A rheumatoid arthritis pulmonary interstitial lesion risk prediction method and system based on self-evolution feature engineering

Abstract

The present application relates to the technical field of medical information processing, in particular to a rheumatoid arthritis interstitial lung disease risk prediction method and system based on self-evolution feature engineering, which establishes a self-adaptive closed-loop control system with model performance as feedback signal, iteratively performs feature generation, screening, evaluation and strategy updating, and dynamically automatically constructs and optimizes the prediction feature set from clinical time series data, medical images and other multi-modal data. The system uses parameterized operators with clear pathological direction to deeply quantify time series dynamics, image texture and multi-modal interaction, and integrates model calibration and attribution explanation framework. The present application can generate high-performance and interpretable risk prediction models, solving the problems of traditional methods relying on static features, being difficult to capture complex time series interaction and lacking clinical interpretability.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, specifically to a method and system for predicting the risk of interstitial lung disease in rheumatoid arthritis based on self-evolutionary feature engineering. Background Technology

[0002] Rheumatoid arthritis (RA) is a common systemic autoimmune disease whose complications severely impact patient prognosis, with interstitial lung disease (ILD) being the most serious and a leading cause of death. Early identification and risk prediction of RA-ILD are crucial for improving patient management and implementing interventions. However, its insidious onset, atypical early symptoms, and significant individual variability in disease progression make accurate prediction before irreversible lung function impairment occurs extremely challenging.

[0003] Currently, the assessment of RA-ILD risk in clinical practice mainly relies on the experience of rheumatologists and pulmonologists, combined with intermittent clinical symptoms, serological markers, and high-resolution CT (HRCT) imaging for comprehensive judgment. This method is highly dependent on expert experience and struggles to effectively integrate and quantify dynamic clinical data from multiple time points and sources, resulting in strong subjectivity and insufficient timeliness in prediction, and failing to achieve automated, dynamic risk stratification for large-scale patients. In recent years, machine learning-based predictive models have provided new insights into these problems. However, most existing models rely on predefined and manually selected static feature sets, such as qualitative descriptions of laboratory indicators or imaging reports at a single time point. These methods fail to fully utilize the large amount of dynamic time-series data generated during long-term follow-up of RA patients and lack in-depth quantitative mining of raw imaging data. More importantly, the complex interaction effects between different data modalities are often ignored, resulting in limited ability of models to capture the multidimensional driving mechanisms of the disease.

[0004] Furthermore, traditional feature engineering processes are highly dependent on domain knowledge and remain fixed, failing to self-optimize based on specific prediction tasks and the data characteristics of different patient subgroups. When faced with complex, high-dimensional, and time-evolving medical data, fixed feature construction strategies easily lead to model performance bottlenecks and decreased generalization ability. Therefore, there is an urgent need for a technical solution that can automatically, dynamically, and adaptively learn from multimodal time-series data and generate optimal predictive features to build more accurate and robust RA-ILD risk prediction models. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting the risk of interstitial lung disease in rheumatoid arthritis based on self-evolving feature engineering. By dynamically adjusting the feature generation and screening strategy through performance indicators in an iterative loop, the system automatically constructs an optimized feature set from multimodal time-series data and outputs a calibrated risk prediction model with clinical interpretability.

[0006] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution: A method for predicting the risk of interstitial lung disease in rheumatoid arthritis based on self-evolutionary feature engineering, wherein the method dynamically generates an optimized prediction feature set by iteratively executing the following steps: Data preparation: Obtain the raw multimodal time-series data of patients in the training set and perform standardized preprocessing to form an initial feature pool; Feature generation and filtering: Based on the control strategy vector of the current iteration round, feature construction operators are selected from the predefined operator library to process the current feature pool to generate new features, and the feature pool is filtered according to the filtering rules in the control strategy vector to obtain the updated feature set; Performance evaluation and feedback: Train the agent model using the updated feature set and evaluate its performance metrics on the validation set, calculating the performance gain of this round relative to the previous round; Strategy decision and update: The performance gain is input into a preset strategy decision function, which outputs an updated control strategy vector for the next iteration based on the magnitude of the performance gain. The control strategy vector is used to dynamically adjust the selection probability of the feature construction operator, the complexity of the feature generation process, and the retention rate of feature selection; the iterative process continues until the preset iteration termination condition is reached, and the final feature set is output.

[0007] Furthermore, in the performance evaluation and feedback step, the performance gain of the k-th iteration... The calculation formula is: in, Let be the performance metric value of the agency model in the k-th round. To prevent division by zero of extremely small constants; In the strategy decision and update step, the strategy decision function updates the control strategy vector according to the following rules: like This increases the probability of selecting operators that contribute positive performance gains and increases the parameter values ​​of feature generation complexity and feature selection retention rate. like This reduces the selection probability preference of the current operator, decreases the parameter values ​​of feature generation complexity and feature selection retention rate, and introduces randomized exploration of operator selection; in, and It is a preset positive threshold, and .

[0008] Furthermore, the control strategy vector Represented as ,in Choose a probability distribution for the operator. Controls the maximum depth of feature generation. The feature selection retention rate; the strategy decision function is adjusted by... , , The value is used to update the control strategy vector.

[0009] Furthermore, the standardized preprocessing in the data preparation step, for the first... For each clinical or laboratory indicator, its raw measurement value Convert to standardized value The conversion formula is: in, and These are the arithmetic mean and standard deviation of the indicator calculated at all time points for all patients in the training set, respectively.

[0010] Furthermore, the operator library includes time-series dynamic feature operators for calculating the weighted trend strength of clinical indicators within the observation time window. The calculation formula is as follows: in, For indicators at a point in time The standardized value, and All of them in the window The mean and standard deviation, The time decay weighting function is... It is a very small constant.

[0011] Furthermore, the time decay weighting function ,in This is the attenuation coefficient; for inflammation-related indicators, The range of values ​​is moon .

[0012] Furthermore, the operator library includes radiomics feature operators for calculating target regions in medical images. Texture heterogeneity index The calculation formula is as follows: in, For the region The total number of voxels within, voxels Standardized image values ​​at the location, The average gray level within the region. As the geometric center of the region, The standard deviation is Gaussian space weighting function.

[0013] Furthermore, the standard deviation of the Gaussian space weighting function The value is set based on the spatial resolution of the image and the typical size of the target pathological structure. .

[0014] Furthermore, the operator library includes multimodal interaction feature operators for processing two standardized features. and Nonlinear combination is performed to generate interactive features. The calculation formula is as follows: in, A scale-sensitive parameter, It is a very small constant.

[0015] On the other hand, the present invention proposes a risk prediction system for interstitial lung disease in rheumatoid arthritis, including a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements the above-mentioned adaptive feature engineering method based on closed-loop feedback and uses the generated final feature set to train a final machine learning model for predicting the risk of interstitial lung disease.

[0016] The beneficial effects of this invention are: This invention establishes an adaptive closed-loop control system with model performance as feedback signal, achieving dynamic optimization and autonomous evolution of the feature engineering process. This fundamentally overcomes the model performance bottleneck and insufficient generalization ability caused by relying on prior static feature sets. The iterative framework of this invention quantitatively evaluates the performance gain of the feature set generated in each iteration on the surrogate model. It is then input into a preset strategy decision function Φ, which dynamically adjusts the control strategy vector accordingly. This vector precisely controls the selection probability distribution of the feature construction operator. Maximum depth of feature generation and screening retention rate .when At that time, the system increases the weight of the currently effective operators and appropriately boosts them. and To deepen local optimization; when At that time, the system will automatically reset. and reduce and This forces a shift towards global exploration. It enables feature engineering to break free from the constraints of fixed paradigms, autonomously discovering and strengthening complex feature combinations with high discriminative power based on specific data distributions and prediction tasks. This achieves a paradigm shift from manually designed features to system-evolved features, significantly improving the adaptability and robustness of the model.

[0017] This invention achieves deep quantitative fusion of multi-source heterogeneous medical data by integrating parameterized feature operators with clear pathological orientation, accurately capturing key dynamic patterns and cross-modal interaction effects in disease evolution. Instead of employing general feature extraction methods, this invention customizes a series of interpretable quantitative operators specifically for the pathophysiological characteristics of RA-ILD. The temporal dynamics operator introduces a time-weighted function ω(tj) with a decay coefficient λ to calculate the weighted trend of the historical trajectory of biomarkers, transforming the clinical concept of recent inflammatory activity into a computable continuous variable. The radiomics operator constructs a texture heterogeneity index H by combining voxel grayscale differences with a Gaussian spatial weighting function Gσ with the lesion's geometric center as the origin and a standard deviation σ = 5 mm. This index specifically responds to the spatially constrained, subtle structural changes characteristic of early ILD. The multimodal interaction operator explicitly models the synergistic or antagonistic relationships between different modal features through nonlinear functions. The parameter γ acts as a scale-sensitive gating mechanism, automatically filtering out clinically significant co-occurring abnormal patterns. The collaborative work of these operators generates a high-information-density feature set that integrates temporal dynamics, spatial texture, and cross-modal correlations, providing the model with deep predictive factors directly related to disease mechanisms.

[0018] Faced with an exponentially large feature space that can be generated by complex combinations of original variables, the closed-loop control system of this invention achieves controlled directional search. The strategy decision function Φ is based on performance feedback. It autonomously switches between two modes: reinforcing the current effective path and randomizing the path to escape local optima. The function in this invention is not a simple heuristic rule, but rather uses a quantitative threshold. , The coordinated adjustment of each component of the policy vector systematically guides the search direction. This allows the system to quickly focus on potential effective feature construction regions while avoiding performance stagnation caused by premature convergence to suboptimal solutions. Thus, without exhaustive search, it obtains a final feature set with high probability and efficiency, possessing both high performance and strong generalization ability, thereby improving the overall robustness and practicality of the method.

[0019] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0021] This embodiment describes a risk prediction system for rheumatoid arthritis-related interstitial lung disease based on self-evolving feature engineering. On the hardware side, it is deployed on a server cluster equipped with high-performance computing units and secure data storage. On the software side, it is implemented as a modular application including a data interface, a processing engine, and a user interface. The core of the implementation lies in the construction of the "adaptive feature engineering engine," which dynamically optimizes the entire process from raw data to predicted features through a closed-loop control system that uses model performance as feedback signals.

[0022] After being anonymized, the raw data enters the preprocessing pipeline. The standardization of numerical clinical and laboratory indicators uses the following formula: The parameters are defined as follows: : indicates the first The patient, the first Items, at a point in time The standardized value.

[0023] : Represents the corresponding original measurement value.

[0024] : indicates the first The arithmetic mean of the indicator is calculated at all time points of all patients in the training set. , This represents the total number of measurements.

[0025] : indicates the first The standard deviation of the indicator is calculated at all time points of all patients in the training set. .

[0026] Eliminating the differences in dimensions and numerical ranges among different medical indicators and converting them into a distribution with a mean of 0 and a standard deviation of 1 provides a stable numerical foundation for subsequent distance- or gradient-based machine learning algorithms.

[0027] In this embodiment, the temporal dynamics feature operator is used to quantify the directional change trend of indicators within a specific observation period. For patients... Indicators At the assessment time point Select time window (For example, (Previous 24 months), the window includes Valid measurement points The calculation formula is as follows: The parameters are defined as follows: : is an indicator At the point of time The measured values ​​(already standardized).

[0028] : is an indicator The mean of all measurements within the window. .

[0029] : is an indicator The standard deviation of all measurements within the window. .

[0030] Time decay weight, .parameter The attenuation coefficient controls the attenuation rate of historical data weights. In this embodiment, a grid search is performed on the validation set to determine the attenuation rate for inflammatory indicators (such as CRP and ESR). The optimal interval is moon This corresponds to a half-life of approximately 5.8 to 8.7 months, highlighting the impact of recent inflammatory activity on ILD risk.

[0031] The minimal constant introduced to prevent the denominator from being zero and to increase numerical stability is usually taken as... .

[0032] Calculated It is a dimensionless scalar. Its absolute value reflects the strength of the trend, and its sign (positive / negative) reflects the direction of the trend (upward / downward).

[0033] In this embodiment, the radiomics feature operator is calculated based on the segmented lung interstitial region Ω from lung CT images (obtained through a pre-trained deep learning segmentation model). The texture heterogeneity index H is defined as: The parameters are defined as follows: : The total number of voxels contained within region Ω.

[0034] x: Represents the voxel coordinate in three-dimensional space .

[0035] I(x): The standardized CT value at voxel x (e.g., linearly mapping Hounsfield units to the [0, 1] interval).

[0036] : Average gray level within region Ω .

[0037] voxels To the geometric center of the region Euclidean distance. Depend on The calculation yielded the result.

[0038] Standard deviation is The one-dimensional Gaussian kernel function, .

[0039] parameter The rate of spatial weight decay is controlled. In this embodiment, the spatial resolution of CT images (e.g., 1mm×1mm×1mm) and the physical size of typical early ILD lesions (5-20mm) are combined to... Set to 5mm. This setting makes the index most sensitive to medium-range texture variations.

[0040] In this embodiment, a multimodal interaction feature operator is used to generate cross-modal nonlinear interaction features. For two standardized features... and Its generation formula is: The parameters are defined as follows: Standardized feature values ​​from different data modalities, for example This could be serum IgG levels. It can be the percentage of ground-glass opacity on HRCT.

[0041] : The constant to prevent division by zero, take .

[0042] Scale-sensitive parameter It controls the decay rate of the effect of the difference between two eigenvalues ​​on the interaction term. In this embodiment, through experiments with various candidate biomarker combinations, it was found that... Setting it to 1.0 will produce interactive features with clinical interpretability and statistical discriminative power in most cases. When When the two eigenvalues ​​are equal, the exponential term is 1; if the difference exceeds approximately 2.3 units, the exponential term decays to below 0.1, significantly suppressing the contribution of the interaction term.

[0043] In this embodiment, the iterative process of the adaptive engine is governed by an explicit feedback control logic. Let the overall performance metric (e.g., AUC) of the lightweight proxy model on the validation set after the k-th iteration be... The performance gain is calculated as follows: Strategy decision function according to Update control policy vector ,in The selection probability distribution for each operator, To the maximum generation depth, The retention rate is used to filter features. The update rules are as follows: like ( ): This is considered significant progress. .in, For this round of matches Operator identifier vectors that make positive contributions This is the learning rate. Meanwhile, , .

[0044] like ( ): This is judged as a slight improvement. Then It has entered a stable utilization stage.

[0045] like If the system is deemed stagnant or degenerate, the exploration mechanism will be triggered. ,in Initially uniform distribution, It is Gaussian noise. Meanwhile, , .

[0046] In this embodiment, the original risk score output by the final stacked ensemble model It needs to be calibrated using PlattScaling to adjust its output probability. Matched to the true incidence rate. The calibration function is: Among them, parameters (Slope) and The intercept is estimated by fitting a logistic regression model to a separate calibration dataset. This calibration dataset is not involved in any of the aforementioned feature engineering or model training. After calibration, Brier scores and calibration curves are used for evaluation to ensure the accuracy of the predicted probabilities.

[0047] In this embodiment, the SHAP (SHapley Additive exPlanations) framework is used for attribution analysis of the final model. For each patient sample, its feature vector... Predicted value Compared with baseline expectations The difference is attributed to each feature. : in, Features The SHAP value is obtained. In this embodiment, the TreeSHAP algorithm is used for efficient computation of the tree ensemble model part, and the DeepSHAP algorithm is used for approximate estimation of the neural network part. A waterfall plot interpretation used to generate feature importance rankings and individual predictions.

[0048] The workflow of this invention in clinical practice begins with secure integration with hospital information systems such as EMR, LIS, and PACS to automatically extract longitudinal clinical data and imaging information of rheumatoid arthritis patients. The system first standardizes and aligns the multi-source heterogeneous data to form a feature pool including basic indicators. Then, a self-evolving feature engineering engine is activated, operating iteratively: in each round, based on the current strategy, operators such as temporal dynamics, radiomics, and multimodal interaction are selected from a pre-set library of pathology-oriented operators to expand and combine the existing feature pool, generating higher-order features such as weighted trend strength of inflammatory indicators or texture heterogeneity indices. Subsequently, through a performance feedback loop, the relative gain of the lightweight proxy model's performance indicators on the independent validation set is used to dynamically adjust the operator selection probability, feature generation depth, and screening intensity in subsequent iterations. This process continues until performance convergence, ultimately outputting a stable feature set adaptively optimized for the prediction task.

[0049] Based on the feature set, the system trains the final ensemble prediction model and performs probability calibration using independent datasets to ensure that its output risk score is consistent with the actual incidence rate. In clinical outpatient settings, when doctors assess patients, the system can access the patient's time-series follow-up data and recent CT images in real time, automatically running the aforementioned feature extraction pipeline. The generated features are then input into the deployed prediction model to calculate the individualized risk probability of the patient developing interstitial lung disease in the future. The system also provides interpretive reports on the model's decisions, such as using SHAP value analysis to identify the main contributing features of this risk assessment, such as the recent upward trend of CRP, the interaction effect of specific image texture features and serological indicators, forming a structured risk stratification and basis summary to assist clinicians in making more accurate follow-up or intervention decisions. The entire system can periodically restart its self-evolution process using newly accumulated clinical data to achieve iterative updates and performance maintenance of the model.

[0050] In summary, this invention proposes an adaptive feature engineering method based on closed-loop feedback and a risk prediction system for rheumatoid arthritis-related interstitial lung disease. By establishing an adaptive closed-loop control system with model performance as feedback signal, iteratively executes feature generation, screening, evaluation, and strategy updates, dynamically and automatically constructing and optimizing the prediction feature set from multimodal data such as clinical time-series data and medical images. The system employs parameterized operators with clear pathological orientation to deeply quantify temporal dynamics, image texture, and multimodal interactions, and integrates a model calibration and attribution explanation framework. This invention can generate high-performance, interpretable risk prediction models, solving the problems of traditional methods relying on static features, difficulty in capturing complex temporal interactions, and lack of clinical interpretability.

[0051] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for predicting the risk of interstitial lung disease in rheumatoid arthritis based on self-evolutionary feature engineering, characterized in that, include: Data preparation: Obtain the raw multimodal time-series data of patients in the training set and perform standardized preprocessing to form an initial feature pool; Feature generation and filtering: Based on the control strategy vector of the current iteration round, feature construction operators are selected from the predefined operator library to process the current feature pool to generate new features, and the feature pool is filtered according to the filtering rules in the control strategy vector to obtain the updated feature set; Performance evaluation and feedback: Train the agent model using the updated feature set and evaluate its performance metrics on the validation set, calculating the performance gain of this round relative to the previous round; Strategy decision and update: Input the performance gain into a preset strategy decision function, and output the updated control strategy vector for the next iteration based on the magnitude of the performance gain; The control strategy vector is used to dynamically adjust the selection probability of the feature construction operator, the complexity of the feature generation process, and the retention rate of feature selection; the iterative process continues until the preset iteration termination condition is reached, and the final feature set is output.

2. The method as described in claim 1, characterized in that, In the performance evaluation and feedback step, the performance gain of the k-th iteration The calculation formula is: in, Let be the performance metric value of the agency model in the k-th round. To prevent division by zero of extremely small constants; In the strategy decision and update step, the strategy decision function updates the control strategy vector according to the following rules: like This increases the probability of selecting operators that contribute positive performance gains and increases the parameter values ​​of feature generation complexity and feature selection retention rate. like This reduces the selection probability preference of the current operator, decreases the parameter values ​​of feature generation complexity and feature selection retention rate, and introduces randomized exploration of operator selection; in, and It is a preset positive threshold, and .

3. The method as described in claim 2, characterized in that, The control strategy vector Represented as ,in Choose a probability distribution for the operator. Controls the maximum depth of feature generation. The feature selection retention rate; the strategy decision function is adjusted by... , , The value is used to update the control strategy vector.

4. The method as described in claim 1, characterized in that, The standardized preprocessing in the data preparation steps, for the first... For each clinical or laboratory indicator, its raw measurement value Convert to standardized value The conversion formula is: in, and These are the arithmetic mean and standard deviation of the indicator calculated at all time points for all patients in the training set, respectively.

5. The method as described in claim 1, characterized in that, The operator library includes time-series dynamic feature operators used to calculate the weighted trend strength of clinical indicators within the observation time window. The calculation formula is as follows: in, For indicators at a point in time The standardized value, and All of them in the window The mean and standard deviation, The time decay weighting function is... It is a very small constant.

6. The method as described in claim 5, characterized in that, The time decay weighting function ,in This is the attenuation coefficient; for inflammation-related indicators, The range of values ​​is moon .

7. The method as described in claim 1, characterized in that, The operator library includes radiomics feature operators for calculating target regions in medical images. Texture heterogeneity index The calculation formula is as follows: in, For the region The total number of voxels within, voxels Standardized image values ​​at the location, The average gray level within the region. As the geometric center of the region, The standard deviation is Gaussian space weighting function.

8. The method as described in claim 7, characterized in that, The standard deviation of the Gaussian space weighting function The value is set based on the spatial resolution of the image and the typical size of the target pathological structure. .

9. The method as described in claim 1, characterized in that, The operator library includes multimodal interaction feature operators for processing two standardized features. and Nonlinear combination is performed to generate interactive features. The calculation formula is as follows: in, A scale-sensitive parameter, It is a very small constant.

10. A risk prediction system for interstitial lung disease in rheumatoid arthritis, characterized in that, It includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the functionality as claimed in claim 1. The method for predicting the risk of interstitial lung disease in rheumatoid arthritis as described in any one of the above methods, and the final machine learning model for predicting the risk of interstitial lung disease is trained using the generated final feature set.

Images