Construction method of glucocorticoid induced diabetes risk prediction model based on LASSO algorithm

Abstract

The invention relates to the technical field of medical data analysis and clinical risk prediction, provides a construction method of a glucocorticoid induced diabetes risk prediction model based on an LASSO algorithm, and belongs to the technical scheme of data processing executed on computing equipment. Clinical data of inpatients receiving systemic glucocorticoid treatment are collected, key variables are screened through LASSO regression, and a Logistic regression model is constructed to achieve risk prediction. The model is composed of four conventional clinical indexes, has good distinction degree and calibration degree, and finally realizes individualized risk assessment through column diagram form output. The method is simple, practical and accurate, and is suitable for SDM prediction and intervention management of clinical high-risk groups.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data analysis and clinical risk prediction, and in particular to a method for constructing a glucocorticoid-induced diabetes risk prediction model based on the LASSO algorithm. Background Art

[0002] Glucocorticoids (GC) are widely used to treat autoimmune diseases, inflammatory diseases, allergic diseases, and certain malignancies due to their potent anti-inflammatory and immunosuppressive effects. However, high-dose or long-term use of glucocorticoids can lead to serious metabolic side effects, among which glucocorticoid-induced hyperglycemia and diabetes (SDM) are particularly common and worrying. SDM refers to new-onset hyperglycemia or diabetes in patients without a history of diabetes due to GC exposure. Its mechanism of occurrence is mainly related to GC-induced pancreatic β-cell dysfunction, increased hepatic gluconeogenesis, and peripheral insulin resistance. It is reported that the use of glucocorticoids can cause hyperglycemia in approximately 40%–50% of patients, making SDM a major challenge in clinical practice.

[0003] Inpatients in dermatology departments often require high-dose or long-term systemic GC therapy due to the severity of their disease, placing them at high risk for SDM. Early identification of these high-risk patients is crucial for timely intervention and prevention of long-term complications. However, current clinical assessment of SDM risk relies primarily on empirical judgment and traditional univariate statistical models, which often struggle to accurately quantify individual risk. Studies have identified several risk factors associated with SDM, such as older age, elevated body mass index (BMI), family history of diabetes, history of hypertension, concomitant use of immunosuppressants, elevated triglycerides, and a high daily GC dose. While these risk factors provide some insight, complex interactions among them make it difficult to accurately predict an individual's probability of developing SDM based solely on a single or limited number of factors. Therefore, there is an urgent need to develop a data-driven approach that integrates multiple clinical indicators to improve the accuracy and individualization of SDM risk prediction.

[0004] Therefore, a method for constructing a glucocorticoid-induced diabetes risk prediction model based on the LASSO algorithm was proposed. Summary of the Invention

[0005] To address the difficulty in accurately predicting SDM risk in existing technologies, the present invention provides an SDM risk prediction model based on a machine learning algorithm and its construction method. This method processes historical clinical data stored in a hospital's electronic medical record (EMR) system offline on a computing device. Using the machine learning algorithm LASSO, four key predictive indicators (such as family history of diabetes) are selected from a pool of up to 26 candidate variables. A logistic regression model is then constructed based on these four indicators. This model accurately predicts the risk of SDM using a concise combination of variables. The model exhibits excellent discrimination and calibration, and intuitively presents individual risk probabilities via a nomogram. This model can be used clinically to identify high-risk patients in advance, guiding physicians to strengthen monitoring and intervention, and thus optimizing patient treatment plans and prognosis.

[0006] The technical solution of the present invention is as follows: a method for constructing a glucocorticoid-induced diabetes risk prediction model based on LASSO, comprising the following steps:

[0007] S1: Data Collection and Preprocessing. Clinical data and follow-up results of inpatient dermatology patients receiving systemic glucocorticoid (GC) treatment were retrieved from the hospital's electronic medical record system. Whether patients developed GC-induced diabetes during treatment was recorded. The collected data were cleaned and preprocessed, including removing missing data and outliers.

[0008] S2: Feature screening (variable selection). The cleaned patient data were input into the machine learning algorithm, and candidate variables were screened by least absolute shrinkage and selection operator (LASSO) regression. K-fold cross-validation was used to determine the optimal lambda parameter, compress the uncorrelated coefficient, and screen out key predictors closely related to the development of glucocorticoid-induced diabetes.

[0009] S3: Model construction. The key predictive factors screened in step S2 are input into a logistic regression model to establish a glucocorticoid-induced diabetes risk prediction model, obtaining a regression equation with the predictive factors as independent variables and the glucocorticoid-induced diabetes risk as the dependent variable.

[0010] S4: Internal Validation: Use cross-validation and bootstrap resampling methods to internally validate the model established in step S3, evaluate the model's stability and generalization ability under different sample partitions, and obtain the model's average performance index.

[0011] S5: Model Evaluation. Model discrimination was evaluated by calculating the area under the receiver operating characteristic (ROC) curve (AUC). Model calibration was assessed using the Hosmer-Lemeshow test and calibration curves. Decision curve analysis (DCA) was used to assess the model's net clinical benefit at different risk thresholds to verify the model's predictive performance and clinical utility. When model discrimination and calibration met the predefined requirements, the model was designated as the final model.

[0012] S6: Output Results. The performance model is visualized as a nomogram, with corresponding scales and scores plotted based on the regression coefficient ratios of each predictor in the logistic regression equation. Based on the individual patient's variable values, the corresponding total score is calculated using the nomogram and mapped to a risk probability or risk score for developing SDM. This output is intended to assist clinicians in identifying high-risk individuals and is not intended to be used as a diagnostic indicator or direct basis for diagnosis.

[0013] Furthermore, the clinical data in S1 include: general demographic information, patient medical history and lifestyle habits, medication use, glucocorticoid treatment-related indicators, and laboratory test indicators.

[0014] Furthermore, in step S2, a total of 26 candidate independent variables including age, BMI, relevant medical history and laboratory indicators were included; λ=λ1se was determined using 10-fold cross-validation, and LASSO regression screening was performed. Ultimately, only four key predictors, including family history of diabetes, use of immunosuppressants, average daily glucocorticoid dose and triglyceride level, were retained, whose regression coefficients were non-zero and served as input features of the subsequent multivariate logistic regression model; the remaining candidate variables were automatically eliminated due to their insignificant predictive contributions.

[0015] Furthermore, the discrimination index is the area under the receiver operating characteristic curve (AUC), and the AUC is required to be no less than 0.85;

[0016] The calibration index was the P value of the Hosmer–Lemeshow test, which was required to be greater than 0.05;

[0017] The clinical benefit index was within the risk threshold range of 0%–100%, and the clinical net benefit curve obtained by decision curve analysis was above the "all intervention" strategy and the "no intervention" strategy.

[0018] Furthermore, the specific mathematical expression of the logistic regression risk prediction model in S3 is:

[0019]

[0020]

[0021] Where P is the probability of a patient developing glucocorticoid-induced diabetes, X1–X4 are the variables corresponding to family history of diabetes, use of immunosuppressants, average daily GC dose, and triglyceride level, respectively, β0 is the intercept term, and β1–β4 are the corresponding regression coefficients.

[0022] Furthermore, the method for constructing the nomogram in S6 includes:

[0023] Step S61: Determine the score scale on the nomogram based on the regression coefficients of each predictor in the logistic regression model. Specifically, the regression coefficients of each predictor are scaled so that factors with greater predictive contributions are assigned a wider score range. For binary variables (such as family history of diabetes and immunosuppressant use), fixed scores are assigned based on their coefficients. For continuous variables (such as average daily GC dose and triglyceride levels), different score scales are drawn based on their numerical ranges and corresponding coefficients.

[0024] Step S62: The process of calculating the predicted risk probability for an individual patient includes the following steps: first, converting the patient's specific values ​​for each predictor into corresponding nomogram scores; second, summing the scores for these four predictors to obtain the patient's total score; and finally, based on this total score, searching the predicted probability corresponding to the total score on the risk scale below the nomogram. This risk scale, pre-drawn based on the risk prediction formula of the logistic regression model, essentially serves as a risk comparison table. This allows clinicians to quickly determine a patient's individualized risk probability for glucocorticoid-induced diabetes without complex calculations, facilitating timely clinical decision-making and intervention.

[0025] Furthermore, in step S1 of the present invention, the collected clinical data are subdivided into the following categories, which are illustrated below to facilitate understanding of the specific meaning of each category:

[0026] General demographic information: age, sex, height, weight, body mass index (BMI), etc.;

[0027] Patient medical history: previous medical history (such as hypertension), family medical history (such as family history of diabetes), etc.;

[0028] Lifestyle habits: smoking history (whether smoking), drinking history (whether drinking), etc.;

[0029] Medication use: current or recent use of immunosuppressants (such as cyclophosphamide, methotrexate, etc.), hypoglycemic drugs, lipid-lowering drugs, diuretics, etc.;

[0030] Glucocorticoid treatment-related indicators: initial glucocorticoid dose, maximum daily dose, average daily dose, cumulative dose, treatment duration, etc.;

[0031] Laboratory examination indicators: routine blood tests, liver function (ALT, AST, GGT, etc.), renal function (urea nitrogen, serum creatinine, cystatin C, etc.), lipid profile (total cholesterol, triglycerides, HDL-C, LDL-C), blood sugar-related indicators (fasting blood glucose, glycosylated hemoglobin), uric acid and uric acid / creatinine ratio, etc.

[0032] Beneficial Effects: This study introduces machine learning algorithms, such as the LASSO, into the field of SDM risk prediction, innovatively constructing a concise model that requires only four common clinical indicators, overcoming the limitations of traditional methods in integrating multiple factors. Compared with existing methods based on experience or single-factor scoring, the proposed model exhibits higher predictive accuracy and individualized specificity, with an AUC of 0.86, demonstrating excellent predictive performance. The model output is presented in the form of a nomogram, allowing medical personnel to intuitively obtain patient risk estimates without complex calculations, facilitating its application in clinical decision-making. Furthermore, the high accuracy and personalized predictive power of the proposed four-indicator model can help clinicians more accurately and earlier identify patients at high risk of SDM in populations receiving glucocorticoid treatment. This provides a strong evidence-based basis for proactively implementing targeted, individualized prevention and intervention measures (such as adjusting treatment regimens, enhancing blood glucose monitoring, and providing lifestyle guidance), potentially reducing the incidence of SDM and improving patient outcomes. Since the model only relies on four conventional indicators and does not require any additional special or expensive tests, it has an extremely low threshold for clinical application and is cost-effective. It can be quickly popularized in medical institutions at all levels and can be widely used in patients who need glucocorticoid treatment. It can help clinicians identify high-risk patients for SDM in advance and guide individualized intervention measures. It has significant clinical significance and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the process of the present invention;

[0034] Figure 2 is a flow chart of the risk prediction method of the present invention;

[0035] Figure 3 It is the LASSO regression coefficient path diagram of the present invention;

[0036] Figure 4 This is the LASSO regression cross validation curve diagram of the present invention;

[0037] Figure 5 This is the LASSO variable screening result diagram of the present invention;

[0038] Figure 6This is the SDM risk prediction nomogram of the present invention;

[0039] Figure 7 : is the ROC curve diagram of the model of the present invention on the entire data set;

[0040] Figure 8 : is the average ROC curve diagram after cross validation of the present invention;

[0041] Figure 9 : is the calibration curve diagram of the model of the present invention;

[0042] Figure 10 : Decision curve analysis (DCA) diagram of the model of the present invention. DETAILED DESCRIPTION

[0043] The present invention is further described below with reference to specific examples, but the protection scope of the present invention is not limited thereto.

[0044] Combine Figures 1-10 Further explanation

[0045] Example 1:

[0046] A method for constructing a glucocorticoid-induced diabetes risk prediction model based on LASSO, comprising the following steps:

[0047] S1: Data collection and preprocessing. First, the hospital electronic medical record system (EMR) was used to extract information on inpatients who received systemic glucocorticoid treatment in the Department of Dermatology, Affiliated Hospital of Xuzhou Medical University from August 2019 to October 2024. The collected data included demographic characteristics, underlying disease history, hormone medication information, and laboratory test indicators. Based on the existing content in the medical records and follow-up records, it was noted whether the patient had glucocorticoid-related dysglycemic events during treatment, which was used for the dependent variable setting in subsequent modeling;

[0048] Subsequently, the raw data were cleaned and preprocessed using statistical software (R 4.4.1). Missing data were graded based on the missingness ratio: missing data with a missingness ratio below 5% were imputed using the median (continuous variables) or mode (categorical variables); missing data between 5% and 20% were imputed using multiple imputation using the mice package in R, with predictive mean matching (PMM) as the imputation method. If the missingness ratio exceeded 20%, the variable was considered for exclusion based on its expected importance in the modeling.

[0049] Finally, the collected data included the following information: general demographic information (such as age, sex, height and weight for calculating body mass index (BMI), etc.); patient medical history and lifestyle habits (such as history of hypertension, smoking history, drinking history, family history of diabetes, etc.); medication use (including whether immunosuppressants are used concomitantly); glucocorticoid treatment-related indicators (including initial GC daily dose, maximum GC daily dose, average GC daily dose and treatment duration); laboratory test indicators (including red blood cell distribution width (RDW), aspartate aminotransferase (AST), alanine aminotransferase (ALT), glutamyl transpeptidase (GGT), blood urea nitrogen (BUN), serum creatinine (SCr), serum uric acid (SUA) and uric acid-creatinine ratio, cystatin C, total cholesterol (TC), triglycerides (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), serum magnesium, etc.).

[0050] S2: Feature Screening (Variable Selection). This paper uses the least absolute shrinkage and selection operator (LASSO) to screen candidate variables. LASSO introduces an L1-norm regularization term into the log-likelihood loss function, imposing a sparsity penalty on the regression coefficient β, shrinking the partially uncorrelated coefficients to zero. Its optimization objective function is as follows:

[0051]

[0052] in

[0053] The probability of predicting positive for the i-th sample;

[0054] The first bracket contains the log-likelihood loss function (commonly used in logistic regression);

[0055] λ is the regularization penalty coefficient, and the optimal value is determined by 10-fold cross validation (cv. glmnet). (lowest mean square error) and . (the simplest model within 1 standard error);

[0056] When λ increases, the penalty becomes stronger, and more is compressed to zero; when λ decreases, the model retains more variables;

[0057] In this example, there are 26 candidate independent variables, including age, BMI, relevant medical history, laboratory indicators, etc., which are selected through 10-fold cross validation. Four variables with non-zero coefficients were obtained: family history of diabetes, use of immunosuppressants, average daily GC dose, and triglyceride level. These four items served as input features for the subsequent multivariate logistic regression.

[0058] Although the remaining variables (such as serum uric acid) are known metabolic indicators, their predictive contribution was not significant after inclusion in the multivariate model and they were automatically excluded in the LASSO screening, indicating that they have limited predictive effect on SDM in this dataset.

[0059] S3: Model construction. The four key predictors screened in S2 were used as independent variables to construct an SDM risk prediction model using multivariate logistic regression. During the construction of the logistic regression model, the regression coefficients of these four variables were estimated to obtain the model's regression equation:

[0060]

[0061] Where P represents the probability of a patient developing SDM, - Corresponding to the above 4 predictive factors (family history of diabetes, use of immunosuppressants, average daily GC dose, and triglyceride level), is the model intercept term, - is the regression coefficient of each factor.

[0062] The results of the logistic regression analysis in this example showed that the four factors of family history of diabetes, use of immunosuppressants, average daily GC dose, and triglyceride level were all highly correlated with the occurrence of SDM, and their regression coefficients were approximately:

[0063] = 2.43 (family history of diabetes, binary variable, 1 for presence, 0 for absence)

[0064] = 1.66 (immunosuppressant use, binary variable, 1 for use and 0 for absence)

[0065] = 3.32 (average daily GC dose in mg / kg)

[0066] = 0.49 (triglyceride level, in mmol / L),

[0067] = -6.12

[0068] The resulting risk prediction model can be expressed as ,

[0069]

[0070] A non-penalized multivariate logistic regression model was constructed based on four variables selected using the LASSO (λ=λ1se) analysis. Results showed that family history of diabetes (OR=11.4, 95% CI 3.69–35.2), concomitant use of immunosuppressants (OR=5.24, 95% CI 2.20–12.5), average daily GC dose (OR=27.5, 95% CI 5.84–130), and triglyceride level (OR=1.63, 95% CI 1.06–2.51) were independent risk factors for SDM (all P<0.05). The model residuals (Deviance=155.07) and AIC=165.07 were good fit.

[0071] The model coefficients indicate that, under the same conditions: having a family history of diabetes or using immunosuppressants will significantly increase the risk of SDM; the greater the average daily GC dose, the higher the risk of SDM; and elevated triglyceride levels are also statistically associated with an increased risk of SDM.

[0072] S4: Internal validation. In order to verify the robustness of the model, the present invention uses a combination of cross-validation and Bootstrap resampling for internal validation. Specifically, in this embodiment, the caret software package is used to randomly divide the original data into 10 parts, 9 parts are used for training models and 1 part for verification each time, and the cycle is repeated 10 times to obtain the cross-validation performance index of the model; at the same time, the Bootstrap sampling is repeated 1000 times, and the data set of each resample is remodeled and the performance is recorded to evaluate the fluctuation of the model under different sampling. During the model validation process, the pROC software package is used to calculate the AUC value of each ROC curve to quantify the prediction performance. The cross-validation results show that the average AUC of the model is approximately 0.839 (95% CI 0.769–0.909), which is very close to the AUC 0.860 obtained on the complete data set, indicating that the model can maintain a high prediction accuracy on different data subsets. The regression coefficient distribution and performance indicators obtained by the Bootstrap method also support the stability of the model, suggesting that the model does not have overfitting caused by specific sample divisions.

[0073] S5: Model evaluation. The final model’s discrimination, calibration, and clinical applicability were comprehensively evaluated. In this example, the discrimination of the model was analyzed by ROC curve, and the area under the curve (AUC) was 0.860, indicating that the model had good discrimination ability for SDM patients and non-SDM patients (e.g. Figure 7The Hosmer-Lemeshow test yielded p = 0.2904, and the calibration curve was plotted to find that the model-predicted probability was very close to the actual observed value, and the curve nearly coincided with the ideal baseline, indicating that the model was well calibrated (e.g. Figure 8 In addition, the present invention uses decision curve analysis to evaluate the clinical benefits of the model. The results show that within a wide range of risk thresholds (e.g., 5% to 100%), using this model for intervention decision-making can achieve higher net benefits than the "all intervention" or "no intervention" strategies, and the net benefit curve of the model within the entire threshold range is above the baseline (e.g., Figure 10 This indicates that the model still has practical value even at higher decision thresholds. Taking all of the above indicators into account, the model of the present invention performs well in terms of discrimination, calibration, and clinical practicality.

[0074] S6: Result output and clinical application. The verified Logistic regression model is made into an intuitive nomogram for clinical application. The nomogram uses a graphical scale to show the contribution of each predictor to the risk: In this embodiment, the nomogram is as follows: Figure 6 As shown, the nomogram includes four sub-items: "Family History of Diabetes," "Use of Immunosuppressants," "Average Daily GC Dose," and "Triglycerides." Each sub-item has a corresponding scale and score below it. To use it, first find the score on the corresponding scale based on the patient's specific value for each indicator. Then, add the four scores to obtain a total score. Finally, read the patient's predicted probability of developing SDM based on the risk scale corresponding to the total score below the nomogram. This allows medical professionals to directly obtain a patient's personalized risk assessment results without complex calculations.

[0075] Example 2:

[0076] Model construction and verification. In this example, inpatients of the Department of Dermatology of the Affiliated Hospital of Xuzhou Medical University were used as the research subjects for a retrospective analysis, and data on 326 patients who received systemic glucocorticoid treatment between August 2019 and October 2024 were collected. According to the inclusion and exclusion criteria (excluding patients with previous glucose metabolism disorders, use of drugs that affect blood sugar, combined with severe hepatopancreatic diseases, and incomplete clinical data), 293 patients were finally included in the statistical analysis. According to the abnormal blood sugar levels recorded in the follow-up and hospitalization records, they were marked as new-onset diabetes events (SDM group), and the remaining 251 cases did not occur (control group). General demographic information, patient medical history and lifestyle habits, drug use, glucocorticoid treatment-related indicators, and laboratory test indicators were collected, and whether SDM occurred was used as the dependent variable.

[0077] In this example, all continuous variables were tested for normality using the Shapiro–Wilk test. The results showed that the p-values ​​for each variable were all less than 0.05, indicating that the null hypothesis was rejected, indicating that none of the collected continuous variables conformed to a normal distribution. Therefore, the median (interquartile range) was used in the subsequent descriptive statistics.

[0078] First, baseline characteristics of patients in the SDM group (n=42) and the non-SDM group (n=251) were compared (see Appendix Table 1). Results showed that the median age of patients in the SDM group was higher than that in the non-SDM group (54 vs. 47 years, p<0.05), the median body mass index (BMI) was slightly higher (24.87 vs. 23.53, p<0.01), and the prevalence of patients with a history of hypertension was also higher (31.0% vs. 14.7%, p<0.05). Notably, a family history of diabetes was significantly higher in the SDM group (38.1%), compared to 4.0% in the non-SDM group (p<0.001). Furthermore, a significantly higher proportion of patients in the SDM group received combined immunosuppressive therapy (57.1%), compared to 15.5% in the non-SDM group (p<0.001).

[0079] In terms of glucocorticoid treatment (see Appendix 2), the median daily GC dose (calculated as methylprednisolone equivalent) of patients in the SDM group was 0.89 mg / kg, which was significantly higher than that of the control group (0.64 mg / kg) (p<0.001), suggesting that higher doses of GC are closely associated with SDM.

[0080] Laboratory test results also showed (see Appendix 3) that the metabolic index risks of patients in the SDM group were higher, among which the median triglyceride level was 1.66mmol / L, which was significantly higher than the 1.20mmol / L in the non-SDM group (p<0.001). At the same time, the levels of urea nitrogen, creatinine, uric acid, total cholesterol, cystatin C and other indicators in the SDM group were higher than those in the control group and the differences were statistically significant (p<0.05).

[0081] The difference in baseline characteristics indicates that patients in the SDM group carry more metabolic risk factors, which provides a basis for variable screening of subsequent models. First, LASSO regression analysis was performed on the above data set according to S2. This example uses the glmnet package of statistical software (R4.4.1) to perform LASSO regression processing on 26 candidate independent variables, and determines the optimal regularization parameter λ of the model (such as Figure 3 As shown in Figure 2 ). Within the range of one standard deviation of λ, four variables with non-zero coefficients were screened, namely family history of diabetes, use of immunosuppressants, average daily GC dose, and triglycerides (e.g. Figure 4 and Figure 5 Then, according to step S3, these four characteristics were incorporated into the multivariate logistic regression model to establish a prediction equation for the risk of SDM. In the process of model construction, the R language statistical module was used to perform logistic regression fitting, and the estimated regression coefficients of each variable were approximately = 2.43, = 1.66, = 3.32, =0.49, the intercept is approximately = -6.12 and the specific mathematical expression of the logistic regression risk prediction model in S3 mentioned above consistent.

[0082] Where P represents the patient's risk of developing glucocorticoid-induced diabetes, e is the base of the natural logarithm, and x is the linear predictor of the logistic regression equation, specifically: x = -6.12 + 2.43 × family history of diabetes + 1.66 × use of immunosuppressants + 3.32 × average daily GC dose + 0.49 × triglyceride level.

[0083] Among them, family history of diabetes and use of immunosuppressants are both binary variables: 1 if present and 0 if not; the unit of average daily GC dose is mg / kg; triglyceride level is measured in mmol / L. The daily glucocorticoid dose (Avg DailyGC Dose) is the average daily dose used by the patient during glucocorticoid treatment. In order to unify the potency differences of different glucocorticoid dosage forms, the present invention converts the dose of all glucocorticoid preparations into methylprednisolone equivalents to improve the consistency of variables and the applicability of the model.

[0084] The ROC curve of the model is as follows Figure 7 As shown, AUC = 0.860, indicating that the model has a high discriminative ability. Through internal 10-fold cross validation, the average AUC = 0.839 (as shown Figure 7 The Hosmer-Lemeshow test p=0.2904>0.05, combined with the calibration curve (as shown in Figure 9 As shown in Figure 2, the predicted risk is very close to the actual observed value, and the model is well calibrated. Decision curve analysis shows that the application of this model within a reasonable risk threshold range can achieve significant clinical net benefits (e.g. Figure 10 The above results show that the risk prediction model constructed in this example has excellent performance and can be reliably used to predict the risk of SDM in inpatients of dermatology department after receiving GC treatment.

[0085] Example 3:

[0086] Construction and application of nomogram. Based on the above Logistic regression equation, the SDM risk prediction nomogram (such as Figure 6 As shown). This embodiment uses the rms software package of the statistical software (R 4.4.1) to generate the corresponding nomogram according to the logistic model, and graphically presents the corresponding relationship between the weight of each predictor and the total risk score and the probability of occurrence. In the nomogram, the value range of each predictor is mapped to a scale of 0 to 100 points according to its logarithmic regression contribution. For binary variables (family history of diabetes and use of immunosuppressants), "yes" and "no" correspond to fixed scores respectively; for continuous variables (average daily GC dose and triglycerides), different score scales correspond to specific values. When using this nomogram, clinicians can follow Figure 6 Based on the patient's specific circumstances, the score is calculated and the risk is predicted.

[0087] For example, consider a 60-year-old male patient with no history of hypertension, nonsmokers, or alcoholics, but with a family history of type 2 diabetes and currently receiving the immunosuppressant cyclophosphamide. He is admitted to the hospital with a diagnosis of pemphigus and requires high-dose GC therapy, with an average daily GC dose of approximately 0.8 mg / kg. Laboratory tests on admission reveal normal fasting blood glucose, but elevated triglycerides of 2.5 mmol / L. Based on this patient's situation, the corresponding scores on the nomogram are: 1 family history of diabetes (approximately 41 points), 1 immunosuppressant use (approximately 27 points), an average daily dose of 0.8 mg / kg (approximately 33 points), and a triglyceride level of 2.5 mmol / L (approximately 16 points), for a total score of approximately 117. This total score indicates that this patient has an approximately 85% risk of developing SDM in the future. Considering this patient to be at high risk, clinical measures should be taken proactively, including close blood glucose monitoring, dietary control, and, if necessary, prophylactic use of hypoglycemic medications.

[0088] To verify the consistency between the nomogram scale and the model formula, the above 60-year-old male case (family history of diabetes = 1, use of immunosuppressants = 1, average daily GC dose = 0.8 mg / kg, TG = 2.5 mmol / L) was substituted into the formula and calculated as follows:

[0089] X = –6.12 + 2.43×1 + 1.66×1 + 3.32×0.8 + 0.49×2.5 ≈ 1.84

[0090] ≈ 0.86

[0091] This means the predicted SDM risk is approximately 86%. This value is consistent with the risk probability read from the nomogram (≈ 85%), indicating that the nomogram scale fully corresponds to the mathematical model and can be used for rapid clinical assessment.

[0092] The nomogram tool provided by this invention allows healthcare professionals to quickly quantify a patient's SDM risk and develop appropriate management strategies for patients at varying risk levels, embodying the principles of precision medicine. It should be noted that those skilled in the art may adjust the specific thresholds and parameters in the above embodiments within the spirit and scope of this invention, for example, by calibrating model coefficients or updating the nomogram scale based on different populations, and such modifications fall within the scope of this invention.

[0093] Appendix 1: Comparison of demographic characteristics of patients in the SDM group and the non-SDM group

[0094] variable Non-SDM (n = 251) SDM (n = 42) P-value Age (years), (median [IQR]) 47.00 [33.00, 62.00] 54.00 [45.50, 62.00] 0.017 Gender, n (%) 1.000 male 97 (38.6) 16 (38.1) female 154 (61.4) 26 (61.9) BMI (kg / m²), (median [IQR]) 23.53 [21.16, 25.71] 24.87 [23.04, 26.78] 0.008 Hypertension, n (%) 37 (14.7) 13 (31.0) 0.018 Family history of diabetes, n (%) 10 (4.0) 16 (38.1) <0.001 Smoking, n (%) 18 (7.2) 5 (11.9) 0.456 Drinking, n (%) 13 (5.2) 4 (9.5) 0.448 Immunosuppressant use, n (%) 39 (15.5) 24 (57.1) <0.001

[0095] Appendix 2: Comparison of glucocorticoid use (converted to methylprednisolone equivalent) between the SDM group and the non-SDM group

[0096] variable Non-SDM (n = 251) SDM (n = 42) P-value Initial GC dose (mg / d) (median [IQR]) 40.00 [32.00, 60.00] 54.00 [40.00, 60.00] 0.298 Maximum daily dose of GC (mg / kg), (median [IQR]) 0.74 [0.57, 1.00] 0.86 [0.68, 1.07] 0.072 Average daily GC dose (mg / kg), (median [IQR]) 0.64 [0.50, 0.84] 0.89 [0.64, 1.29] <0.001 GC treatment duration (days), (median [IQR]) 8.00 [6.00, 11.00] 6.50 [5.00, 9.75] 0.254

[0097] Appendix 3: Comparison of laboratory test indicators between the SDM group and the non-SDM group

[0098] variable Non-SDM (n = 251) SDM (n = 42) P-value Red blood cell distribution width (fl), (median [IQR]) 43.30 [41.10, 46.65] 45.65 [41.32, 48.98] 0.144 Aspartate aminotransferase (U / L), (median [IQR]) 19.00 [14.00, 29.00] 16.50 [14.00, 27.25] 0.420 Alanine aminotransferase (U / L), (median [IQR]) 23.00 [14.00, 44.50] 20.50 [14.00, 38.75] 0.529 Glutamyl transpeptidase (U / L), (median [IQR]) 23.00 [14.00, 39.00] 31.50 [21.25, 50.50] 0.004 Urea nitrogen (mmol / L), (median [IQR]) 4.59 [3.79, 6.00] 6.78 [4.67, 8.73] <0.001 Serum creatinine (µmol / L), (median [IQR]) 54.00 [44.00, 63.50] 63.00 [49.75, 83.50] 0.002 Serum uric acid, (median [IQR]) 235.00 [182.00, 305.00] 293.00 [239.50, 370.75] <0.001 Uric acid to creatinine ratio, (median [IQR]) 4.23 [3.39, 5.40] 4.57 [3.75, 5.62] 0.352 Cystatin C (mg / L), (median [IQR]) 0.93 [0.78, 1.09] 1.17 [0.87, 1.40] <0.001 Total cholesterol (mmol / L), (median [IQR]) 4.12 [3.46, 4.94] 5.15 [4.05, 5.74] <0.001 Triglycerides (mmol / L), (median [IQR]) 1.20 [0.88, 1.64] 1.66 [1.14, 2.49] <0.001 High-density lipoprotein cholesterol (mmol / L), (median [IQR]) 1.31 [1.06, 1.65] 1.07 [0.83, 1.47] 0.005 Low-density lipoprotein cholesterol (mmol / L), (median [IQR]) 2.50 [2.03, 2.96] 2.94 [2.15, 3.48] 0.052 Serum magnesium (mmol / L), (median [IQR]) 0.93 [0.86, 0.96] 0.92 [0.84, 0.99] 0.607

Claims

1. A method for constructing a glucocorticoid-induced diabetes risk prediction model based on the LASSO algorithm, characterized in that: The following steps are involved: S1. Data collection and preprocessing: The historical clinical data and follow-up records of dermatology inpatients receiving systemic glucocorticoid treatment were read from the hospital's electronic medical record system. Whether the patients developed glucocorticoid-induced diabetes during treatment was recorded. The collected data were cleaned and preprocessed, including removing missing data and outliers. S2. Feature screening: The cleaned patient data were input into the machine learning algorithm, and candidate variables were screened by least absolute shrinkage and selection operator regression. The optimal lambda parameter was determined by K-fold cross-validation, the uncorrelated coefficient was compressed, and four key predictors highly correlated with glucocorticoid-induced diabetes were screened by LASSO regression. S3. Model construction: The key predictive factors screened in S2 are input into a logistic regression model to establish a glucocorticoid-induced diabetes risk prediction model, thereby obtaining a regression equation with the predictive factors as independent variables and the risk of glucocorticoid-induced diabetes as the dependent variable; S4, internal validation: Use cross-validation and bootstrap resampling methods to internally validate the model established in S3, evaluate the stability and generalization ability of the model under different sample partitions, and obtain the average performance index of the model; S5. Model evaluation: The model obtained in step S3 is evaluated for discrimination, calibration, and clinical benefit in sequence; when the evaluation results meet the preset performance standards, the model is marked as a performance-compliant model; S6. Result output: The performance compliance model is visualized as a nomogram. The corresponding scales and scores are drawn according to the regression coefficient ratios of each predictor in the logistic regression model. The nomogram is used to predict the risk of glucocorticoid-induced diabetes in individual patients, and the probability value or risk score is output for reference by clinicians in decision-making.

2. The method for constructing a glucocorticoid-induced diabetes risk prediction model based on the LASSO algorithm according to claim 1, characterized in that: The clinical data in S1 include: general demographic information, patient medical history and lifestyle habits, medication use, glucocorticoid treatment-related indicators, and laboratory test indicators.

3. The method for constructing a glucocorticoid-induced diabetes risk prediction model based on the LASSO algorithm according to claim 1, characterized in that: In step S2, four key predictors, including family history of diabetes, use of immunosuppressants, average daily glucocorticoid dose, and triglyceride level, were retained through LASSO regression screening, and their regression coefficients were non-zero, serving as input features for the subsequent multivariate logistic regression model.

4. The method for constructing a glucocorticoid-induced diabetes risk prediction model based on the LASSO algorithm according to claim 3, characterized in that: The specific mathematical expression of the logistic regression risk prediction model in S3 is: X=β0+β1X1+β2X2+β3X3+β4X4 Where P is the probability of a patient developing glucocorticoid-induced diabetes, X1–X4 are the variables corresponding to family history of diabetes, use of immunosuppressants, average daily GC dose, and triglyceride level, respectively, β0 is the intercept term, and β1–β4 are the corresponding regression coefficients.

5. The method for constructing a risk prediction model according to claim 4, characterized in that: The method for constructing the nomogram in S6 includes: S61. Determine the score scale of each predictor in the nomogram according to the regression coefficient of each predictor in the logistic regression model; S62. Add the scores of the four predictive factors for an individual patient to obtain a total score, and calculate the predicted probability of the patient developing glucocorticoid-induced diabetes based on the total score using the risk scale below the nomogram.

Images