Real-time risk early warning and intervention method and system for radioactive oral mucositis based on multi-parameter dynamic monitoring

Abstract

The present application relates to the technical field of early warning and intervention of radioactive oral mucositis, and discloses a radioactive oral mucositis real-time risk early warning and intervention method and system based on multi-parameter dynamic monitoring, comprising the following steps: S1: a multi-parameter dynamic acquisition module dynamically acquires patient-related parameters; S2: a data processing module pre-processes the collected related data; S3: the processed data is used to calculate the probability of the patient developing severe radioactive oral mucositis by a dynamic joint prediction model in a model calculation module, and a corresponding risk level is generated; S4: a risk early warning push module generates corresponding early warning information and sends the early warning information to a medical care end, and sends corresponding personalized intervention prompts to a patient end; S5: an intervention feedback module collects feedback data, and S1-S4 are repeated to perform real-time risk early warning and intervention on the patient developing severe radioactive oral mucositis. The present application can improve prediction accuracy, intervene in a timely manner, and reduce the occurrence of severe radioactive oral mucositis.

Description

Technical Field

[0001] This invention relates to the field of radiation-induced oral mucositis early warning and intervention technology, specifically to a method and system for real-time risk early warning and intervention of radiation-induced oral mucositis based on multi-parameter dynamic monitoring. Background Technology

[0002] Radiation-induced oral mucositis (RIOM) is one of the most common and serious complications in cancer patients undergoing radiotherapy. Severe RIOM can cause oral / pharyngeal pain, difficulty eating, interruption of radiotherapy, and even endanger life.

[0003] Currently, clinical risk assessment for radiation-induced oral mucositis mainly relies on traditional predictive models, but these models have significant limitations: First, they are limited by static assessment. Existing models are mostly built based on baseline data (such as pre-radiotherapy demographics and tumor stage), without considering the dynamic changes in risk factors during radiotherapy (such as real-time fluctuations in anxiety levels, nutritional status, and blood routine indicators), leading to inaccurate prediction results. Second, they have limited parameters. Most current models only include physiological or treatment-related data (such as radiotherapy dose), ignoring the synergistic effects of key multidimensional factors such as the patient's psychological state (such as anxiety and depression) and the oral microenvironment (such as pH value), resulting in inaccurate prediction results. Third, they have delayed warnings. Traditional models cannot update risk assessment results in real time, making it difficult to trigger timely intervention before radiation-induced oral mucositis occurs, leading to delays in clinical intervention. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, one of the objectives of this invention is to provide a real-time risk warning and intervention method for radiation-induced oral mucositis based on multi-parameter dynamic monitoring. This method enables multi-parameter dynamic monitoring, improves prediction accuracy, allows for real-time dynamic monitoring, guides clinical intervention, and enables timely intervention to reduce the occurrence of severe radiation-induced oral mucositis, reduce adverse events such as treatment interruption, increased pain, and malnutrition, and improve patients' quality of life.

[0005] The technical solution adopted in this invention is as follows: a method for real-time risk warning and intervention of radiation-induced oral mucositis based on multi-parameter dynamic monitoring, comprising the following steps:

[0006] S1: The multi-parameter dynamic acquisition module collects patients' sociodemographic data, health-related data, physiological-related data, disease-related data, treatment-related data, and psychological and nutritional-related data in real time through the electronic medical record interface, medical staff terminal, and patient terminal.

[0007] S2: The data processing module preprocesses the collected patient sociodemographic data, health-related data, physiological-related data, disease-related data, treatment-related data, and psychological and nutritional-related data.

[0008] S3: The processed data is input into the model calculation module, and the dynamic joint prediction model, which includes a linear mixed-effects longitudinal sub-model and a Cox regression survival sub-model, and realizes data association and collaborative modeling through the posterior distribution of the random effects of the samples, is used to predict the probability of patients developing severe radiation-induced oral mucositis in the future and generate the corresponding risk level.

[0009] S4: The risk warning push module generates corresponding warning information based on the risk level predicted by the model calculation module, and sends the warning information to the medical staff terminal and corresponding personalized intervention prompts to the patient terminal;

[0010] S5: The intervention feedback module collects feedback data after intervention through the medical staff and patient terminals. The feedback data is stored in the storage module for later query by medical staff. Repeating S1-S4, it provides real-time risk warning and intervention for patients with severe radiation-induced oral mucositis.

[0011] The second objective of this invention is to provide a real-time risk warning and intervention system for radiation-induced oral mucositis based on multi-parameter dynamic monitoring. The system employs a real-time risk warning and intervention method for radiation-induced oral mucositis based on multi-parameter dynamic monitoring, including a multi-parameter dynamic acquisition module, a data processing module, a model calculation module, a risk warning push module, an intervention feedback module, and a storage module.

[0012] The multi-parameter dynamic acquisition module is used to collect patients' sociodemographic data, health-related data, physiological-related data, disease-related data, treatment-related data, and psychological and nutritional-related data in real time through the electronic medical record interface, medical staff terminal, and patient terminal.

[0013] The data processing module is used to preprocess the collected parameters, and the model calculation module is used to dynamically predict the probability of a patient developing severe radiation-induced oral mucositis within the next 1 to 3 months based on the preprocessed parameters, and generate the corresponding risk level.

[0014] The risk warning push module is used to generate corresponding warning information based on the risk level predicted by the model calculation module, and send the warning information to the medical staff terminal and corresponding personalized intervention prompts to the patient terminal;

[0015] The intervention feedback module is used to collect feedback data after intervention through medical staff and patients, and the storage module is used to store the feedback data for easy retrieval by medical staff later.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] This invention covers all aspects of disease progression through multi-dimensional dynamic data and combines dynamic model algorithms to more accurately identify high-risk patients. Real-time monitoring and graded early warning ensure that high-risk patients are intervened within 24 hours, while medium-risk patients receive focused attention. This avoids treatment delays caused by the "late warning" of traditional technologies. It can monitor dynamically at any time, guide clinical intervention, and intervene in a timely manner, reducing the incidence of severe radiation-induced oral mucositis, reducing adverse events such as treatment interruption, increased pain, and malnutrition, and improving patients' quality of life.

[0018] The linear mixed-effects longitudinal sub-model and the Cox regression survival sub-model form a "two-way feedback" through shared random effects. The Cox regression survival sub-model uses the dependent variable of the linear mixed-effects longitudinal sub-model to update the risk, while the linear mixed-effects longitudinal sub-model uses event information from the Cox regression survival sub-model to optimize trajectory prediction. Ultimately, this achieves a deep integration of "baseline variables + longitudinal variables". This invention can not only predict outcomes before treatment begins, but also remeasure longitudinal variables on the longitudinal timeline as time progresses during treatment. After adding the data from the longitudinal variables, the outcome can be dynamically predicted. Attached Figure Description

[0019] Figure 1 This is a flowchart of the real-time risk warning and intervention method for radiation-induced oral mucositis based on multi-parameter dynamic monitoring, as per the present invention. Detailed Implementation

[0020] Typical embodiments embodying the features and advantages of the present invention will be specifically described in the following description. It should be understood that the present invention can have various variations in different embodiments without departing from the scope of the present invention, and the descriptions and illustrations herein are for illustrative purposes only and not intended to limit the present invention.

[0021] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] Explanation: Baseline variables include indicators of longitudinal variables. Longitudinal variables are called longitudinal variables because they are repeatedly measured and added to the longitudinal model during the radiotherapy period, thus enabling dynamic monitoring. (Traditional predictive models are generally based on baseline variables and only predict before treatment begins; they lack dynamic monitoring capabilities during treatment.)

[0023] This embodiment discloses a method for real-time risk warning and intervention of radiation-induced oral mucositis based on multi-parameter dynamic monitoring, such as... Figure 1 As shown, it includes the following steps:

[0024] S1: The multi-parameter dynamic acquisition module collects patients' sociodemographic data, health-related data, physiological-related data, disease-related data, treatment-related data, and psychological and nutritional-related data in real time through the electronic medical record interface, medical staff terminal, and patient terminal.

[0025] In S1, sociodemographic data includes gender, age, marital status, employment status, medical insurance coverage, and average monthly household income per capita; health-related data includes body mass index, oral hygiene score, smoking history, alcohol consumption history, and diabetes history; physiological data includes oral pH, white blood cell count, neutrophil percentage, and red blood cell count; disease-related data includes time to diagnosis (months), cancer type, tumor type, and clinical stage; treatment-related data includes radiotherapy dose parameters (total dose, average maximum oral dose) and radiotherapy regimen (whether concurrent chemoradiotherapy, whether induction chemoradiotherapy, and whether sequential radiotherapy); and psychological and nutritional data includes anxiety and depression levels, nutritional risks, perceived social support, and coping styles.

[0026] In this embodiment, the electronic medical record interface is used to connect to the hospital information system to collect patient disease-related data and treatment-related data. The medical staff end is used for medical staff to enter patient physiological data, and the patient end is used for patients to fill in sociodemographic data, health-related data, and self-assessment of psychological and nutritional data.

[0027] In this embodiment, relevant parameters are dynamically collected, and the specific collection frequency can be customized according to clinical needs. For example, treatment-related data are synchronized in real time as the radiotherapy plan is adjusted, while health-related data, psychological and nutritional data are self-assessed by patients weekly, and physiological data are collected and entered weekly.

[0028] In this embodiment, oral hygiene scores are assessed using an oral hygiene status scale, anxiety and depression levels are assessed using a hospital anxiety and depression scale, nutritional risk is assessed using a nutrition risk screening (NRS2002) scale, perceived social support is assessed using a perceived social support scale, and coping styles are assessed using a simplified coping style questionnaire (SCSQ).

[0029] In this embodiment, patients obtain corresponding scores by self-assessing the corresponding scale.

[0030] For example, oral hygiene status scales assess patients' awareness of the importance of maintaining oral hygiene and their knowledge of related knowledge such as the prevention of oral diseases, including brushing methods, selection of oral cleaning tools, flossing methods, denture care, and factors that affect oral hygiene.

[0031] The oral hygiene status scale has a score range of 1 to 3 for each item, representing good, fair, and poor, respectively. The total score is the sum of the scores for all items, ranging from 12 to 36. A higher score indicates a poorer oral hygiene condition and a greater need for oral hygiene care.

[0032]

[0033] Oral Hygiene Status Scale

[0034] S2: The data processing module preprocesses the collected patient sociodemographic data, health-related data, physiological-related data, disease-related data, treatment-related data, and psychological and nutritional-related data.

[0035] Specifically, S2 includes the following steps:

[0036] S21: Based on whether the collected data is repeatedly collected over time, the collected data is divided into baseline variables and longitudinal variables;

[0037] In this embodiment, baseline variables refer to the variables collected for the first time at the start of the intervention / follow-up (before the start of radiotherapy). Their core function is to predict the outcome variable (whether severe oral mucositis occurs) based on individual baseline differences. Baseline variables include sociodemographic data, health-related data, physiological-related data, treatment-related data, disease-related data, and psychological and nutritional-related data.

[0038] Baseline variables, such as age (continuous variable, unit: years), sex (binary: male=1, female=0), body mass index (continuous variable, unit: kg / m²), smoking history (binary: yes=1, no=0), history of diabetes (binary: yes=1, no=0), tumor TNM stage (multi-classification: T0-T4, N0-N3), and radiotherapy regimen (e.g., whether concurrent chemoradiotherapy is used, binary: yes=1, no=0), are determined when the radiotherapy plan is developed and belong to "static baseline information" during the monitoring process, which will not change with the monitoring time.

[0039] Longitudinal variables refer to variables that are repeatedly measured over time during the monitoring process (such as anxiety, depression, and nutritional risk, which need to be repeatedly assessed weekly during radiotherapy and 1 / 2 / 3 months after radiotherapy).

[0040] Longitudinal variables include oral hygiene scores, physiological data, and psychological and nutritional data.

[0041] S22: The data processing module performs initial logarithmic transformation on the vertical variables;

[0042] In this embodiment, a logarithmic transformation is performed on the longitudinal variables (e.g., anxiety, depression, nutritional risk). Since these longitudinal variables are skewed, the transformation is more consistent with the normality assumption of the linear model. At the same time, the logarithmic transformation can convert the hazard ratio (HR) into a more interpretable exponential form.

[0043] S23: The data processing module then performs data cleaning, missing value imputation, data standardization, and data type conversion standardization on the baseline variables and the initially processed longitudinal variables. Data cleaning is used to correct the deviation of the original data, identify and process outliers and erroneous values ​​in the original data, missing value imputation is used to make the original data complete, data standardization is used to make the data scale consistent, and data type conversion is used to make the data match the model input format.

[0044] In this embodiment, the specific rules for data cleaning are as follows:

[0045] First, outlier verification and handling identify data that exceeds clinically reasonable ranges or contains logical contradictions (e.g., the normal range for oral pH is 5.5-7.5; if a data point shows pH=9.0, it is considered an outlier; the normal range for white blood cell count is 4-10×10). 9 / L, if "0.5×10" appears 9 If the error message " / L" is missing and there is no clinical medical order explanation, it may be an input error. Specifically, the clinical threshold method involves setting a reasonable range based on medical guidelines or clinical consensus (e.g., referring to the WHO classification of oral mucositis, the pain score range is 0-10 points, and anything exceeding this range is marked as abnormal); the statistical method involves identifying statistical outliers using the "3σ principle" (if the data exceeds the mean ± 3 times the standard deviation) or the "interquartile range" (if it exceeds the range of Q1-1.5IQR to Q3+1.5IQR); the handling strategy is to correct confirmed input errors (e.g., "pH=9.0" should actually be "6.0") by providing a pop-up reminder on the healthcare provider's end; and to correct clinically rare but real data (e.g., a sudden drop in white blood cell count to 0.8×10 after chemotherapy). 9 / L), retain the original value and label it with "special clinical circumstances" to avoid accidentally deleting key risk signals.

[0046] Second, duplicate or logically contradictory data is cleaned. Duplicate data is deduplicated by "timestamp + data source priority" (retaining the scores after review by medical staff, or taking the average of multiple submissions). For example, due to network delays, the hospital anxiety and depression scale was submitted twice at the same time in the patient's mini-program (with scores of 8 and 10 respectively). Logically contradictory data is automatically checked by the system to prompt medical staff to confirm the validity of the data (to avoid the error of "discharged patients still being included in real-time risk calculation").

[0047] In this embodiment, real-time risk warning and intervention for radiation-induced oral mucositis relies on multi-dimensional data (such as complete blood count, oral pH, and hospital anxiety and depression scale scores). If any type of data is missing (e.g., patients miss completing the daily scale, or medical staff miss recording indicators), direct removal will lead to a reduction in sample size or missing input dimensions for the model. Therefore, it is necessary to fill in the missing values ​​using scientific methods. The specific rules for missing value filling are as follows:

[0048] First, targeted fill-in strategies based on clinical scenarios;

[0049]

[0050] Second, prioritize clinical rationality while retaining missing markers: for example, when filling in the "nutritional risk score", it is necessary to refer to the patient's daily eating ability (subjective feeling data) and weight change (objective indicator) to avoid pure statistical filling that is detached from clinical reality; all filled data should be labeled with "fill source" (such as "linear interpolation" or "multiple interpolation") so that "real data" and "filled data" can be distinguished during subsequent model optimization, reducing the impact of filling errors on parameters.

[0051] In this embodiment, the dimensions and scales of the multi-source data vary greatly (e.g., "anxiety score" ranges from 0 to 21; "oral pH" ranges from 4.0 to 8.0). If directly input into the model, large-scale data (e.g., D50) will mask the influence of small-scale data (e.g., pH), leading to parameter weight bias. Data standardization is used to solve the problem of inconsistent data scales.

[0052] In this embodiment, the specific rules for data standardization are as follows:

[0053] 1. Continuous physiological / therapeutic parameters (such as oral pH, radiotherapy dose D50, white blood cell count)

[0054] Z-Score standardization is used, and the calculation method is "(original data - mean of the variable in the training set) ÷ standard deviation of the variable in the training set".

[0055] Example: If the mean of the oral pH training set is 6.2 and the standard deviation is 0.5, and a patient's original oral pH is 5.8, the standardized result is (5.8-6.2)÷0.5=-0.8;

[0056] Applicable scenarios: Used as input for the Cox regression survival sub-model to eliminate the influence of different dimensions (such as pH is 1-8, D50 is 30-60Gy) on the calculation of hazard ratio.

[0057] 2. Discrete subjective rating parameters (such as anxiety score, nutritional risk score, and oral hygiene score)

[0058] The Min-Max normalization method is used, and the calculation method is "(original data - theoretical minimum value of the variable) ÷ (theoretical maximum value of the variable - theoretical minimum value)".

[0059] Example: If the theoretical range of anxiety scores is 0-21, and a patient's original score is 10.5, the normalized result is (10.5-0)÷(21-0)=0.5;

[0060] Applicable scenarios: Used as input for longitudinal sub-models of linear mixed effects, mapping dynamic scores to the 0-1 interval to facilitate quantitative analysis of score changes.

[0061] 3. Binary / multi-category parameters after classification coding (e.g., "intervention type")

[0062] Range standardization is adopted, and the calculation method is "(coded data - coded minimum value) ÷ (coded maximum value - coded minimum value)" (for binary classification, the minimum value is 0 and the maximum value is 1; for multi-class independent coding, the minimum value is 0 and the maximum value is 1).

[0063] Applicable scenarios: Used for the fusion of shared random effects features in joint models to ensure that the weights of categorical and continuous variables are consistent.

[0064] 4. Time-series dynamic parameters (such as follow-up time, time series values ​​of longitudinal variables)

[0065] The time standardization is adopted, and the calculation method is "(current time - radiotherapy start time) ÷ (follow-up end time - radiotherapy start time)" (the radiotherapy start time is recorded as 0, and the follow-up end time in this embodiment is 3 months after radiotherapy, recorded as 90 days).

[0066] Example: The standardized result of the 15-day follow-up time after radiotherapy is (15-0)÷(90-0)≈0.167;

[0067] Applicable scenarios: Used as input for "follow-up time" in longitudinal sub-models to unify the time dimension for different patients and avoid interference from differences in follow-up periods.

[0068] In this embodiment, the original data types (such as text, categorical, and continuous) need to be converted into a numerical format that the model can recognize. The specific data type conversion rules are as follows:

[0069] First, categorical data encoding: binary variables (such as "whether concurrent chemotherapy" or "whether psychological intervention was received") are encoded using 0-1 (yes = 1, no = 0); multi-category variables (such as "intervention type": A = oral care, B = nutritional supplementation, C = psychological counseling) are encoded using one-hot encoding, which converts them into three binary variables (A = [1,0,0], B = [0,1,0], C = [0,0,1]) to avoid the model misjudging categorical variables as "ordered relationships" (such as mistakenly interpreting A = 1 and B = 2 as "the intensity of option B is twice that of option A").

[0070] Second, the text data is structured. If the medical staff enters "oral mucosal status" as a text description (such as "mild congestion of the left buccal mucosa"), key information is extracted and encoded through natural language processing (NLP) (such as "congestion" = 1, "ulcer" = 2, "no abnormality" = 0), and transformed into numerical variables that the model can recognize.

[0071] This embodiment preprocesses the collected data, unifying the data format, correcting data biases, and filling data gaps, transforming the multi-source heterogeneous raw data (oral pH value, hospital anxiety and depression scale assessment score) into "clean, consistent, and usable" data that meets the model input requirements.

[0072] Specifically, to ensure data quality, the model is cleaned and filled to reduce the interference of raw data bias on the model and avoid "missing high-risk patients due to data errors"; to ensure model fairness, the model is normalized / standardized to eliminate differences in data scale, so that risk factors of different dimensions such as "oral pH" and "anxiety score" can participate in risk calculation fairly; and to support real-time dynamic calculation, all processing steps are designed as "automated processes" (such as cleaning, filling and normalization within 10 seconds after receiving data in real time), matching the system's need for "real-time risk updates" and providing immediate data support for intervention triggering within the next 24 hours.

[0073] S3: The processed data is input into the model calculation module, and the dynamic joint prediction model, which includes a linear mixed-effects longitudinal sub-model and a Cox regression survival sub-model, and realizes data association and collaborative modeling through the posterior distribution of the random effects of the samples, is used to predict the probability of patients developing severe radiation-induced oral mucositis in the next 1 to 3 months and generate the corresponding risk level.

[0074] In S3, the dynamic joint prediction model includes a COX regression survival sub-model and a linear mixed effects longitudinal sub-model. The COX regression survival sub-model and the linear mixed effects longitudinal sub-model are combined to construct the dynamic joint prediction model by using the shared random effects model of the JMbayes2 package in RStudio 4.3.1.

[0075] In S3, the construction of the Cox regression survival sub-model and the linear mixed effects longitudinal sub-model includes the following steps:

[0076] S31: Multiple relevant potential risk factors leading to severe radiation-induced oral mucositis were screened from baseline variables using a LASSO regression model;

[0077] The LASSO regression model identifies multiple potential risk factors associated with severe radiation-induced oral mucositis from baseline variables. Specific steps include:

[0078] First, the regularization parameter that minimizes the cross-validation error is determined through five-fold cross-validation, ensuring that the model balances good fit and generalization. Specifically:

[0079] The regularization parameter λ ranges from 0 to 10 (increasing by 0.01 steps, e.g., 0, 0.01, 0.02...10). The samples are randomly divided into 5 groups. Each time, 4 groups are used as the training set and substituted into the LASSO objective function containing λ for optimization. The LASSO regression model parameters (regression coefficients) corresponding to different λ values ​​under this round of training set are fitted. The model is used to predict the outcome of the remaining 1 test set ("whether severe RIOM occurs", binary classification: yes=1, no=0). The mean squared error of the test set in this round is calculated according to the formula: cross-validation mean squared error = (1 / number of test set samples) × Σ(actual value of test set - model predicted value)². For each λ value, the mean squared error of the cross-validation over 5 rounds (i.e., CV-MSE) is calculated. The CV-MSE corresponding to all λ values ​​is compared, and the λ with the smallest CV-MSE is selected as the optimal regularization parameter.

[0080] Example: Suppose that when λ=0.05, the mean squared errors of the test set for 5 rounds of cross-validation are 0.12, 0.11, 0.13, 0.12, and 0.11, with an average of 0.118; when λ=0.1, the average mean squared error for 5 rounds is 0.132; then we choose λ=0.05 (corresponding to a smaller CV-MSE) as the optimal value and substitute it into the LASSO objective function to complete the variable selection.

[0081] Second, the selected regularization parameters are substituted into the objective function of the LASSO regression model to calculate the regression coefficients of each candidate variable. Baseline variables with non-zero regression coefficients are retained, ultimately yielding several relevant potential risk factors, specifically:

[0082] The objective function is constructed with the goal of minimizing the sum of the model prediction error and the absolute values ​​of the constraint parameters, as shown in formula (1), to obtain the regression coefficients of the corresponding candidate variables:

[0083] (1),

[0084] In formula (1), For sample size, such as the 294 nasopharyngeal carcinoma patients in this protocol, For logistic regression loss function, For the first The outcome variable for each sample, here "whether severe radiation-induced oral mucositis occurred", is binary: yes = 1, no = 0, and is used for preliminary association analysis in the variable screening stage of the LASSO regression model. For model intercept, , for A vector of regression coefficients for candidate variables. To measure the prediction error of the model, , For the first A vector of candidate variable values ​​for each sample. for transpose, This is the regularization term in the LASSO regression model. For regularization parameters, The summation of the absolute values ​​of the regression coefficients of all candidate variables. For the first regression coefficients of candidate variables The absolute value of.

[0085] S32: The steps for training the Cox regression survival sub-model and the linear mixed-effects longitudinal sub-model are as follows:

[0086] First, the samples are divided into training and test sets. In this embodiment, the samples are randomly split in a 7:3 ratio (e.g., out of 294 patients, 206 are in the training set and 88 are in the test set) to ensure that the two groups are consistent in their baseline characteristics (e.g., age) (verified by chi-square test or t-test, P>0.05).

[0087] Second, in the Cox regression survival sub-model, multiple relevant potential risk factors screened by the LASSO regression model are used as independent variables, and the time of first diagnosis of severe radiation oral mucositis and the outcome event of first diagnosis of severe radiation oral mucositis are used as dependent variables. The dependent variables are used as labels. In this embodiment, if severe radiation oral mucositis does not occur during the monitoring period, the data in which severe radiation oral mucositis does not occur during the monitoring period is marked as truncated data, and the corresponding time is the truncated time.

[0088] In the linear mixed-effects longitudinal sub-model, the covariates in the time and baseline variables at each monitoring point are used as independent variables, and the measured values ​​of the longitudinal variables at each monitoring point are used as dependent variables (the log-transformed values ​​of anxiety, depression, and nutritional risk are used as dependent variables).

[0089] In this embodiment, the covariates in the baseline variables refer to variables that are determined at the "intervention / follow-up start time (before the start of radiotherapy)" and do not change over time during the follow-up process.

[0090] Third, the Cox regression survival sub-model and the linear mixed effects longitudinal sub-model were trained respectively. In the Cox regression survival sub-model, the independent variables were used as input features and the dependent variable was used as labels. The regression coefficients of each variable were calculated by minimizing the negative log-likelihood function, and the hazard ratios were obtained through the regression coefficients. The hazard rates of each variable were calculated by the maximum likelihood estimation method and the regression coefficients, as shown in formulas (2), (3), and (4).

[0091] (2),

[0092] (3),

[0093] (4),

[0094] In formulas (2), (3) and (4), For logistic regression loss function, For sample size, As the dependent variable, the first In this embodiment, to determine whether a target event occurs in a given sample individual, when severe radiation-induced oral mucositis occurs, The value is 1, when severe radiation-induced oral mucositis does not occur. =0, The baseline cumulative risk function at time... The value of , These are the regression coefficients, i.e., the weights to be determined. This is the transpose of the eigenvector. For the first The feature vector (value vector of independent variables) of each sample. For risk ratio, For the sample in time The risk rate, As the benchmark risk rate, For feature vectors;

[0095] In the linear mixed-effects longitudinal sub-model, the time and covariates in the baseline variables at each monitoring time are used as inputs, and the measured values ​​of the longitudinal variables at each monitoring time are used as labels (outcomes). These are substituted into the expression of the linear mixed-effects longitudinal sub-model to train the linear mixed-effects longitudinal sub-model, as shown in formula (6).

[0096] (6),

[0097] In formula (6), For the first The first patient Longitudinal variable values ​​from each monitoring session (such as log-transformed values ​​of anxiety and depression scale scores). For sample size, , , All are fixed effects weights. This refers to the monitoring period (also known as the follow-up period). The initial values ​​of the vertical variables, , Let be the posterior distribution of the random effect. For random intercept, For random slope, This is the error term;

[0098] Fourth, the test set is fed into the trained linear mixed-effects longitudinal sub-model and the Cox regression survival sub-model, respectively. The linear mixed-effects longitudinal sub-model is evaluated using the root mean square error (RMSE); the Cox regression survival sub-model is evaluated using the consistency index (C-index). Specifically, the following steps are included:

[0099] Optimization of the longitudinal sub-model for linear mixed effects:

[0100] When the root mean square error (RMSE) between the predicted outcome of the test set obtained by the linear mixed effects longitudinal sub-model and the actual outcome of the test set is less than or equal to the threshold, the linear mixed effects longitudinal sub-model stops optimizing; when the RMSE between the predicted outcome of the test set obtained by the linear mixed effects longitudinal sub-model and the actual outcome of the test set is greater than the threshold, the linear mixed effects longitudinal sub-model starts optimizing until the RMSE is less than or equal to the threshold.

[0101] In this embodiment, the threshold is 2.

[0102] The residual distribution plot is calculated based on the difference between the predicted results of the test set and the actual outcome of the linear mixed-effects longitudinal sub-model. If heteroscedasticity exists, a weighting function is introduced into the linear mixed-effects longitudinal sub-model, as shown in Equation (7), to correct the uneven error distribution and thus optimize the model, as shown in Equation (8).

[0103] (7),

[0104] (8),

[0105] In formulas (7) and (8), For the first The first patient The weight value of each follow-up (monitoring) visit. For the first The first patient The residual variance estimate of the follow-up (obtained by calculating the residuals after fitting the initial model); As a weighted error term, the error distribution at different follow-up points is made more uniform by adjusting the weights, thereby eliminating the influence of heteroscedasticity on the model fitting effect; For fixed-effects intercept, This is the fixed-effects coefficient for follow-up time. The fixed effects coefficients of the baseline covariates. For the first The first patient Follow-up time, For the first The first patient's One baseline covariate, For random intercept, For random slope, This is the original error term.

[0106] If the residual analysis indicates that the variable changes non-linearly over time, a quadratic time term should be added to formula (8). ) and the interaction term of "time × baseline covariate" (e.g. × By adjusting the model structure to adapt to nonlinear relationships, the optimized complete formula is as follows:

[0107] (9),

[0108] In formula (9), The coefficient of the quadratic term over time is used to capture the non-linear trend of longitudinal variables over time (such as the trajectory of anxiety scores that first rise and then fall). The coefficient of the interaction term "time × baseline covariate" is used to characterize the effect of baseline covariates (such as BMI, baseline oral pH) on the rate of change of longitudinal variables (e.g., the rate of increase in nutritional risk score is slower in patients with high BMI).

[0109] COX regression survival sub-model optimization:

[0110] When the consistency index between the predicted outcome of the test set obtained by the Cox regression survival sub-model and the actual outcome of the test set is greater than or equal to the threshold, the Cox regression survival sub-model stops optimizing; when the consistency index between the predicted outcome of the test set obtained by the Cox regression survival sub-model and the actual outcome of the test set is less than the threshold, the Cox regression survival sub-model starts optimizing until the consistency index is greater than or equal to the threshold.

[0111] In this embodiment, the threshold is 0.7.

[0112] Using the Schönfeld residual test, if the residuals of a certain variable are significantly correlated with time (P<0.05), it indicates that the variable violates the PH hypothesis and is the main source of error. Therefore, a time-dependent covariate (such as X) should be introduced. ij ×log(t)), and at the same time, supplement the key risk factors for clinical validation (such as "dynamic changes in oral pH" and "neutrophil / lymphocyte ratio"). Re-screen variables through the LASSO regression model, remove redundant variables (variables with regression coefficients close to 0), and segment continuous covariates (such as D50 dose) (such as dividing them into 4 groups according to quartiles). The corrected result is shown in formula (10):

[0113] (10)

[0114] In formula (10), The dynamic risk rate of the i-th patient at time t. As the benchmark risk rate, It is the vector of regression coefficients of the basic risk factors (variables that meet the PH assumption). It is the regression coefficient vector of time-dependent covariates. The basic risk factor vector that conforms to the PH hypothesis (such as baseline oral pH and perceived social support). This represents a vector of risk factors that violate the PH hypothesis (such as D50 dose and white blood cell count). This is the natural logarithm of the follow-up time (a unit of measurement for a unified time dimension).

[0115] In S32, the linear mixed effects longitudinal sub-model and the Cox regression survival sub-model achieve data association and collaborative modeling through the posterior distribution of the random effects of the samples. The Cox regression survival sub-model uses the dependent variable of the linear mixed effects longitudinal sub-model to update the risk, and the linear mixed effects longitudinal sub-model uses the event information of the Cox regression survival sub-model to optimize trajectory prediction.

[0116] The linear mixed-effects longitudinal sub-model is updated via Bayesian posterior and uses the event-corrected trajectory prediction from the Cox regression survival sub-model, specifically including the following steps:

[0117] First, the prior distribution of individual random effects is a multivariate normal distribution, as shown in formulas (11) and (12).

[0118] (11),

[0119] (12),

[0120] In formulas (11) and (12), For the first Individual patients share a random effects vector. The random intercept term reflects the first... Individual differences in the initial levels of longitudinal variables for each patient (such as anxiety score, nutritional risk score) (e.g., before the start of radiotherapy, patient A's baseline anxiety score was naturally higher than patient B's, and this difference was caused by...). Quantization), and the fixed effects intercept in the longitudinal sub-model of linear mixed effects. Collaboration forms the basis for initial level predictions of longitudinal variables; The random slope term reflects the first... Individual differences in the rate of change of longitudinal variables over follow-up time for each patient (e.g., patient A's anxiety score increased by 0.3 points per week during radiotherapy, while patient B's increased by 0.1 points per week; this rate difference is due to...). (Quantization), and the time-fixed effects coefficient in the linear mixed-effects longitudinal sub-model Collaboratively capture the dynamic trajectory of longitudinal variables. The vector transpose symbol is used to transpose column vectors. Convert to row vector This satisfies the matrix operation rules, ensuring that the random effects vector can be linearly combined with other parameters in the model (such as fixed effects coefficients and covariate vectors). This is the random effects covariance matrix, used to characterize the random intercept. With random slope The variance and the covariance between them;

[0121] Second, the posterior distribution of the random effects is updated using Bayes' theorem, as shown in formula (13).

[0122] (13)

[0123] In formula (13), Let be the posterior distribution of the random effect. The measured values ​​of longitudinal variables for each patient (e.g., anxiety scores at each follow-up point). As the dependent variable, Let be the likelihood function of the Cox regression survival sub-model. For the first The baseline risk factor vector for each patient (e.g., oral pH, D50 dose). Let be the likelihood function of the longitudinal sub-model of the linear mixed effects. For the i-th patient, the baseline covariate vector (e.g., BMI, smoking history) is used. This represents the prior distribution of the random effect;

[0124] In this embodiment, the posterior distribution defined by the Markov Chain Monte Carlo (MCMC) sampling approximation formula (13) is executed using the update() function of the JMbayes2 package (the number of iterations is set to 10,000, with the first 2,000 iterations discarded as a warm-up period). In each iteration, the posterior distribution of the random effects is updated according to the following rules, as shown in formula (15).

[0125] (15)

[0126] In formula (15), For the number of iterations, The value is taken for the random effect vector in the (s+1)th iteration; This is the specific sampling form of the posterior distribution of formula (13), where , The first The fixed effects coefficients and random effects covariance matrix estimates of each iteration are known parameters of the sampling process. The sampling distribution of formula (15) is essentially the conditional distribution of formula (13). Through multiple iterations, the true form of the posterior distribution is gradually approximated. Finally, the posterior mean is obtained by calculating the mean in the later stages of the iteration. Simultaneously, based on the residual distribution of the training set, future time points t are derived. ∗ weight value and prediction error estimate Substitute them together into formula (14) and complete the correction of the future trajectory of the longitudinal variable based on the optimized linear mixed effect longitudinal sub-model (formula 9).

[0127] When the Cox regression survival submodel observed that "the patient did not develop severe radiation-induced oral mucositis () =0), the posterior mean of the random effect is downgraded, as shown in formula (16).

[0128] (16)

[0129] In formula (16), This represents the final updated posterior distribution of the random effects. The random effects prior mean before the update. To adjust the coefficient, it is set to 0.2 in this embodiment. This is the partial derivative of the likelihood function of the Cox regression survival submodel with respect to random effects;

[0130] When the COX regression survival submodel observed "severe radiation-induced oral mucositis has occurred ( =1), the posterior mean of the random effect is increased, as shown in formula (17).

[0131] (17)

[0132] In formula (17), This represents the final updated posterior distribution of the random effects.

[0133] Third, based on the updated random effects posterior distribution, the predicted future trajectories of the longitudinal variables in the linear mixed effects longitudinal sub-model are corrected, as shown in Equation (14).

[0134] (14)

[0135] In formula (14), For the first Patients at a future time point Adjusted predicted values ​​of longitudinal variables (anxiety score / nutritional risk score) at week 4 post-radiotherapy (e.g., week 4 post-radiotherapy). For the first One patient The transpose of the fixed-effects covariate vector at time points (such as follow-up time and baseline oral pH) has a vector dimension that matches the fixed-effects term in formula (9), specifically as follows: , This represents the fixed effects coefficient vector of the longitudinal sub-model of linear mixed effects, i.e. ,in For intercept, For time coefficient, For the coefficient of the quadratic term of time, The coefficient of the interaction term. The coefficients of each baseline covariate (i.e., those in formula (9)) (corresponding to formula (9)). For the first One patient The transpose of the random effects design matrix at time points takes values ​​in the range [1, t*], corresponding to the random intercepts. and random slope , and the random effects vector in formula (11) =( , ) T match, For event information based on the Cox regression survival sub-model ( The updated posterior mean of the random effects, whose distribution is defined by formula (13), is obtained by MCMC sampling of formula (15) and is used to correct individual-specific trajectory bias. For a future time point t ∗ The estimated value of the prediction error, , For a future time point t ∗ The residual variance estimate;

[0136] The Cox regression survival sub-model uses the posterior mean obtained from the linear mixed-effects longitudinal sub-model as the random effect and calculates the hazard ratio. It also integrates the random effect into the hazard rate formula, clarifying the relationship between the hazard ratio and the hazard rate, as detailed below:

[0137] ① Calculation of the risk ratio of the associated risk rate

[0138] (18)

[0139] In formula (18), The change in baseline risk factors that conforms to the PH assumption. To violate the PH assumption, the change in the variable This represents the change in the posterior mean of the random effect.

[0140] ② Risk rate of the integrated random effects posterior mean

[0141] The risk rate is absolute risk, while the risk ratio is relative risk. The conversion relationship between the two is as follows:

[0142] (19)

[0143] (20)

[0144] In formula (19), the benchmark risk rate is used. Risk ratio Multiplying these values ​​yields the real-time risk rate for the target patient, which can be used to classify patients into low, medium, and high risk levels.

[0145] In formula (20), For the first The patient in time The dynamic risk rate As the benchmark risk rate, The vector of regression coefficients for baseline risk factors that conform to the proportional risk (PH) assumption. For the time-dependent covariate regression coefficient vector of the variable that violates the PH hypothesis, This represents a vector of risk factors that violate the PH hypothesis (such as D50 dose and white blood cell count). This is the vector of random effect weight coefficients (estimated through the training set, reflecting the strength of the impact of random effects on survival risk). The individual dynamic trajectory contribution term is passed to the longitudinal sub-model of linear mixed effects, enabling collaborative modeling of the two sub-models.

[0146] In this embodiment, the Cox regression survival sub-model monitors in real time whether the patient develops severe RIOM (δi=0 / 1), and updates b using a Bayesian formula.i The posterior distribution, if δ i =0 (Not occurred): Downgrade b associated with high risk i (e.g., the slope of the anxiety rise), making the anxiety trajectory predicted by the longitudinal sub-model smoother, if δ i =1 (occurred): Upgrade risk-related b i (e.g., the slope of nutritional risk increase), making the trajectory predicted by the longitudinal sub-model of linear mixed effects more closely match the characteristics of high risk, the corrected b i The COX regression survival sub-model will be re-input to further optimize risk prediction, forming a closed loop of "trajectory fitting → risk prediction → trajectory correction → risk re-prediction".

[0147] For example, using the hospital anxiety and depression scale score as the longitudinal variable, the linear mixed-effects longitudinal sub-model fits the "trajectory of anxiety level changes over follow-up time (monitoring time)" in real time, and outputs the "individual-specific fluctuation parameter of the patient's anxiety level" through shared random effects (such as the rate of sudden increase in anxiety in a patient during the 3rd week of radiotherapy); the Cox regression survival sub-model takes the above "anxiety fluctuation parameter" as input, combines it with baseline variables such as oral pH, and dynamically adjusts the risk rate of severe radiation-induced oral mucositis for the patient; feedback optimization: if the survival sub-model observes that "the patient did not develop severe radiation-induced oral mucositis", its shared random effects will reverse the prediction of the anxiety trajectory of the longitudinal sub-model (such as adjusting the predicted value of subsequent anxiety level), forming a closed-loop optimization.

[0148] S33: Input the processed data into the model calculation module, and use the trained linear mixed-effects longitudinal sub-model in the model calculation module to predict the values ​​of multiple corresponding risk factors for the patient in the next 1 to 3 months. Then, use the trained Cox regression survival sub-model in the model calculation module to predict the risk rate of the patient developing severe radiation oral mucositis in the next 1 to 3 months. When the risk rate is less than or equal to 30%, a low risk level is generated; when the risk rate is greater than 31% and less than 60%, a medium risk level is generated; and when the risk rate is greater than or equal to 60%, a high risk level is generated.

[0149] S4: The risk warning push module generates corresponding warning information based on the risk level predicted by the model calculation module, and sends the warning information to the medical staff terminal and corresponding personalized intervention prompts to the patient terminal;

[0150] In this embodiment, when the risk level predicted by the model calculation module is low risk, the risk warning push module generates low risk warning information, sends the risk level to the medical staff, and sends routine care tips to the patient (such as "It is recommended to rinse your mouth with warm water twice today to keep your mouth clean").

[0151] When the model calculation module predicts a medium-risk level, the risk warning push module generates a medium-risk warning message and sends the risk level to the medical staff. At the same time, it pushes a reminder to the medical staff, which includes the risk rate (e.g., "45% risk of severe radiation-induced oral mucositis in the next month"), key fluctuation parameters (e.g., "oral pH value has decreased by 0.3 compared to yesterday"), and suggests that medical staff conduct a key assessment at the next consultation. It also sends enhanced nursing guidance to the patient (e.g., "It is recommended to use mouthwash to clean the mouth and monitor the oral mucosa for redness and swelling daily").

[0152] When the model calculation module predicts a high-risk level, the risk warning push module generates a high-risk warning message and sends the risk level to the medical staff terminal. At the same time, it immediately triggers an emergency warning on the medical staff terminal (such as a pop-up window and SMS notification). It also pushes a reminder to the medical staff terminal, which includes the risk rate ("72% risk of severe radiation-induced oral mucositis in the next month") and key fluctuation parameters. It also automatically links to the intervention plan template (such as "It is recommended to complete a detailed oral mucosal assessment within 24 hours and start nutritional support intervention"). An emergency intervention prompt is sent to the patient terminal, which includes real-time medical guidance (such as "Please go to the radiotherapy department nurse station this afternoon for oral care") and a symptom self-check list (such as "If ulcers or severe pain occur, please contact the medical staff immediately through the patient terminal").

[0153] S5: The intervention feedback module collects feedback data after intervention through the medical staff and patient terminals. The feedback data is stored in the storage module for later review by medical staff. Repeating S1-S4, it provides real-time risk warning and intervention for patients with severe radiation-induced oral mucositis.

[0154] In this embodiment, the feedback data includes intervention data, objective outcome indicator data, and subjective feeling data.

[0155] Intervention data (details of intervention plans entered by medical staff) include oral care: nursing frequency (e.g., "rinse with alkaline mouthwash 3 times a day"), nursing tools (e.g., "soft-bristled toothbrush + alcohol-free mouthwash"), and drug intervention (e.g., "recombinant human epidermal growth factor gel application, twice a day").

[0156] Psychological interventions include: intervention methods (e.g., "cognitive behavioral therapy, twice a week, 30 minutes each time"), intervention duration (e.g., "for 4 weeks"), and intervention personnel (e.g., "psychotherapist qualification level").

[0157] Nutritional supplements: type of nutritional preparation (e.g., "high-protein nutritional liquid (200ml / day) + B vitamins"), frequency of supplementation (e.g., "oral nutritional supplement 3 times a day"), dietary guidance (e.g., "liquid diet, avoid spicy and irritating foods").

[0158] Other interventions: radiotherapy plan adjustment (e.g., "change the dose fractionation method from conventional to superfractionation"), drug adjuvant (e.g., "use of amifostine, intravenous injection 30 minutes before radiotherapy").

[0159] Objective outcome data included those related to oral mucositis: WHO classification (e.g., "from grade 3 to grade 1 after intervention"), changes in oral pH (e.g., "from 5.8 to 6.5"), and oral hygiene scores (e.g., "from 8 to 12").

[0160] Physiological parameters: blood routine indicators (changes in white blood cell count, neutrophil percentage, and red blood cell count), weight change (such as "weekly weight change rate"), and nutritional indicators (serum albumin and prealbumin levels).

[0161] Subjective experience data includes (collected by the patient's self-assessment interaction module) pain level: Numerical pain score (NRS, such as "from 7 points to 3 points"), and eating-related pain (such as "the frequency of pain during eating changed from pain every time eating to occasional pain").

[0162] Feeding ability: Feeding difficulty score (e.g., "from being unable to eat orally to being able to eat semi-liquid"), swallowing function score (e.g., "from level 3 to level 1 in Kubota water drinking test");

[0163] Quality of life: Changes in scores for dimensions such as "oral pain", "eating" and "social functioning" on the EORTCQLQ-H&N35 scale, and changes in scores on anxiety / depression scales (such as HADS) (e.g., "anxiety score dropped from 15 to 8").

[0164] This embodiment discloses a real-time risk warning and intervention system for radiation-induced oral mucositis based on multi-parameter dynamic monitoring. The system employs a real-time risk warning and intervention method for radiation-induced oral mucositis based on multi-parameter dynamic monitoring, including a multi-parameter dynamic acquisition module, a data processing module, a model calculation module, a risk warning push module, an intervention feedback module, and a storage module.

[0165] The multi-parameter dynamic acquisition module is used to collect patients' sociodemographic data, health-related data, physiological-related data, disease-related data, treatment-related data, and psychological and nutritional-related data in real time through the electronic medical record interface, medical staff terminal, and patient terminal.

[0166] The data processing module is used to preprocess the collected parameters, and the model calculation module is used to dynamically predict the probability of a patient developing severe radiation-induced oral mucositis within the next 1 to 3 months based on the preprocessed parameters, and generate the corresponding risk level.

[0167] The risk warning push module is used to generate corresponding warning information based on the risk level predicted by the model calculation module, and send the warning information to the medical staff terminal and corresponding personalized intervention prompts to the patient terminal;

[0168] The intervention feedback module is used to collect feedback data after intervention through medical staff and patients, and the storage module is used to store the feedback data for easy retrieval by medical staff later.

[0169] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method for real-time risk warning and intervention of radiation-induced oral mucositis based on multi-parameter dynamic monitoring, characterized in that, Includes the following steps: S1: The multi-parameter dynamic acquisition module collects patients' sociodemographic data, health-related data, physiological-related data, disease-related data, treatment-related data, and psychological and nutritional-related data in real time through the electronic medical record interface, medical staff terminal, and patient terminal. S2: The data processing module preprocesses the collected patient sociodemographic data, health-related data, physiological-related data, disease-related data, treatment-related data, and psychological and nutritional-related data. S3: The processed data is input into the model calculation module, and the dynamic joint prediction model, which includes a linear mixed-effects longitudinal sub-model and a Cox regression survival sub-model, and realizes data association and collaborative modeling through the posterior distribution of the random effects of the samples, is used to predict the probability of patients developing severe radiation-induced oral mucositis in the future and generate the corresponding risk level. S4: The risk warning push module generates corresponding warning information based on the risk level predicted by the model calculation module, and sends the warning information to the medical staff terminal and corresponding personalized intervention prompts to the patient terminal; S5: The intervention feedback module collects feedback data after intervention through the medical staff and patients. The feedback data is stored in the storage module for later query by medical staff.

2. The method for real-time risk warning and intervention of radiation-induced oral mucositis based on multi-parameter dynamic monitoring according to claim 1, characterized in that: In S3, the dynamic joint prediction model includes a COX regression survival sub-model and a linear mixed effects longitudinal sub-model. The COX regression survival sub-model and the linear mixed effects longitudinal sub-model are combined by sharing a random effects model to jointly construct the dynamic joint prediction model.

3. The method for real-time risk warning and intervention of radiation-induced oral mucositis based on multi-parameter dynamic monitoring according to claim 1, characterized in that: In S3, the training steps for the COX regression survival sub-model are as follows: First, the samples are divided into training and testing sets; Second, multiple relevant potential risk factors screened by the LASSO regression model were used as independent variables, and the time of first diagnosis of severe radiation-induced oral mucositis and the outcome event of first diagnosis of severe radiation-induced oral mucositis were used as dependent variables, with the dependent variables serving as labels. Third, using the independent variables as input features and the dependent variable as labels, the regression coefficients of each variable are calculated by minimizing the negative log-likelihood function. The risk ratio is then obtained from the regression coefficients, and the risk rate of each variable is calculated using the maximum likelihood estimation method and the regression coefficients, as shown in formulas (2), (3), and (4). （2）， （3）， （4）， In formulas (2), (3) and (4), For logistic regression loss function, For sample size, As the dependent variable, The baseline cumulative risk function at time... The value of , These are the regression coefficients, i.e., the weights to be determined. This is the transpose of the eigenvector. For the first The feature vector of each sample For risk ratio, For the sample in time The risk rate, As the benchmark risk rate, For feature vectors; Fourth, the test set is fed into the trained Cox regression survival sub-model. When the consistency index between the predicted outcome obtained by the test set through the Cox regression survival sub-model and the actual outcome of the test set is greater than or equal to the threshold, the Cox regression survival sub-model stops optimizing. When the consistency index between the predicted outcome obtained by the test set through the Cox regression survival sub-model and the actual outcome of the test set is less than the threshold, the Cox regression survival sub-model starts optimizing until the consistency index is greater than or equal to the threshold.

4. The method for real-time risk warning and intervention of radiation-induced oral mucositis based on multi-parameter dynamic monitoring according to claim 3, characterized in that: The optimization of the COX regression survival sub-model includes the following: If the residual of a variable is significantly correlated with time through the Schönfeld residual test, it indicates that the variable violates the PH hypothesis and is the main source of error. Time-dependent covariates are introduced, and key risk factors for clinical validation are added. Variables are re-screened through the LASSO regression model, redundant variables are removed, and continuous covariates are segmented. The corrected result is shown in formula (10): （10）， In formula (10), The dynamic risk rate of the i-th patient at time t. As the benchmark risk rate, It is the regression coefficient vector of the basic risk factors. It is the regression coefficient vector of time-dependent covariates. To form a basic risk factor vector that conforms to the PH assumption, For the risk factor vector that violates the PH assumption, This is the natural logarithm of the follow-up time.

5. The method for real-time risk warning and intervention of radiation-induced oral mucositis based on multi-parameter dynamic monitoring according to claim 3, characterized in that: In S3, multiple potential risk factors for severe radiation-induced oral mucositis were identified from baseline variables using a LASSO regression model, including the following steps: First, the regularization parameter that minimizes the cross-validation error is determined through five-fold cross-validation, ensuring that the model balances fit and generalization. Second, the selected regularization parameters are substituted into the objective function of the LASSO regression model to calculate the regression coefficients of each candidate variable, as shown in formula (1). Baseline variables with non-zero regression coefficients are retained, and finally, multiple related potential risk factors are obtained. The objective function is constructed with the goal of minimizing the sum of the model prediction error and the absolute values ​​of the constraint parameters, as shown in formula (1), to obtain the regression coefficients of the corresponding candidate variables: （1）， In formula (1), For sample size, For logistic regression loss function, For the first The outcome variable for each sample. For model intercept, for A vector of regression coefficients for candidate variables. To measure the prediction error of the model, For the first A vector of candidate variable values ​​for each sample. for transpose, This is the regularization term in the LASSO regression model. For regularization parameters, The summation of the absolute values ​​of the regression coefficients of all candidate variables. For the first regression coefficients of candidate variables The absolute value of.

6. The method for real-time risk warning and intervention of radiation-induced oral mucositis based on multi-parameter dynamic monitoring according to claim 1, characterized in that: In S3, the training steps for the linear mixed-effects longitudinal sub-model are as follows: First, the samples are divided into training and testing sets; Second, in the longitudinal sub-model of linear mixed effects, the covariates in the time and baseline variables at each monitoring time are used as independent variables, and the measured values ​​of the longitudinal variables at each monitoring time are used as dependent variables. Third, the covariates in the time and baseline variables at each monitoring point are used as inputs, and the measured values ​​of the longitudinal variables at each monitoring point are used as labels. These are substituted into the expression of the linear mixed-effects longitudinal sub-model to train the linear mixed-effects longitudinal sub-model, as shown in formula (6). （6）， In formula (6), For the first The first patient The longitudinal variable values ​​of this monitoring, For sample size, , , All are fixed effects weights. For monitoring time, The initial values ​​of the vertical variables, , Let be the posterior distribution of the random effect. For random intercept, For random slope, This is the error term; Fourth, the test set is fed into the trained linear mixed effects longitudinal sub-model. When the root mean square error (RMSE) between the predicted outcome obtained by the test set through the linear mixed effects longitudinal sub-model and the actual outcome of the test set is less than or equal to the threshold, the linear mixed effects longitudinal sub-model stops optimizing. When the RMSE between the predicted outcome obtained by the test set through the linear mixed effects longitudinal sub-model and the actual outcome of the test set is greater than the threshold, the linear mixed effects longitudinal sub-model starts optimizing until the RMSE is less than or equal to the threshold.

7. The method for real-time risk warning and intervention of radiation-induced oral mucositis based on multi-parameter dynamic monitoring according to claim 6, characterized in that: The optimization of the longitudinal sub-model for linear mixed effects includes the following: The residual distribution plot is calculated based on the difference between the predicted results of the test set and the actual outcome of the linear mixed-effects longitudinal sub-model. If heteroscedasticity exists, a weighting function is introduced into the linear mixed-effects longitudinal sub-model, as shown in Equation (7), to correct the uneven error distribution and thus optimize the model, as shown in Equation (8). （7）， （8）， In formulas (7) and (8), For the first The first patient The weighting of each follow-up visit. For the first The first patient Residual variance estimates for each follow-up period; This is the weighted error term; For fixed-effects intercept, This is the fixed-effects coefficient for follow-up time. The fixed effects coefficients of the baseline covariates. For the first The first patient Follow-up time, For the first The first patient's One baseline covariate, For random intercept, For random slope, This is the original error term; If the residual analysis indicates that the variables change nonlinearly over time, a time quadratic term and a "time × baseline covariate" interaction term are added to formula (8) to adjust the model structure to adapt to the nonlinear relationship. The optimized complete formula is as follows: （9）， In formula (9), The coefficient of the quadratic term over time is used to capture the nonlinear trend of the longitudinal variable over time. These are the interaction term coefficients, used to characterize the influence of baseline covariates on the rate of change of longitudinal variables.

8. The method for real-time risk warning and intervention of radiation-induced oral mucositis based on multi-parameter dynamic monitoring according to claim 1, characterized in that: In S3, the linear mixed-effects longitudinal sub-model and the Cox regression survival sub-model achieve data association and collaborative modeling through the posterior distribution of the random effects of the samples. The Cox regression survival sub-model uses the dependent variable of the linear mixed-effects longitudinal sub-model to update the risk, and the linear mixed-effects longitudinal sub-model uses event information from the Cox regression survival sub-model to optimize trajectory prediction, including the following steps: First, the prior distribution of individual random effects is a multivariate normal distribution, as shown in formulas (11) and (12). （11）， （12）， In formulas (11) and (12), For the first Individual patients share a random effects vector. The random intercept term reflects the first... Individual differences in the initial levels of longitudinal variables for each patient, and the fixed-effects intercept in the linear mixed-effects longitudinal sub-model. Collaboration forms the basis for initial level predictions of longitudinal variables; The random slope term reflects the first... Individual differences in the rate of change of longitudinal variables over follow-up time for each patient, and the time-fixed effect coefficient in the linear mixed-effects longitudinal sub-model. Collaboratively capture the dynamic trajectory of longitudinal variables. The symbol for vector transpose. The random effects covariance matrix; Second, the posterior distribution of the random effects is updated using Bayes' theorem, as shown in formula (13). （13）， In formula (13), Let be the posterior distribution of the random effect. The series of measured values ​​of longitudinal variables for each patient. As the dependent variable, Let be the likelihood function of the Cox regression survival sub-model. For the first Baseline risk factor vector for each patient Let be the likelihood function of the longitudinal sub-model of the linear mixed effects. Let i be the baseline covariate vector for the i-th patient. This represents the prior distribution of the random effect; The Markov chain Monte Carlo sampling method, defined by formula (13), is executed using the update() function of the JMbayes2 package. In each iteration, the posterior distribution of the random effects is updated according to the following rules, as shown in formula (15). （15）， In formula (15), For the number of iterations, The value is taken for the random effect vector in the (s+1)th iteration; This is the specific sampling form of the posterior distribution of formula (13), where , The first The estimated values ​​of the fixed effects coefficients and random effects covariance matrix for each iteration; When the Cox regression survival submodel observed that the patient did not develop severe radiation-induced oral mucositis, the posterior mean of the random effects was downregulated, as shown in formula (16). （16）， In formula (16), This represents the final updated posterior distribution of the random effects. The random effects prior mean before the update. To adjust the coefficient, This is the partial derivative of the likelihood function of the Cox regression survival submodel with respect to random effects; When the Cox regression survival submodel observed severe radiation-induced oral mucositis, the posterior mean of the random effects was upregulated, as shown in formula (17). （17）， In formula (17), This represents the final updated posterior distribution of the random effects; Third, based on the updated random effects posterior distribution, the predicted future trajectories of the longitudinal variables in the linear mixed effects longitudinal sub-model are corrected, as shown in Equation (14). （14）， In formula (14), For the first Patients at a future time point The predicted values ​​after correction for the longitudinal variables, For the first One patient Transpose of the fixed-effects covariate vector at time points. , For the first One patient Design matrix transpose for random effects at time points. This represents the posterior mean of the random effects after updating event information based on the Cox regression survival sub-model. For a future time point t ∗ The estimated value of the prediction error, , For a future time point t ∗ The residual variance estimate; The Cox regression survival sub-model uses the posterior mean obtained in the linear mixed-effects longitudinal sub-model as the random effect and calculates the hazard ratio. At the same time, it integrates the random effect into the hazard rate formula, clarifying the correlation logic between the hazard ratio and the hazard rate, as shown in formulas (18), (19), and (20). （18）， In formula (18), The change in baseline risk factors that conforms to the PH assumption. To violate the PH assumption, the change in the variable This represents the change in the posterior mean of the random effect; （19）， （20）， In formula (19), the benchmark risk rate is used. Risk ratio Multiplying these values ​​yields the real-time risk rate for the target patient, which can be used to classify patients into low, medium, and high risk levels. In formula (20), For the first The patient in time The dynamic risk rate As the benchmark risk rate, This is the vector of regression coefficients for baseline risk factors that conform to the proportional risk assumption. For the time-dependent covariate regression coefficient vector of the variable that violates the PH hypothesis, For the risk factor vector that violates the PH assumption, This is the vector of weight coefficients for random effects. The individual dynamic trajectory contribution term is passed to the longitudinal sub-model of linear mixed effects, enabling collaborative modeling of the two sub-models.

9. The method for real-time risk warning and intervention of radiation-induced oral mucositis based on multi-parameter dynamic monitoring according to claim 1, characterized in that: In S4, when the risk level predicted by the model calculation module is low risk, the risk warning push module generates low risk warning information, sends the risk level to the medical staff, and sends routine nursing reminders to the patient. When the risk level predicted by the model calculation module is medium risk, the risk warning push module generates medium risk warning information, sends the risk level to the medical staff, pushes a reminder to the medical staff, and sends enhanced nursing guidance prompts to the patient. When the model calculation module predicts a high-risk level, the risk warning push module generates a high-risk warning message and sends the risk level to the medical staff. At the same time, it immediately triggers an emergency warning on the medical staff and pushes a reminder to the medical staff, and sends an emergency intervention prompt to the patient.

10. A real-time risk warning and intervention system for radiation-induced oral mucositis based on multi-parameter dynamic monitoring, characterized in that: The method for real-time risk warning and intervention of radiation-induced oral mucositis based on multi-parameter dynamic monitoring as described in any one of claims 1-9 includes a multi-parameter dynamic acquisition module, a data processing module, a model calculation module, a risk warning push module, an intervention feedback module, and a storage module. The multi-parameter dynamic acquisition module is used to collect patients' sociodemographic data, health-related data, physiological-related data, disease-related data, treatment-related data, and psychological and nutritional-related data in real time through the electronic medical record interface, medical staff terminal, and patient terminal. The data processing module is used to preprocess the collected parameters, and the model calculation module is used to dynamically predict the probability of a patient developing severe radiation-induced oral mucositis within the next 1 to 3 months based on the preprocessed parameters, and generate the corresponding risk level. The risk warning push module is used to generate corresponding warning information based on the risk level predicted by the model calculation module, and send the warning information to the medical staff terminal and corresponding personalized intervention prompts to the patient terminal; The intervention feedback module is used to collect feedback data after intervention through medical staff and patients, and the storage module is used to store the feedback data for easy retrieval by medical staff later.

Images