A colorectal cancer risk assessment system based on gut microbiota diversity

Abstract

This invention relates to the fields of bioinformatics and medical artificial intelligence data processing technology, specifically a colorectal cancer risk assessment system based on gut microbiome diversity. The system includes: a multimodal time-series data acquisition module for acquiring the microbiome time-series characteristic data, sampling timestamp sequences, and current medical resource load parameters of the target object; a compliance entropy calculation module for calculating the longitudinal compliance entropy characterizing the target object's willingness to upload data continuously; an artifact interference identification module for generating physiological disturbance probability values ​​and denoised pathological evolution feature vectors; a dual-constraint risk assessment module for calculating the initial pathological risk index of the target object; and a dynamic decision-making scheduling module for determining the final dynamic intervention strategy for the target object. This invention significantly reduces the false positive rate and solves the technical pain point of traditional methods in distinguishing between real pathological evolution and environmental interference.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and medical artificial intelligence data processing technology, specifically a colorectal cancer risk assessment system based on intestinal microbial diversity. Background Technology

[0002] Colorectal cancer risk assessment systems are a crucial link in the early screening and prevention of gastrointestinal tumors, directly impacting the efficiency of medical resource allocation and patient survival rates. These systems primarily involve the collection of microbiome temporal characteristics, pathological evolution analysis, and clinical decision scheduling, achieving risk warnings by monitoring dynamic changes in the gut microbiota. A key research focus in the field of smart healthcare is how to accurately identify pathological features in a high-dimensional dynamic monitoring space, balance sensitivity and specificity, and achieve optimal intervention under medical resource constraints.

[0003] Existing technologies primarily rely on single cross-sectional pathological features or static rules, lacking quantitative consideration of longitudinal compliance during data collection, making it difficult to predict the risk of data chain disruptions caused by user behavior. Furthermore, the gut microbiota exhibits highly non-Euclidean structural characteristics and transient physiological fluctuations, making it difficult for existing methods to effectively remove artifacts caused by non-pathological factors, resulting in a high false positive rate. In addition, traditional models lack dynamic perception of the burden on medical resources, and the passive response mode using fixed thresholds easily leads to resource strain due to blind referrals or missed diagnoses due to user dropout. Moreover, the heterogeneity and sparse sampling characteristics of data from multiple sources severely restrict the accuracy and robustness of cross-scenario assessments. Therefore, a solution is urgently needed to address the problems existing in current technologies. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a colorectal cancer risk assessment system based on gut microbiome diversity. Specifically, the technical solution of this invention includes:

[0005] The multimodal time-series data acquisition module is used to acquire the microbiome time-series characteristic data, sampling timestamp sequences, and current medical resource load status parameters of the target object;

[0006] The compliance entropy calculation module is used to calculate the longitudinal compliance entropy, which represents the target object's willingness to upload data continuously, based on the distribution pattern of the sampling timestamp sequence.

[0007] The artifact interference identification module is used to process the microbiome time-series feature data using a time-series graph neural network model, separate transient physiological disturbance artifact signals, and generate physiological disturbance probability values ​​and denoised pathological evolution feature vectors.

[0008] The dual-constraint risk assessment module is used to calculate the initial pathological risk index of the target object by combining the pathological evolution feature vector with the preset clinical evidence-based decision-making rules.

[0009] The dynamic decision-making and scheduling module is used to determine the final dynamic intervention strategy for the target object based on the initial pathological risk index, the physiological disturbance probability value, the longitudinal compliance entropy, and the medical resource load status parameters.

[0010] The final dynamic intervention strategy includes one of the following: an immediate referral instruction, a high-frequency home sampling instruction, or a low-frequency observation instruction.

[0011] Preferably, the compliance entropy calculation module calculates the longitudinal compliance entropy, which characterizes the target object's willingness to upload data continuously, based on the distribution pattern of the sampling timestamp sequence, including:

[0012] Calculate the time interval between adjacent time points in the sampling timestamp sequence to generate a sampling interval sequence;

[0013] Based on the mean and variance of the sampling interval sequence, an irregularity model of the time series is constructed.

[0014] Irregularity is mapped to information entropy value, and this information entropy value is defined as the longitudinal compliance entropy, where a higher entropy value indicates a higher risk of future data supply disruption for the target object.

[0015] Preferably, the artifact interference identification module uses a time-series graph neural network model to process the microbiome time-series feature data and separate transient physiological perturbation artifact signals, including:

[0016] Construct a dynamic graph structure containing microbial nodes and co-occurrence relationship edges;

[0017] Spatial dependence features of microbial communities are captured by graph convolutional layers, and temporal dependence features are captured by recurrent neural network layers.

[0018] The fluctuation amplitude of the calculated feature within a short time window is determined to be the physiological disturbance artifact signal when the fluctuation amplitude exceeds a preset nonlinear threshold and lacks long-term trend support.

[0019] Preferably, the dual-constraint risk assessment module combines the pathological evolution feature vector with preset clinical evidence-based decision-making rules to calculate the initial pathological risk index of the target object, including:

[0020] The pathological evolution feature vector is input into a pre-trained risk prediction sub-model, which outputs a data-driven soft risk probability.

[0021] The clinical evidence-based decision-making rules are invoked to perform hard boundary constraint verification on the soft risk probability;

[0022] When the soft risk probability triggers a Boolean hard boundary constraint, the soft risk probability is weighted and corrected to generate the initial pathological risk index.

[0023] Preferably, the dynamic decision-making and scheduling module determines the final dynamic intervention strategy for the target object based on the initial pathological risk index, the physiological disturbance probability value, the longitudinal compliance entropy, and the medical resource load status parameters, including:

[0024] Construct a game-theoretic loss function that includes both resource protection and risk capture objectives;

[0025] The state of the longitudinal compliance entropy is determined when the initial pathological risk index is higher than a preset high-risk threshold and the physiological disturbance probability value is higher than a preset interference threshold.

[0026] In response to the high entropy of the longitudinal compliance, in order to maintain the continuity of the data chain, the high-frequency home sampling instruction is output as the final dynamic intervention strategy to smooth the short-term risk assessment.

[0027] Preferably, the dynamic decision-making scheduling module, in determining the final dynamic intervention strategy for the target object, further includes:

[0028] In response to the initial pathological risk index being higher than a preset high-risk threshold and the physiological disturbance probability value being lower than a preset interference threshold, the immediate referral instruction is directly output.

[0029] In response to the medical resource load status parameter indicating resource saturation and the initial pathological risk index being in the critical range, the threshold for issuing the immediate referral instruction is increased, and the high-frequency home sampling instruction is issued simultaneously to increase monitoring density.

[0030] Preferably, the training process of the time-series graph neural network model includes:

[0031] Construct a hybrid loss function that includes a compliance regularization term;

[0032] The longitudinal compliance entropy is introduced as a weighting factor into the hybrid loss function;

[0033] During training, higher penalty weights are assigned to the prediction errors of samples with high longitudinal compliance entropy, in order to force the model to prioritize learning robust features under sparse data conditions.

[0034] Preferred options also include:

[0035] The data heterogeneity alignment module is used after the multimodal time-series data acquisition module acquires the data:

[0036] Identify the source device type of the microbiome time-series feature data;

[0037] Based on a preset baseline drift matrix, data from different source device types are mapped to a unified feature vector space to eliminate systematic bias caused by differences in sampling environment.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. This invention introduces longitudinal compliance entropy to quantitatively assess the continuity of user data uploads and uses it as a regularization weight for model training. This not only provides early warning of the risk of data chain breakage caused by irregular user behavior, but also forces the temporal graph neural network to prioritize learning robust features that do not rely on dense sampling during training. Thus, it can maintain the stability and accuracy of risk assessment even in real-world scenarios with poor user compliance or sparse data.

[0040] 2. This invention utilizes a temporal graph neural network to construct a dynamic graph structure containing microbial nodes and co-occurrence relationships, which can accurately separate transient physiological disturbance artifacts caused by non-pathological factors; by comparing short-term fluctuation amplitudes with long-term trend support, the system effectively filters random co-occurrence noise, significantly reduces the false positive rate, and solves the technical pain point of traditional methods in distinguishing between real pathological evolution and environmental interference.

[0041] 3. This invention constructs a ternary game decision-making mechanism that includes pathological risk, compliance, and resource load; the system combines clinical evidence-based rules to impose hard boundary constraints on the probability of soft risks, ensuring the safety of decision-making; at the same time, it dynamically adjusts the referral threshold according to the medical resource load status, and intelligently switches to a high-frequency home monitoring mode when resources are saturated; this mechanism avoids resource congestion caused by blind referrals and prevents the risk of missed diagnosis due to medical resource constraints.

[0042] 4. The present invention sets up a data heterogeneity alignment module, which can automatically identify the source device type of microbial data and use the baseline drift matrix to map data from different sources to a unified feature vector space. This design effectively eliminates the systematic bias caused by different sequencing device sensitivity and sampling environment, ensures the consistency of time series data analysis across devices and scenarios, and improves the universality of the system in multi-center applications. Attached Figure Description

[0043] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0044] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0046] Example 1:

[0047] Please see Figure 1 A colorectal cancer risk assessment system based on gut microbiome diversity includes:

[0048] The multimodal time-series data acquisition module is used to acquire the microbiome time-series characteristic data, sampling timestamp sequences, and current medical resource load status parameters of the target object;

[0049] The compliance entropy calculation module is used to calculate the longitudinal compliance entropy, which represents the target object's willingness to upload data continuously, based on the distribution pattern of the sampling timestamp sequence.

[0050] The artifact interference identification module is used to process the microbiome time-series feature data using a time-series graph neural network model, separate transient physiological disturbance artifact signals, and generate physiological disturbance probability values ​​and denoised pathological evolution feature vectors.

[0051] The dual-constraint risk assessment module is used to calculate the initial pathological risk index of the target object by combining the pathological evolution feature vector with the preset clinical evidence-based decision-making rules.

[0052] The dynamic decision-making and scheduling module is used to determine the final dynamic intervention strategy for the target object based on the initial pathological risk index, the physiological disturbance probability value, the longitudinal compliance entropy, and the medical resource load status parameters.

[0053] The final dynamic intervention strategy includes one of the following: an immediate referral instruction, a high-frequency home sampling instruction, or a low-frequency observation instruction.

[0054] This embodiment details the overall architecture and core data flow logic of the system, aiming to address the issue of neglecting longitudinal compliance and transient physiological artifacts during data acquisition when relying solely on pathological features for risk assessment. The system activates a multimodal time-series data acquisition module to construct a high-dimensional dynamic monitoring space. This module acquires real-time time-series microbiome characteristic data of the target object. This refers to the bacterial community abundance matrix obtained based on 16S rRNA or metagenomic sequencing, while simultaneously acquiring the sampling timestamp sequence. and current medical resource load parameters The compliance entropy calculation module intervenes, based on the sampling timestamp sequence. The distribution pattern is analyzed to calculate the vertical compliance entropy, which represents the target object's willingness to upload data continuously. The system quantifies and predicts the risk of data supply disruption; the artifact interference identification module performs denoising and fidelity preservation operations, uses a time-series graph neural network model to process the microbiome time-series feature data, separates transient physiological disturbance artifact signals caused by non-pathological factors, and generates physiological disturbance probability values. and the denoised pathological evolution feature vector ;

[0055] Based on this, the dual-constraint risk assessment module combines the pathological evolution feature vector with preset clinical evidence-based decision-making rules to calculate the initial pathological risk index of the target object. To balance sensitivity and safety, the dynamic decision-making and scheduling module determines the final dynamic intervention strategy for the target population based on the initial pathological risk index, the probability value of physiological disturbance, the longitudinal compliance entropy, and the medical resource load status parameters. To achieve optimal intervention under resource constraints;

[0056] This embodiment introduces longitudinal compliance entropy as a moderating factor for risk assessment, establishing a ternary game mechanism of user behavior, pathological risk, and medical resources in the colorectal cancer screening system. This mechanism not only focuses on pathological risk but also on compliance and resource load, thereby effectively avoiding resource exhaustion caused by blind referrals in extreme scenarios where medical resources are saturated, while also preventing the risk of missed diagnoses due to user dropout.

[0057] Example 2:

[0058] The compliance entropy calculation module, based on the distribution pattern of the sampling timestamp sequence, calculates the longitudinal compliance entropy, which characterizes the target object's willingness to upload data continuously, including:

[0059] Calculate the time interval between adjacent time points in the sampling timestamp sequence to generate a sampling interval sequence;

[0060] Based on the mean and variance of the sampling interval sequence, an irregularity model of the time series is constructed.

[0061] Irregularity is mapped to information entropy value, and this information entropy value is defined as the longitudinal compliance entropy, where a higher entropy value indicates a higher risk of future data supply disruption for the target object.

[0062] This embodiment provides a detailed description of the implementation of the compliance entropy calculation module, aiming to quantify the irregularity of user sampling; the module calculates the sampling timestamp sequence. The time interval between adjacent time points is used to generate a sampling interval sequence. ,in, The time difference between two adjacent samples is given. Based on the mean and variance of the sampling interval sequence, a longitudinal compliance entropy calculation model is constructed. To strictly adhere to the limitation of calculation based on mean and variance in this invention, this embodiment adopts a parametric estimation method, assuming that the fluctuation of the sampling interval follows a statistical distribution determined by the mean and variance, i.e., a Gaussian distribution under the maximum entropy principle, and the differential entropy of this distribution combined with the coefficient of variation is defined as the longitudinal compliance entropy. Considering the domain constraint of the logarithmic function and the consistency of physical dimensions, a normalization factor is introduced, and its calculation formula is as follows:

[0063]

[0064] in, The source is the model calculation output, and its physical meaning is the longitudinal compliance entropy, with the unit being Nat;

[0065] The term originates from the differential entropy derived from the mean and variance, characterizing the uncertainty of the sampling interval distribution;

[0066] The source is calculated from historical user data, and its physical meaning is the standard deviation of the sampling interval sequence, in days;

[0067] The source is a preset minimum value, and its physical meaning is the minimum variance cutoff threshold. Its value is set to half of the minimum recording precision of the sampling timestamp. In this embodiment, the sampling precision is days, so the value is 0.5, which aims to prevent the logarithmic function from diverging and the denominator from being zero due to the user's completely regular sampling.

[0068] The source is a preset constant, and its physical meaning is a time normalization factor, which is taken as 1 day and used as the denominator to eliminate the dimension problem of the logarithmic function.

[0069] The data is derived from historical user data and its physical meaning is the variance of the sampling interval sequence, expressed in square days.

[0070] The data is calculated from historical user data, and its physical meaning is the mean of the sampling interval sequence, in days.

[0071] The source is a preset value, and its physical meaning is the coefficient of variation penalty factor, which is used to enhance the sensitivity to relative fluctuation amplitude. The unit is the weighting coefficient.

[0072] The source is a preset constant, and its physical meaning is a smoothing term to prevent the denominator from being zero;

[0073] This embodiment maps discrete time interval statistical characteristics to continuous information entropy values ​​through the above calculations. The entropy value reflects the degree of disorder in the sampling behavior, while the coefficient of variation reflects the relative irregularity. Through this calculation, the system can accurately identify high-entropy users with highly disordered sampling behavior, thereby providing early warning of the risk of data link breakage and providing quantitative basis for subsequent decision-making modules.

[0074] Example 3:

[0075] The artifact interference identification module uses a time-series graph neural network model to process the time-series feature data of the microbiome, separating transient physiological perturbation artifact signals, including:

[0076] Construct a dynamic graph structure containing microbial nodes and co-occurrence relationship edges;

[0077] Spatial dependence features of microbial communities are captured by graph convolutional layers, and temporal dependence features are captured by recurrent neural network layers.

[0078] The fluctuation amplitude of the calculated feature within a short time window is determined to be the physiological disturbance artifact signal when the fluctuation amplitude exceeds a preset nonlinear threshold and lacks long-term trend support.

[0079] This embodiment addresses the non-Euclidean structural characteristics of microbial community data, clarifying the construction and computational details of a time-series graph neural network model, particularly supplementing the generation logic and probability quantification formulas for graph structures; the module constructs a dynamic graph structure. , among which, nodes represent Each microbial taxonomic unit; to quantify co-occurrence relationships, the system calculates the Pearson correlation coefficient of microbial abundance sequences within a sliding window. The length of the sliding window is set to For example, 30 days, to ensure the sample size meets the statistical significance requirement; and an adjacency matrix is ​​constructed based on the threshold. The calculation formula is as follows:

[0080]

[0081] in, The threshold is set to 0.3. The value is selected based on the statistical significance threshold P<0.05 of the Pearson correlation coefficient with a sample size of 30, which is used to filter out statistically insignificant random co-occurrence noise.

[0082] Using graph convolutional layers Capture spatial dependency features, define For microbiome time-series characteristic data At any moment The feature slice matrix is ​​given by the formula: and output Input is used to a GRU recurrent neural network layer to capture time-dependent features; to ensure the computability and uniqueness of the formula, the following definitions are added: For example, a non-linear activation function. subscript Used to distinguish the variance sign from those in the foregoing embodiments; To add self-loops to the adjacency matrix, It is the identity matrix; for The degree matrix, whose elements ; The learnable weight matrix of the graph convolutional layer is used to map node features to the hidden layer space. Before this, in order to eliminate the influence of differences in the abundance scale of different microbial communities on the threshold determination, the Min-Max normalization method is used to map the microbiome time series feature data to the [0,1] interval.

[0083] Based on this, the system defines the fluctuation quantization and probability generation logic: for feature vectors Specific dimensions in , recorded as Calculate features in a short time window For example, the fluctuation range within 3 days And in long time windows For example, the slope of a linear trend over 14 days To maintain consistency of sign, in the following formulas Referential eigenvectors For scalar values ​​in this dimension, the system pre-performs validity checks for sparse data scenarios: if within a short time window... If the actual number of samples collected is less than 2, then directly set... If no valid fluctuation characteristics are observed during this period, the error of zero denominator in variance calculation can be avoided; otherwise, the following calculation is performed:

[0084]

[0085] in, Representation of features In a short time window The arithmetic mean within, i.e. ; Representation of features In long window The arithmetic mean within;

[0086] To meet the requirements of generating physiological perturbation probability values ​​in the embodiments. The system utilizes the requirements. The function maps physical features to probabilities:

[0087]

[0088] in, The nonlinear fluctuation threshold is set at three times the standard deviation of historical stable periods and is used as a baseline to define abnormal fluctuations.

[0089] Trend significance threshold, with a value of 0.05, is used to ignore weak background drift;

[0090] : Scaling factor, with values ​​of 10.0 and 5.0 respectively; this factor is obtained by fitting a logistic regression model to the labeled dataset, aiming to map feature differences to the probability interval [0,1], so that when fluctuations are significant and there is no trend support, It rapidly approaches 1.

[0091] Example 4:

[0092] The dual-constraint risk assessment module combines the pathological evolution feature vector with preset clinical evidence-based decision-making rules to calculate the initial pathological risk index of the target object, including:

[0093] The pathological evolution feature vector is input into a pre-trained risk prediction sub-model, which outputs a data-driven soft risk probability.

[0094] The clinical evidence-based decision-making rules are invoked to perform hard boundary constraint verification on the soft risk probability;

[0095] When the soft risk probability triggers a Boolean hard boundary constraint, the soft risk probability is weighted and corrected to generate the initial pathological risk index.

[0096] This embodiment employs a combined hardware and software strategy to address the lack of interpretability of purely data-driven models in medical scenarios; it integrates pathological evolution feature vectors... Input a pre-trained deep learning risk prediction sub-model, output a data-driven soft risk probability. The system invokes preset clinical evidence-based decision rules, such as determining excessive abundance of *Fusobacterium nucleatum*, to perform hard boundary constraint verification on the soft risk probability. In response to the soft risk probability triggering a Boolean hard boundary constraint (i.e., the rule is deemed true), the system performs a weighted correction on the soft risk probability, generating an initial pathological risk index. The calculation formula is as follows:

[0097]

[0098] in, The source is the weighted calculation output, the physical meaning is the initial pathological risk index, and the unit is a dimensionless index;

[0099] The source is model prediction, the physical meaning is soft risk probability, and the unit is probability value;

[0100] The source is the rule base definition, and the physical meaning is the risk assignment corresponding to the rule; in this embodiment, the normalized assignment method is adopted: when a Boolean hard boundary constraint is triggered, such as when the relative abundance of Fusobacterium nucleatum is >15%, When not triggered That is, it does not increase additional risk;

[0101] The source is a preset value, and its physical meaning is the confidence weight; its value ranges from 0.6 to 0.8, and 0.7 is preferred in this embodiment. This value is determined based on the AUC performance on the validation set: when the uncertainty of the model prediction is high, it should be appropriately reduced. To give clinical rules a higher decision-making weight;

[0102] This embodiment introduces hard boundary constraints to ensure that, under any circumstances, high-risk indications clearly defined in medical guidelines have the highest priority; this mechanism prevents artificial intelligence models from ignoring obvious clinically critical values ​​due to training data bias, thereby ensuring the safety and compliance of medical decisions.

[0103] Example 5:

[0104] The dynamic decision-making and scheduling module determines the final dynamic intervention strategy for the target population based on the initial pathological risk index, the probability value of physiological disturbance, the longitudinal compliance entropy, and the medical resource load status parameters, including:

[0105] Construct a game-theoretic loss function that includes both resource protection and risk capture objectives;

[0106] The state of the longitudinal compliance entropy is determined when the initial pathological risk index is higher than a preset high-risk threshold and the physiological disturbance probability value is higher than a preset interference threshold.

[0107] In response to the high entropy of the longitudinal compliance entropy, in order to maintain the continuity of the data chain, the high-frequency home sampling instruction is output as the final dynamic intervention strategy to smooth the short-term risk assessment.

[0108] The dynamic decision-making and scheduling module determines the final dynamic intervention strategy for the target object, and also includes:

[0109] In response to the initial pathological risk index being higher than a preset high-risk threshold and the physiological disturbance probability value being lower than a preset interference threshold, the immediate referral instruction is directly output.

[0110] In response to the medical resource load status parameter indicating resource saturation and the initial pathological risk index being in the critical range, the threshold for issuing the immediate referral instruction is increased, and the high-frequency home sampling instruction is issued simultaneously to increase monitoring density.

[0111] This embodiment details the implementation logic of the dynamic decision-making and scheduling module, focusing on supplementing the quantitative definition of high-entropy states and the complete logical closed loop; the module constructs a game loss function, the calculation formula of which is:

[0112]

[0113] in, The pre-defined risk weight for missed diagnoses. The pre-defined weighting of medical resource usage costs; based on the principle of no harm in medical ethics and health economics assessment, the following settings are established. In this embodiment, take This means that the system penalty for missing a cancer diagnosis is 10 times the resource cost of an endoscopy, thus ensuring that the system is highly risk-sensitive when making decisions. This is the resource load sensitivity coefficient, with a value of 0.5, used to adjust the nonlinear effect of load on the threshold. As decision variables, Referral by representative Representative observations were made, and the dynamic referral threshold was derived from this. The baseline threshold is defined as follows: ;

[0114] The system executes the following strictly defined logical branches:

[0115] High-entropy state definition and handling: The system calculates the statistical distribution of longitudinal compliance entropy from the historical database and sets the 75th percentile as the high-entropy threshold. When the system is in the cold start phase or the sample size is insufficient, the preset... This serves as a temporary baseline until a sufficient sample size is accumulated, at which point it switches to dynamic statistics. Accumulating a sufficient sample size means that the number of valid users recorded by the system exceeds the preset minimum statistical sample size, such as 100 people. and At that time, the system checks compliance: if This indicates a high-entropy state, where the system determines the user faces an extremely high risk of detachment and outputs high-frequency home sampling commands to maintain the data link through high-frequency interaction; if This indicates a low-entropy state, suggesting that the data upload pattern is reliable, and the system determines the current probability value of physiological disturbance. An abnormally high level may be due to a sudden physiological change rather than a sampling artifact. In this case, a resampling instruction will be output, requiring the target object to undergo a second confirmation sampling within 24 hours.

[0116] Dynamic referral criteria: If and If the system determines that the pathological risk-benefit exceeds the resource cost and there is no artifact interference, it will output an immediate referral instruction.

[0117] Resource saturation protection: Response to resource load The rise led to For those falling within the critical interval For the target individuals, the system determines that although the risk is higher than the baseline, it does not meet the referral criteria allowed by current resources. Therefore, it outputs a high-frequency home sampling command to increase the monitoring density and postpones the referral.

[0118] The source is ROC curve optimization, and the physical meaning is the physiological disturbance judgment threshold, such as 0.7; The source is group statistics, and the physical meaning is the threshold for judging high compliance entropy;

[0119] This embodiment clarifies The acquisition method and the mutual exclusion conditions of logical branches ensure the executableness of the decision logic at the code level, and solve the problem of the lack of quantitative standards for high-entropy states in the original description.

[0120] Example 6:

[0121] The training process of a time-series graph neural network model includes:

[0122] Construct a hybrid loss function that includes a compliance regularization term;

[0123] The longitudinal compliance entropy is introduced as a weighting factor into the hybrid loss function;

[0124] During training, higher penalty weights are assigned to the prediction errors of samples with high longitudinal compliance entropy, in order to force the model to prioritize learning robust features under sparse data conditions.

[0125] This embodiment discloses a special training method for a time-series graph neural network model, designed to adapt to the realities of sparse data and irregular sampling; it constructs a hybrid loss function that includes a compliance regularization term; and it incorporates longitudinal compliance entropy. The hybrid loss function is introduced as a weighting factor; to avoid the weights flipping to negative values ​​due to negative entropy (i.e., in high compliance scenarios), thus disrupting the optimization direction, this embodiment adopts an exponential weighting strategy, calculated as follows:

[0126]

[0127] in, The source is the calculation result, and the physical meaning is the total loss value after compliance weighting, with the unit being a numerical value;

[0128] The source is the comparison between the model output and the label; the physical meaning is the basic prediction error, such as cross-entropy loss; the unit is numerical value.

[0129] The source is the compliance calculation module, and its physical meaning is the longitudinal compliance entropy corresponding to the sample. The unit is dimensionless.

[0130] The source is the exponential function, and its physical meaning is to ensure that the compliance weights are always positive and to prevent gradient reversal.

[0131] The source is a preset value, and its physical meaning is the compliance penalty weight coefficient; in this embodiment... The value is set to 2.0; this value is determined during the hyperparameter search phase of model training, with the aim of magnifying the gradient contribution of high-entropy samples in the loss function. This forces the model to prioritize learning the sparse robust features implicit in such samples; during training, the system assigns a higher penalty weight to the prediction error of samples with high longitudinal compliance entropy based on the above formula.

[0132] This embodiment uses a weighted penalty mechanism to force the model to minimize prediction error under conditions of sparse data and highly uncertain time intervals, i.e., a high-entropy state. This guides the model to prioritize learning those robust biological features that do not rely on dense sampling. This enables the model to maintain high prediction stability when faced with user data that is sampled very irregularly in the real world.

[0133] Example 7:

[0134] This system also includes a data heterogeneity alignment module, used after the multimodal time-series data acquisition module acquires the data:

[0135] Identify the source device type of the microbiome time-series feature data;

[0136] Based on a preset baseline drift matrix, data from different source device types are mapped to a unified feature vector space to eliminate systematic bias caused by differences in sampling environment.

[0137] This embodiment describes the execution logic of the data heterogeneity alignment module, which aims to solve the problem of multi-device data fusion. After the acquisition of multimodal time-series data, the module identifies the source device type of the microbiome time-series feature data from the metadata; based on a preset baseline drift matrix... A linear transformation is used to map data from different device types to a unified feature vector space. The calculation formula is as follows:

[0138]

[0139] in, The source is the computational output, the physical meaning is a unified spatial feature vector, and the unit is the eigenvalue;

[0140] The source is the original data collection, the physical meaning is the original feature vector, and the unit is the feature value;

[0141] The source is a standard comparison experiment; the specific steps for obtaining it are: using... Microbial community standards were sequenced on target device A and reference device B, respectively, to obtain feature vectors. and Considering that the deviation between different sequencing devices is mainly reflected in the linear scaling of the amplification efficiency of specific bacterial species, this embodiment will... The constraints are diagonal matrices; the solution is obtained using the least squares method or direct comparison method. Thus, the optimal linear transformation matrix can be calculated. The physical meaning is the baseline sensitivity drift factor between devices;

[0142] This embodiment eliminates systematic biases caused by differences in sampling equipment and environment through baseline drift correction, ensuring that the changes observed in time series data are real biological changes rather than technical noise, thereby further improving the accuracy of risk assessment across device scenarios.

[0143] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A colorectal cancer risk assessment system based on gut microbial diversity, characterized by, The method comprises the following steps: A multi-modal time series data acquisition module is used to obtain microbiome time series feature data of a target object, a sampling timestamp sequence, and a current medical resource load state parameter; An adherence entropy calculation module is used to calculate longitudinal adherence entropy representing the continuousness of the target object's data uploading intention based on the distribution law of the sampling timestamp sequence; An artifact interference identification module is used to process the microbiome time series feature data using a time series graph neural network model, separate out transient physiological disturbance artifact signals, and generate physiological disturbance probability values and denoised pathological evolution feature vectors; A double-constraint risk assessment module is used to combine the pathological evolution feature vectors and preset clinical evidence-based decision rules to calculate an initial pathological risk index of the target object; A dynamic decision scheduling module is used to determine a final dynamic intervention strategy for the target object based on the initial pathological risk index, the physiological disturbance probability values, the longitudinal adherence entropy, and the medical resource load state parameter; The final dynamic intervention strategy comprises one of an immediate referral instruction, a high-frequency home sampling instruction, or a low-frequency observation instruction; The double-constraint risk assessment module combines the pathological evolution feature vectors and preset clinical evidence-based decision rules to calculate an initial pathological risk index of the target object, which comprises the following steps: The pathological evolution feature vectors are input into a pre-trained risk prediction sub-model to output data-driven soft risk probabilities; The clinical evidence-based decision rules are called to perform hard boundary constraint verification on the soft risk probabilities; When the soft risk probabilities exceed the hard boundary constraint, the soft risk probabilities are weighted and corrected to generate the initial pathological risk index; the calculation formula of the initial pathological risk index is as follows: ； wherein, : initial pathological risk index, unit is dimensionless index; : confidence weight, value range is 0.6 to 0.8; : soft risk probability, unit is probability value; : risk assignment corresponding to the rule; when the soft risk probability triggers the hard boundary constraint of the rule, = 1.0, not triggered = .

2. The system for assessing the risk of colorectal cancer based on gut microbial diversity according to claim 1, wherein, The adherence entropy calculation module calculates longitudinal adherence entropy representing the continuousness of the target object's data uploading intention based on the distribution law of the sampling timestamp sequence, which comprises the following steps: The time intervals between adjacent time points in the sampling timestamp sequence are calculated to generate a sampling interval sequence; Based on the mean and variance of the sampling interval sequence, an irregularity model of the time sequence is constructed; The irregularity is mapped to an information entropy value, and the information entropy value is defined as the longitudinal adherence entropy; the higher the entropy value, the higher the future data supply risk of the target object.

3. The system for assessing the risk of colorectal cancer based on gut microbial diversity according to claim 1, wherein, The artifact interference identification module uses a time series graph neural network model to process the microbiome time series feature data to separate out transient physiological disturbance artifact signals, which comprises the following steps: A dynamic graph structure containing microorganism nodes and co-occurrence relationship edges is constructed; Spatial dependence features of the microbiota are captured through a graph convolution layer, and time dependence features are captured through a recurrent neural network layer; The fluctuation amplitude of the features within a short time window is calculated; when the fluctuation amplitude exceeds a preset nonlinear threshold and lacks long-term trend support, the corresponding feature fluctuation is determined as the physiological disturbance artifact signal.

4. The system for assessing the risk of colorectal cancer based on gut microbial diversity according to claim 1, wherein, The dynamic decision scheduling module determines a final dynamic intervention strategy for the target object based on the initial pathological risk index, the physiological disturbance probability values, the longitudinal adherence entropy, and the medical resource load state parameter, which comprises the following steps: Construct a game-theoretic loss function that includes both resource protection and risk capture objectives; The state of the longitudinal compliance entropy is determined when the initial pathological risk index is higher than a preset high-risk threshold and the physiological disturbance probability value is higher than a preset interference threshold. In response to the high entropy of the longitudinal compliance, in order to maintain the continuity of the data chain, the high-frequency home sampling instruction is output as the final dynamic intervention strategy to smooth the short-term risk assessment.

5. The system for assessing the risk of colorectal cancer based on gut microbial diversity according to claim 4, wherein The dynamic decision-making and scheduling module determines the final dynamic intervention strategy for the target object, and also includes: In response to the initial pathological risk index being higher than a preset high-risk threshold and the physiological disturbance probability value being lower than a preset interference threshold, the immediate referral instruction is directly output. In response to the medical resource load status parameter indicating resource saturation and the initial pathological risk index being in the critical range, the threshold for issuing the immediate referral instruction is increased, and the high-frequency home sampling instruction is issued simultaneously to increase monitoring density.

6. The system for assessing the risk of colorectal cancer based on gut microbial diversity according to claim 3, wherein, The training process of the time-series graph neural network model includes: Construct a hybrid loss function that includes a compliance regularization term; The longitudinal compliance entropy is introduced as a weighting factor into the hybrid loss function; During training, higher penalty weights are assigned to the prediction errors of samples with high longitudinal compliance entropy, in order to force the model to prioritize learning robust features under sparse data conditions.

7. The system for assessing the risk of colorectal cancer based on gut microbial diversity according to claim 1, wherein Also includes: The data heterogeneity alignment module is used after the multimodal time-series data acquisition module acquires the data: Identify the source device type of the microbiome time-series feature data; Based on a preset baseline drift matrix, data from different source device types are mapped to a unified feature vector space to eliminate systematic bias caused by differences in sampling environment.

Images