression analysis were used to select variables. Finally, six independent predictors were selected, including: immediate blood glucose, glycated hemoglobin, indirect bilirubin, carbon dioxide combining power, white blood cell count and co-infection status. Multivariate logistic regression was used, and the inclusion criterion was P<0.05. The formula for the multivariate logistic regression model is as follows: ln[P / (1-P)]=β0+β1X1+β2X2+β3X3+β4X4+β5X5+β6X6, In the formula, P is the probability of DKA, β0 is the intercept term, β1 to β6 are the regression coefficients corresponding to glycated hemoglobin, co-infection status, indirect bilirubin, carbon dioxide combining power, white blood cell count and immediate blood glucose, respectively, and X1 to X6 are the values of the corresponding independent variables. S6, Nomogram Model Construction: The model construction utilizes multivariate logistic regression analysis, with the occurrence status of DKA as the dependent variable and six routine clinical laboratory indicators as independent variables. The model formula is established and the nomogram prediction model is constructed using R language. In the nomogram prediction model, each independent variable corresponds to a certain number of score points based on its value, and the total score range is 0 to 220 points. S7, Risk Probability Calculation: Based on the total score, the risk probability is calculated using R language and the following formula to calculate the probability of DKA occurring: The linear forecast value LP = -11.1196577360942 + 0.167802877788283 × Total Points; Risk probability Risk = 1 / (1 + exp(-LP)); In the formula, LP represents the linear predicted value, Total Points represents the total risk score of DKA, and Risk represents the probability of DKA occurring. S8, Risk Stratification: Risk stratification is based on maximizing the Youden index and clinical expert consensus, setting four risk stratification thresholds: low risk (probability less than 11.49%), medium risk (probability 11.49% to 44%), high risk (probability 44% to 80%), and very high risk (probability greater than or equal to 80%). S9, Model Validation: The Bootstrap resampling method is used to internally validate the nonograph prediction model to evaluate its discrimination and calibration. S10, Integrated Decision Making: Combining the quantitative risk probability of the nomogram prediction model with the qualitative judgment results of DKA precipitating factors and symptoms, outputting comprehensive risk warnings and clinical decision-making suggestions.
[0007] Preferably, the univariate analysis in step S3 includes: using an independent samples t-test or Mann-Whitney U test for continuous variables, and a chi-square test or Fisher's exact test for categorical variables to compare the inter-group differences of each indicator between the T1DM-DKA group and the T1DM-non-DKA group; further performing univariate logistic regression analysis on the indicators that passed the inter-group difference test, with DKA status as the dependent variable and each clinical indicator as the independent variable, calculating the odds ratio (OR) and its 95% confidence interval for each indicator; and including indicators with a p-value less than a preset threshold α as candidate predictors, where α = 0.1.
[0008] Preferably, the six independent predictors and their odds ratios (ORs) in step S5 are as follows, and the data analysis is based on Graphpad Prism 10.1.2 software: HbA1c: OR=2.162, 95% confidence interval is 1.840 to 2.608; Co-infection: OR=3.988, 95% confidence interval 1.563 to 10.52; Indirect bilirubin (IBIL): OR = 1.150, 95% confidence interval: 1.070 to 1.240; Carbon dioxide binding affinity (CO2CP): OR = 0.7437, 95% confidence interval: 0.6749 to 0.8078; White blood cell count (WBC): OR = 1.356, 95% confidence interval 1.204 to 1.555; Immediate blood glucose (Glu): OR = 1.125, 95% confidence interval: 1.070 to 1.190.
[0009] Preferably, in step S6, the scoring criteria for the six predictor variables of the nomogram prediction model are determined using linear interpolation: glycated hemoglobin score: starting at 4% corresponding to 0 points, increasing by 3.25 points for each additional 1%, and 26% corresponding to 71.5 points; Co-infection status score: 0 points for no, 7.6 points for yes; Indirect bilirubin score: 0 points are given starting from 0 μmol / L, and linear interpolation is used within the range of 0 to 40 μmol / L, with 21.8 points given for 40 μmol / L. Carbon dioxide binding capacity score: 0 points are given starting from 35 mmol / L, and the score is calculated by reverse linear interpolation within the range of 35 to 0 mmol / L, with 44.7 points for 0 mmol / L. White blood cell count score: 0 points are given as the starting point of 0 × 10⁹ per liter, and linear interpolation is used within the range of 0 to 50 × 10⁹ per liter. 50 × 10⁹ per liter corresponds to 36.7 points. Immediate blood glucose score: 0 points are assigned starting from 0 mmol / L, and linear interpolation is used within the range of 0 to 70 mmol / L, with 37.5 points corresponding to 70 mmol / L. The formula for calculating the total score is as follows: Total Points=P_HbA1c+P_Infection+P_IBIL+P_CO2CP+P_WBC+P_Glu, In the formula, P_HbA1c represents the number of score points corresponding to HbA1c, P_Infection represents the number of score points corresponding to co-infection, P_IBIL represents the number of score points corresponding to indirect bilirubin, P_CO2CP represents the number of score points corresponding to carbon dioxide combining power, P_WBC represents the number of score points corresponding to white blood cell count, and P_Glu represents the number of score points corresponding to immediate blood glucose.
[0010] Preferably, the coefficients in the risk probability calculation formula in step S7 are derived from the parameter estimation of multifactor logistic regression, and the complete calculation formulas for the linear predicted value LP and the risk probability Risk are as follows: LP=-11.1196577360942+0.167802877788283×Total Points, Risk = 1 / (1 + exp(-LP)), The risk probability ranges from 0 to 1, corresponding to a DKA occurrence probability of 0% to 100%.
[0011] Preferably, the risk stratification threshold in step S8 is determined after internal verification using the Bootstrap resampling method, which is repeated 1000 times, and the optimal cut point is determined based on the principle of maximizing the Youden exponent. The risk stratification includes four risk levels: low risk (DKA probability less than 11.49%), medium risk (DKA probability 11.49% to 44%), high risk (DKA probability 44% to 80%), and very high risk (DKA probability greater than or equal to 80%). The area under the receiver operating characteristic (AUC) curve for model validation was 0.988, with a 95% confidence interval of 0.984 to 0.993, a sensitivity of 93.4%, and a specificity of 94.9%.
[0012] A risk assessment system for ketoacidosis in children with type 1 diabetes is developed as an interactive web-based assessment tool using HTML, CSS, and JavaScript. It employs a responsive layout design to support desktop and mobile access, and can be deployed on Windows, Linux servers, or cloud platforms. Doctors or parents of the child can access and use it in real-time via a browser. The data acquisition and preprocessing module is used to acquire routine clinical laboratory indicators of the children and perform missing value processing, wherein the missing value processing adopts the grouped median imputation method. The risk probability calculation module is used to calculate the probability of DKA occurrence based on the nomogram model; The trigger and symptom verification module provides an interactive interface for users to select DKA triggers and typical symptoms. The integrated decision-making module combines model calculations of risk probabilities and qualitative verification results to generate the final risk level and clinical recommendations. The visualization output module is used to generate risk probability progress bars, risk level labels, score details, and structured clinical recommendations.
[0013] Preferably, the triggering factors and symptom verification module includes the following DKA triggering factor verification items: new-onset diabetes, i.e., first diagnosis of type 1 diabetes, insulin interruption or missed injection, overeating or improper diet, expired or improperly stored insulin, and other stress factors including surgery, trauma, and emotional excitement. The following symptoms and signs are checked and divided into two levels: severe symptoms and moderate symptoms. Severe symptoms include deep and rapid breathing or Kussmaul breathing or sighing breathing, breath with the smell of rotten apple or acetone, altered consciousness or lethargy or drowsiness or confusion or coma, obvious dehydration or dry mouth or sunken eyes or poor skin turgor or decreased urine output. Moderate symptoms include nausea or vomiting or abdominal pain, increased thirst or urination or worsening weight loss, fatigue or weakness or decreased activity, poor mental state or reluctance to speak or slowed reaction time.
[0014] Preferably, the decision logic of the integrated decision module includes three levels of judgment rules: The first layer is the symptom warning rule: when a child has any serious symptom or sign, regardless of the model's predicted probability, the system will trigger a high-priority warning and recommend immediate investigation of DKA. The second layer is the trigger superposition rule: when the child has clear triggers for DKA such as insulin interruption or missed injection, new-onset diabetes, or co-infection, the system increases the risk level by one level based on the model probability and outputs a corresponding high-priority warning. The third layer consists of model-driven rules: when the model predicts a high risk (≥44%) or an extremely high risk (≥80%), the system outputs the corresponding risk level and emergency clinical treatment recommendations; when the model predicts a low risk (<11.49%) and the child has no obvious symptoms or signs, the system recommends routine monitoring.
[0015] The risk probability calculation module supports the automatic filling function of normal reference values for selected indicators. When HbA1c, IBIL, CO2CP, and WBC indicators are not detected, the system automatically fills them with normal reference values of 5.5%, 6.0 μmol / L, 25.0 mmol / L, and 7.5 × 10⁹ per liter, respectively, and marks the indicator as an automatic filling value to the user.
[0016] Preferably, the output content of the visualization output module includes: The risk probability value is displayed as a percentage indicating the probability of DKA occurring. The risk probability progress bar uses a color gradient bar to visually display the position of the risk probability within the range of 0% to 100%. Risk level labels are displayed as color-coded badges indicating low, medium, high, or very high risk levels. The scoring details are displayed in a grid format, showing the score points for each of the six indicators and the total score. Structured clinical recommendations automatically generate corresponding levels of clinical treatment recommendations based on risk level and symptoms / signs.
[0017] Compared with the prior art, the present invention has the following technical effects: 1. This invention selects six clinically available laboratory indicators as predictive factors, including immediate blood glucose, glycated hemoglobin, white blood cell count, indirect bilirubin, carbon dioxide combining power, and co-infection status. These indicators can be routinely tested in medical institutions at all levels without relying on special equipment or technology, ensuring the model's wide applicability and scalability in different levels of medical institutions and its good applicability at the grassroots level.
[0018] 2. This invention employs a multi-factor logistic regression model combined with a nomogram visualization approach. Compared to black-box machine learning algorithms, the logistic regression model offers superior interpretability and transparency. The nomogram transforms the complex regression equation into an intuitive graphical scoring tool, allowing clinicians to clearly understand the independent contribution of each predictor to DKA risk, thereby improving the model's acceptance and credibility in clinical practice. The odds ratios (ORs) of all six predictors are statistically significant and align with medical theory, further validating the model's scientific validity.
[0019] 3. This invention develops an interactive web-based risk assessment tool based on HTML / CSS / JavaScript, which organically integrates three major functions: model prediction, symptom assessment, and cause verification. It achieves full automation from data input to risk assessment report output, is easy to operate, and responds quickly. It can run directly in web browsers on various terminals such as desktop computers, laptops, tablets, and smartphones without the need to install additional software, which greatly reduces the threshold for use and realizes deep integration of the predictive model with the clinical workflow.
[0020] 4. This invention innovatively designs a symptom coverage mechanism. When the model's predicted probability does not reach the high-risk threshold but the patient has already developed severe DKA-related symptoms (such as Kussmaul breathing, altered consciousness, etc.), the system automatically triggers a symptom alert, prompting clinicians to be vigilant and conduct further examinations. This effectively compensates for clinical emergencies that may be missed by purely quantitative models, achieving a complementary advantage between quantitative prediction and qualitative judgment, and significantly improving the safety and clinical applicability of risk assessment. Attached Figure Description
[0021] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation
[0022] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.
[0023] This invention uses routine clinical laboratory indicators as predictive factors and employs multivariate logistic regression to construct a quantitative prediction model for the risk of DKA. The model results are visualized through nomograms, and an interactive web-based risk assessment tool is developed. This tool integrates a checklist of DKA triggers and a checklist of symptoms and signs, forming a technical solution that organically combines quantitative risk prediction with qualitative clinical judgment.
[0024] A method for constructing an early warning model for type 1 diabetic ketoacidosis in children includes the following steps: Step S1, Data Collection and Screening: Clinical data of patients with type 1 diabetes who visited or were hospitalized in the Department of Endocrinology at Shanxi Provincial Children's Hospital were retrospectively collected. Inclusion criteria were: age ≤18 years, meeting the diagnostic criteria for T1DM established by the World Health Organization (WHO) or the International Society for Childhood and Adolescent Diabetes (ISPAD). Collected clinical data included general demographic information, clinical symptoms and signs, and clinical examination indicators; among which, the clinical examination indicators included immediate blood glucose (Glu, unit mmol / L, i.e., blood glucose value measured at any time point, not limited to fasting state), glycated hemoglobin (HbA1c, unit %), and white blood cell count (WBC, unit ×10⁻¹⁰). 9 The diagnostic criteria for co-infection include: indirect bilirubin (IBIL, μmol / L), carbon dioxide combining power (CO2CP, mmol / L), and co-infection status (binary variable, 0 = none, 1 = present). In this invention, "co-infection status" is defined as: within 72 hours of admission, meeting the clinical diagnosis of an infectious disease (including but not limited to pneumonia, bronchitis, urinary tract infection, skin and soft tissue infection, sepsis, and intracranial infection), and having corresponding anti-infective treatment orders recorded; this condition is considered as having co-infection (coded as 1), otherwise it is considered as not having co-infection (coded as 0). The diagnostic criteria for the above definition refer to the "Guidelines for the Diagnosis and Treatment of Community-Acquired Pneumonia in Children" and the ISPAD Clinical Practice Guidelines.
[0025] Step S2, Data Preprocessing: The collected clinical data are cleaned and preprocessed. Variables with a missing value ratio exceeding a preset threshold are deleted. For continuous variables with a missing value ratio within an acceptable range, the grouped median imputation method is used to handle missing values. That is, the samples are divided into DKA group and non-DKA group according to whether DKA has occurred, the median of the variable in each group is calculated, and the missing value of the corresponding group is imputed using the group median. For discrete variables, the within-group mode imputation method is used to handle missing values.
[0026] Step S3, Predictor Screening: Univariate logistic regression analysis is used to initially screen potential predictors associated with the occurrence of DKA, with a significance level of α=0.05. Variables with P-values less than α in the univariate analysis are included in multivariate logistic regression analysis. Backward stepwise regression (with a rejection criterion of P>0.10) is used to screen variables, and finally, the variables with statistical significance in the model are retained as predictors of the early warning model for DKA.
[0027] Univariate analysis methods included: independent samples t-test or Mann-Whitney U test for continuous variables, and chi-square test or Fisher exact test for categorical variables to compare the differences between the T1DM-DKA group and the T1DM-non-DKA group in each indicator; for the indicators that passed the inter-group difference test, univariate logistic regression analysis was further performed, with DKA status as the dependent variable and each clinical indicator as the independent variable, to calculate the odds ratio (OR) and its 95% confidence interval for each indicator; indicators with a p-value less than the preset threshold α (0.1) were included as candidate predictors.
[0028] Step S4, Construction of the multivariate logistic regression model: Based on the predictors selected in step S3, a multivariate logistic regression model is constructed. The mathematical expression of the model is: ln[P / (1-P)]=β0+β1X1+β2X2+β3X3+β4X4+β5X5+β6X6, In the formula, P is the probability of DKA occurrence, β0 is the intercept term, β1 to β6 are the regression coefficients corresponding to glycated hemoglobin, co-infection status, indirect bilirubin, carbon dioxide combining power, white blood cell count and immediate blood glucose, respectively, and X1 to X6 are the values of the corresponding independent variables; X1 represents glycated hemoglobin (HbA1c), X2 represents co-infection (0=no, 1=yes), X3 represents indirect bilirubin (IBIL), X4 represents carbon dioxide combining power (CO2CP), X5 represents white blood cell count (WBC), and X6 represents immediate blood glucose (Glu).
[0029] The six predictors and their odds ratios (ORs) are as follows: HbA1c: OR = 2.162, 95% confidence interval: 1.840 to 2.608; co-infection: OR = 3.988, 95% confidence interval: 1.563 to 10.52; indirect bilirubin (IBIL): OR = 1.150, 95% confidence interval: 1.070 to 1.240; carbon dioxide combining power (CO2CP): OR = 0.7437, 95% confidence interval: 0.6749 to 0.8078; white blood cell count (WBC): OR = 1.356, 95% confidence interval: 1.204 to 1.555; immediate blood glucose (Glu): OR = 1.125, 95% confidence interval: 1.070 to 1.190.
[0030] Based on clinical data from 732 children with type 1 diabetes mellitus (T1DM) (304 in the DKA group and 428 in the non-DKA group), the maximum likelihood estimation method was used to fit the model and the parameter estimates are shown in Table 1. Table 1. Parameter estimation of the multivariate logistic regression model variable regression coefficient β Standard error OR value 95% CI intercept -8.888 1.642 0.00014 — HbA1c 0.7712 0.08833 2.162 1.840-2.608 Co-infection 1.383 0.4839 3.988 1.563-10.52 Indirect bilirubin 0.1400 0.03757 1.150 1.070-1.240 CO2CP -0.2961 0.04562 0.7437 0.6749-0.808 WBC 0.3047 0.06422 1.356 1.204-1.555 Glu 0.1182 0.02706 1.125 1.070-1.190
[0031] As shown in Table 1, the regression coefficients of the six predictors were all statistically significant in the multivariate model (P<0.05), indicating that, after controlling for other variables, each predictor was an independent risk factor or protective factor for the occurrence of DKA. Among them, the odds ratio for co-infection (OR=3.988, 95%CI: 1.563-10.52) was the largest, indicating that co-infection is the strongest risk factor for DKA, and the risk of DKA in children with co-infection is about 4 times that of children without infection; the odds ratio for HbA1c (OR=2.162, 95%CI: 1.840-2.608) was close to 2.0, indicating that for every 1 percentage point increase in HbA1c, the risk of DKA increases by about 1 time; the regression coefficient for CO2CP was negative (β=-0.2961, OR=0.7437), indicating that CO2CP is a protective factor against DKA, and the lower the CO2CP level, the higher the risk of DKA, which is consistent with the pathophysiological mechanism of DKA leading to metabolic acidosis.
[0032] Step S5, Nomogram Visualization Model Construction: Based on the multifactor logistic regression model constructed in Step S4, the model is visualized using a nomogram method. The nomogram is constructed as follows: the value range of each predictor is mapped to a linear score, the scores of each predictor are summed to obtain the total score, and the total score is mapped to the predicted probability of DKA occurrence through a logistic function.
[0033] For continuous predictors, the formula for calculating their nomogram score Si is: Sᵢ=(Xᵢ-Xᵢ_min) / (Xᵢ_max-Xᵢ_min)×Sᵢ_max, where Xi is the actual value of the predictor; Xi_min and Xi_max are the minimum and maximum values of the predictor in the training data, respectively; and Si_max is the maximum nomogram score corresponding to the predictor.
[0034] For the binary variable co-infection, a value of 0 (no infection) results in a score of 0, while a value of 1 (infection) is assigned a fixed score. The total score is calculated as follows: Total Points = S_HbA1c + S_Infection + S_IBIL + S_CO2CP + S_WBC + S_Glu, where Total Points is the sum of the nomogram scores for the six predictors, with a theoretical range of 0 to 220 points.
[0035] The mapping from total score to predicted probability is achieved using the following linear prediction formula and Logistic function: LP=-11.1196577360942+0.167802877788283×Total Points; Risk = 1 / (1+e^(-LP)), where LP is the linear prediction value; Risk is the predicted probability of DKA occurring; and e is the natural constant.
[0036] The scoring criteria for the six predictor variables in the nomogram prediction model were determined using linear interpolation. Glycated hemoglobin (HbA1c) score: Starting from 4% corresponds to 0 points, and each additional 1% increases the score by 3.25 points, with 26% corresponding to 71.5 points. The score calculation formula is: S_HbA1c=(HbA1c-4)×3.25; Co-infection status score: 0 points for no, 7.6 points for yes; Indirect bilirubin (IBIL) score: 0 points are given starting from 0 μmol / L, and linear interpolation is performed within the range of 0 to 40 μmol / L. 40 μmol / L corresponds to 21.8 points. The score calculation formula is: S_IBIL=IBIL×21.8 / 40; Carbon dioxide binding capacity (CO2CP) score: 0 points are given starting at 35 mmol / L, and the score is calculated by inverse linear interpolation within the range of 35 to 0 mmol / L, with 44.7 points corresponding to 0 mmol / L. The score calculation formula is: S_CO2CP=(35-CO2CP)×44.7 / 35; White blood cell count (WBC) score: 0 × 10 9 / L is the starting point, corresponding to 0 points, from 0 to 50 × 10 9 Linear interpolation within the range of / L, 50×10 9 / L corresponds to 36.7 points, and the score calculation formula is: S_WBC=WBC×36.7 / 50; Immediate blood glucose (Glu) score: 0 points are given starting from 0 mmol / L, and linear interpolation is used within the range of 0 to 70 mmol / L. 70 mmol / L corresponds to 37.5 points. The score calculation formula is: S_Glu=Glu×37.5 / 70.
[0037] The formula for calculating the total score is: Total Points = P_HbA1c + P_Infection + P_IBIL + P_CO2CP + P_WBC + P_Glu, where P_HbA1c represents the number of score points corresponding to HbA1c, P_Infection represents the number of score points corresponding to co-infection, P_IBIL represents the number of score points corresponding to indirect bilirubin, P_CO2CP represents the number of score points corresponding to carbon dioxide combining power, P_WBC represents the number of score points corresponding to white blood cell count, and P_Glu represents the number of score points corresponding to immediate blood glucose.
[0038] The coefficients in the risk probability calculation formula are derived from the parameter estimates of multivariate logistic regression. The complete formulas for calculating the linear predictor LP and the risk probability Risk are as follows: LP=-11.1196577360942+0.167802877788283×Total Points; Risk = 1 / (1+exp(-LP)), where the risk probability ranges from 0 to 1, corresponding to a DKA probability of 0% to 100%.
[0039] The risk assessment system also includes a data integrity management mechanism: the system divides all input indicators into two categories: mandatory indicators and optional indicators. Mandatory indicators are immediate blood glucose (Glu) and co-infection status; these two indicators contribute most critically to the prediction of DKA risk and must be entered by the user. Optional indicators are HbA1c, IBIL, CO2CP, and WBC. If the user does not fill in these values, the system automatically fills them with their respective normal reference values (HbA1c = 5.5%, IBIL = 6.0 μmol / L, CO2CP = 25.0 mmol / L, WBC = 7.5 × 10⁻⁶). 9 / L). The system ensures the transparency and reliability of the autofill function through the following methods: ① All autofilled values are clearly marked with "(Auto)" in the scoring details display area; ② An "Autofill Notification" prompt box is displayed at the top of the results report, listing the names of all autofilled indicators; ③ The prompt text clearly states, "The predicted results are for reference only. It is recommended to complete relevant examinations to improve the accuracy of the assessment." The above mechanisms ensure that clinicians can clearly distinguish between complete data assessment and missing data assessment, avoiding misjudgments caused by autofill.
[0040] Step S6, Model Performance Evaluation and Validation: The model is internally validated using the Bootstrap resampling method, with B = 1000 Bootstrap iterations. Evaluation metrics include: Area Under the Receiver Operating Characteristic (AUC) to assess the model's discriminative power; Hosmer-Lemeshow goodness-of-fit test to assess the model's calibration; Tjur R² to assess the model's interpretability; and Decision Curve Analysis (DCA) to assess the model's clinical applicability.
[0041] The validation results of this invention based on 732 pediatric patients with type 1 diabetes mellitus are shown in Table 2: Table 2 Model Performance Evaluation Indicators Standard numerical values Remark AUC 0.988 95% CI: 0.984-0.993 Sensitivity 93.4% 284 / 304 cases of DKA were correctly classified Specificity 94.9% 406 / 428 cases of non-DKA were correctly classified Positive predictive value 92.8% - Negative predictive value 95.3% - Overall correct classification rate 94.3% 690 / 732 cases Tjur R² 0.7814 - AICc (Select Model) 285.9 - AICc (Intercept-only model) 1008 -
[0042] The model's AUC was 0.988 (95% CI: 0.984–0.993), close to the ideal value of 1.0, indicating that the model has excellent ability to distinguish between DKA patients and non-DKA patients. The Tjur R² was 0.7814, indicating that the model can explain 78.14% of the variation in DKA, demonstrating strong explanatory power. The AICc of the selected model (285.9) was significantly lower than that of the intercept-only model (1008), indicating that the model including six predictive factors had a significantly better fit than the empty model. At the classification cutoff point of 0.5, the model's sensitivity was 93.4%, specificity was 94.9%, positive predictive value was 92.8%, and negative predictive value was 95.3%, demonstrating good overall discriminative performance.
[0043] Step S7, Determining the Risk Stratification Threshold: Based on the Youden Index (J = Sensitivity + Specificity - 1), determine the optimal cutoff value for DKA risk stratification. Calculate the Youden Index at different probability thresholds, select the probability corresponding to the maximum value of the Youden Index as the risk stratification cutoff value, and make appropriate adjustments based on clinical expert opinions and actual application scenarios. Finally, the risk of DKA is divided into the following four levels.
[0044] Low risk: Predicted probability Risk < 11.49%, the risk of DKA is low. Routine monitoring of blood glucose and ketones is recommended, and the current treatment plan should be maintained. Medium risk: 11.49% ≤ Risk < 44%, the risk of DKA is moderate. It is recommended to increase the frequency of blood glucose monitoring, assess whether there are any DKA triggers, and perform blood gas analysis and blood ketone testing if necessary. High risk: 44% ≤ Risk < 80%, the risk of DKA is high. It is recommended to immediately perform arterial blood gas analysis and blood ketone testing, actively find and eliminate DKA triggers, and adjust the insulin treatment plan. Very high risk: Risk ≥ 80%, the risk of DKA is extremely high. It is recommended to immediately initiate the DKA diagnosis and treatment process, complete relevant examinations, and prepare for emergency treatment.
[0045] This invention also provides a risk assessment system for type 1 diabetic ketoacidosis in children. The system is built as an interactive web application based on HTML5 / CSS3 / JavaScript technologies, and can run in web browsers on various terminal devices without requiring additional software installation. The system includes the following functional modules.
[0046] Data acquisition and preprocessing module: Used to receive clinical laboratory indicator data input by the user, including immediate blood glucose (Glu, unit mmol / L), glycated hemoglobin (HbA1c, unit %), and white blood cell count (WBC, unit ×10). 9The module displays the values of indirect bilirubin (IBIL, μmol / L), carbon dioxide combining power (CO2CP, mmol / L), and the status of co-infection (yes / no). It provides a numerical input interface, allowing users to set reasonable value range limits and input validity checks for each indicator, and displays the normal reference range for each indicator to assist users in judging the reasonableness of the input data.
[0047] The risk probability calculation module is used to calculate the predicted probability of DKA occurrence based on the clinical laboratory indicator data received by the data input module, by calling pre-stored nomogram score mapping rules and Logistic regression model parameters. The module's calculation process is as follows: First, the input clinical laboratory indicator values are converted into corresponding nomogram scores by calling the score calculation functions of each predictor; then, the six scores are added together to obtain the total score (Total Points); finally, the linear predictive value (LP) and the predicted probability (Risk) are calculated using formulas. The score calculation functions for each predictor in the model calculation module of the risk assessment system are implemented as follows.
[0048] The glycated hemoglobin (HbA1c) scoring function `getHbA1cPoints(v)` returns 0 if v ≤ 4, 71.5 if v ≥ 26, and (v-4) × 3.25 otherwise. The co-infection scoring function `getInfectionPoints(v)` returns 7.6 if v = 1, and 0 otherwise. The indirect bilirubin (IBIL) scoring function `getIBILPoints(v)` returns IBIL (μmol / L) values if v ≤ 0, 21.8 if v ≥ 40, and v × 21.8 / 40 otherwise. The carbon dioxide combining power scoring function `g`... etCO2CPPoints(v): Input parameter v is the CO2CP value (mmol / L). If v≤0, it returns 44.7; if v≥35, it returns 0; otherwise, it returns (35-v)×44.7 / 35. White blood cell count score function getWBCPoints(v): Input parameter v is the WBC value (×10^9 / L). If v≤0, it returns 0; if v≥50, it returns 36.7; otherwise, it returns v×36.7 / 50. Instant blood glucose score function getGluPoints(v): Input parameter v is the Glu value (mmol / L). If v≤0, it returns 0; if v≥70, it returns 37.5; otherwise, it returns v×37.5 / 70.
[0049] The risk probability calculation module supports automatic filling of normal reference values for optional indicators. When HbA1c, IBIL, CO2CP, and WBC are not detected, the system calculates them at concentrations of 5.5%, 6.0 μmol / L, 25.0 mmol / L, and 7.5 × 10⁻⁶, respectively. 9The system automatically populates the normal reference value for HbA1c, and indicates to the user that this is an automatically populated value. Furthermore, the system is developed as an interactive web-based assessment tool using HTML, CSS, and JavaScript. It is important to emphasize that the auto-fill function is applicable when an indicator is temporarily missing due to clinical limitations (such as the inability to obtain HbA1c results promptly in emergency situations). In this case, the system uses the normal reference value as an approximate estimate, and the accuracy of the output risk probability is lower than in scenarios with complete data input. When outputting results, the system will prominently display all automatically populated items in the scoring details area and display a mandatory prompt at the top of the results report: "The following indicators were not tested and have been automatically populated with normal reference values. The predicted results are for reference only; it is recommended to complete relevant examinations as soon as possible." The clinical value of this function lies in providing clinicians with preliminary risk references when data is incomplete, rather than replacing complete testing. Users are obligated to complete the tests when conditions permit to obtain more accurate assessment results.
[0050] The Triggering Factors and Symptoms Verification Module is used to collect and assess information on patients' DKA-related symptoms and signs, as well as information on potential triggering factors that may lead to DKA. The module provides a list of symptoms and signs, including: polydipsia, polyuria, weight loss, nausea and vomiting, abdominal pain, rapid and deep breathing (Kussmaul breathing), rotten apple odor on exhaled breath (acetone odor), fatigue, lethargy, altered consciousness (drowsiness / coma), and signs of severe dehydration (poor skin turgor, sunken eyes, dry lips). Users select the patient's current symptoms using checkboxes, and the system categorizes the symptoms into two types for comprehensive assessment: severe warning signs (Kussmaul breathing, rotten apple odor on exhaled breath, altered consciousness, signs of severe dehydration) and general symptoms.
[0051] The module also provides a checklist of DKA triggers, including: newly diagnosed diabetes (not yet diagnosed or first-time discovery), overeating or irregular eating habits, insulin interruption or missed injections, expired or improperly stored insulin, other stress factors (such as infection, trauma, surgery, emotional excitement, etc.), and no obvious trigger. Users select the patient's current triggers using checkboxes, and the system performs a comprehensive assessment based on the trigger type.
[0052] The triggering factors and symptom verification module includes the following DKA triggering factor verification items: newly diagnosed type 1 diabetes, insulin interruption or missed injections, overeating or improper diet, expired or improperly stored insulin, other stress factors including surgery, trauma, and emotional excitement; and the following symptom and sign verification items, divided into two levels: severe symptoms and moderate symptoms. Severe symptoms include deep and rapid breathing or Kussmaul breathing or sighing breathing, breath with a rotten apple or acetone odor, altered consciousness or lethargy or drowsiness or confusion or coma, significant dehydration or dry mouth or sunken eyes or poor skin turgor or decreased urine output; moderate symptoms include nausea or vomiting or abdominal pain, polydipsia or polyuria or worsening weight loss, fatigue or weakness or decreased activity, poor mental state or reluctance to speak or slowed reaction.
[0053] The integrated decision-making module is used to integrate the quantitative prediction probability obtained from the integrated model calculation module, the qualitative symptom assessment results obtained from the symptom assessment module, and the qualitative causal assessment results obtained from the causal verification module, and outputs the final DKA risk assessment report.
[0054] The integrated decision-making module works as follows: First, it determines the basic risk level (low, medium, high, or very high risk) based on the predicted probability Risk. Then, it checks the output of the symptom assessment module. If there are severe warning symptoms and the basic risk level is low or medium (i.e., Risk < 44%), the symptom coverage mechanism is triggered, and symptom warning information is added to the risk assessment report, prompting clinicians that "the patient has developed severe DKA-related symptoms, and it is recommended to immediately check blood gas analysis and blood ketones," and that the patient be managed as a high-risk patient. At the same time, it checks the output of the trigger verification module. If there are serious triggers such as new-onset diabetes or insulin interruption, they are highlighted in the report. Finally, it outputs a complete risk assessment report containing the basic risk level, symptom warning information, trigger information, and corresponding clinical recommendations.
[0055] The decision-making logic of the integrated decision-making module includes three layers of judgment rules. The first layer is the symptom warning rule: when a child has any serious symptom or sign, regardless of the model's predicted probability, the system triggers a high-priority warning and recommends immediate investigation for DKA. The second layer is the triggering factor superposition rule: when a child has a clear triggering factor for DKA, such as insulin interruption or missed injection, new-onset diabetes, or co-infection, the system increases the risk level by one level based on the model probability and outputs a corresponding high-priority warning. The third layer is the model-driven rule: when the model predicts a high risk (greater than or equal to 44%) or an extremely high risk (greater than or equal to 80%), the system outputs the corresponding risk level and emergency clinical treatment suggestions; when the model predicts a low risk (less than 11.49%) and the child has no obvious symptoms or signs, the system recommends routine monitoring.
[0056] The visualization output module generates a risk probability progress bar, risk level labels, score details, and structured clinical recommendations based on the DKA risk assessment report. The risk probability numerical display shows the probability of DKA occurrence as a percentage; the risk probability progress bar uses a color gradient to visually indicate the position of the risk probability within the range of 0% to 100%; the risk level labels use color-coded badges to indicate low, medium, high, or very high risk levels; the score details display uses a grid to show the score points for each of the six indicators and the total score; and the structured clinical recommendations automatically generate corresponding clinical treatment suggestions based on the risk level and symptoms / signs.
[0057] The entire system adopts a responsive layout design, supporting access on both desktop and mobile devices. It can be deployed on Windows, Linux servers, or cloud platforms, allowing doctors or parents of sick children to access and use it in real time via a browser. The system employs a pure front-end architecture with a front-end and back-end separation. All computational logic is executed locally in the user's browser's JavaScript engine, eliminating the need to transmit patient data to the server, thus fully protecting patient privacy and data security. Furthermore, the nomogram prediction model has been internally validated.
[0058] The receiver operating characteristic (ROC) curve area under the curve (AUC) was 0.988, with a sensitivity of 93.4%, a specificity of 94.9%, a positive predictive value of 92.8%, a negative predictive value of 95.3%, an accuracy of 94.3%, a Tjur R² of 0.7814, a Hosmer-Lemeshow test statistic of 4.132, and a p-value of 0.8405.
[0059] The present invention will be further verified through practical examples below.
[0060] Example 1: Construction and Validation of the DKA Early Warning Model This embodiment details the complete construction and validation process of an early warning model for ketoacidosis in children with type 1 diabetes, including data acquisition, data preprocessing, predictor selection, construction of a multivariate logistic regression model, visualization of nomograms, model performance evaluation, and determination of risk stratification thresholds.
[0061] Step 1: Data collection. Clinical data of patients with type 1 diabetes who visited the Department of Endocrinology of a children's hospital in a certain province from January 2018 to December 2024 were retrospectively collected. The inclusion criteria were: (1) age ≤ 18 years; (2) meeting the diagnostic criteria for T1DM established by WHO or ISPAD, namely, having symptoms of hyperglycemia (polydipsia, polyuria, polyphagia, weight loss) and random blood glucose ≥ 11.1 mmol / L or fasting blood glucose ≥ 7.0 mmol / L, low or negative C-peptide level, and requiring insulin therapy to maintain life; (3) complete clinical data, including data on the six predictive factors required by this invention. The exclusion criteria were: (1) type 2 diabetes, gestational diabetes or other special types of diabetes; (2) complicated with severe liver and kidney dysfunction, malignant tumors or other serious systemic diseases; (3) clinical data were severely missing, and key predictive factor data could not be obtained.
[0062] The final sample size included in the analysis was 732 cases, of which 304 cases (41.5%) were in the DKA group and 428 cases (58.5%) were in the non-DKA group. The diagnostic criteria for DKA were based on the ISPAD guidelines: blood glucose >11.1 mmol / L, and venous blood pH <7.30 or blood bicarbonate <15 mmol / L, with the presence of ketoacidosis (serum β-hydroxybutyrate ≥3 mmol / L) or ketonuria (urine ketone bodies ≥++).
[0063] The six predictive factors collected include: (1) Immediate blood glucose (Glu), unit mmol / L, normal reference value 3.9-6.1mmol / L; (2) Glycated hemoglobin (HbA1c), unit %, normal reference value 4%-6%, reflecting the average blood glucose level over the past 2-3 months; (3) White blood cell count (WBC), unit ×10^9 / L, normal reference value 4-10×10^9 / L; (4) Indirect bilirubin (IBIL), unit μmol / L, normal reference value 0-15μmol / L; (5) Carbon dioxide combining power (CO2CP), unit mmol / L, normal reference value 22-29 mmol / L, reflecting the acid-base balance in the body; (6) Co-infection status, which is a binary variable, 0 indicates no co-infection and 1 indicates co-infection.
[0064] Step Two: Data Preprocessing. The collected raw data underwent data cleaning and preprocessing. First, variables with a missing rate exceeding 30% were removed; in this study's dataset, the missing rate for 25-hydroxyvitamin D (25OHD) exceeded 30% and was therefore removed. For the remaining variables with missing values, the median imputation method was used: the 732 samples were divided into a DKA group (304 cases) and a non-DKA group (428 cases) based on whether DKA occurred. The median for each continuous variable within each group was calculated, and the median from the DKA group was used to imput missing values within the DKA group, and the median from the non-DKA group was used to imput missing values within the non-DKA group. For the binary variable of co-infection, the mode imputation method was used to handle missing values. After the above preprocessing, the data for all six predictors in the 732 samples were complete and without missing values.
[0065] Step 3: Predictor screening. First, a univariate logistic regression analysis was performed, using the occurrence of DKA (0 = not occurred, 1 = occurred) as the dependent variable. Six independent univariate logistic regressions were conducted with immediate blood glucose, glycated hemoglobin, white blood cell count, indirect bilirubin, carbon dioxide combining power, and co-infection as independent variables, with a significance level of α = 0.05. The univariate analysis results showed that all six variables were significantly associated with the risk of DKA (P < 0.05).
[0066] All six variables were included in a multivariate logistic regression analysis. Backward stepwise regression was used for variable selection, with a p-value > 0.10 as the exclusion criterion. The multivariate analysis results showed that, controlling for other variables, the regression coefficients of all six predictors were statistically significant (p < 0.05), indicating that all six variables were independent predictors of DKA.
[0067] The parameter estimation results showed that the odds ratio for co-infection was the highest (OR=3.988, 95%CI: 1.563-10.52), indicating that co-infection is the strongest risk factor for DKA. Children with co-infection have approximately four times the risk of developing DKA compared to children without infection. This is consistent with the pathophysiological mechanism of DKA, which is caused by increased stress hormones and insulin resistance due to infection. The odds ratio for HbA1c was 2.162 (95%CI: 1.840-2.608), suggesting that for every 1 percentage point increase in HbA1c, the risk of DKA approximately doubles, reflecting the close relationship between long-term poor glycemic control and DKA. The regression coefficient for CO2CP was negative (β=-0.2961, OR=0.7437), making it the only protective factor in the model. Its odds ratio was less than 1, indicating that lower CO2CP levels are associated with a higher risk of DKA. This is entirely consistent with the pathophysiological changes caused by DKA leading to metabolic acidosis and decreased CO2CP, thus validating the medical rationale of the model from the opposite perspective. The odds ratios of IBIL (OR=1.150), WBC (OR=1.356), and Glu (OR=1.125) were all greater than 1 and the 95% confidence interval did not include 1.0, suggesting that all three are independent risk factors for DKA.
[0068] Step 4: Nonograph Visualization Model Construction. Based on the parameter estimation results of the multi-factor Logistic Regression model, the nonograph method is used to visualize the model. The nonograph construction process is as follows: For continuous predictors, their value range is linearly mapped to a score interval. Taking HbA1c as an example, the value range is 4% to 26%, corresponding to a nonograph score of 0 to 71.5 points, calculated as S_HbA1c=(HbA1c-4)×3.25. For the binary variable of co-infection, no infection scores 0 points, and infection scores 7.6 points. The scores of each predictor are summed to obtain the total score (Total Points), which theoretically ranges from 0 to 220 points. The mapping formula from total score to predicted probability is LP=-11.1196577360942+0.167802877788283×Total Points, Risk=1 / (1+e^(-LP)). Taking a Total Points score of 78.4 as an example, LP = -11.1197 + 0.1678 × 78.4 = 2.0346, and Risk = 1 / (1 + e^(-2.0346)) = 88.5%. Using a nomogram, users only need to connect the values of each predictor on the corresponding scale to read the score, sum the scores, and then quickly obtain the probability of DKA occurrence by referring to the corresponding probability scale, without needing to perform complex logarithmic calculations.
[0069] To clarify the fate of candidate variables not included in the final model in the univariate analysis, the analysis results of some representative variables that were removed are explained below: Total bilirubin (TBIL) was correlated with DKA in the univariate analysis, but failed to meet the inclusion criteria in the multivariate stepwise regression (P<0.05); and TBIL and IBIL were highly correlated (r>0.9), showing strong collinearity. Furthermore, IBIL is more directly associated with DKA pathophysiologically. To avoid information redundancy, IBIL was retained while TBIL was removed. Triglycerides (TG) were one of the candidate variables in the univariate analysis of this study. Data analysis showed that the median triglyceride level in the DKA group was 2.54 mmol / L (interquartile range 1.50–4.41 mmol / L), while the median in the non-DKA group was 0.74 mmol / L (interquartile range 0.61–1.02 mmol / L). The DKA group had significantly higher levels than the non-DKA group, which is consistent with the pathophysiological mechanism of increased lipolysis due to insulin deficiency. However, in the multivariate stepwise regression, triglycerides were excluded because they did not meet the inclusion criteria (P<0.05), indicating that their predictive information could be covered by other included variables (such as HbA1c, immediate blood glucose, etc.). Blood urea nitrogen (BUN) was excluded because it did not meet the inclusion criteria (P<0.05) in the multivariate analysis. 25-hydroxyvitamin D (25(OH)D) was excluded because it did not meet the inclusion criteria in the multivariate analysis and was removed by the stepwise regression algorithm. Although the above-mentioned excluded variables were correlated with DKA in the univariate analysis, they did not show independent predictive value in the multivariate analysis controlling for other variables, and therefore were not included in the final model. This screening process was based on pre-set objective statistical criteria (inclusion P<0.05, exclusion P>0.10) and a collinearity diagnostic threshold (VIF≥5). All decisions were data-driven, and there was no selective reporting behavior.
[0070] Step 5: Model performance evaluation. The model was internally validated using the Bootstrap resampling method (B=1000 times). The 1000 Bootstrap validation results showed that the corrected AUC of the model was 0.988 (95% CI: 0.984-0.993), close to the ideal value of 1.0, indicating that the model has excellent discriminative ability. At the classification cutoff point of 0.5, the model's sensitivity was 93.4% (correctly classifying 284 / 304 DKA cases), specificity was 94.9% (correctly classifying 406 / 428 non-DKA cases), positive predictive value was 92.8%, negative predictive value was 95.3%, and the overall correct classification rate was 94.3% (690 / 732 cases).
[0071] The Tjur R² was 0.7814, indicating that the model could explain 78.14% of the variation in DKA, demonstrating strong explanatory power. The AICc of the selected model was 285.9, significantly lower than the AICc (1008) of the intercept-only model, with a difference of 722.1, indicating that the model including six predictors had a significantly better fit than the empty model. The Hosmer-Lemeshow goodness-of-fit test statistic was 4.132 (P=0.8405), suggesting a good model fit. Decision curve analysis showed that, over a wide range of threshold probabilities, the net benefit of using this model for DKA risk screening was higher than the two extreme strategies of "all screening" and "no screening," indicating that the model has good clinical practical value.
[0072] Step Six: Determining Risk Stratification Thresholds. The optimal cutoff values for risk stratification were determined based on the Youden index (J = Sensitivity + Specificity - 1). Calculations showed the following cutoff values: Level 1 (distinguishing between low and medium-high risk) was 11.49%, Level 2 (distinguishing between medium and high risk) was 44%, and Level 3 (distinguishing between high and very high risk) was 80%. Combining clinical expert opinions, a four-tiered risk stratification standard was finalized: Low risk (Risk < 11.49%): routine monitoring recommended; Medium risk (11.49% ≤ Risk < 44%): enhanced monitoring and assessment of triggering factors recommended; High risk (44% ≤ Risk < 80%): immediate blood gas analysis and blood ketone testing recommended; Very high risk (Risk ≥ 80%): immediate initiation of the DKA treatment process recommended.
[0073] The working principle of the above technical solution is as follows: Multivariate Logistic Regression is a generalized linear regression model suitable for risk prediction of binary dependent variables. The model establishes a linear prediction equation through logarithmic odds (Logit transformation), mapping the linear combination of each predictor factor to a probability value between 0 and 1. This invention is based on clinical data from 732 pediatric patients with type 1 diabetes mellitus (T1DM), using maximum likelihood estimation to fit the model parameters and selecting six routine clinical laboratory indicators with independent predictive value. A nomogram is a graphical representation of the Logistic Regression model, mapping the value of each predictor factor to an intuitive score scale, allowing clinicians to quickly obtain risk probabilities without understanding complex logarithmic calculations, significantly lowering the barrier to model use. Bootstrap resampling validation evaluates the model's internal generalization ability through random sampling with replacement, correcting the optimistic bias in model performance estimation on a single dataset and ensuring the robustness of the model evaluation results.
[0074] The early warning model for DKA constructed in this invention achieves an AUC of 0.988, a sensitivity of 93.4%, and a specificity of 94.9%, demonstrating excellent overall performance. Furthermore, all six predictive factors are routine clinical laboratory indicators, indicating good applicability at the primary care level. Nomogram visualization transforms the complex regression model into an intuitive clinical scoring tool, significantly improving the model's interpretability and clinical usability. The four-level risk stratification standard provides differentiated clinical management recommendations for children at different risk levels, contributing to precise prevention and early intervention of DKA.
[0075] Example 2: Technical Implementation of a Risk Assessment System This embodiment details the technical implementation scheme of the interactive webpage risk assessment system built based on the above-mentioned DKA early warning model.
[0076] (1) Data Acquisition and Preprocessing Module: A responsive HTML5 layout design is adopted, providing data input controls for six predictors in the top area of the page. For the five continuous variables (Glu, HbA1c, WBC, IBIL, CO2CP), a numeric input box with type=number is provided, with min and max attributes set to limit the value range, and an oninput event listener is added for real-time validity verification; for the binary variable of co-infection, a radio button group (Yes / No) is provided. The right side of the input interface displays the normal reference range of each indicator to help clinicians judge the rationality of the input data. The interface uses a gradient progress bar to display the deviation of the current input value from the normal range in real time, providing intuitive visual feedback.
[0077] (2) Risk Probability Calculation Engine: Implemented using native JavaScript, the core calculation logic is encapsulated into six score calculation functions and one probability calculation function. The glycated hemoglobin scoring function getHbA1cPoints(v) is implemented as follows: if v≤4, it returns 0; if v≥26, it returns 71.5; otherwise, it returns (v-4)×3.25. The co-infection scoring function getInfectionPoints(v) is implemented as follows: if v=1, it returns 7.6; otherwise, it returns 0. The indirect bilirubin scoring function, carbon dioxide combining power scoring function, white blood cell count scoring function, and immediate blood glucose scoring function are implemented using similar boundary judgment and linear interpolation logic. The total score calculation function sumPoints calls the above six scoring functions and adds the return values. The probability calculation function calcProb(totalPoints) calculates the linear prediction value using the formula LP=-11.1196577360942+0.167802877788283×totalPoints, and then returns 1 / (1+Math.exp(-LP)).
[0078] (3) The trigger and symptom verification module includes a symptom assessment interface and a trigger verification interface.
[0079] Symptom assessment interface: Provides an assessment checklist of 11 DKA-related symptoms and signs, using HTML checkboxes for multiple selection. The system categorizes symptoms into two types: severe warning symptoms (Kussmaul breathing, rotten apple odor on exhaled breath, altered consciousness, signs of severe dehydration) and general symptoms (excessive thirst, polyuria, weight loss, nausea and vomiting, abdominal pain, fatigue, lethargy). As the user selects symptoms, a JavaScript event listener updates the symptom selection status in real time, providing input for subsequent symptom coverage mechanism judgments.
[0080] Trigger Check Interface: Provides a checklist of 6 DKA triggers, using HTML checkboxes for multiple selection. Triggers include: new-onset diabetes, overeating or irregular eating habits, insulin interruption or missed injections, expired or improperly stored insulin, other stress factors, and no obvious trigger. The system specifically marks "new-onset diabetes" and "insulin interruption or missed injections" as high-risk triggers and highlights them in the risk assessment report.
[0081] (5) Integrated Decision-Making Module and Visual Output Module: This module integrates the model calculation results, symptom assessment results, and trigger verification results to generate a structured DKA risk assessment report. The report includes the following parts: basic risk level (displayed with color coding: low risk - green, medium risk - yellow, high risk - orange, very high risk - red), predicted probability percentage, symptom warning information (if the symptom coverage mechanism is triggered), trigger hint information, and graded clinical recommendations. The implementation logic of the symptom coverage mechanism is as follows: if the Risk output by calcProb() is <0.44 and the symptoms array contains any serious warning symptom, then a symptom warning label and recommendation text are added to the report.
[0082] Responsive layout adaptation: CSS3 media queries are used to achieve multi-device adaptation, optimizing the interface layout and control sizes for desktop (≥1024px), tablet (768-1023px), and mobile (<768px) respectively, ensuring a good user experience across different devices. The system can be deployed on Windows, Linux servers, or cloud platforms and supports real-time access via a browser.
[0083] Automatic fill function: For optional indicators (HbA1c, IBIL, CO2CP, WBC), when the user does not enter a value, the system automatically fills in the normal reference values: HbA1c defaults to 5.5%, IBIL defaults to 6.0 μmol / L, CO2CP defaults to 25.0 mmol / L, and WBC defaults to 7.5 × 10⁻⁶.9 / L. The system marks the indicator with a green "(Automatic)" mark in the scoring details display area to indicate to the user that the indicator is automatically filled in, and displays an automatic fill notification in the results area to remind the user that undetected indicators are filled in with normal reference values. The predicted results are for reference only, and it is recommended to improve the accuracy of the assessment by completing relevant checks.
[0084] The risk assessment system adopts a pure front-end architecture with a front-end and back-end separation. All calculation logic is executed locally in the user's browser's JavaScript engine, eliminating the need to transmit patient data to the server and fully protecting patient privacy and data security. The model calculation engine achieves millisecond-level response through pre-stored regression coefficients and interpolation lookup tables. When the user modifies any indicator value on the input interface, the system recalculates the risk probability and updates the risk level display in real time through an event-driven mechanism, providing instant feedback. The symptom coverage mechanism adds conditional judgment logic to the risk output stage, cross-validating qualitative symptom information with quantitative probability predictions. When the two are inconsistent (i.e., low probability but severe symptoms), the system prioritizes the warning effect of symptom information, reminding clinicians not to rely solely on model probabilities and ignore clinical signs.
[0085] The risk assessment system is user-friendly and responsive, completing the entire process from data input to risk assessment report output within tens of seconds, significantly improving the efficiency of DKA risk assessment. Its pure front-end architecture ensures that patient data remains locally, meeting stringent requirements for medical data privacy. The innovative symptom coverage mechanism compensates for the potential omission of clinical emergencies by purely quantitative models, achieving a complementary advantage between quantitative prediction and qualitative judgment, thus enhancing the safety and clinical reliability of risk assessment. Its responsive layout allows the system to be used on various terminal devices, facilitating flexible application by clinicians in different work scenarios (clinics, wards, emergency rooms).
[0086] Example 3: Typical Clinical Applications of Risk Assessment Systems This embodiment describes a typical application process of a risk assessment system in clinical practice, demonstrating the complete usage process from data input to result output.
[0087] Step 1: Data Input. Enter the following six indicators in the system data input interface: Glu=18.5 mmol / L, HbA1c=11.2%, WBC=15.3×10^9 / L, IBIL=8.5μmol / L, CO2CP=18 mmol / L, and co-infection=Yes (bronchopneumonia). Clinical Scenario: A 10-year-old male child with type 1 diabetes mellitus (T1DM) presented with fever and cough for 2 days, accompanied by nausea and vomiting for 1 day. The child has a 3-year history of T1DM and is usually treated with insulin. In the past 2 days, due to poor appetite, the parents have reduced the insulin dosage on their own. Physical examination upon presentation: slightly lethargic, slightly dry lips, skin turgor is acceptable, and breathing is slightly rapid. Emergency blood test results: immediate blood glucose 18.5 mmol / L, glycated hemoglobin 11.2%, white blood cell count 15.3×10^9 / L, indirect bilirubin 8.5 μmol / L, carbon dioxide combining power 18 mmol / L, chest X-ray showed bronchopneumonia.
[0088] Step 2: Model Calculation. The system's calculation engine performs the following calculations: S_HbA1c = (11.2 - 4) × 3.25 = 23.4 points; S_Infection = 7.6 points; S_IBIL = 8.5 × 21.8 / 40 = 4.6 points; S_CO2CP = (35 - 18) × 44.7 / 35 = 21.7 points; S_WBC = 15.3 × 36.7 / 50 = 11.2 points; S_Glu = 18.5 × 37.5 / 70 = 9.9 points. Total Points = 23.4 + 7.6 + 4.6 + 21.7 + 11.2 + 9.9 = 78.4 points. LP = -11.1197 + 0.1678 × 78.4 = 2.0346, Risk = 1 / (1 + e^(-2.0346)) = 88.5%.
[0089] Step 3: Symptom assessment. Select the three general symptoms: nausea and vomiting, slightly rapid breathing, and slightly poor mental state. Do not select the severe warning symptoms, and the symptom coverage mechanism will not be triggered.
[0090] Step 4: Trigger verification. Select infection (bronchopneumonia) and reduced insulin dosage (belonging to the category of insulin interruption). The system will identify two high-risk triggers.
[0091] Step 5: Risk Report Output. System output: Baseline risk level is extremely high (predicted probability 88.5%); Triggering factors: The patient has two high-risk triggers: infection and insulin interruption; Clinical recommendations: Immediately initiate the DKA diagnosis and treatment process, perform arterial blood gas analysis and blood ketone testing, actively control the infection, resume adequate insulin therapy, and prepare for emergency treatment. Although the child has not yet shown typical severe DKA symptoms, the model's predicted probability has exceeded the extremely high risk threshold of 80%, and two high-risk triggers are present, so the risk of DKA should be highly suspected.
[0092] In this typical application scenario, the child's HbA1c (11.2%) was significantly higher than the normal range, indicating poor long-term glycemic control; CO2CP (18 mmol / L) was lower than the lower limit of the normal range (22 mmol / L), indicating a tendency for mild metabolic acidosis; WBC (15.3×10^9 / L) was significantly elevated, indicating an inflammatory response caused by infection; and the concurrent infection (bronchopneumonia) contributed 7.6 points as a binary variable to the fixed score. The abnormalities of the above four indicators collectively pushed up the total score (78.4 points), making the predicted probability reach an extremely high risk level of 88.5%. The system's risk output module comprehensively presents the quantitative predicted probability and qualitative precipitating information (infection + insulin interruption), providing users with comprehensive and reliable decision support information.
[0093] As can be seen from the above typical application process, the risk assessment system of this invention can complete the quantitative assessment of DKA risk within seconds based on routine clinical laboratory indicators, accurately identifying the child as being in an extremely high-risk state (88.5%), and promptly indicating the presence of two high-risk triggers: infection and insulin interruption, providing clear action guidance for clinicians. Although the system's symptom coverage mechanism was not triggered in this scenario (because the baseline risk level was already extremely high), its design concept ensures that no clinical emergency will be missed due to a low model probability under any circumstances. In summary, the technical solution of this invention has significant practical value and promising prospects for widespread application in early warning and clinical decision support for DKA.
[0094] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Those skilled in the art can make various modifications and variations to the present invention without departing from its spirit and scope. For example, the specific number and type of predictive factors can be appropriately adjusted according to different medical scenarios and data availability; the risk stratification threshold can be personalized according to different clinical application scenarios and population characteristics; the specific implementation technology of the risk assessment system is not limited to HTML / CSS / JavaScript, but can also employ other suitable software development technologies. Therefore, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention also intends to include these modifications and variations.
Claims
1. A method for constructing an early warning model for type 1 diabetic ketoacidosis in children, characterized in that: Includes the following steps, S1, Data Acquisition: Extract routine clinical laboratory indicators from the medical information system for children with type 1 diabetes. These indicators include: complete biochemical test indicators, blood gas analysis indicators, glycated hemoglobin, co-infection status, 25-hydroxyvitamin D, liver function test indicators, blood lipid indicators, complete blood count indicators, kidney function indicators, and urine test indicators. S2, Data Preprocessing: Data preprocessing is performed on routine clinical laboratory indicators. Based on the principle of clinical applicability, laboratory indicators that can be quickly obtained in routine outpatient examinations are selected. Missing values are handled using the grouped median imputation method. After grouping by DKA status, the missing values of the corresponding groups are imputed using the median of each group. Samples with missing indicator rates exceeding the preset threshold are identified and deleted. S3, Candidate Variable Screening: Clinical characteristic analysis was performed on body mass index, co-infection status, glycated hemoglobin, total bilirubin, direct bilirubin, indirect bilirubin, total cholesterol, immediate blood glucose, triglycerides, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, blood urea nitrogen, carbon dioxide combining power, pH value, 25-hydroxyvitamin D, C-peptide, white blood cell count, neutrophil percentage, absolute neutrophil count, lymphocyte percentage, absolute lymphocyte count, monocyte percentage, absolute monocyte count, red blood cell count, hematocrit, hemoglobin, mean corpuscular volume, mean corpuscular hemoglobin, mean corpuscular hemoglobin concentration, red blood cell distribution width coefficient of variation, platelet count, C-reactive protein, urinary creatinine, urinary β2-microglobulin, and urinary microalbumin, and Graphpad was used for analysis. Prism 10.1.2 software used univariate logistic regression analysis to screen candidate variables for the indicators, and initially obtained candidate indicators such as glycated hemoglobin, co-infection status, total bilirubin, indirect bilirubin, total cholesterol, triglycerides, urea nitrogen, carbon dioxide combining power, white blood cell count, and 25-hydroxyvitamin D. The results of the univariate logistic regression analysis were obtained using the following formula: ln[P / (1-P)]=β0+β1X, In the formula, P represents the probability of a child developing diabetic ketoacidosis; β0 represents the intercept term; β1 represents the regression coefficient of the independent variable X; and X represents the clinical indicator to be analyzed. S4, Collinearity Diagnosis: Collinearity diagnosis is performed on the above candidate indicators, the variance inflation factor (VIF) is calculated, and collinear variables with VIF ≥ 5 are eliminated; The collinearity analysis results are obtained using the following formula: VIFj = 1 / (1-Rj²), where VIFj is the variance inflation factor of the j-th independent variable, and Rj² is the coefficient of determination obtained by linear regression with the j-th independent variable as the dependent variable and the remaining independent variables as predictors. When VIFj≥5, the independent variable is considered to have multicollinearity and is removed. S5, Predictor identification: Interaction term analysis was performed on the retained candidate indicators. Stepwise regression and multivariate logistic regression analysis were used to select variables. Finally, six independent predictors were selected, including: immediate blood glucose, glycated hemoglobin, indirect bilirubin, carbon dioxide combining power, white blood cell count and co-infection status. Multivariate logistic regression was used, and the inclusion criterion was P<0.
05. The formula for the multivariate logistic regression model is as follows: ln[P / (1-P)]=β0+β1X1+β2X2+β3X3+β4X4+β5X5+β6X6, In the formula, P is the probability of DKA, β0 is the intercept term, β1 to β6 are the regression coefficients corresponding to glycated hemoglobin, co-infection status, indirect bilirubin, carbon dioxide combining power, white blood cell count and immediate blood glucose, respectively, and X1 to X6 are the values of the corresponding independent variables. S6, Nomogram Model Construction: The model construction utilizes multivariate logistic regression analysis, with the occurrence status of DKA as the dependent variable and six routine clinical laboratory indicators as independent variables. The model formula is established and the nomogram prediction model is constructed using R language. In the nomogram prediction model, each independent variable corresponds to a certain number of score points based on its value, and the total score range is 0 to 220 points. S7, Risk Probability Calculation: Based on the total score, the risk probability is calculated using R language and the following formula to calculate the probability of DKA occurring: The linear forecast value LP = -11.1196577360942 + 0.167802877788283 × Total Points; Risk probability Risk = 1 / (1 + exp(-LP)); In the formula, LP represents the linear predicted value, Total Points represents the total risk score of DKA, and Risk represents the probability of DKA occurring. S8, Risk Stratification: Risk stratification is based on maximizing the Youden index and clinical expert consensus, setting four risk stratification thresholds: low risk (probability less than 11.49%), medium risk (probability 11.49% to 44%), high risk (probability 44% to 80%), and very high risk (probability greater than or equal to 80%). S9, Model Validation: The Bootstrap resampling method is used to internally validate the nomogram prediction model to evaluate its discrimination and calibration. S10, Integrated Decision Making: Combining the quantitative risk probability of the nomogram prediction model with the qualitative judgment results of DKA precipitating factors and symptoms, outputting comprehensive risk warnings and clinical decision-making suggestions.
2. The method for constructing an early warning model for type 1 diabetic ketoacidosis in children according to claim 1, characterized in that: The univariate analysis in step S3 includes: using independent samples t-test or Mann-Whitney U test for continuous variables, and chi-square test or Fisher exact test for categorical variables to compare the inter-group differences of each indicator between the T1DM-DKA group and the T1DM-non-DKA group; further performing univariate logistic regression analysis on the indicators that passed the inter-group difference test, with DKA status as the dependent variable and each clinical indicator as the independent variable, calculating the odds ratio (OR) and its 95% confidence interval for each indicator; and including indicators with P-values less than the preset threshold α as candidate predictors.
3. The method for constructing an early warning model for type 1 diabetic ketoacidosis in children according to claim 1, characterized in that: The six independent predictors and their odds ratios (ORs) in step S5 are as follows. Data analysis was performed using GraphpadPrism 10.1.2 software: HbA1c: OR=2.162, 95% confidence interval is 1.840 to 2.608; Co-infection: OR=3.988, 95% confidence interval 1.563 to 10.52; Indirect bilirubin (IBIL): OR = 1.150, 95% confidence interval: 1.070 to 1.240; Carbon dioxide binding affinity (CO2CP): OR = 0.7437, 95% confidence interval: 0.6749 to 0.8078; White blood cell count (WBC): OR = 1.356, 95% confidence interval 1.204 to 1.555; Immediate blood glucose (Glu): OR = 1.125, 95% confidence interval: 1.070 to 1.
190.
4. The method for constructing an early warning model for type 1 diabetic ketoacidosis in children according to claim 1, characterized in that: In step S6, the scoring criteria for the six predictor variables of the nomogram prediction model are determined using linear interpolation: Glycated hemoglobin score: 0 points for 4%, 3.25 points for each additional 1%, and 71.5 points for 26%; Co-infection status score: 0 points for no, 7.6 points for yes; Indirect bilirubin score: 0 points are given starting from 0 μmol / L, and linear interpolation is used within the range of 0 to 40 μmol / L, with 21.8 points given for 40 μmol / L. Carbon dioxide binding capacity score: 0 points are given starting from 35 mmol / L, and the score is calculated by inverse linear interpolation within the range of 35 to 0 mmol / L, with 44.7 points corresponding to 0 mmol / L. White blood cell count score: 0 points are given as the starting point of 0 × 10⁹ per liter, and linear interpolation is used within the range of 0 to 50 × 10⁹ per liter. 50 × 10⁹ per liter corresponds to 36.7 points. Immediate blood glucose score: 0 points are assigned starting from 0 mmol / L, and the score is linearly interpolated within the range of 0 to 70 mmol / L, with 37.5 points assigned to 70 mmol / L. The formula for calculating the total score is as follows: Total Points=P_HbA1c+P_Infection+P_IBIL+P_CO2CP+P_WBC+P_Glu, In the formula, P_HbA1c represents the number of score points corresponding to HbA1c, P_Infection represents the number of score points corresponding to co-infection, P_IBIL represents the number of score points corresponding to indirect bilirubin, P_CO2CP represents the number of score points corresponding to carbon dioxide combining power, P_WBC represents the number of score points corresponding to white blood cell count, and P_Glu represents the number of score points corresponding to immediate blood glucose.
5. The method for constructing an early warning model for type 1 diabetic ketoacidosis in children according to claim 1, characterized in that: The coefficients in the risk probability calculation formula in step S7 are derived from the parameter estimation of multifactor logistic regression. The complete calculation formulas for the linear predicted value LP and the risk probability Risk are as follows: LP=-11.1196577360942+0.167802877788283×Total Points, Risk = 1 / (1 + exp(-LP)), The risk probability ranges from 0 to 1, corresponding to a DKA occurrence probability of 0% to 100%.
6. The method for constructing an early warning model for type 1 diabetic ketoacidosis in children according to claim 1, characterized in that: In step S8, the risk stratification threshold is determined after internal verification using the Bootstrap resampling method. The sampling is repeated 1000 times, and the optimal cut point is determined based on the principle of maximizing the Youden exponent. The risk stratification includes four risk levels: low risk (DKA probability less than 11.49%), medium risk (DKA probability 11.49% to 44%), high risk (DKA probability 44% to 80%), and very high risk (DKA probability greater than or equal to 80%). The area under the receiver operating characteristic (AUC) curve for model validation was 0.988, with a 95% confidence interval of 0.984 to 0.993, a sensitivity of 93.4%, and a specificity of 94.9%.
7. A risk assessment system based on the method for constructing an early warning model for type 1 diabetic ketoacidosis in children as described in any one of claims 1-6, characterized in that: include, The data acquisition and preprocessing module is used to acquire routine clinical laboratory indicators of the children and perform missing value processing, wherein the missing value processing adopts the grouped median imputation method. The risk probability calculation module is used to calculate the probability of DKA occurrence based on the nomogram model; The trigger and symptom verification module provides an interactive interface for users to select DKA triggers and typical symptoms. The integrated decision-making module combines model calculations of risk probabilities and qualitative verification results to generate the final risk level and clinical recommendations. The visualization output module is used to generate risk probability progress bars, risk level labels, score details, and structured clinical recommendations.
8. The ketoacidosis risk assessment system for children with type 1 diabetes as described in claim 7, characterized in that: The trigger and symptom verification module includes the following DKA trigger verification items: new-onset diabetes, i.e., first diagnosis of type 1 diabetes, insulin interruption or missed injection, overeating or improper diet, expired or improperly stored insulin, and other stress factors including surgery, trauma, and emotional excitement. The following symptoms and signs are checked and divided into two levels: severe symptoms and moderate symptoms. Severe symptoms include deep and rapid breathing or Kussmaul breathing or sighing breathing, breath with the smell of rotten apple or acetone, altered consciousness or lethargy or drowsiness or confusion or coma, obvious dehydration or dry mouth or sunken eyes or poor skin turgor or decreased urine output. Moderate symptoms include nausea or vomiting or abdominal pain, increased thirst or urination or worsening weight loss, fatigue or weakness or decreased activity, poor mental state or reluctance to speak or slowed reaction time.
9. A ketoacidosis risk assessment system for children with type 1 diabetes according to claim 7, characterized in that: The decision-making logic of the comprehensive decision-making module includes three levels of judgment rules: The first layer is the symptom warning rule: when a child has any serious symptom or sign, regardless of the model's predicted probability, the system will trigger a high-priority warning and recommend immediate investigation of DKA. The second layer is the trigger superposition rule: when the child has clear triggers for DKA such as insulin interruption or missed injection, new-onset diabetes, or co-infection, the system increases the risk level by one level based on the model probability and outputs a corresponding high-priority warning. The third layer consists of model-driven rules: when the model predicts a high risk (≥44%) or an extremely high risk (≥80%), the system outputs the corresponding risk level and emergency clinical treatment recommendations; when the model predicts a low risk (<11.49%) and the child has no obvious symptoms or signs, the system recommends routine monitoring.
10. A ketoacidosis risk assessment system for children with type 1 diabetes according to claim 7, characterized in that: The output content of the visualization output module includes: The risk probability value is displayed as a percentage indicating the probability of DKA occurring. The risk probability progress bar uses a color gradient bar to visually display the position of the risk probability within the range of 0% to 100%. Risk level labels are displayed as color-coded badges indicating low, medium, high, or very high risk levels. The scoring details are displayed in a grid format, showing the score points for each of the six indicators and the total score. Structured clinical recommendations automatically generate corresponding levels of clinical treatment recommendations based on risk level and symptoms / signs.
