Dynamic risk assessment method for senile osteoporosis based on multi-modal data and interaction quantification

Abstract

The invention provides a multi-modal data and interaction quantification-based senile osteoporosis dynamic risk assessment method. The method comprises the steps of screening effective variables in candidate risk factors; collecting structured data and unstructured data, and cleaning and standardizing the data to obtain multi-mode data of a patient; calculating a statistical weight and a clinical weight to obtain a multi-dimensional weight, calculating a dynamic total risk integral of an effective variable, and quantifying a variable interaction effect; and carrying out risk level layering and clinical intervention suggestion to obtain a dynamic risk assessment result, and then carrying out internal cross validation and external data set validation. According to the method, data full coverage, weight self-adaption and risk dynamic interaction analysis can be achieved, scientificity, sensitivity and clinical friendliness of osteoporosis screening are greatly improved, the method is low in cost and high in prediction precision, risk grading is associated with specific prevention and treatment measures, and a prevention and treatment integrated information closed loop is achieved.

Description

Technical Field

[0001] This invention relates to the field of bone risk assessment technology, and in particular to a dynamic risk assessment method for osteoporosis in the elderly based on multimodal data and interaction quantification. Background Technology

[0002] The diagnosis and risk assessment of osteoporosis have long relied on dual-energy X-ray absorptiometry (DXA) and quantitative CT (QCT). While these techniques are considered the gold standard, they have significant limitations: strong equipment dependence (limited to large medical institutions), radiation exposure risks (especially for elderly patients), and low availability (difficult to cover primary care hospitals). Traditional risk assessment tools such as OSTA and FRAX integrate a small number of clinical parameters through simplified linear models, possessing some universality, but their predictive efficacy is limited (AUC values ​​are generally below 0.8), and they cannot dynamically integrate multidimensional data (such as lifestyle habits, biochemical indicators, and medication history) and the interactions between quantitative variables. In recent years, AI models have attempted to combine imaging features (such as trabecular bone texture analysis), but limited by the processing capabilities of high-dimensional unstructured data and the complexity of manual feature extraction, the model's generalization and clinical applicability remain unsatisfactory. Existing technologies have the following limitations:

[0003] 1) Limited data dimensions and insufficient multimodal fusion: Existing AI models are mostly based on a single data source (clinical questionnaires or imaging data, etc.), lacking systematic integration of multimodal data (serum vitamin D levels, history of hormone drug use, calcium intake and other lifestyle indicators). This limitation makes it difficult for the model to comprehensively capture the multifactorial pathogenesis of osteoporosis, thus restricting the accuracy of risk stratification.

[0004] 2) Poor model generalization and neglect of variable interactions: Traditional machine learning methods rely on manual feature engineering, which makes it difficult to effectively handle high-dimensional unstructured data; in addition, existing tools do not quantify the synergistic effects between variables (such as the interaction between age and menopause time), resulting in systematic bias in risk prediction.

[0005] 3) Data imbalance and lack of validation procedures: The sample size of osteoporosis patients is much smaller than that of the normal population, and the data imbalance causes the model to be biased towards the majority class classification (high specificity but low sensitivity). At the same time, most models lack rigorous internal and external validation procedures, and their clinical application value is questionable.

[0006] 4) Lack of dynamic adjustment: Traditional methods rely on expert experience to assign weights without combining data-driven statistical weights and clinical evidence, resulting in insufficient scientific rigor and flexibility in the calculation of risk scores. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a dynamic risk assessment method for osteoporosis in the elderly based on multimodal data and interaction quantification. This method can achieve full data coverage, adaptive weighting, and dynamic interactive risk analysis, greatly improving the scientific rigor, sensitivity, and clinical friendliness of osteoporosis screening. Furthermore, this method is low-cost, highly accurate in prediction, and associates risk grading with specific prevention and control measures, thus achieving an integrated information loop for prevention and control.

[0008] To achieve the above objectives, the present invention provides the following solution: a dynamic risk assessment method for osteoporosis in the elderly based on multimodal data and interaction quantification, comprising:

[0009] Fourteen candidate risk factors were screened, and the stepwise discriminant method was used to further screen the candidate risk factors to obtain effective variables.

[0010] Structured and unstructured data of patients are collected in the hospital's HIS-LIS system. The structured and unstructured data are then cleaned and standardized to obtain multi-modal patient data.

[0011] Calculate the statistical and clinical weights of the effective variables to obtain multidimensional weights. Based on the multidimensional weights and the patient multimodal data, calculate the dynamic total risk score of the effective variables and quantify the interaction of variables.

[0012] Risk levels are stratified and clinical intervention recommendations are made based on the dynamic total risk score to obtain dynamic risk assessment results. The dynamic risk assessment results are then subjected to internal cross-validation and external dataset validation to optimize the dynamic risk assessment process.

[0013] Optionally, 14 candidate risk factors are screened, and the candidate risk factors are further screened using a stepwise discriminant method to obtain effective variables, including:

[0014] Based on meta-analysis, cohort studies, and cross-sectional studies, and with reference to domestic and international osteoporosis diagnosis and treatment guidelines, 14 candidate risk factors were screened from the PubMed, Embase, and CNKI databases.

[0015] Using forward selection combined with Wilks ， The Lambda method and F-test are used to evaluate the candidate risk factors one by one to obtain the current set. Then, the backward elimination method is used to remove redundant variables in the current set.

[0016] Repeat the forward selection method and the backward elimination method until no redundant variables are eliminated, and obtain the effective variables.

[0017] Optionally, the 14 candidate risk factors include age, day exposure time, serum calcium concentration, family history, calcium intake, time to menopause, history of fragility fracture, sex, BMI, bone mineral density, serum vitamin D3, regular exercise, smoking, and alcohol consumption; the effective variables include age, BMI, serum vitamin D3, time to menopause, history of hormone drug use, and calcium intake.

[0018] Optionally, structured and unstructured patient data are collected from the hospital's HIS-LIS system. The structured and unstructured data are then cleaned and standardized to obtain multi-modal patient data, including:

[0019] The hospital's HIS-LIS system interface is used to collect patients' basic information, past medical history, medical orders and laboratory test results in real time. NLP methods are used to parse the electronic medical record text and extract key information and disease cycle from chief complaint, personal history and menstrual history to obtain patient information.

[0020] Based on the patient information, missing, outlier, and duplicate values ​​are cleaned, and data is completed using the mean, median, or predictive modeling to obtain the first data.

[0021] Based on the effective variables, extract the corresponding fields and records from the first data, convert text-type numbers into numerical types, convert categorical variables into dummy variables, complete the data extraction operation, and obtain the second data;

[0022] The second data is processed using preset missing value handling rules and logical verification mechanisms to obtain patient multimodal data.

[0023] Optionally, the missing value handling rules include a first rule for defaulting to age 50 when age cannot be obtained, a second rule for defaulting to BMI 24 when height or weight cannot be obtained, a third rule for defaulting to serum vitamin D3 35 ng / mL when serum vitamin D3 cannot be obtained, a fourth rule for defaulting to premenopausal status when female menopause time cannot be obtained, a fifth rule for defaulting to premenopausal status when hormone medication history cannot be obtained, and a sixth rule for defaulting to calcium intake less than 1000 mg / d when calcium intake cannot be obtained.

[0024] The logic verification mechanism includes a contradiction handling mechanism for determining whether a patient is under 40 years old and has been menopausal for more than 10 years; if so, a manual review is performed. It also includes an outlier removal mechanism for determining whether a patient has a BMI of less than 10 or greater than 40, serum vitamin D3 greater than 150 ng / mL, and calcium intake greater than 10,000 mg / d; if so, a manual review is performed.

[0025] Optionally, the statistical weights and clinical weights of the effective variables are calculated to obtain multi-dimensional weights, including:

[0026] The information entropy of the effective variables is calculated using the modified entropy weight method to obtain the statistical weights of the effective variables; the expression for calculating the statistical weights is as follows:

[0027] ;

[0028] in, The information entropy weight for each variable,

[0029] Based on literature and consensus, the effective variables were weighted to obtain initial clinical weights, which were then standardized to obtain the final clinical weights. The expression for calculating the clinical weights is as follows:

[0030] ;

[0031] in, Clinical weights for each variable, As the initial clinical weight, For all initial clinical weights.

[0032] Optionally, based on the multidimensional weights and the patient multimodal data, the dynamic total risk score of the effective variables is calculated, and the interaction effects of the variables are quantified, including:

[0033] Based on epidemiological studies and clinical consensus, the multidimensional weights, and the patient multimodal data, variable grading thresholds and corresponding variable grading scores are set. Then, based on the variable grading scores, the basic integral of the effective variable is calculated. The expression for calculating the basic integral is:

[0034] ;

[0035] in, Rank and score the variables;

[0036] The preset synergy threshold is 2.58. Using logistic regression, the univariate OR and joint OR values ​​of the effective variables are calculated to obtain the interaction integral. When the interaction integral is greater than 2.58, a significant synergy exists. A synergy correction term is calculated when both variables have a rank score greater than or equal to 2. When the interaction integral is less than 2.58, no synergy exists. The expression for the synergy correction term is:

[0037] , ;

[0038] in, The integral of the interaction;

[0039] The dynamic total risk integral is calculated based on the basic integral and the synergistic effect correction term; the expression for the dynamic total risk integral is as follows:

[0040] .

[0041] Optionally, the risk level stratification includes low risk, medium risk, high risk level 1, and high risk level 2.

[0042] This invention discloses the following technical effects by providing a dynamic risk assessment method for osteoporosis in the elderly based on multimodal data and interaction quantification:

[0043] 1) Multimodal Integration of Core Variables: Traditional scoring tools rely solely on basic patient data, with limited parameter dimensions, making it difficult to comprehensively capture the multifactorial pathogenesis of osteoporosis. This invention uses a stepwise discriminant analysis to screen for six core risk factors (age, BMI, serum vitamin D3, menopausal time, history of hormone medication use, and calcium intake), covering multimodal data from clinical, biochemical, and lifestyle factors, significantly improving predictive efficacy, especially increasing the sensitivity for identifying early-stage osteoporosis patients to 85%.

[0044] 2) Dynamic Weighting: Existing methods typically rely on fixed weights, such as the expert-based weighting in the FRAX model, without incorporating data-driven statistical analysis. This invention introduces an improved entropy weighting method to calculate the information entropy of each variable and combines it with weighting based on clinical literature. Through this dual-weighting mechanism, the model reduces subjective bias while meeting actual clinical needs, thereby significantly improving the clarity of weight interpretation.

[0045] 3) Quantitative Correction for Interaction Effects: Traditional tools often overlook the synergistic effects between variables. For example, the longer the postmenopausal period in female patients, the older the corresponding age. When there is a linear relationship between two variables, the incidence of osteoporosis in these patients will significantly increase. To more accurately assess this interaction, this invention introduces a dynamic risk integral, identifies significant synergistic variables through logistic regression, and triggers a correction term only in the intermediate-to-high-risk stratification, thereby more precisely quantifying the impact of the interaction between variables on disease risk.

[0046] 4) Rigorous Validation Process: To address the shortcomings of traditional model validation, this model employs cross-validation and multi-center external validation, and its superiority is verified by comparing it with the FRAX tool. Furthermore, the Hosmer-Lemeshow test demonstrates a high degree of consistency between the predicted risk and the actual risk, with the calibration curve slope approaching 1.0, significantly outperforming existing methods.

[0047] 5) Low cost: Compared to relying on expensive equipment such as DXA and QCT (cost per test > 500 RMB), this model automatically collects data through the Hospital Information System (HIS) and combines NLP technology to analyze unstructured medical records (such as hormone drug usage records), achieving non-invasive screening with zero equipment dependence and zero radiation risk. Primary healthcare institutions only need basic medical information to complete the assessment, significantly improving screening efficiency.

[0048] 6) Tiered Decision-Making Closed Loop: Traditional risk assessment tools only output probability values ​​and lack direct intervention guidance. This model maps the total risk score to a four-level stratification (low risk, intermediate risk, high risk level 1, and high risk level 2) and associates it with specific clinical measures (such as high risk level 2 requiring hospitalization) to ensure timely and effective intervention.

[0049] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention;

[0052] Figure 2 A detailed flowchart illustrating the dynamic risk assessment method for osteoporosis in the elderly provided in this embodiment of the invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] Example

[0056] like Figure 1 , Figure 2 As shown, this invention provides a method for dynamic risk assessment of osteoporosis in the elderly based on multimodal data and interaction quantification, including:

[0057] 1. Fourteen candidate risk factors were screened, and the stepwise discriminant method was used to further screen these factors to obtain effective variables. Specifically, this included:

[0058] 1.1 Based on meta-analysis, cohort studies, and cross-sectional studies, and with reference to domestic and international osteoporosis diagnosis and treatment guidelines, 14 candidate risk factors were screened from the PubMed, Embase, and CNKI databases.

[0059] Candidate variables were selected based on a systematic literature review, covering authoritative international journals and clinical guidelines from the past decade, such as the "Chinese Guidelines for the Diagnosis and Treatment of Primary Osteoporosis (2022 Edition)," and were screened through the following process:

[0060] Literature search: A literature search was conducted in PubMed and Embase databases using "osteoporosis risk factors" as the keyword. Simultaneously, the same keywords were used to search CNKI, Wanfang, and the Chinese Medical Database. The search period was approximately 10 years, and the included literature types were meta-analyses, cohort studies, and cross-sectional studies. Following the PRISMA protocol, duplicate, non-clinical studies, and low-quality literature (Newcastle-Ottawa Scale score <6) were excluded, ultimately retaining 48 high-quality studies.

[0061] Variable extraction: Risk factors consistently reported in ≥3 studies and with an odds ratio (OR) > 1.5 were extracted to form an initial candidate variable pool. Specifically, the 14 candidate risk factors included: age, duration of sunlight exposure, serum calcium concentration, family history, calcium intake, time to menopause, history of fragility fractures, sex, BMI, bone mineral density, serum vitamin D3, regular exercise, smoking, and alcohol consumption.

[0062] 1.2 Using forward selection combined with Wilks ， The Lambda method and F-test are used to evaluate the candidate risk factors one by one to obtain the current set. Then, the backward elimination method is used to remove redundant variables from the current set.

[0063] Assume the initial model is an empty set. Unselected variable set

[0064] Forward selection: Univariate discriminant ability assessment: for each variable Calculate its Wilks' Lambda:

[0065] ;

[0066] in:

[0067] ;

[0068] ;

[0069] The mean of the group. This is the population mean. Choose to make... Find the smallest variable X* and calculate its partial F-statistic:

[0070] ;

[0071] in, This represents the number of variables in the current model. If F... enter >F critical (1, If X* is added to model S, then X* will be added to model S.

[0072] Backward elimination: For each variable in the current model Calculate the results after removal j, and calculate:

[0073] ;

[0074] If F remove <F enter (1, If ), then remove. .

[0075] 1.3 Repeat the forward selection method and the backward elimination method until no redundant variables are eliminated or can be added, to obtain the effective variables S={X1, X2, X3, X4, X5, X6}.

[0076] X1 = Age;

[0077] X2 = BMI;

[0078] X3 = Serum Vitamin D3;

[0079] X4 = menopausal duration;

[0080] X5 = History of hormone medication use;

[0081] X6 = Calcium intake.

[0082] 2. Collect structured and unstructured patient data from the hospital's HIS-LIS system. Clean and standardize the structured and unstructured data to obtain multi-modal patient data. Specifically, this includes:

[0083] 2.1 The hospital's HIS-LIS system interface is used to collect patients' basic information, medical history, medical orders and laboratory test results in real time. NLP methods are used to parse the electronic medical record text and extract key information and disease cycle from chief complaint, personal history and menstrual history to obtain patient information.

[0084] 2.2 Based on the patient information, missing values, outliers, and duplicate values ​​are cleaned, and data is completed using the mean, median, or predictive modeling to obtain the first data.

[0085] 2.3 Based on the effective variables, extract the corresponding fields and records from the first data, convert text-type numbers into numerical types, convert categorical variables into dummy variables, complete the data extraction operation, and obtain the second data.

[0086] 2.4 Using preset missing value handling rules and logical verification mechanisms, the second data is processed to obtain patient multimodal data.

[0087] The missing value handling rules include:

[0088] First rule: When age cannot be obtained, the default age is 50 years old, X1 = 1 point;

[0089] Second rule: If height or weight cannot be obtained, the default BMI is 24, x2 = 1 point;

[0090] Third rule: If serum vitamin D3 cannot be obtained, the patient's serum vitamin D3 is assumed to be 35 ng / mL, X3 = 1 point;

[0091] Fourth rule: If the time of menopause for a woman cannot be obtained, the female patient is assumed to be premenopausal. Female patients x4 = 1 point, male patients x4 = 0 points.

[0092] Fifth rule: If a history of hormone medication use cannot be obtained, it is assumed that the patient has not used hormone medication, X5 = 1 point;

[0093] Rule 6: If calcium intake cannot be obtained, the patient's calcium intake is assumed to be less than 1000 mg / day, X6 = 2 points.

[0094] The logical verification mechanism includes:

[0095] Conflict handling mechanism: used to determine whether the patient is under 40 years old and has been menopausal for more than 10 years. If so, manual review is performed, prompting the doctor to manually confirm or modify the information.

[0096] Outlier removal mechanism: Determine if the patient has a BMI less than 10 or greater than 40, serum vitamin D3 greater than 150 ng / mL, or calcium intake greater than 10000 mg / d. If so, a manual review is performed, prompting the doctor to manually confirm or modify the data.

[0097] 3. Calculate the statistical weight and clinical weight of the effective variables to obtain multi-dimensional weights. Based on the multi-dimensional weights and the patient multi-model data, calculate the dynamic total risk score of the effective variables and quantify the interaction of variables.

[0098] 3.1 Calculate multi-dimensional weights, including:

[0099] The information entropy of the effective variables is calculated using the modified entropy weight method to obtain the statistical weights of the effective variables; the expression for calculating the statistical weights is as follows:

[0100] ;

[0101] in, The information entropy weights for each variable are determined by information entropy, which measures the disorder of the data distribution; the smaller the entropy value, the higher the weight. ) / ( ) is a standardized value (used to eliminate dimensional differences and compress each variable into the [0,1] interval).

[0102] The information entropy and statistical weight W of each factor were calculated based on real-world data. s The values ​​are shown in Table 1 below:

[0103] Table 1. Factors - Information Entropy - Statistical Weights

[0104]

[0105] Based on literature and consensus, the effective variables were weighted to obtain the initial clinical weights, as shown in Table 2 below:

[0106] Table 2 Initial Clinical Weighting Table

[0107]

[0108] The initial clinical weights are then standardized to obtain the clinical weights; the calculation expression for the clinical weights is:

[0109] ;

[0110] in, Clinical weights for each variable, As the initial clinical weight, For all initial clinical weights.

[0111] 3.2 Calculate the dynamic total risk integral and quantify the interaction of variables, including:

[0112] Based on epidemiological studies and clinical consensus, the multidimensional weights, and the patient multimodal data, variable grading thresholds and corresponding variable grading scores were set, as shown in Table 3 below:

[0113] Table 3. Variable grading thresholds and corresponding variable grading scores

[0114]

[0115] Then, based on the variable grading threshold and the variable grading score, the basic integral of the effective variable is calculated; the expression for calculating the basic integral is:

[0116]

[0117] in, Assign scores to the corresponding variables; .

[0118] Example: A patient is 80 years old (X1) and is at high risk (3 points). W s1 =0.24, W c1 =0.17, the fundamental integral S of X1 base1 =0.24×0.17×3=0.1224.

[0119] With a preset synergy threshold of 2.58, logistic regression was used to calculate the univariate OR and joint OR values ​​of the effective variables, yielding the interaction integral:

[0120] ;

[0121] A significant synergistic effect exists when the interaction integral is greater than 2.58. A synergistic effect correction term is calculated when both variables have a rank score greater than or equal to 2. No synergistic effect exists when the interaction integral is less than 2.58. inter =0; the threshold of 2.58 is a strict cutoff value for the standard normal distribution at the significance level of α=0.01.

[0122] The calculation expression for the synergistic effect correction term is as follows:

[0123] ,

[0124] in, For the interaction integral; when Synergistic effects are only considered when both variables score at or above 2 (medium risk).

[0125] The dynamic total risk integral is calculated based on the basic integral and the synergistic effect correction term; the expression for the dynamic total risk integral is as follows:

[0126] .

[0127] Example: Suppose that the patient's age (X1) and menopausal time (X4) are both high-risk (3 points), and that they influence each other. Assume that logistic regression yields... =3.1, S inter =3.1×min(3,3)=3.1×3=9.3.

[0128] 4. Based on the dynamic total risk score, risk level stratification and clinical intervention recommendations are made to obtain dynamic risk assessment results. Then, the dynamic risk assessment results are subjected to internal cross-validation and external dataset validation to optimize the dynamic risk assessment process.

[0129] The risk level stratification includes:

[0130] Low risk: S total ≤0.4;

[0131] Intermediate risk: 0.4 < S total ≤1.2;

[0132] High risk: S total >1.2;

[0133] High-risk level one: S total ϵ (1.2, 1.6];

[0134] High risk level 2: S total >1.6;

[0135] Clinical intervention recommendations are shown in Table 4 below:

[0136] Table 4 Clinical Intervention Recommendations

[0137]

[0138] Example 1

[0139] I. Model Validation and Performance Evaluation

[0140] 1.1 Data Preparation and Processing

[0141] Dataset Description: This dataset contains 2591 orthopedic inpatient medical records from three tertiary hospitals between 2018 and 2023 (1295 cases in the disease group and 1296 cases in the non-disease group). The patients were aged 45-78 years, with 52.6% (1364 cases) being male and 47.4% (1227 cases) being female. Gold Standard: All samples were diagnosed by dual-energy X-ray absorptiometry (DXA) (T-score ≤ -2.5 was considered diseased).

[0142] Data preprocessing:

[0143] Missing values ​​were handled: Vitamin D3 (missing rate 3.2%) was filled using multiple imputation.

[0144] Data balancing: To address the issue of a slightly low sample size in the disease group, SMOTE (synthetic minority oversampling) was used to bring the ratio of the two types of samples closer to 1:1.

[0145] Variable standardization: Z-score standardization is performed on continuous variables such as age and BMI to ensure model convergence stability.

[0146] 1.2 Data Partitioning and Model Training

[0147] Partitioning strategy:

[0148] Training set (70%): 1814 cases (907 cases in the disease group and 907 cases in the non-disease group), used for weight calculation and threshold optimization.

[0149] Validation set (30%): 777 cases (388 cases in the disease group and 389 cases in the non-disease group), independently assessing the model's generalization.

[0150] Stratified random sampling: Stratify by gender and age (±5 years) to avoid selection bias.

[0151] 1.3 Performance Evaluation Indicators and Standards

[0152] Discrimination:

[0153] AUC-ROC curve: Target AUC > 0.85 (better than FRAX's 0.75).

[0154] Sensitivity: Among all patients diagnosed with the disease, the model correctly classifies ≥85% of them as high-risk.

[0155] Specificity: Among all unaffected individuals, the proportion of those correctly classified as non-high-risk (low-risk + intermediate-risk) by the model is ≥75%.

[0156] Calibration:

[0157] Hosmer-Lemeshow test: requires chi-square value < 15.5 (degrees of freedom = 8, p > 0.05).

[0158] Calibration curve: Brier score < 0.15 indicates that the predicted probability is highly consistent with the actual risk.

[0159] Classification accuracy:

[0160] Accuracy: The clinical applicability of the model is assessed through decision curve analysis (DCA).

[0161] F1 score: A target value ≥ 0.80, balancing sensitivity and specificity. (Based on a classification strategy of high risk = positive)

[0162] 1.4 Internal Validation (Cross-Validation)

[0163] K-fold cross-validation (K=5):

[0164] The training set was randomly divided into 5 subsets (363 cases per fold) to ensure that the proportion of sick / non-sick cases was consistent in each fold.

[0165] The model was repeatedly trained and the metrics were recorded, with one subset used as the validation set and the remaining four subsets used as the training set.

[0166] Calculated average performance: AUC: 0.84 (95% CI: 0.81–0.87), sensitivity: 87.2 ± 2.1%, specificity: 79.5 ± 1.8%.

[0167] 1.5 External Validation

[0168] Dataset source: The external dataset comes from 3 tertiary hospitals (a total of 1024 cases), and the variable definitions are consistent with the training set.

[0169] Validation results compared with traditional tools: Since the sensitivity and specificity of the FRAX model vary depending on fracture type, cut-off value, and study population, the comparison conditions here are: the sensitivity and specificity of FRAX in predicting major osteoporotic fractures (MOFs) with a cut-off value of 3% when bone mineral density test results are unavailable, as shown in Table 5 below:

[0170] Table 5 External Validation Results

[0171] index training set Validation set FRAX AUC 0.83 0.86 0.6-0.63 Sensitivity 88.36% 88.2% 57%-85% Specificity 78.55% 79.19% 34%-79%

[0172] 1.6 Conclusion

[0173] The model has an AUC of >0.8 on both the training and validation sets, which is significantly better than the traditional tool FRAX, and the calibration is good (p>0.05).

[0174] II. System Evaluation Steps

[0175] 1. Read and preprocess patient information.

[0176] 1.1 Collecting Patient Information

[0177] The real-time patient diagnosis and treatment information to be obtained through the interface from the Hospital Information System (HIS) and Laboratory Information Management System (LIS) is shown in Table 6 below:

[0178] Table 6. Sources of Patient Information

[0179]

[0180] 1.2 Unstructured Data Parsing

[0181] The pre-trained language model BERT was used to parse electronic medical record text, extracting key information (such as menopausal time and disease stage) and other structured data from free text such as chief complaint, past medical history, personal history, and menstrual history in the electronic medical record admission records. Patients diagnosed with "osteoporosis" were excluded.

[0182] 1.3 Handling Missing Values

[0183] Age: If the age cannot be obtained, the default age is 50 years old, X1=1 point.

[0184] BMI: If the patient's height or weight cannot be obtained, the default BMI is 24, x2 = 1 point.

[0185] Serum vitamin D3: If unavailable, the default value is 35 ng / mL, X3 = 1 point.

[0186] Menopausal time: If unavailable, the patient is assumed to be premenopausal. Female patients x4 = 1 point, male patients x4 = 0 points.

[0187] History of hormone medication use: If this information is unavailable, it is assumed that the patient has not used hormone medications, X5=1 point.

[0188] Calcium intake: If not available, the patient’s calcium intake is assumed to be ≤1000mg / d, X6=2 points.

[0189] 1.4 Logical Verification

[0190] Handling contradictory values: If the patient's age X1 is less than 40 years but the time since menopause is greater than 10 years, the manual review process is triggered, prompting the doctor to manually confirm or modify the value.

[0191] Outlier removal: When BMI < 10 or > 40, serum vitamin D3 > 150 ng / mL, or calcium intake > 10000 mg / d, the manual review process is triggered, prompting the doctor to manually confirm or modify the data.

[0192] 1.5 Real-world case studies

[0193] Assuming a female patient, the specific indicators are shown in Table 7 below:

[0194] Table 7 Comprehensive Risk Assessment Table

[0195]

[0196] 2. Calculate basic integrals and determine...

[0197] 2.1 Determine the rank score X of each factor based on the parsed electronic medical record text. grade ; =3, , =2, =3, =3, =3.

[0198] 2.2 According to formula S base =W c (Clinical weighting score) × W s (Statistical score) × X grade (Ranking); S of each risk factor base =0.2049+0.0654+0.0480+0.0993+0.0221+0.0726=0.5123.

[0199] 2.3 Calculate λ using Logistic Regression jk Value, if λ jk If λ > 2.58, proceed to the next step of the calculation. jk If the integral is ≤2.58, then the interaction integral is directly denoted as S. inter =0.

[0200] The six core variables were paired, resulting in C(6,2) = 15 pairs. Pairs with λ > 2.58 were retained, and insignificant terms were removed. The results are shown in Table 8 below.

[0201] Table 8 Variable Screening Table

[0202]

[0203] Age (X1) and time to menopause (X4): derived by logistic regression. =3.1>2.58, correction term S inter14 = ×min(3,3)=3.1×3=9.3;

[0204] BMI (X2) and calcium intake (X6): derived from logistic regression. =2.9>2.58, correction term S inter26 = ×min(3,3) = 2.9 × 3 = 8.7;

[0205] Serum vitamin D3 (X3) and hormones (X5): derived by logistic regression. =2.6>2.58, correction term S inter35 = ×min(2,3)=2.6×2=5.2.

[0206] 3. Calculate the interaction integral and the total score.

[0207] 3.1 Calculate the interaction integral of each pair of variables according to the formula.

[0208] , ( )

[0209] Various S inter =9.3+8.7+5.2=23.2.

[0210] 3.2 Calculate the patient's total score according to the formula.

[0211]

[0212] Denominator Adjustment Item =3.1=2.9+2.6=8.6

[0213] S total = = ≈2.470

[0214] 4. Assess the risk level and output the results.

[0215] Based on the patient's total score S total The results, along with suggested intervention measures, are displayed in Table 9 below:

[0216] Table 9. Comparison of Intervention Measures

[0217]

[0218] S total ≈2.470 → High Risk Level II (Fracture Risk > 30%)

[0219] Recommended measures: inpatient intervention, multidisciplinary management (orthopedics, endocrinology, nutrition), and early prevention of fractures.

[0220] Therefore, this invention provides a dynamic risk assessment method for osteoporosis in the elderly based on multimodal data and interaction quantification. This method can achieve full data coverage, adaptive weighting, and dynamic interactive risk analysis, greatly improving the scientific rigor, sensitivity, and clinical friendliness of osteoporosis screening. Furthermore, this method is low-cost, highly accurate in prediction, and associates risk grading with specific prevention and control measures, thus achieving an integrated information loop for prevention and control.

[0221] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0222] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A dynamic risk assessment method for osteoporosis in the elderly based on multimodal data and interaction quantification, characterized in that, include: Fourteen candidate risk factors were screened, and the stepwise discriminant method was used to further screen the candidate risk factors to obtain effective variables. Structured and unstructured data of patients are collected in the hospital's HIS-LIS system. The structured and unstructured data are then cleaned and standardized to obtain multi-modal patient data. Calculate the statistical and clinical weights of the effective variables to obtain multidimensional weights. Based on the multidimensional weights and the patient multimodal data, calculate the dynamic total risk score of the effective variables and quantify the interaction of variables. Risk levels are stratified and clinical intervention recommendations are made based on the dynamic total risk score to obtain dynamic risk assessment results. The dynamic risk assessment results are then subjected to internal cross-validation and external dataset validation to optimize the dynamic risk assessment process.

2. The method for dynamic risk assessment of osteoporosis in the elderly based on multimodal data and interaction quantification as described in claim 1, characterized in that, Fourteen candidate risk factors were screened, and the stepwise discriminant method was used to further screen these factors to obtain effective variables, including: Based on meta-analysis, cohort studies, and cross-sectional studies, and with reference to domestic and international osteoporosis diagnosis and treatment guidelines, 14 candidate risk factors were screened from the PubMed, Embase, and CNKI databases. Using forward selection combined with Wilks ， The Lambda method and F-test are used to evaluate the candidate risk factors one by one to obtain the current set. Then, the backward elimination method is used to remove redundant variables in the current set. Repeat the forward selection method and the backward elimination method until no redundant variables are eliminated, and obtain the effective variables.

3. The method for dynamic risk assessment of osteoporosis in the elderly based on multimodal data and interaction quantification as described in claim 2, characterized in that, The 14 candidate risk factors include age, day exposure time, serum calcium concentration, family history, calcium intake, time to menopause, history of fragility fracture, sex, BMI, bone mineral density, serum vitamin D3, regular exercise, smoking, and alcohol consumption; the effective variables include age, BMI, serum vitamin D3, time to menopause, history of hormone drug use, and calcium intake.

4. The method for dynamic risk assessment of osteoporosis in the elderly based on multimodal data and interaction quantification as described in claim 3, characterized in that, Structured and unstructured data of patients are collected in the hospital's HIS-LIS system. The structured and unstructured data are then cleaned and standardized to obtain multi-modal patient data, including: The hospital's HIS-LIS system interface is used to collect patients' basic information, past medical history, medical orders and laboratory test results in real time. NLP methods are used to parse the electronic medical record text and extract key information and disease cycle from chief complaint, personal history and menstrual history to obtain patient information. Based on the patient information, missing, outlier, and duplicate values ​​are cleaned, and data is completed using the mean, median, or predictive modeling to obtain the first data. Based on the effective variables, extract the corresponding fields and records from the first data, convert text-type numbers into numerical types, convert categorical variables into dummy variables, complete the data extraction operation, and obtain the second data; The second data is processed using preset missing value handling rules and logical verification mechanisms to obtain patient multimodal data.

5. The method for dynamic risk assessment of osteoporosis in the elderly based on multimodal data and interaction quantification as described in claim 4, characterized in that: The missing value handling rules include: a first rule for defaulting to age 50 when age cannot be obtained; a second rule for defaulting to BMI 24 when height or weight cannot be obtained; a third rule for defaulting to serum vitamin D3 35 ng / mL when serum vitamin D3 cannot be obtained; a fourth rule for defaulting to premenopausal status when female menopause time cannot be obtained; a fifth rule for defaulting to premenopausal status when hormone medication history cannot be obtained; and a sixth rule for defaulting to calcium intake less than 1000 mg / d when calcium intake cannot be obtained. The logic verification mechanism includes a contradiction handling mechanism for determining whether a patient is under 40 years old and has been menopausal for more than 10 years; if so, a manual review is performed. It also includes an outlier removal mechanism for determining whether a patient has a BMI of less than 10 or greater than 40, serum vitamin D3 greater than 150 ng / mL, and calcium intake greater than 10,000 mg / d; if so, a manual review is performed.

6. The method for dynamic risk assessment of osteoporosis in the elderly based on multimodal data and interaction quantification as described in claim 5, characterized in that, Calculate the statistical and clinical weights of the effective variables to obtain multidimensional weights, including: The information entropy of the effective variables is calculated using the modified entropy weight method to obtain the statistical weights of the effective variables; the expression for calculating the statistical weights is as follows: ； in, The information entropy weight for each variable, Based on literature and consensus, the effective variables were weighted to obtain initial clinical weights, which were then standardized to obtain the final clinical weights. The expression for calculating the clinical weights is as follows: ； in, Clinical weights for each variable, As the initial clinical weight, For all initial clinical weights.

7. The method for dynamic risk assessment of osteoporosis in the elderly based on multimodal data and interaction quantification as described in claim 6, characterized in that, Based on the multidimensional weights and the patient multimodal data, the dynamic total risk integral of the effective variables is calculated, and the interaction effects of the variables are quantified, including: Based on epidemiological studies and clinical consensus, the multidimensional weights, and the patient multimodal data, variable grading thresholds and corresponding variable grading scores are set. Then, based on the variable grading scores, the basic integral of the effective variable is calculated. The expression for calculating the basic integral is: ； in, Rank and score the variables; The preset synergy threshold is 2.

58. Using logistic regression, the univariate OR and joint OR values ​​of the effective variables are calculated to obtain the interaction integral. When the interaction integral is greater than 2.58, a significant synergy exists. A synergy correction term is calculated when both variables have a rank score greater than or equal to 2. When the interaction integral is less than 2.58, no synergy exists. The expression for the synergy correction term is: ， ； in, The integral of the interaction; The dynamic total risk integral is calculated based on the basic integral and the synergistic effect correction term; the expression for the dynamic total risk integral is as follows: 。 8. The method for dynamic risk assessment of osteoporosis in the elderly based on multimodal data and interaction quantification according to claim 7, characterized in that, The risk level stratification includes low risk, medium risk, high risk level 1, and high risk level 2.

Images