An artificial intelligence-based safety evaluation method for pharmaceutical clinical trials

Abstract

The application discloses a kind of based on artificial intelligence's pharmaceutical clinical test safety evaluation method, comprising the following steps: collecting heterogeneous clinical survival data with deletion characteristics and time series drift characteristics, and constructing original survival data set;Based on propensity score model and survival outcome model, implement deletion correction, obtain individual non-compliance risk score;Through stratified weighting rule and two-side order-preserving prediction mechanism, calibrate and update individual risk prediction interval;Calculate individual risk upper and lower bound probability, define three-domain risk threshold rule, generate real-time risk decision result through three-domain decision mapping and dynamically evaluate overall safety trend.The application improves the robustness and interpretability of clinical safety risk decision.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, and in particular to an artificial intelligence-based method for safety assessment of pharmaceutical clinical trials. Background Technology

[0002] Pharmaceutical clinical trials are a crucial step in evaluating drug safety and efficacy. Their core task is to quantitatively assess and dynamically monitor potential safety risks based on multidimensional clinical data from individual subjects. With the increasing prevalence of multicenter collaborative clinical trials, clinical data exhibits high dimensionality, heterogeneity, and censoring characteristics. Subject outcomes are often influenced by multiple factors, leading to increased complexity and uncertainty in safety risk assessment. Currently used safety assessment methods mainly include survival analysis models based on statistical regression and risk prediction models based on machine learning. Models such as random forests, support vector machines, and deep neural networks can uncover the relationship between features and risks under nonlinear conditions. However, in the presence of censored temporal features and heterogeneous multicenter data, their output results are difficult to interpret, and risk scores lack causal robustness.

[0003] In multicenter pharmaceutical clinical trials, differences in sample size, inclusion criteria, and follow-up periods across research centers result in significantly non-independent and identically distributed data distributions, making models prone to risk bias when predicting across centers. Existing methods generally lack effective stratified weighting and interval calibration mechanisms, making it difficult to ensure consistency in risk estimation results across different centers. Furthermore, in situations where censored data is widespread, survival analysis methods often employ censored sample removal or parameter assumption compensation, leading to systematic biases in risk estimation and affecting the reliability of clinical safety assessments. While current AI-based safety assessment systems possess high fitting capabilities, they cannot quantify the uncertainty of prediction results, making it difficult to provide interpretable and coverage-assured risk intervals for regulatory review, and also unable to achieve near real-time updates of individual-level safety status during clinical trials.

[0004] Therefore, how to provide an artificial intelligence-based method for safety assessment of pharmaceutical clinical trials is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an artificial intelligence-based method for safety assessment of pharmaceutical clinical trials. Addressing the problem that existing clinical trial safety assessments, under conditions of multi-center heterogeneous data and censored time series, cannot obtain individual-level risk assessment results with causal robustness, interpretability, and uncertainty quantification, this invention proposes an AI assessment method based on dual-model censoring correction, hierarchical weighting, and two-sided order-preserving prediction. This invention offers the advantages of guaranteed interval coverage, high causal robustness, and interpretability for review in risk prediction results.

[0006] According to an embodiment of the present invention, a method for safety assessment of pharmaceutical clinical trials based on artificial intelligence includes: Heterogeneous clinical survival data with censoring and temporal drift characteristics were collected from multiple pharmaceutical clinical trial research centers to generate the original survival dataset; Based on the original survival dataset, a propensity score model and a survival outcome model were constructed respectively. The survival outcome model was then censored using the propensity score model to obtain the individual non-consistency risk score. Based on the heterogeneity and distribution differences of clinical data from different research centers, a stratified weighting rule was defined, and a two-sided order-preserving prediction mechanism was used to independently calibrate the individual non-compliance risk scores of different research centers to generate individual risk prediction intervals. A sliding window mechanism is adopted to continuously update the calibration parameters of the order-preserving prediction mechanism on both sides online, generating a real-time updated individual risk prediction range; Based on the real-time updated individual risk prediction interval, the upper and lower bound probabilities of risk for each subject are calculated to obtain individual risk assessment indicators; Based on individual risk assessment indicators, three-domain risk threshold rules are defined, and real-time individual risk decision results are obtained through three-domain decision mapping. Based on real-time decision-making results regarding individual risks, the overall safety trend of clinical trials is dynamically assessed, and corresponding safety warnings are automatically triggered.

[0007] Optionally, the original survival dataset specifically includes: Multi-source heterogeneous data, including individual baseline characteristics, medication regimens, follow-up records, outcome events, and censoring markers, were collected from multiple pharmaceutical clinical trial research centers. Unified encoding and feature alignment processing are performed on multi-source heterogeneous data to map the variable naming, measurement units and data formats of different centers to a unified feature template; Based on the timestamp information of the data collection time, the follow-up records of each subject were sorted and completed in chronological order. For time segments with missing or abnormal data, a hybrid correction strategy based on nearest neighbor imputation and survival time consistency constraints was used to complete the data. A deletion indication matrix is ​​established based on deletion identifiers and event occurrence status. Deletion samples are labeled with deletion type, and deletion information is structured and encoded. After completing the structured preprocessing of the data from each center, the data is indexed and merged according to individual identifiers, and a unified multidimensional time-series representation structure is formed using center identifier, individual number, survival time, censoring identifier, feature vector, and outcome label to form the original survival dataset.

[0008] Optionally, the construction of the propensity score model and the survival outcome model respectively is as follows: Based on individual baseline characteristics and medication regimens in the original survival dataset, a data input vector for a propensity score model is constructed. Based on the processed data input vector, a propensity scoring model is constructed using a multilayer perceptron neural network. The propensity scoring model includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is the same as the feature dimension of the data input vector. The hidden layer contains two fully connected neural network layers. The number of nodes in the hidden layer decreases layer by layer. The output layer is a logistic regression unit that outputs an individual probability propensity score. Based on the survival time, censoring markers, feature vectors, and outcome events contained in the original survival dataset, construct the data input vector for the survival outcome model; Based on the data input vector of the survival outcome model, a deep recurrent neural network is used to construct the survival outcome model. The model structure of the survival outcome model includes an input layer, a feature encoding layer, a bidirectional long short-term memory network layer, and a survival risk output layer. The number of nodes in the input layer corresponds to the feature dimension of the data input vector. The feature encoding layer performs nonlinear feature mapping on the input vector. The bidirectional long short-term memory network layer extracts the long-term dependency features of the input vector. The survival risk output layer uses a fully connected network structure to map the initial risk probability of the individual's survival event. The inverse probability weighting factor for each subject is calculated based on the individual propensity score, and the initial risk probability is weighted and corrected to generate the censored survival risk probability. The individual's non-compliance risk score is calculated based on the survival risk probability after censoring correction.

[0009] Optionally, the hierarchical weighting rule and the two-sided order-preserving prediction mechanism are specifically as follows: The hierarchical weighting rule is as follows: based on the heterogeneity, sample size and censoring ratio of data from each pharmaceutical clinical trial research center, multiple data levels are defined, and different weight coefficients are assigned to each data level. The two-sided order-preserving prediction mechanism is as follows: calculate the upper and lower quantile boundaries of the non-compliance risk score of each individual in each data level, and map the quantile boundaries of each data level to a unified standard quantile space to form standardized quantile anchors; The upper and lower bounds of the individual non-compliance risk score intervals within each data level are calibrated using quantile anchor points to maintain order. A hierarchical weighting rule is used to weight and fuse the upper and lower bounds of each data level after order preservation calibration to obtain the fused individual risk prediction range.

[0010] Optionally, the calibration parameters of the two-sided order-preserving prediction mechanism are updated online, specifically as follows: The newly acquired survival data collected continuously during pharmaceutical clinical trials is periodically extracted using a preset fixed-length time window to form a sliding data subset; Within each sliding window period, the original survival dataset is updated based on the newly added sliding data subset, and the upper and lower quantile boundaries of the individual non-compliance risk scores are recalculated to generate the latest quantile anchor points. Based on the latest calculated quantile anchor points, the upper and lower bounds of the individual non-compliance risk score intervals of each data level in the current sliding data subset are recalibrated using monotonic order preservation. The calibrated upper and lower bounds are then re-integrated using hierarchical weighted fusion to update the individual risk prediction intervals for each subject. The individual risk prediction interval obtained in the current sliding window period is compared with the individual risk prediction interval obtained in the previous sliding window period. Based on the degree and direction of the change in the upper and lower bounds of the individual risk prediction intervals in the two periods, the calibration parameters of the order-preserving prediction mechanism on both sides of the next sliding window period are adjusted.

[0011] Optionally, the calculation of the upper and lower bound probabilities of risk for each individual subject is specifically as follows: Based on the upper and lower bounds of the individual risk prediction interval, a unified probability density distribution function is constructed according to the numerical distribution of the individual risk prediction intervals of all subjects. Substitute the upper and lower bounds of the individual risk prediction interval for each subject into the probability density distribution function to calculate the probability density values ​​corresponding to the upper and lower bounds of the risk prediction interval for each subject. The upper bound probability density value of the risk prediction interval for each individual subject is integrated, and the result is used as the upper bound probability of the individual risk. The lower bound probability density value of the risk prediction interval for each individual subject is integrated, and the result is used as the lower bound probability of the individual risk. The calculated upper and lower bound probabilities of risk for each subject are used as the maximum and minimum probability estimates of the individual's risk occurrence, respectively, and are used as indicators for individual risk assessment.

[0012] Optionally, the three-domain risk threshold rule is specifically as follows: Based on the overall numerical distribution of the upper bound probability and lower bound probability of individual risk for all subjects, a first risk threshold and a second risk threshold are set, and the subjects are divided into three non-overlapping decision domains: the safety domain, the concern domain, and the risk domain. The rule for defining the safe domain is as follows: when the upper bound probability of an individual's risk is lower than the first risk threshold, it indicates that the individual has a low risk of experiencing a survival event, and the individual is classified into the safe domain. The rules for defining the domain of concern are as follows: when the upper bound probability of an individual's risk is higher than or equal to the first risk threshold and the lower bound probability of an individual's risk is lower than the second risk threshold, it indicates that there is uncertainty in the risk of an individual experiencing a survival event. The individual is then classified into the domain of concern and risk monitoring is implemented for the individual. The rule for defining the risk domain is as follows: when the lower bound probability of an individual's risk is higher than or equal to the second risk threshold, it indicates that the individual has a high and certain risk of experiencing a survival event. The individual is then classified into the risk domain, and corresponding safety warnings and risk interventions are triggered.

[0013] Optionally, obtaining the real-time individual risk decision result through three-domain decision mapping specifically involves: Based on the upper bound probability and lower bound probability of individual risk for each subject, the numerical distance between the corresponding individual risk probability and the first risk threshold and the second risk threshold is calculated to obtain the distance difference. The mapping strength function for risk decision-making is defined based on the magnitude and direction of the distance difference; The membership scores of each subject to the safety domain, concern domain, and risk domain are calculated based on the mapping strength function, and the domain to which the individual's real-time risk decision-making result belongs is determined. When the highest membership score of a subject corresponds to the safety domain, the individual's real-time risk decision-making result is determined as a safe state. When the maximum membership score corresponds to the domain of concern, the individual risk real-time decision-making result is determined as a state of concern, and real-time risk monitoring is triggered; When the maximum membership score corresponds to the risk domain, the individual risk real-time decision result is determined as the risk status, and real-time safety warnings and risk intervention measures are triggered. The real-time decision results are smoothed over time to obtain real-time decision results for individual risks.

[0014] Optionally, the dynamic assessment of the overall safety trend of the clinical trial specifically includes: Based on the real-time risk decision results of all subjects in each assessment period, the number of subjects classified into the safety domain, concern domain, and risk domain, as well as their proportion in the total number of subjects, are counted to obtain the proportion of individuals in each decision domain. Calculate the numerical change in the proportion of individuals in each decision domain between the current evaluation period and the previous evaluation period to obtain the trend of the proportion change in each decision domain. A safety trend index is defined based on the trend of percentage change, and a safety trend index that characterizes the overall safety trend change of clinical trials is formed by weighted calculation. By comparing the current safety trend index with the preset safety trend threshold, when the safety trend index is lower than the preset safety trend threshold, it is determined that the overall safety trend of the clinical trial has deteriorated, and an overall risk trend warning is triggered.

[0015] The beneficial effects of this invention are: (1) This invention constructs a propensity score model and a survival outcome model, and generates an individual non-compliance risk score based on a dual-model censoring correction mechanism. This achieves accurate quantification of individual-level risk and effective correction of censored data, effectively improving the causal robustness of safety assessment in pharmaceutical clinical trials and enhancing the accuracy and reliability of risk prediction.

[0016] (2) This invention realizes real-time adjustment and dynamic optimization of individual risk prediction intervals through hierarchical weighted rules, two-sided order-preserving prediction mechanism and sliding window online update technology, which significantly improves the robustness and timeliness of risk prediction under multi-center heterogeneous data conditions, and shows better adaptability in real-time safety monitoring of pharmaceutical clinical trials.

[0017] (3) In terms of risk prediction uncertainty quantification and review interpretability, this invention effectively solves the problem of insufficient reliability of safety decision threshold under multi-center censoring time series conditions by defining three-domain risk threshold rules and three-domain decision mapping mechanism. It breaks through the limitations of lack of causal robustness and review interpretability of risk prediction results, achieves significant progress in pharmaceutical clinical trial safety assessment technology, and effectively improves the quality of clinical trial safety review decision. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the overall process of a pharmaceutical clinical trial safety assessment method based on artificial intelligence proposed in this invention. Figure 2 This is a structural diagram of the stratified weighting rule and the two-sided order-preserving prediction mechanism in the artificial intelligence-based pharmaceutical clinical trial safety assessment method proposed in this invention; Figure 3 This is a diagram illustrating the three-domain risk threshold rules and real-time decision mapping process in an artificial intelligence-based pharmaceutical clinical trial safety assessment method proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 An artificial intelligence-based method for safety assessment in pharmaceutical clinical trials, comprising: Heterogeneous clinical survival data with censoring and temporal drift characteristics were collected from multiple pharmaceutical clinical trial research centers to generate the original survival dataset; Based on the original survival dataset, a propensity score model and a survival outcome model were constructed respectively. The survival outcome model was then censored using the propensity score model to obtain the individual non-consistency risk score. Based on the heterogeneity and distribution differences of clinical data from different research centers, a stratified weighting rule was defined, and a two-sided order-preserving prediction mechanism was used to independently calibrate the individual non-compliance risk scores of different research centers to generate individual risk prediction intervals. A sliding window mechanism is adopted, with a preset window length as the period, to continuously update the calibration parameters of the order-preserving prediction mechanism on both sides online, generating a real-time updated individual risk prediction range; Based on the real-time updated individual risk prediction interval, the upper and lower bound probabilities of risk for each subject are calculated to obtain individual risk assessment indicators; Based on the definition of individual risk assessment indicators, risk threshold rules are defined in three domains: safety, concern, and risk. Real-time decision-making results for individual risk are obtained through three-domain decision mapping. Based on real-time decision-making results regarding individual risks, the overall safety trend of clinical trials is dynamically assessed, and corresponding safety warnings are automatically triggered.

[0021] In this embodiment, the original survival dataset specifically refers to: Multi-source heterogeneous data, including individual baseline characteristics, medication regimens, follow-up records, outcome events, and censoring markers, were collected from multiple pharmaceutical clinical trial research centers. Unified encoding and feature alignment processing are performed on multi-source heterogeneous data to map the variable naming, measurement units and data formats of different centers to a unified feature template; Based on the timestamp information of the data collection time, the follow-up records of each subject were sorted and completed in chronological order. For time segments with missing or abnormal data, a hybrid correction strategy based on nearest neighbor imputation and survival time consistency constraints was used to complete the data. The hybrid correction strategy based on nearest neighbor imputation and survival time consistency constraints refers to the following steps when data loss or anomalies are found in the follow-up records of subjects. The strategy involves calculating the similarity between the feature vector of the missing data point and other complete data individuals, selecting the nearest neighbor data with the smallest feature distance for preliminary imputation, and verifying and correcting the imputation results based on the difference in survival time between the corresponding subjects at the effective observation time points before and after. For missing or abnormal data segments in the follow-up records of each subject, based on the valid records corresponding to the upstream and downstream time nodes of the missing segment, the Euclidean distance between the feature vectors of the complete time window before and after the missing segment and the feature vectors of other individuals in the corresponding time period is calculated, and the individuals with the smallest distance are selected as the nearest neighbors. The preliminary interpolation results are generated by averaging or weighting the feature data of the time segments corresponding to the nearest neighbor individuals. After interpolation is completed, the interpolated data is verified. Specifically, by comparing the difference in lifetime between the interpolated data points and the existing valid data points in the upstream and downstream, the rationality of the interpolated data in the overall lifetime time series is judged. If a logical conflict is found, the interpolated data points are corrected in the lifetime time direction. A deletion indication matrix is ​​established based on deletion identifiers and event occurrence status. Deletion samples are labeled with deletion types, and deletion information is structured and encoded through the joint distribution of deletion time and survival time. The aforementioned censoring marker refers to a status marker used in the follow-up of pharmaceutical clinical trials to indicate whether the clinical outcome event of interest occurred in the subject or whether the clinical outcome event was not observed due to loss to follow-up, withdrawal midway, observation termination, etc. It is expressed in a binary manner, with 1 being recorded when the clinical outcome event occurred and 0 being recorded when it did not occur (censoring). The method for establishing the censoring indicator matrix is ​​as follows: the follow-up records of all subjects in the original survival dataset are uniformly identified, and the survival data corresponding to each subject is clearly distinguished into two states: event occurrence and event non-occurrence. The non-occurrence of events is further refined into censoring situations where the event did not occur due to the end of the observation and events could not be observed due to dropout or loss to follow-up. The matrix is ​​structured and encoded in the form of a two-dimensional matrix. The rows of the matrix represent different subjects, and the columns of the matrix represent various time nodes or observation windows of the follow-up. Each element is filled with the state identifier corresponding to the subject at that time, i.e., the event has occurred, the event has not occurred and the follow-up is normal, and the event has not occurred and the event has been censored, thus forming a structured and standardized censoring indicator matrix. After completing the structured preprocessing of the data from each center, the data is indexed and merged according to individual identifiers, and a unified multidimensional time-series representation structure is formed using center identifier, individual number, survival time, censoring identifier, feature vector, and outcome label to form the original survival dataset.

[0022] In this embodiment, the construction of the propensity score model and the survival outcome model respectively is specifically as follows: Based on individual baseline characteristics and medication regimens in the original survival dataset, a data input vector for a propensity score model is constructed. The baseline features and medication regimens recorded by each subject at the beginning of the trial were extracted from the original survival dataset. The extracted features were then subjected to data type conversion and numerical normalization. Categorical features were vectorized using one-hot encoding, and numerical features were normalized. The processed feature data were then concatenated in a uniform order to form a data input vector with fixed dimensions, consistent data structure, and suitable for input to the bias scoring model. Based on the processed data input vector, a propensity scoring model is constructed using a multilayer perceptron neural network. The propensity scoring model includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is the same as the feature dimension of the data input vector. The hidden layer contains two fully connected neural network layers. The number of nodes in the hidden layer decreases layer by layer. The output layer is a logistic regression unit, which outputs an individual propensity score for accepting a specific treatment plan. The process of constructing the propensity scoring model is as follows: The processed data input vector is used as the input signal of the input layer. Each feature dimension corresponds to a node in the input layer. The signal is passed from the input layer to the hidden layer. The hidden layer uses two fully connected neural network layers to perform non-linear feature transformation. The number of nodes in each hidden layer decreases layer by layer. A non-linear activation function is used to achieve multiple non-linear combinations and feature mappings of the input features. The feature mapping vector output by the hidden layer is passed to the logistic regression unit of the output layer. The individual probability propensity score between 0 and 1 is obtained by transforming it through the sigmoid function. Based on the survival time, censoring markers, feature vectors, and outcome events contained in the original survival dataset, construct the data input vector for the survival outcome model; Survival time, censoring flag, feature vector, and outcome event features for each subject were extracted from the original survival dataset. Survival time is the continuous observation duration from the start of the experiment to the occurrence or censoring time of the event. Censoring flag indicates whether the event has been observed or censored. Feature vector includes the individual's baseline features and medication regimen features. Outcome event indicates the occurrence status of the event of interest in the experiment. Each extracted feature was standardized and coded. Continuous variables were numerically normalized, and categorical variables were one-hot encoded. Survival time, censoring flag, feature vector, and outcome event features were concatenated in a predetermined order to form a complete data input vector. Based on the data input vector of the survival outcome model, a deep recurrent neural network is used to construct the survival outcome model. The model structure of the survival outcome model includes an input layer, a feature encoding layer, a bidirectional long short-term memory network layer, and a survival risk output layer. The number of nodes in the input layer corresponds to the feature dimension of the data input vector. The feature encoding layer performs nonlinear feature mapping on the input vector. The bidirectional long short-term memory network layer extracts the long-term dependency features of the input vector. The survival risk output layer uses a fully connected network structure to map the initial risk probability of the individual's survival event. The process of constructing the survival outcome model is as follows: The processed data input vector is input into the model's input layer. Each input node corresponds to a feature dimension in the data input vector. The input layer passes the feature signal to the feature encoding layer. The feature encoding layer performs feature transformation and high-dimensional mapping on the input features through a non-linear activation function. The output features of the feature encoding layer are input into a bidirectional long short-term memory network layer. The bidirectional long short-term memory network layer performs cyclic operations along the forward and backward directions of the data sequence. It captures and stores long-term temporal dependency information in the individual data sequence through a gating mechanism. The hidden states of the forward and backward outputs of the bidirectional long short-term memory network layer are concatenated to form a comprehensive temporal feature expression. The comprehensive temporal feature expression is then passed to the survival risk output layer. A fully connected network structure is used to perform non-linear mapping on the input comprehensive temporal features to generate the initial risk probability of the individual survival event. The inverse probability weighting factor for each subject is calculated based on the individual propensity score, and the initial risk probability is weighted and corrected to generate the censored survival risk probability. The method for calculating the inverse probability weighting factor is as follows: The individual propensity score for each subject to receive a specific treatment plan, output by the propensity score model, is used as the basis. For each individual subject's actual treatment status, if the individual subject actually received a specific treatment plan, the inverse probability weighting factor is calculated as the reciprocal of the individual's probability propensity score. That is, the smaller the individual's probability propensity score, the larger the inverse probability weighting factor. If the subject did not actually receive the specific treatment plan, the inverse probability weighting factor is calculated as the reciprocal of the result after subtracting the individual's probability propensity score from 1. That is, the larger the individual's probability propensity score, the larger the inverse probability weighting factor. The process of weighted correction of the initial risk probability is as follows: The initial risk probability is multiplied one by one with the inverse probability weighting factor of the corresponding individual subject, so that the initial risk probability of the individual is adjusted according to the size of the weighting factor. Normalization is used to ensure that all weighted risk probability values ​​are within a reasonable probability range, eliminating the bias caused by censored data and confounding factors on the initial risk probability estimate, and thus obtaining the individual survival risk probability corrected for censoring bias. Calculate the individual's non-compliance risk score based on the censored survival risk probability. The calculation logic of the individual non-compliance risk score is as follows: obtain the weighted and corrected survival risk probability of each subject, take the actual observed survival outcome event status of the subject as the reference benchmark, quantify the degree of difference between the individual's survival risk probability and the actual observed outcome event, calculate the numerical difference between the two, and map the obtained difference to the non-negative continuous score domain through monotonic transformation. The obtained continuous score is defined as the individual non-compliance risk score.

[0023] In this embodiment, the hierarchical weighted rule and the two-sided order-preserving prediction mechanism are specifically as follows: The hierarchical weighting rule is as follows: based on the heterogeneity, sample size and censoring ratio of data from each pharmaceutical clinical trial research center, multiple data levels are defined, and different weight coefficients are assigned to each data level. Data layers with low heterogeneity, large sample size, and low censoring rate are assigned higher weight coefficients, while data layers with high heterogeneity, small sample size, and high censoring rate are assigned lower weight coefficients, so that the role of individual non-compliance risk scores in the fusion process is positively correlated with data quality. The two-sided order-preserving prediction mechanism is as follows: calculate the upper and lower quantile boundaries of the non-compliance risk score of each individual in each data level, and map the quantile boundaries of each data level to a unified standard quantile space to form standardized quantile anchors; The calculation logic for the upper and lower quantile boundaries is as follows: The non-compliance risk scores of all subjects in each data level are arranged in ascending order to form an increasing ordered numerical sequence. Based on the pre-set confidence level, the quantile positions corresponding to the lower and upper bounds in the non-compliance risk score sequence are determined respectively. The lower quantile boundary is the value corresponding to the lower percentile position in the risk score distribution, and the upper quantile boundary is the value corresponding to the higher percentile position in the risk score distribution. The upper and lower bounds of the individual non-compliance risk score intervals within each data level are calibrated using quantile anchors to maintain a monotonic order within the standard quantile space. Based on the obtained standardized quantile anchors, the upper and lower bounds of the non-compliance risk scores of all subjects in each data level are mapped to the corresponding standard anchor positions in the standard quantile space. A monotonically increasing mapping function is used to calibrate the score interval. The monotonically increasing mapping function satisfies the requirement of strictly maintaining a monotonically increasing relationship of values ​​within the standard quantile space. That is, the higher the original risk score value, the higher the standard anchor point position corresponding to the mapped value. During calibration, the monotonically increasing mapping function is applied sequentially to the upper and lower bounds of the non-compliance risk scores of each individual in each data level, so that the calibrated upper and lower bounds always maintain a monotonically ordered order in the standardized quantile space, without numerical crossover or position reversal. A hierarchical weighting rule is used to weight and fuse the upper and lower bounds of each data level after order preservation calibration to obtain the fused individual risk prediction range.

[0024] In this embodiment, the online updating of the calibration parameters for the two-sided order-preserving prediction mechanism specifically involves: The newly acquired survival data collected continuously during pharmaceutical clinical trials is periodically extracted using a preset fixed-length time window to form a sliding data subset; Within each sliding window period, the original survival dataset is updated based on the newly added sliding data subset, and the upper and lower quantile boundaries of the non-compliance risk scores of individuals in each research center are recalculated to generate the latest quantile anchor points. Based on the latest calculated quantile anchor points, the upper and lower bounds of the individual non-compliance risk score intervals of each data level in the current sliding data subset are recalibrated using monotonic order preservation. The calibrated upper and lower bounds are then re-integrated using hierarchical weighted fusion to update the individual risk prediction intervals for each subject. The individual risk prediction interval obtained in the current sliding window period is compared with the individual risk prediction interval obtained in the previous sliding window period. Based on the degree and direction of the change in the upper and lower bounds of the individual risk prediction intervals in the two periods, the calibration parameters of the order-preserving prediction mechanism on both sides of the next sliding window period are adjusted. Calculate the numerical difference between the upper and lower bounds of the individual risk prediction interval for each subject in the current sliding window period and the corresponding upper and lower bounds of the interval in the previous sliding window period, and determine the magnitude and trend direction of the change in the individual risk prediction interval in two consecutive periods. Set a change threshold, and judge whether the change in the individual risk range exceeds the change threshold based on the actual calculated change range. If the change range exceeds the threshold, it indicates that the individual risk prediction range has changed significantly. At this time, increase the adjustment step size of the calibration parameters of the order-preserving prediction mechanism on both sides. If the change does not exceed the threshold, it indicates that the individual risk prediction range is relatively stable. In this case, the adjustment step size of the calibration parameter should be reduced.

[0025] In this embodiment, the calculation of the upper and lower bound probabilities of the risk for each individual subject is specifically as follows: Based on the upper and lower bounds of the individual risk prediction interval, a unified probability density distribution function is constructed according to the numerical distribution of the individual risk prediction intervals of all subjects. The construction of the unified probability density distribution function is specifically as follows: Based on the upper and lower bounds of the risk prediction interval updated in real time for all subjects, a non-parametric kernel density estimation method is used to model the overall data distribution characteristics of the upper and lower bounds of the risk prediction interval. In the specific implementation, a Gaussian kernel function is used as the kernel function, and distance weights are assigned to data points near each risk interval value point. The bandwidth parameter is selected according to the overall data distribution. The probability density distribution function is obtained by calculating and normalizing the kernel function weighted sum of all data points in the risk prediction interval; Substitute the upper and lower bounds of the individual risk prediction interval for each subject into the probability density distribution function to calculate the probability density values ​​corresponding to the upper and lower bounds of the risk prediction interval for each subject. The calculation process for the probability density value is as follows: Substitute the upper and lower bounds of the risk prediction interval for each individual subject into the probability density distribution function; Based on the kernel density estimation principle used by the probability density distribution function, at each upper and lower bound value to be calculated, the overall probability density of the value point location is obtained by summing the distance-weighted contribution values ​​of all data points calculated by the kernel function. The above calculations were performed on the upper and lower bounds of the risk prediction interval for each individual subject to obtain the probability density values ​​corresponding to each upper and lower bound. The upper bound probability density value of the risk prediction interval for each individual subject is integrated. The integration is performed from the upper bound value of the individual risk prediction interval to the upper limit of the maximum value of the probability density distribution function, and the result is used as the upper bound probability of the individual risk. The lower bound probability density value of the risk prediction interval for each individual subject is integrated. The integration is performed from the lower limit of the minimum value of the probability density distribution function to the lower bound value of the individual risk prediction interval, and the result is used as the lower bound probability of the individual risk. The calculated upper and lower bound probabilities of risk for each subject are used as the maximum and minimum probability estimates of the individual's risk occurrence, respectively, and are used as indicators for individual risk assessment.

[0026] In this embodiment, the three-domain risk threshold rule is specifically as follows: Based on the overall numerical distribution of the upper bound probability and lower bound probability of individual risk for all subjects, a first risk threshold and a second risk threshold are set, and the subjects are divided into three non-overlapping decision domains: the safety domain, the concern domain, and the risk domain. The method for setting the first risk threshold and the second risk threshold is as follows: The upper bound probabilities of individual risks for all subjects are sorted from smallest to largest to obtain the numerical distribution sequence of the upper bound probabilities of individual risks. Based on the predetermined safety domain risk tolerance standard, the value corresponding to the 25th percentile in the sorting sequence is selected as the first risk threshold, indicating that individuals below the threshold have a lower risk. The lower bound probabilities of individual risks for all subjects are sorted from smallest to largest to obtain the numerical distribution sequence of the lower bound probabilities of individual risks. Based on the predetermined risk domain intervention criteria, the value corresponding to the 75th percentile in the ranking sequence is selected as the second risk threshold, indicating that individuals above the threshold have a high and certain risk. The determined first and second risk thresholds were evaluated and verified by clinical experts, and the two thresholds were fine-tuned and optimized based on the actual needs of clinical safety monitoring. The rule for defining the safe domain is as follows: when the upper bound probability of an individual's risk is lower than the first risk threshold, it indicates that the individual has a low risk of experiencing a survival event, and the individual is classified into the safe domain. The rules for defining the domain of concern are as follows: when the upper bound probability of an individual's risk is higher than or equal to the first risk threshold and the lower bound probability of an individual's risk is lower than the second risk threshold, it indicates that there is uncertainty in the risk of an individual experiencing a survival event. The individual is then classified into the domain of concern and risk monitoring is implemented for the individual. The rule for defining the risk domain is as follows: when the lower bound probability of an individual's risk is higher than or equal to the second risk threshold, it indicates that the individual has a high and certain risk of experiencing a survival event. The individual is then classified into the risk domain, and corresponding safety warnings and risk interventions are triggered.

[0027] In this embodiment, obtaining real-time individual risk decision results through three-domain decision mapping specifically involves: Based on the upper bound probability and lower bound probability of individual risk for each subject, the numerical distance between the corresponding individual risk probability and the first risk threshold and the second risk threshold is calculated to obtain the distance difference. The mapping strength function for risk decision-making is defined based on the magnitude and direction of the distance difference. The mapping strength function assigns a larger decision weight when the upper or lower bound probability of an individual risk is far from the risk threshold, and a smaller decision weight when the upper or lower bound probability of an individual risk is close to the risk threshold. The membership score of each subject to the safe domain, concern domain, and risk domain is calculated based on the mapping strength function, and the domain to which the individual's real-time risk decision-making result belongs is determined by the magnitude of the membership score. When the highest membership score of a subject corresponds to the safety domain, the individual's real-time risk decision-making result is determined as a safe state. When the maximum membership score corresponds to the domain of concern, the individual risk real-time decision-making result is determined as a state of concern, and real-time risk monitoring is triggered; When the maximum membership score corresponds to the risk domain, the individual risk real-time decision result is determined as the risk status, and real-time safety warnings and risk intervention measures are triggered. The method for calculating the membership degree score is as follows: Based on the distance differences between the upper and lower bound probabilities of individual risk and the first and second risk thresholds, respectively, the mapping strength function is substituted into the mapping calculation. The initial mapping weights of individuals in the three domains of safety, concern, and risk are obtained through the mapping strength function. The initial mapping weights of individuals in each domain are normalized, and the membership scores of individuals to the three decision domains are calculated respectively. The real-time decision results are smoothed over time. Specifically, the real-time decision results are combined with the real-time decision results of multiple consecutive periods, and the stability of the current real-time decision results is corrected by using a time-weighted voting mechanism to prevent frequent fluctuations in individual risk decision results and obtain individual risk real-time decision results.

[0028] In this embodiment, the dynamic assessment of the overall safety trend of the clinical trial specifically refers to: Based on the real-time risk decision results of all subjects in each assessment period, the number of subjects classified into the safety domain, concern domain, and risk domain, as well as their proportion in the total number of subjects, are counted to obtain the proportion of individuals in each decision domain. Calculate the numerical change in the proportion of individuals in each decision domain between the current evaluation period and the previous evaluation period to obtain the trend of the proportion change in each decision domain. The safety trend index is defined based on the trend of percentage change. The specific calculation method of the safety trend index is as follows: the increase of the safety domain percentage is given a positive contribution, the increase of the risk domain percentage is given a negative contribution, and the change of the concern domain percentage is given a neutral weighted contribution according to the direction and magnitude. The safety trend index is formed by weighted calculation to represent the overall safety trend change of the clinical trial. By comparing the current safety trend index with the preset safety trend threshold, when the safety trend index is lower than the preset safety trend threshold, it is determined that the overall safety trend of the clinical trial has deteriorated, and an overall risk trend warning is triggered. Example 1

[0029] To verify the feasibility of this invention in practice, it was applied to the design and actual operation of a pharmaceutical clinical trial safety assessment system. This was to verify whether, under conditions of multi-center heterogeneous data and censoring, the invention could achieve near real-time quantitative prediction of individual-level safety risks and stable monitoring of overall safety trends. The implementation environment included 1680 pharmaceutical clinical trial samples from different research centers, covering various drug classes and dosing regimens. Clinical data included individual baseline characteristics, medication regimens, follow-up records, outcome events, and censoring markers, with a censoring rate of approximately 27.3%. In this application scenario, survival analysis primarily relied on the Cox proportional hazards model and its censored extension model. Due to the non-independent and identically distributed characteristics of the data in a multi-center distribution, the model exhibited significant bias in cross-center predictions; furthermore, the removal of censored samples caused information loss, leading to an underestimation of risk. In addition, while existing neural network-based risk prediction models can fit complex feature relationships, they cannot robustly adjust for censored data, and the prediction results lack interval interpretability, failing to support regulatory review and individual safety intervention needs.

[0030] In the specific implementation of this invention, clinical trial data uploaded by various centers are first aggregated through a multi-source heterogeneous data acquisition module. The system performs unified encoding and feature alignment on the collected multi-source data, mapping variable naming, units of measurement, and data formats from different centers to a unified template. Then, individual follow-up records are time-series-completed based on timestamp information. For missing or abnormal time periods, a hybrid correction strategy based on nearest neighbor imputation and survival time consistency constraints is used to complete the data, forming a structured original survival dataset. Next, the system constructs a propensity score model and a survival outcome model, calculating individual non-compliance risk scores through a dual-model censoring correction mechanism. The propensity score model adopts a three-layer multilayer perceptron structure, with 52 baseline and medication characteristic variables as input dimensions, 128 and 64 hidden layer nodes respectively, ReLU activation function, and a Sigmoid function output layer to generate an individual's propensity score for receiving treatment. The survival outcome model adopts a bidirectional long short-term memory network structure, with an input layer dimension corresponding to a time window step size of 12, 128 hidden layer units, and an output layer that calculates the initial risk probability of survival events through a fully connected layer. The inverse probability weighting factor for each subject is calculated based on the results of the two models, and the risk probability output by the survival outcome model is corrected accordingly to obtain the censored survival risk probability.

[0031] The system assigns weights to data from different centers based on their heterogeneity using a hierarchical weighting rule. Centers with low heterogeneity and large sample sizes have a weight of 0.35, those with moderate heterogeneity have a weight of 0.25, and those with high heterogeneity have a weight of 0.15. Within each stratum, the upper and lower quantile boundaries of the non-compliance risk score are calculated. Each quantile boundary is transformed to the standard quantile space using a quantile mapping function, and a monotonically increasing mapping function is used to perform monotonically order-preserving calibration of the upper and lower bounds to ensure that the risk intervals between individuals maintain a logical order. After calibration, the upper and lower bounds of the risk intervals at each stratum are weighted and merged to generate preliminary individual risk prediction intervals. During the real-time evaluation phase, the system uses a sliding window mechanism for online updates, with a window length of 72 hours. After each sliding period, the quantile anchor points are recalculated and the calibration parameters are updated. If the change in the upper and lower bounds of the risk interval exceeds 5% over two consecutive window periods, the calibration step size is adaptively adjusted to improve dynamic response speed; if the change is less than 1%, the step size is reduced to improve smoothness.

[0032] In the risk quantification phase, the system establishes a unified probability density distribution function for the risk prediction interval data of all subjects, and uses Gaussian kernel density estimation to smooth the upper and lower bound values, with a bandwidth parameter set to 0.08. Subsequently, the upper and lower bound values ​​of each individual are substituted into the probability density distribution function to calculate the corresponding probability density, and integration is performed along the upper and lower bounds of the interval to obtain the upper and lower bound probabilities of the individual's risk. Based on the risk probability distribution of all individuals, the system sets the first risk threshold as the 25th percentile (0.42) of the upper bound probability and the second risk threshold as the 75th percentile (0.73) of the lower bound probability, dividing individuals into a safe domain, a domain of concern, and a risk domain. For each subject, the system calculates the distance difference between the risk probability and the threshold, and calculates the weight of the mapping strength function based on the magnitude of the difference, further calculating the membership score. Finally, the real-time risk decision status of the individual is determined, and a weighted voting mechanism is used on the time series to smooth the output results to prevent periodic fluctuations.

[0033] The system was validated on three independent multicenter clinical datasets. Results showed that the proposed method maintained a risk interval coverage of 95.7% even with censoring exceeding 25%, an improvement of approximately 14.2% compared to the model without censoring correction. In the intercenter risk consistency test, the proposed method achieved an interval intersection-union ratio (IUU) of 0.86, higher than the Cox model (0.62) and the random survival forest model (0.73). The system-generated three-domain risk classification achieved an accuracy of 91.4%, a sensitivity of 88.9%, and a specificity of 93.7% on the validation set. In real-time sliding window updates, the average stability coefficient of individual risk intervals reached 0.94, indicating robust performance under dynamic conditions.

[0034] The table below shows a comparison of the censored individual risk prediction results with the actual clinical observation risk distribution in samples from different centers.

[0035] Table 1 Comparison of Individual Risk Prediction Ranges and Actual Risk Statuses from Different Research Centers

[0036] As shown in Table 1, the prediction results of the method of this invention exhibit high consistency across centers, with interval coverage exceeding 0.95 and average prediction error controlled within 0.06. The interval coverage of center C-03 reaches 0.97, with a consistency coefficient of 0.89, indicating that the model maintains stable risk prediction capabilities even under heterogeneous data conditions across centers. Through the sliding window online update mechanism, the system effectively captures the changing trends of safety risks over time, avoiding risk shifts caused by time-series drift. Compared with traditional risk models based on single-point estimation, the individual risk interval prediction method provided by this invention can quantify risk uncertainty, providing a more interpretive basis for safety decisions in pharmaceutical clinical trials. Practical application results show that when the system detects that the proportion of individuals in the risk domain exceeds 30% for two consecutive periods, it automatically triggers a safety trend warning, prompting the monitoring department to initiate intervention measures, thereby significantly reducing the incidence of clinical trial safety events.

[0037] The results of this embodiment verify the applicability and stability of the method of the present invention in a heterogeneous environment of multi-center clinical data. It demonstrates that the present invention can accurately quantify individual risk intervals and assess overall safety trends in real time through dual-model censoring correction, hierarchical weighted order-preserving prediction, and a sliding window online update mechanism. This method has good promotional value and promising technological application prospects in application scenarios such as safety supervision of pharmaceutical clinical trials, early warning of adverse drug reactions, and risk control of personalized treatment.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for safety assessment of pharmaceutical clinical trials based on artificial intelligence, characterized in that, include: Heterogeneous clinical survival data with censoring and temporal drift characteristics were collected from multiple pharmaceutical clinical trial research centers to generate the original survival dataset; Based on the original survival dataset, a propensity score model and a survival outcome model were constructed respectively. The survival outcome model was then censored using the propensity score model to obtain the individual non-consistency risk score. Based on the heterogeneity and distribution differences of clinical data from different research centers, a stratified weighting rule was defined, and a two-sided order-preserving prediction mechanism was used to independently calibrate the individual non-compliance risk scores of different research centers to generate individual risk prediction intervals. A sliding window mechanism is adopted to continuously update the calibration parameters of the order-preserving prediction mechanism on both sides online, generating a real-time updated individual risk prediction range; Based on the real-time updated individual risk prediction interval, the upper and lower bound probabilities of risk for each subject are calculated to obtain individual risk assessment indicators; Based on individual risk assessment indicators, three-domain risk threshold rules are defined, and real-time individual risk decision results are obtained through three-domain decision mapping. Based on real-time decision-making results regarding individual risks, the overall safety trend of clinical trials is dynamically assessed, and corresponding safety warnings are automatically triggered.

2. The method for safety assessment of pharmaceutical clinical trials based on artificial intelligence according to claim 1, characterized in that, The original survival dataset is specifically as follows: Multi-source heterogeneous data, including individual baseline characteristics, medication regimens, follow-up records, outcome events, and censoring markers, were collected from multiple pharmaceutical clinical trial research centers. Unified encoding and feature alignment processing are performed on multi-source heterogeneous data to map the variable naming, measurement units and data formats of different centers to a unified feature template; Based on the timestamp information of the data collection time, the follow-up records of each subject were sorted and completed in chronological order. For time segments with missing or abnormal data, a hybrid correction strategy based on nearest neighbor imputation and survival time consistency constraints was used to complete the data. A deletion indication matrix is ​​established based on deletion identifiers and event occurrence status. Deletion samples are labeled with deletion type, and deletion information is structured and encoded. After completing the structured preprocessing of the data from each center, the data is indexed and merged according to individual identifiers, and a unified multidimensional time-series representation structure is formed using center identifier, individual number, survival time, censoring identifier, feature vector, and outcome label to form the original survival dataset.

3. The method for safety assessment of pharmaceutical clinical trials based on artificial intelligence according to claim 1, characterized in that, The construction of the propensity score model and the survival outcome model, respectively, is as follows: Based on individual baseline characteristics and medication regimens in the original survival dataset, a data input vector for a propensity score model is constructed. Based on the processed data input vector, a propensity scoring model is constructed using a multilayer perceptron neural network. The propensity scoring model includes an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is the same as the feature dimension of the data input vector. The hidden layer contains two fully connected neural network layers. The number of nodes in the hidden layer decreases layer by layer. The output layer is a logistic regression unit that outputs an individual probability propensity score. Based on the survival time, censoring markers, feature vectors, and outcome events contained in the original survival dataset, construct the data input vector for the survival outcome model; Based on the data input vector of the survival outcome model, a deep recurrent neural network is used to construct the survival outcome model. The model structure of the survival outcome model includes an input layer, a feature encoding layer, a bidirectional long short-term memory network layer, and a survival risk output layer. The number of nodes in the input layer corresponds to the feature dimension of the data input vector. The feature encoding layer performs nonlinear feature mapping on the input vector. The bidirectional long short-term memory network layer extracts the long-term dependency features of the input vector. The survival risk output layer uses a fully connected network structure to map the initial risk probability of the individual's survival event. The inverse probability weighting factor for each subject is calculated based on the individual propensity score, and the initial risk probability is weighted and corrected to generate the censored survival risk probability. The individual's non-compliance risk score is calculated based on the survival risk probability after censoring correction.

4. The method for safety assessment of pharmaceutical clinical trials based on artificial intelligence according to claim 1, characterized in that, The hierarchical weighting rule and the two-sided order-preserving prediction mechanism are as follows: The hierarchical weighting rule is as follows: based on the heterogeneity, sample size and censoring ratio of data from each pharmaceutical clinical trial research center, multiple data levels are defined, and different weight coefficients are assigned to each data level. The two-sided order-preserving prediction mechanism is as follows: calculate the upper and lower quantile boundaries of the non-compliance risk score of each individual in each data level, and map the quantile boundaries of each data level to a unified standard quantile space to form standardized quantile anchors; The upper and lower bounds of the individual non-compliance risk score intervals within each data level are calibrated using quantile anchor points to maintain order. A hierarchical weighting rule is used to weight and fuse the upper and lower bounds of each data level after order preservation calibration to obtain the fused individual risk prediction range.

5. The method for safety assessment of pharmaceutical clinical trials based on artificial intelligence according to claim 1, characterized in that, The calibration parameters for the two-sided order-preserving prediction mechanism are updated online, specifically as follows: The newly acquired survival data collected continuously during pharmaceutical clinical trials is periodically extracted using a preset fixed-length time window to form a sliding data subset; Within each sliding window period, the original survival dataset is updated based on the newly added sliding data subset, and the upper and lower quantile boundaries of the individual non-compliance risk scores are recalculated to generate the latest quantile anchor points. Based on the latest calculated quantile anchor points, the upper and lower bounds of the individual non-compliance risk score intervals of each data level in the current sliding data subset are recalibrated using monotonic order preservation. The calibrated upper and lower bounds are then re-integrated using hierarchical weighted fusion to update the individual risk prediction intervals for each subject. The individual risk prediction interval obtained in the current sliding window period is compared with the individual risk prediction interval obtained in the previous sliding window period. Based on the degree and direction of the change in the upper and lower bounds of the individual risk prediction intervals in the two periods, the calibration parameters of the order-preserving prediction mechanism on both sides of the next sliding window period are adjusted.

6. The method for safety assessment of pharmaceutical clinical trials based on artificial intelligence according to claim 1, characterized in that, The calculation of the upper and lower bound probabilities of risk for each individual subject is specifically as follows: Based on the upper and lower bounds of the individual risk prediction interval, a unified probability density distribution function is constructed according to the numerical distribution of the individual risk prediction intervals of all subjects. Substitute the upper and lower bounds of the individual risk prediction interval for each subject into the probability density distribution function to calculate the probability density values ​​corresponding to the upper and lower bounds of the risk prediction interval for each subject. The upper bound probability density value of the risk prediction interval for each individual subject is integrated, and the result is used as the upper bound probability of the individual risk. The lower bound probability density value of the risk prediction interval for each individual subject is integrated, and the result is used as the lower bound probability of the individual risk. The calculated upper and lower bound probabilities of risk for each subject are used as the maximum and minimum probability estimates of the individual's risk occurrence, respectively, and are used as indicators for individual risk assessment.

7. The method for safety assessment of pharmaceutical clinical trials based on artificial intelligence according to claim 1, characterized in that, The three-domain risk threshold rule is as follows: Based on the overall numerical distribution of the upper bound probability and lower bound probability of individual risk for all subjects, a first risk threshold and a second risk threshold are set, and the subjects are divided into three non-overlapping decision domains: the safety domain, the concern domain, and the risk domain. The definition rule for the safe domain is: when the upper bound probability of an individual's risk is lower than the first risk threshold, it indicates that the individual has a low risk of experiencing a survival event, and the individual is classified into the safe domain. The rules for defining the domain of concern are as follows: when the upper bound probability of an individual's risk is higher than or equal to the first risk threshold and the lower bound probability of an individual's risk is lower than the second risk threshold, it indicates that there is uncertainty in the risk of an individual experiencing a survival event. The individual is then classified into the domain of concern and risk monitoring is implemented for the individual. The rule for defining the risk domain is as follows: when the lower bound probability of an individual's risk is higher than or equal to the second risk threshold, it indicates that the individual has a high and certain risk of experiencing a survival event. The individual is then classified into the risk domain, and corresponding safety warnings and risk interventions are triggered.

8. The method for safety assessment of pharmaceutical clinical trials based on artificial intelligence according to claim 1, characterized in that, The method of obtaining real-time individual risk decision results through three-domain decision mapping is as follows: Based on the upper bound probability and lower bound probability of individual risk for each subject, the numerical distance between the corresponding individual risk probability and the first risk threshold and the second risk threshold is calculated to obtain the distance difference. The mapping strength function for risk decision-making is defined based on the magnitude and direction of the distance difference; The membership scores of each subject to the safety domain, concern domain, and risk domain are calculated based on the mapping strength function, and the domain to which the individual's real-time risk decision-making result belongs is determined. When the highest membership score of a subject corresponds to the safety domain, the individual's real-time risk decision-making result is determined as a safe state. When the maximum membership score corresponds to the domain of concern, the individual risk real-time decision-making result is determined as a state of concern, and real-time risk monitoring is triggered; When the maximum membership score corresponds to the risk domain, the individual risk real-time decision result is determined as the risk status, and real-time safety warnings and risk intervention measures are triggered. The real-time decision results are smoothed over time to obtain real-time decision results for individual risks.

9. The method for safety assessment of pharmaceutical clinical trials based on artificial intelligence according to claim 1, characterized in that, The dynamic assessment of the overall safety trend of clinical trials specifically includes: Based on the real-time risk decision results of all subjects in each assessment period, the number of subjects classified into the safety domain, concern domain, and risk domain, as well as their proportion in the total number of subjects, are counted to obtain the proportion of individuals in each decision domain. Calculate the numerical change in the proportion of individuals in each decision domain between the current evaluation period and the previous evaluation period to obtain the trend of the proportion change in each decision domain. A safety trend index is defined based on the trend of percentage change, and a safety trend index that characterizes the overall safety trend change of clinical trials is formed by weighted calculation. By comparing the current safety trend index with the preset safety trend threshold, when the safety trend index is lower than the preset safety trend threshold, it is determined that the overall safety trend of the clinical trial has deteriorated, and an overall risk trend warning is triggered.

Images