Rare disease patient treatment scheme evaluation and recommendation method
By employing multimodal data fusion and group-based strategy optimization, the problems of difficulty in multimodal feature fusion and sample scarcity in rare disease data scenarios were solved, enabling personalized treatment plan recommendations for rare disease patients and improving the accuracy of risk prediction and clinical application value.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DALIAN UNIV OF TECH
- Filing Date
- 2025-12-22
- Publication Date
- 2026-05-05
AI Technical Summary
In existing technologies, temporal misalignment and modal heterogeneity of multimodal data make feature fusion difficult, and models cannot uniformly model image, text, structured and genetic features. Existing reinforcement learning algorithms rely on value network estimation, which requires a large number of interactive samples to converge, making them unsuitable for rare disease data scenarios with scarce samples.
By integrating multimodal data, preprocessing and aligning data, extracting features, using a cross-modal Transformer model and reinforcement learning (RL) framework, and combining the Monte Carlo Dropout method, we achieve unified representation and risk prediction of multimodal data. We also use a group-relative strategy to optimize the GRPO algorithm to generate individualized treatment plans.
It significantly improves the completeness of the status characteristics of rare disease patients and the accuracy of risk prediction, overcomes the overfitting problem of traditional reinforcement learning, achieves stable learning and interpretability under small sample size, has adaptive closed-loop update capability, and improves the scientific nature and clinical application value of individualized treatment for rare diseases.
Smart Images
Figure CN121983331A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence and intelligent medical diagnosis and treatment systems, specifically to a method for evaluating and recommending treatment plans for patients with rare diseases. Background Technology
[0002] In recent years, with the continuous development of medical informatics, gene testing, and intelligent image analysis technologies, clinicians have access to increasingly richer data types in the management of rare disease patients, covering medical images (CT, MRI, echocardiography), structured electronic health records (EHR), text-based diagnostic records, and gene testing results. These data constitute a complete trajectory of patient disease progression over time, containing a wealth of implicit patterns related to risk evolution, treatment response, and long-term prognosis, serving as crucial drivers for precision medicine and intelligent decision-making. Existing research has made some progress in multimodal clinical data analysis, including medical image segmentation, aortic diameter measurement, gene mutation association analysis, risk stratification, and prognostic assessment. However, most of these studies remain at the level of specific tasks (such as lesion detection and risk scoring), failing to achieve cross-modal, temporal, and individualized systematic modeling. Aortic-related rare diseases, exemplified by Marfan syndrome, exhibit significant individual variability and temporal dynamics in their disease progression: the rate of aortic dilation, blood pressure fluctuations, genotype, and medication adherence all contribute to determining the timing of surgery and treatment strategies. Traditional manual assessment methods rely primarily on physician experience and discrete indicators, making it difficult to capture the implicit dynamic relationships that depend on long-term factors. With the rapid development of deep learning technology, Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), BioBERT (Bio-Language Model), and Tab Transformers have demonstrated outstanding performance in medical data analysis; simultaneously, Reinforcement Learning (RL) methods show great potential in policy optimization and personalized recommendations. Nevertheless, existing methods still have serious limitations in recommending treatment plans for rare diseases: on the one hand, the temporal misalignment and modal heterogeneity of multimodal data make feature fusion difficult, and models cannot uniformly model image, text, structured, and genetic features; on the other hand, existing reinforcement learning algorithms (such as Proximal Policy Optimization, PPO) rely on value network estimation, requiring a large number of interaction samples to converge, making them unsuitable for rare disease data scenarios with scarce samples. Therefore, how to achieve cross-modal fusion, temporal risk prediction, and controllable reinforcement learning policy optimization under limited sample conditions has become a key technical problem that urgently needs to be solved in the field of intelligent medicine.
[0003] The ViT in this invention is from Dosovitskiy A. An image is worth 16x16 words: Transformers for image recognition at scale[J]. arXiv preprint arXiv:2010.11929, 2020.
[0004] The CNN is from LeCun Y, Bottou L, Bengio Y, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 2002, 86(11): 2278-2324.
[0005] BioBERT is from Lee J, Yoon W, Kim S, et al. BioBERT: a pre-trained biomedical language representation model for biomedical text mining[J]. Bioinformatics, 2020, 36(4): 1234-1240.
[0006] TabTransformer is from Huang X, Khetan A, Cvitkovic M, et al. Tabtransformer: Tabular data modeling using contextual embeddings[J]. arXiv preprint arXiv:2012.06678, 2020.
[0007] Reinforcement Learning is from Sutton R S, Barto A G. Reinforcement learning: An introduction[M]. Cambridge: MIT press, 1998, and Proximal Policy Optimization is from Schulman J, Wolski F, Dhariwal P, et al. Proximal policy optimization algorithms[J]. arXiv preprint arXiv:1707.06347, 2017.
[0008] GRPO comes from Shao Z, Wang P, Zhu Q, et al. Deepseekmath: Pushing thelimits of mathematical reasoning in open language models[J]. arXiv preprintarXiv:2402.03300, 2024,.
[0009] Monte Carlo Dropout comes from Gal Y, Ghahramani Z. Dropout as a bayesianapproximation: Representing model uncertainty in deep learning[C] / / international conference on machine learning. PMLR, 2016: 1050-1059.
[0010] The method of representing gene modalities using Embedding layers comes from Sanjaya P, Maljanen K, Katainen R, et al. Mutation-Attention (MuAt): deep representation learning of somatic mutations for tumour typing and subtyping[J]. Genome medicine, 2023,15(1): 47. Summary of the Invention
[0011] The purpose of this invention is to address the problems in existing technologies where temporal misalignment and modal heterogeneity of multimodal data make feature fusion difficult, and models cannot uniformly model image, text, structured, and genetic features; existing reinforcement learning algorithms rely on value network estimation, require a large number of interactive samples to converge, and are not suitable for rare disease data scenarios with scarce samples.
[0012] To address the above problems, this invention provides a method for evaluating and recommending treatment plans for patients with rare diseases, including: Step 1: Multimodal data acquisition and integration. The acquired data includes three categories: structured data, semi-structured data, and unstructured data. The acquired data is then integrated, and the integrated multimodal data is uniformly indexed and stored in five sub-tables: image, EHR, text, gene, and metadata. Step 2: Data preprocessing and alignment. A uniform time grid is set to divide the multimodal data and obtain observation data. Each observation data is mapped to the most recent time according to the timestamp information. Step 3 involves feature extraction from the multimodal data, and after concatenating the unified features of each modality, the dimensionality is reduced through a linear projection layer to obtain a unified multimodal fusion sequence. Step 4 is based on the cross-modal Transformer model. The multimodal fusion sequence is taken as input and the output is the time-series state matrix of the whole sequence. The risk prediction is performed on the output vector corresponding to the time step in the time-series state matrix of the whole sequence through the prediction function. Step 5 is based on the reinforcement learning (RL) framework, where the output vector, i.e., the patient's state, is used as the state input to obtain the treatment plan as the action; Step 6 uses the Monte Carlo Dropout method to perform forward propagation on the same input to obtain multiple sets of prediction results. The average prediction value of the multiple sets of prediction results is the final risk probability prediction.
[0013] In the preferred approach, step 1, multimodal data acquisition and integration, includes: Step 1-1: Multimodal data acquisition; The collected data includes three categories: Structured data: Electronic health record (EHR) data, including blood pressure, heart rate, medication history, surgical records, and blood biochemistry indicators; Semi-structured data: refers to gene data, including information on mutation sites in FBN1, TGFBR2, and SMAD3; Unstructured data: Imaging data, including CT images, MRI images, and echocardiograms; text data, including doctor's diagnostic reports, medical records, and discharge summaries; Steps 1-2: Multimodal data integration; Image data in DICOM format is uniformly converted to NIfTI format; EHR data tables are standardized with timestamps and units, namely blood pressure in mmHg and diameter in mm, and processed into structured EHR data tables. The text data is encoded, converted to UTF-8 format, and the record time field is extracted. The FBN1, TGFBR2, and SMAD3 mutation site information from the gene data, after variant annotation and pathogenicity grading, is used to generate a structured table containing fields for patient ID, gene name, mutation type, and pathogenicity level. This table is then linked to the main table of the EHR data table using the patient's unique ID (UID) as the primary key. The linking method is as follows: The integrated multimodal data is indexed by UID and stored in five sub-tables: image, EHR, text, gene, and metadata.
[0014] In the preferred embodiment, step 2, data preprocessing and alignment, includes: Step 2-1: Set a uniform time grid; The entire follow-up period was divided into: Let there be discrete time points, denoted as . ,in Indicates the first Each time step For time step index; For each modal data record, select the timestamp closest to the current time step. As observational data; For the observation data in each modality, map it to the most recent time step based on the timestamp information. The mapping method is as follows: Based on the principle of minimizing time difference, the difference between the observation time and the center point of each time step is calculated for each record, and it is assigned to the nearest time step; if an observation is missing at a certain time step, a dual-strategy interpolation mechanism is used to complete the data, specifically as follows: For blood pressure, heart rate, and blood biochemical indicators in structured continuous variable EHR data, the forward imputation method was used; Spline interpolation was used to measure the aortic diameter, wall thickness, flow velocity, and compliance obtained from aortic CT images, MRI images, and echocardiography. Anomaly detection is performed, and the 3σ principle and sliding window statistics are used to output a time-series feature matrix in a uniform format with the shape of [number of patients × time step × feature dimension].
[0015] In the preferred approach, step 3, multimodal feature representation learning, includes: The image modality uses a visual transformer (ViT) to divide the aortic CT image region into several image blocks. After passing through a multi-layer attention mechanism, the image features are output, as shown in the formula: in, In time step Extracted image features, Image modal feature dimension; The text modality uses the medical language model BioBERT to extract semantic embeddings and output text features, as shown in the formula: in, In time step Extracted image features, For text modal features; The structured EHR modality, after time standardization and unit unification, is processed using TabTransformer. Combined with class embedding and an attention layer, it outputs EHR features, as shown in the formula: in, In time step Extracted EHR features, EHR modal feature dimension; Vectorization of gene modalities is achieved by constructing a learnable mutation embedding matrix, as shown in the formula: in, Represents the total number of all possible mutation sites, in each row. The embedding representation corresponding to the i-th mutation site; for a sample at time step The mutation set is given by the formula: in, This indicates that the sample is at time step The index of the i-th mutation site corresponds to the specific mutation event in the genome. This indicates the number of mutations detected in the sample at each time step; a lookup table is performed: the index of each mutation site is... Mapped to the corresponding embedding vector The mutation vector sequence is obtained, and the formula is: Mean pooling is applied to the mutation vector sequence to map the variable-length mutation set to a fixed-dimensional gene modality feature vector, as shown in the formula: in, In time step Extracted gene features, For gene modality features; After unifying and stitching the features of each modality, a linear projection layer is used to reduce the dimensionality to a unified state, as shown in the formula: in, Indicates a linear projection layer. This represents a vector concatenation operation. This indicates that images, text, EHRs, and gene modalities are at time steps. The comprehensive representation, where d is the dimension of the comprehensive representation; Ultimately, for Perform the above operations separately to obtain the multimodal fusion sequence. .
[0016] In the preferred approach, step 4, which integrates modeling and time-series risk prediction, includes: Based on the cross-modal Transformer model, multimodal fusion sequences are As input, the output is the time-series state matrix of the entire sequence, as shown in the formula: in, , represents the set of patient state representations obtained after time-series modeling; each For matrix H corresponding to time step The output vector is used to calculate the future time window through a linear classification layer. The probability prediction function for the occurrence of internal aortic events is as follows: in, To predict probabilities, For the Sigmoid function, The model parameters are trained using the cross-entropy loss function, as shown in the formula: in, It indicates whether an aortic dilatation / dissection event will occur within the future window, and the output results include individual risk scores and time-series risk curves.
[0017] In the preferred approach, step 5 treatment is recommended to include: Based on a reinforcement learning (RL) framework, the output vector... That is, the patient's state is used as the state input, the treatment plan is used as the action 'a', and the reward function is defined as follows: in, Risk of non-intervention For treatment costs, As a penalty for violating clinical guidelines, As a weighting factor; Based on the Group Relative Policy Optimization (GRPO) algorithm, each round starts from the policy network. Sample G candidate actions The formula for calculating the group mean and standard deviation is: in, The group mean Standard deviation, Let be the numerical stability constant, and the standardization dominance formula be defined as follows: The policy loss function is: in, , indicating the current strategy Compared to the previous strategy Under the same state h, for the same action The probability ratio, This is the cutting factor.
[0018] In the preferred embodiment, the uncertainties and explanations of step 6 include: Step 6-1: Use the Monte Carlo Dropout method to perform M forward propagations on the same input to obtain multiple sets of prediction results. Average forecast value For the final risk probability prediction, the confidence interval ,in and The prediction results are as follows The 2nd and 97th percentiles are used as the prediction confidence interval. If the confidence interval span is greater than 0.1, it is marked as a high uncertainty sample. Step 6-2: For the image modality, the gradient-weighted class activation mapping (Grad-CAM) technique is used to visualize the region of interest of the model on the aortic image. A heat map is generated by the gradient-weighted feature map and superimposed on the grayscale image to display the key lesion area. For structured EHR and text modality, attention weighting analysis is used to calculate feature importance and generate key variable contribution descriptions. For recommendation strategies, decision explanations are generated through counterfactual reasoning, specifically including: Let the current optimal treatment strategy be... Alternative treatment strategies are Calculate the risk probability predictions for both. and The risk difference is defined as ,like This indicates that the alternative reduces the risk of future aortic events in patients.
[0019] The beneficial effects of this invention are as follows: Addressing the issues of limited sample size, diverse modalities, and fragmented information in rare diseases, this invention constructs a multimodal data fusion system, achieving unified representation of image, EHR, text, and genetic data, significantly improving the completeness of patient status characteristics and the accuracy of risk prediction. In the treatment recommendation stage, the Group Relative Policy Optimization (GRPO) algorithm is introduced to stably learn the optimal treatment plan with a small sample size, overcoming the problems of overfitting and unstable training in traditional reinforcement learning. Regarding clinical usability, the system combines Monte Carlo Dropout and a multimodal interpretation mechanism to output risk confidence intervals and visual interpretation reports, achieving transparency and verifiability of results. Furthermore, the system possesses adaptive closed-loop update capabilities, continuously collecting follow-up and feedback data for online optimization, maintaining long-term model stability and transferability over time. In summary, this invention surpasses existing technologies in data fusion, policy optimization, model interpretability, and system self-learning capabilities, significantly enhancing the scientific rigor and clinical application value of personalized treatment for rare diseases. Attached Figure Description
[0020] Figure 1 This is the overall architecture diagram; Figure 2 This is a diagram of the multimodal data acquisition and integration architecture; Figure 3 This is a flowchart of data preprocessing and time alignment; Figure 4 Multimodal feature representation network architecture diagram; Figure 5 This is a flowchart of time-series risk prediction; Figure 6 This is a flowchart of the GRPO strategy; Figure 7 This is the flowchart for the Uncertainty and Explanation module. Detailed Implementation
[0021] Example 1: This invention relates to a method for evaluating and recommending treatment plans for patients with rare diseases based on multimodal data fusion and group-based relative strategy optimization, with Marfan syndrome as a typical application scenario. This method models the rare disease treatment process as a dynamic state-action decision-making system, achieving a closed loop of risk prediction and individualized recommendation through joint learning of multimodal data. The entire process consists of seven modules: a multimodal data acquisition module, a data preprocessing and alignment module, a feature representation learning module, a fusion modeling and temporal risk prediction module, a treatment recommendation module, an uncertainty and interpretation module, and a clinical data feedback and validation module. This invention designs the entire process from patient enrollment to recommendation plan generation as an end-to-end repeatable computational system, with the overall architecture as follows: Figure 1 As shown.
[0022] Step 1: Multimodal data acquisition and integration; The purpose of this step is to establish a complete multimodal data warehouse for patients with rare diseases. The collected data includes three categories: (1) structured data, which comes from quantifiable time series variables in electronic medical records (EHRs), such as blood pressure, heart rate, medication history, surgical records, and blood biochemical indicators, with clear fields and units; (2) semi-structured data, which refers to gene data (such as FBN1, TGFBR2, SMAD3 mutation site information), whose data format usually contains fixed fields but the content length and description structure are variable; (3) unstructured data, including imaging data (CT, MRI, echocardiography) and free text medical records (doctor's diagnosis report, medical record, discharge summary), which need to be structured through feature extraction or language model before they can participate in unified modeling.
[0023] Step 1-1: Multimodal data acquisition; Imaging data, including: aortic CT images, MRI images, and cardiac ultrasound; EHR data include: blood pressure, heart rate, medication history, surgical records, and blood biochemical indicators; Text data includes: doctor's diagnosis reports, medical records, and discharge summaries; Genetic data, including: FBN1, TGFBR2, and SMAD3 mutation site information; Steps 1-2: Multimodal data integration; Image data in DICOM format is uniformly converted to NIfTI format; EHR data tables are standardized with timestamps and units (blood pressure unit mmHg, diameter unit mm), and processed into structured EHR data tables; The text data is encoded, converted to UTF-8 format, and the record time field is extracted. The mutation site information of FBN1, TGFBR2, and SMAD3 in the gene data is annotated and pathogenicity graded to generate a structured table containing fields such as patient ID, gene name, mutation type, and pathogenicity grade. The table is then linked to the EHR main table using the patient's unique ID (UID) as the primary key.
[0024] The multimodal data that has undergone preliminary verification, formatting, and validation are stored in five sub-tables—image, EHR, text, gene, and metadata—using UID as a unified index, forming a rare disease multimodal data warehouse that supports time series modeling and cross-modal analysis.
[0025] For example, a patient's aortic CT images (3 times), MRI images (2 times), EHR data (24 records), text data (7 pieces), and genetic data (1 piece, FBN1 mutation positive) are all uniformly numbered with UID-A001 and imported into the database, generating five sub-tables corresponding to images, EHR, text, genes, and metadata.
[0026] Figure 2 It demonstrates the relationship between data acquisition and structure mapping: data from different sources are input into the database through a unified interface, and each modality forms a corresponding structure table or embedded file.
[0027] Step 2: Data preprocessing and alignment; This step aims to achieve time synchronization and outlier correction for cross-modal data, ensuring comparability of different modalities on a unified time axis. For example... Figure 3 The diagram shows the alignment process.
[0028] Step 2-1: Set a uniform time grid; The entire follow-up period was divided into: Let there be discrete time points, denoted as . ,in Indicates the first Each time step For time step index; For each modality of data (image data, EHR data, text data, genetic data), select the timestamp closest to the current time step. As observations (if multiple observations exist within the same time window, a weighted average is taken, with the weights calculated using the sampling time distance), they are used as observation data. For the observation data in each modality, it is mapped to the nearest time step based on the timestamp information. The mapping process is based on the principle of minimizing time difference. For each record, the difference between its observation time and the center point of each time step is calculated, and the record is assigned to the nearest time step. If a time step is missing an observation, a dual-strategy interpolation mechanism is used to complete the data. For structured continuous variables (blood pressure, heart rate, and blood biochemical indicators in EHR data), forward fill (FF) was used. Spline interpolation (SI) was used to achieve smooth transitions for imaging measurements (aortic diameter, wall thickness, flow velocity, and compliance obtained from aortic CT images, MRI images, and echocardiography). Simultaneously, anomaly detection is performed: using the 3σ principle and sliding window statistics, abnormal heart rate, aortic diameter, and blood pressure data points are removed; finally, a time-series feature matrix with a uniform format is output, with the shape: [number of patients × time step × feature dimension]; For example: A patient's blood pressure record was missing for two weeks between March 2, 2024 and April 13, 2024. The system automatically used forward padding to complete the data to 118 / 76 mmHg; the aortic CT measurement value was in the... and The diameters are 44mm and 46mm respectively, and the system uses spline interpolation for completion. The time is 45mm; abnormal data with heart rate observations exceeding 130 beats / min are automatically identified and removed; Step 3: Multimodal feature representation learning; This module extracts high-dimensional feature vectors for each modality using a multimodal deep neural network. The overall network architecture is as follows: Figure 4 As shown.
[0029] The image modality uses a VisionTransformer (ViT) to divide the aortic CT image region into several image patches. After passing through a multi-layer attention mechanism, the image features are output, as shown in the formula: in, In time step Extracted image features, Image modal feature dimension; The text modality uses the medical language model BioBERT to extract semantic embeddings and output text features, as shown in the formula: in, In time step Extracted image features, For text modal features; The structured EHR modality, after time standardization and unit unification, is processed using TabTransformer. Combined with class embedding and an attention layer, it outputs EHR features, as shown in the formula: in, In time step Extracted EHR features, EHR modal feature dimension; Vectorization of gene modalities is achieved by constructing a learnable mutation embedding matrix, as shown in the formula: in, Represents the total number of all possible mutation sites, in each row. The embedding representation corresponding to the i-th mutation site; for a sample at time step The mutation set is given by the formula: in, This indicates that the sample is at time step The index of the i-th mutation site corresponds to the specific mutation event in the genome. This indicates the number of mutations detected in the sample at each time step; a lookup table is performed: the index of each mutation site is... Mapped to the corresponding embedding vector The mutation vector sequence is obtained, and the formula is: Mean pooling is applied to the mutation vector sequence to map the variable-length mutation set to a fixed-dimensional gene modality feature vector, as shown in the formula: in, In time step Extracted gene features, For gene modality features; After unifying and stitching the features of each modality, a linear projection layer is used to reduce the dimensionality to a unified state, as shown in the formula: in, Indicates a linear projection layer. This represents a vector concatenation operation. This indicates that images, text, EHRs, and gene modalities are at time steps. The comprehensive representation, where d is the dimension of the comprehensive representation; Ultimately, for Perform the above operations separately to obtain the multimodal fusion sequence. .
[0030] For example, ViT extracts 768-dimensional image features, BioBERT generates 768-dimensional text features, TabTransformer outputs 256-dimensional EHR features, and gene embeddings are 64-dimensional; after stitching, a projection layer reduces the dimensions to a unified 512-dimensional representation. .
[0031] Step 4: Integrate modeling with time-series risk prediction; The multimodal fusion feature sequence obtained in step 3 As input, a cross-modal Transformer model is used to model intermodal semantic interactions and temporal dependencies. The Transformer, through a multi-layer self-attention mechanism and a feedforward network, models the dynamic relationships between different time steps, outputting the temporal state matrix of the entire sequence. in, , represents the set of patient state representations obtained after time-series modeling; each For matrix H corresponding to time step The output vector represents the comprehensive patient state representation generated by the model at that moment after incorporating historical information. (The process involves obtaining the temporal state vector.) Then, the future time window is calculated through a linear classification layer. The probability prediction function for the occurrence of internal aortic events is as follows: in, To predict probabilities, For the Sigmoid function, The model parameters are trained using the cross-entropy loss function, as shown in the formula: in, This indicates whether aortic dilatation or dissection will occur within the future window. Outputs include individual risk scores and time-series risk curves. The prediction process is as follows: Figure 5 As shown. For example, the patient's fusion vector sequence As input, the model predicts the risk of aortic events within 6 months. The risk curve shows an upward trend month by month. The results are then used to make intervention decisions in the treatment recommendation module.
[0032] Step 5: Treatment Recommendation Module (Based on Group-Relative Strategy Optimization): The core innovation of this invention is to generate individualized treatment strategies based on risk status.
[0033] The system is based on the reinforcement learning (RL) framework, and the patient's state... As the state input, the treatment plan is taken as action 'a', and the reward function is defined as follows: in, Risk of non-intervention Treatment costs (drug side effects, surgical fees, and the risk of potential long-term complications), As a penalty for violating clinical guidelines, As a weighting factor; In terms of algorithm implementation, the structure design of the policy gradient method Proximal Policy Optimization (PPO) is referenced, and the idea of Group Relative Policy Optimization (GRPO) proposed in recent years is used to improve policy learning in the context of small sample rare diseases. The GRPO algorithm improves policy learning in each round from the policy network. Sample G candidate actions The formula for calculating the group mean and standard deviation is: in, The group mean Standard deviation, Let be the numerical stability constant, and the standardization dominance formula be defined as follows: The policy loss function is: in, , indicating the current strategy Compared to the previous strategy Under the same state h, for the same action The probability ratio, The pruning coefficients are used. This method does not require a value network (Critic) and maintains stable learning even in small-sample rare disease environments. The output is the optimal treatment plan. This includes recommended medication dosages, follow-up appointment intervals, and timing of surgery.
[0034] Strategy optimization process as follows Figure 6 As shown. For example, for a given patient, the system generates G=10 candidate treatment combinations (different... Calculate the corresponding dosage of beta-blockers, follow-up visit frequency, and whether / when surgery is performed. mean Standard deviation Ultimately optimal strategy For "medium dose" "By administering the blocker, following an 8-week follow-up appointment, and recommending elective ascending aortic replacement surgery within six months," the predicted risk decreased to 0.21. ).
[0035] Step 6: Uncertainty and Explanation Module; This is designed to enhance the transparency and usability of the system in clinical applications, providing physicians with interpretable risk predictions and treatment guidelines. The process is as follows: Figure 7 As shown.
[0036] Step 6-1: First, use the Monte Carlo Dropout method to assess the uncertainty of the risk prediction results; During the prediction phase, a Dropout layer is enabled, and the same input is forward-propagated M times to obtain multiple sets of prediction results. Average forecast value For the final risk probability prediction, the confidence interval ,in and The prediction results are as follows The 2nd and 97.5th percentiles are used as the prediction confidence interval. If the confidence interval span is greater than 0.1, it is marked as a high uncertainty sample.
[0037] Step 6-2: The system then performs a multimodal interpretability analysis; For the imaging modality, Grad-CAM (gradient-weighted class activation mapping) technology is used to visualize the region of interest of the model on the aortic image. A heat map is generated by gradient-weighted feature map and superimposed on the grayscale image to display the key lesion area. For structured EHRs and text modalities, feature importance is calculated using SHAP (SHapley Additive exPlanations) or attention weighting analysis to generate descriptions of the contributions of key variables. For example, the system can identify the positive contributions of "elevated blood pressure" and "increased aortic diameter" to the risk score; For reinforcement learning recommendation strategies, counterfactual reasoning is used to generate decision explanations to assess the impact of different treatment actions on risk changes; Let the current optimal treatment strategy be... Alternative treatment strategies are Calculate the risk probability predictions for both. and The risk difference is defined as ,like This indicates that the alternative reduces the risk of future aortic events in patients.
[0038] For example, the system could output an explanation: "If the dose of a beta-blocker is increased by 10%, the risk decreases by approximately 15% ( Ultimately, the system integrates the uncertainty analysis results, Grad-CAM image heatmaps, EHR feature importance ranking, and counterfactual inference results into a visualized explanatory report. The report includes prediction confidence levels, key risk influencing factors, and actionable treatment recommendations, which can be directly referenced by physicians, achieving transparency, verifiability, and clinical interpretability of the model results. For example, for a specific patient, the system assesses the uncertainty of its risk prediction model using... Sub-Monte Carlo Dropout sampling yields the prediction set. Calculate the average value Confidence interval The interval span is .
[0039] The system therefore determined the sample to have a high confidence level. During the interpretability analysis phase, Grad-CAM visualization results showed that the model responded most strongly in the ascending aortic root region, suggesting that the model focuses on locally dilated areas; SHAP analysis results indicated that "systolic blood pressure increase of 15 mmHg," "aortic diameter increase of 2.5 mm," and "previous surgical history" were the top three contributing factors to increased risk; counterfactual inference showed that when 10% increase in receptor blocker dosage (alternative) The system recalculates the risk. , and the optimal solution The difference is This indicates that increasing the dosage can reduce the risk of aortic events by approximately 3%.
[0040] Step 7: System closed-loop update and instance verification; In clinical use, the system's recommendations and physician decisions form a two-way feedback mechanism. New follow-up data is fed back into the database for continuous model training and online optimization. To verify effectiveness, three typical patients were selected for experimental validation: Example 1: Patient A, aortic dilation rate of 3.2 mm / year, the system recommended a medium-dose β-blocker and a follow-up visit in 8 weeks, reducing the predicted risk from 0.31 to 0.21; Example 2: Patient B, with poor blood pressure control and FBN1 mutation, the system recommended combination therapy and early surgery, reducing the predicted risk by 28%; Example 3: Patient C, in a stable condition, the system suggested maintaining the current treatment plan and outputting a high confidence level (95% CI of 0.22–0.25). If the consistency rate between the system's recommendations and expert opinions exceeds 85%, the system is considered to be able to assist physicians in making individualized treatment decisions.
[0041] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the present invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope. All such changes and modifications fall within the scope of the present invention as claimed, which is defined by the appended claims and their equivalents.
Claims
1. A method for evaluating and recommending treatment plans for patients with rare diseases, characterized in that, include: Step 1: Multimodal data acquisition and integration. The acquired data includes three categories: structured data, semi-structured data, and unstructured data. The collected data were integrated, and the integrated multimodal data were indexed in a unified manner and stored in five sub-tables: image, EHR, text, gene, and metadata. Step 2: Data preprocessing and alignment. A uniform time grid is set to divide the multimodal data and obtain observation data. Each observation data is mapped to the most recent time according to the timestamp information. Step 3 involves feature extraction from the multimodal data, and after concatenating the unified features of each modality, the dimensionality is reduced through a linear projection layer to obtain a unified multimodal fusion sequence. Step 4 is based on the cross-modal Transformer model. The multimodal fusion sequence is taken as input and the output is the time-series state matrix of the whole sequence. The risk prediction is performed on the output vector corresponding to the time step in the time-series state matrix of the whole sequence through the prediction function. Step 5 is based on the reinforcement learning (RL) framework, where the output vector, i.e., the patient's state, is used as the state input to obtain the treatment plan as the action; Step 6 uses the Monte Carlo Dropout method to perform forward propagation on the same input to obtain multiple sets of prediction results. The average prediction value of the multiple sets of prediction results is the final risk probability prediction.
2. The method for evaluating and recommending treatment plans for patients with rare diseases according to claim 1, characterized in that, Step 1, multimodal data acquisition and integration, includes: Step 1-1: Multimodal data acquisition; The collected data includes three categories: Structured data: Electronic health record (EHR) data, including blood pressure, heart rate, medication history, surgical records, and blood biochemistry indicators; Semi-structured data: refers to gene data, including information on mutation sites in FBN1, TGFBR2, and SMAD3; Unstructured data: Imaging data, including CT images, MRI images, and echocardiograms; text data, including doctor's diagnostic reports, medical records, and discharge summaries; Steps 1-2: Multimodal data integration; Image data in DICOM format is uniformly converted to NIfTI format; EHR data tables are standardized with timestamps and units, namely blood pressure in mmHg and diameter in mm, and processed into structured EHR data tables. The text data is encoded, converted to UTF-8 format, and the record time field is extracted. The FBN1, TGFBR2, and SMAD3 mutation site information from the gene data, after variant annotation and pathogenicity grading, is used to generate a structured table containing fields for patient ID, gene name, mutation type, and pathogenicity level. This table is then linked to the main table of the EHR data table using the patient's unique ID (UID) as the primary key. The linking method is as follows: The integrated multimodal data is indexed by UID and stored in five sub-tables: image, EHR, text, gene, and metadata.
3. The method for evaluating and recommending treatment plans for patients with rare diseases according to claim 1, characterized in that, Step 2, data preprocessing and alignment, includes: Step 2-1: Set a uniform time grid; The entire follow-up period was divided into: Let there be discrete time points, denoted as . ,in Indicates the first Each time step For time step index; For each modal data record, select the timestamp closest to the current time step. As observational data; For the observation data in each modality, map it to the most recent time step based on the timestamp information. The mapping method is as follows: Based on the principle of minimizing time difference, the difference between the observation time and the center point of each time step is calculated for each record, and it is assigned to the nearest time step; if an observation is missing at a certain time step, a dual-strategy interpolation mechanism is used to complete the data, specifically as follows: For blood pressure, heart rate, and blood biochemical indicators in structured continuous variable EHR data, the forward imputation method was used; Spline interpolation was used to measure the aortic diameter, wall thickness, flow velocity, and compliance obtained from aortic CT images, MRI images, and echocardiography. Anomaly detection is performed, and the 3σ principle and sliding window statistics are used to output a time-series feature matrix in a uniform format with the shape of [number of patients × time step × feature dimension].
4. The method for evaluating and recommending treatment plans for patients with rare diseases according to claim 1, characterized in that, Step 3, multimodal feature representation learning, includes: The image modality uses a visual transformer (ViT) to divide the aortic CT image region into several image blocks. After passing through a multi-layer attention mechanism, the image features are output, as shown in the formula: in, In time step Extracted image features, Image modal feature dimension; The text modality uses the medical language model BioBERT to extract semantic embeddings and output text features, as shown in the formula: in, In time step Extracted image features, For text modal features; The structured EHR modality, after time standardization and unit unification, is processed using TabTransformer. Combined with class embedding and an attention layer, it outputs EHR features, as shown in the formula: in, In time step Extracted EHR features, EHR modal feature dimension; Vectorization of gene modalities is achieved by constructing a learnable mutation embedding matrix, as shown in the formula: in, Represents the total number of all possible mutation sites, in each row. The embedding representation corresponding to the i-th mutation site; for a sample at time step The mutation set is given by the formula: in, This indicates that the sample is at time step The index of the i-th mutation site corresponds to the specific mutation event in the genome. This indicates the number of mutations detected in the sample at each time step; a lookup table is performed: the index of each mutation site is... Mapped to the corresponding embedding vector The mutation vector sequence is obtained, and the formula is: Mean pooling is applied to the mutation vector sequence to map the variable-length mutation set to a fixed-dimensional gene modality feature vector, as shown in the formula: in, In time step Extracted gene features, For gene modality features; After unifying and stitching the features of each modality, a linear projection layer is used to reduce the dimensionality to a unified state, as shown in the formula: in, Indicates a linear projection layer. This represents a vector concatenation operation. This indicates that images, text, EHRs, and gene modalities are at time steps. The comprehensive representation, where d is the dimension of the comprehensive representation; Ultimately, for Perform the above operations separately to obtain the multimodal fusion sequence. .
5. The method for evaluating and recommending treatment plans for patients with rare diseases according to claim 1, characterized in that, Step 4, which integrates modeling and time-series risk prediction, includes: Based on the cross-modal Transformer model, multimodal fusion sequences are As input, the output is the time-series state matrix of the entire sequence, as shown in the formula: in, , represents the set of patient state representations obtained after time-series modeling; each For matrix H corresponding to time step The output vector is used to calculate the future time window through a linear classification layer. The probability prediction function for the occurrence of internal aortic events is as follows: in, To predict probabilities, For the Sigmoid function, The model parameters are trained using the cross-entropy loss function, as shown in the formula: in, It indicates whether an aortic dilatation / dissection event will occur within the future window, and the output results include individual risk scores and time-series risk curves.
6. The method for evaluating and recommending treatment plans for patients with rare diseases according to claim 1, characterized in that, Step 5 treatment recommendations include: Based on a reinforcement learning (RL) framework, the output vector... That is, the patient's state is used as the state input, the treatment plan is used as the action 'a', and the reward function is defined as follows: in, Risk of non-intervention For treatment costs, As a penalty for violating clinical guidelines, As a weighting factor; Based on the Group Relative Policy Optimization (GRPO) algorithm, each round starts from the policy network. Sample G candidate actions The formula for calculating the group mean and standard deviation is: in, The group mean Standard deviation, Let be the numerical stability constant, and the standardization dominance formula be defined as follows: The policy loss function is: in, , indicating the current strategy Compared to the previous strategy Under the same state h, for the same action The probability ratio, This is the cutting factor.
7. The method for evaluating and recommending treatment plans for patients with rare diseases according to claim 1, characterized in that, Step 6 Uncertainties and Explanations include: Step 6-1: Use the Monte Carlo Dropout method to perform M forward propagations on the same input to obtain multiple sets of prediction results. Average forecast value For the final risk probability prediction, the confidence interval ,in and The prediction results are as follows The 2nd and 97th percentiles are used as the prediction confidence interval. If the confidence interval span is greater than 0.1, it is marked as a high uncertainty sample. Step 6-2: For the image modality, the gradient-weighted class activation mapping (Grad-CAM) technique is used to visualize the region of interest of the model on the aortic image. A heat map is generated by the gradient-weighted feature map and superimposed on the grayscale image to display the key lesion area. For structured EHR and text modality, attention weighting analysis is used to calculate feature importance and generate key variable contribution descriptions. For recommendation strategies, decision explanations are generated through counterfactual reasoning, specifically including: Let the current optimal treatment strategy be... Alternative treatment strategies are Calculate the risk probability predictions for both. and The risk difference is defined as ,like This indicates that the alternative reduces the risk of future aortic events in patients.