Search

CN-121983331-A - Rare patient treatment regimen evaluation and recommendation method

CN121983331ACN 121983331 ACN121983331 ACN 121983331ACN-121983331-A

Abstract

The invention discloses a rare patient treatment scheme evaluation and recommendation method which comprises the steps of collecting data including structured data, semi-structured data and unstructured data, uniformly indexing the integrated multi-modal data, setting uniform time grids to divide the multi-modal data, mapping each observation data to the latest time according to timestamp information, carrying out feature extraction on the multi-modal data, uniformly splicing all modal features, reducing dimensions through a linear projection layer to be uniform to obtain a multi-modal fusion sequence, carrying out risk prediction on output vectors corresponding to time steps in a time sequence state matrix of the full sequence by taking the multi-modal fusion sequence as input through a prediction function, carrying out forward propagation on the same input by taking the treatment scheme as a dynamic Monte Carlo method to obtain a plurality of prediction results, achieving the purposes of improving the integrity and the risk prediction precision of the state features of a patient, overcoming the problem that the traditional reinforcement learning is easy to be fitted and the training is unstable, and realizing the transparency and verifiability of the results.

Inventors

  • JIN BO
  • Yao Houxiong
  • ZHANG LIANG

Assignees

  • 大连理工大学

Dates

Publication Date
20260505
Application Date
20251222

Claims (7)

  1. 1. A method of rare patient treatment regimen evaluation and recommendation comprising: Step 1, multi-modal data acquisition and integration, wherein the acquired data comprises three types of structured data, semi-structured data and unstructured data, the acquired data is integrated, the integrated multi-modal data is uniformly indexed and respectively stored into five sub-tables of images, EHR, texts, genes and meta-information; Step 2, preprocessing and aligning data, setting uniform time grids to divide multi-mode data, obtaining observation data, and mapping each observation data to the latest time according to timestamp information; Step 3, extracting characteristics of the multi-mode data, and uniformly splicing the characteristics of all modes and reducing the dimension of the linear projection layer to be uniform to obtain a multi-mode fusion sequence; Step 4, based on a cross-modal converter model, taking a multi-modal fusion sequence as input, outputting a time sequence state matrix of a full sequence, and carrying out risk prediction on an output vector corresponding to a time step in the time sequence state matrix of the full sequence through a prediction function; Step 5, based on the reinforcement learning RL framework, outputting vectors, namely patient states, as state inputs, and obtaining a treatment scheme as an action; And step 6, adopting a Monte Carlo dropouout method to forward propagate the same input to obtain a plurality of groups of prediction results, wherein the average prediction value of the plurality of groups of prediction results is the final risk probability prediction.
  2. 2. The rare patient treatment regimen evaluation and recommendation method of claim 1, wherein step 1 multi-modality data acquisition and integration comprises: step 1-1, multi-mode data acquisition; the collected data includes three types: The structured data is electronic medical record EHR data, including blood pressure, heart rate, medical history, operation record and blood biochemical index; Semi-structural data, namely gene data comprising mutation point information of FBN1, TGFBR2 and SMAD 3; Unstructured data, including image data, including CT images, MRI images, cardiac ultrasound, text data, including doctor diagnostic reports, disease course records, discharge nodules; Step 1-2, multi-mode data integration; the DICOM format of the image data is uniformly converted into NIfTI format; the EHR data table is standardized in time stamp and unit, namely blood pressure unit mmHg and diameter unit mm, and is processed into a structured EHR data table; the text data is coded, converted into UTF-8 format and the recording time field is extracted; The mutation point location information of the FBN1, the TGFBR2 and the SMAD3 of the gene data is subjected to mutation annotation and pathogenicity grading to generate a structured table containing patient number, gene name, mutation type and pathogenicity grade fields, and then the structured table is associated with the main table of the EHR data table by taking the unique patient number UID as a main key, and the association method comprises the following steps: And respectively storing the integrated multi-mode data into five sub-tables of images, EHR, texts, genes and meta-information by taking the UID as a unified index.
  3. 3. The rare patient treatment regimen evaluation and recommendation method of claim 1, wherein step 2 data pre-processing and alignment comprises: Step 2-1, setting a unified time grid; dividing the whole follow-up period into Discrete time points, noted as Wherein Represent the first The number of time steps is one, Is a time step index; selecting a timestamp for each record of modality data that is closest to the current time step As observation data; for the observation data in each modality, it is mapped to the most recent time step according to the timestamp information The mapping method is that based on the minimum time difference principle, the difference value between the observation time and the center point of each time step is calculated for each record and is distributed to the time step closest to the observation time, if the observation value is missing in a certain time step, a double-strategy interpolation mechanism is adopted to carry out data complement, and the specific method is as follows: a forward filling method is used for blood pressure, heart rate and blood biochemical indexes in the structured continuous variable EHR data; The spline interpolation method is used for measuring the aorta internal diameter, wall thickness, flow velocity and compliance obtained by the aorta CT image, the nuclear magnetic resonance MRI image and the heart ultrasound; And performing abnormality detection, and outputting a timing characteristic matrix in a unified format by using a3 sigma principle and sliding window statistics, wherein the shape is [ patient number multiplied by time step multiplied by characteristic dimension ].
  4. 4. The rare patient treatment regimen evaluation and recommendation method of claim 1, wherein step 3 multi-modal characterization learning comprises: the image mode uses a vision converter ViT to divide the aortic CT image area into a plurality of image blocks, and outputs image characteristics after passing through a multi-layer attention mechanism, and the formula is as follows: Wherein, the To be in a time step The extracted image features are used to extract the image features, Is the characteristic dimension of the image mode; The text modality uses a medical language model BioBERT to extract semantic embedding, output text features, and the formula is: Wherein, the To be in a time step The extracted image features are used to extract the image features, Is a text modal feature dimension; The structural EHR mode processed by time standardization and unit unification adopts TabTransformer, and combines category embedding and attention layer output EHR characteristics, and the formula is as follows: Wherein, the To be in a time step The extracted EHR characteristics are used to determine, The EHR mode feature dimension; vectorization of gene modes is achieved by constructing a leachable mutation embedding matrix, and the formula is as follows: Wherein, the Representing the total number of all possible mutation sites, each row An embedded representation of the mutation site corresponding to the ith mutation site, at time steps for a sample Is given by: Wherein, the Indicating that the sample is in time steps The index of the ith mutation site that occurs, corresponds to a specific mutation event in the genome, Representing the number of mutations detected by the sample at time steps, performing a look-up table mapping by indexing each mutation site Mapping to corresponding embedded vectors Obtaining a mutation vector sequence, wherein the formula is as follows: carrying out average pooling mean pooling on the mutation vector sequence, and mapping the mutation set with variable length into a gene modal feature vector with fixed dimension, wherein the formula is as follows: Wherein, the To be in a time step The characteristics of the extracted genes, Is the characteristic dimension of the gene mode; After the mode characteristics are spliced, the dimension reduction of the linear projection layer is unified, and the formula is as follows: Wherein, the A linear projection layer is represented and, The vector concatenation operation is represented by a vector, Representing images, text, EHR, and gene modality over time D is the dimension of the comprehensive characterization; Finally, to The above operations are respectively carried out to obtain a multi-mode fusion sequence 。
  5. 5. The rare patient treatment regimen evaluation and recommendation method of claim 1, wherein step 4 of fusing modeling and temporal risk prediction comprises: based on a trans-modal converter model, the multi-modal fusion sequence is obtained As an input, the output is a time sequence state matrix of full sequence, and the formula is: Wherein, the Representing a patient state representation set obtained by time sequence modeling, each For corresponding time steps in matrix H By means of a linear classification layer to calculate a future time window The probability of occurrence risk of the internal aortic event is calculated by the following predictive function formula: Wherein, the In order to predict the probability of a probability, As a function of the Sigmoid, Training by adopting a cross entropy loss function as a model parameter, wherein the formula is as follows: Wherein, the Indicating whether an aortic dilation/dissection event has occurred within a future window, the output results include individual risk scores and a time-series risk profile.
  6. 6. The rare patient treatment regimen evaluation and recommendation method of claim 1, wherein step 5 treatment recommendation comprises: Based on the reinforcement learning RL framework, output vectors I.e. patient status as status input, treatment regimen as action a, reward function is defined as follows: Wherein, the In order to be at risk of non-intervention, In order to treat the cost of the treatment, In order to violate the penalty term of the clinical guideline, Is a weight factor; group relative strategy optimization GRPO algorithm based on each round of slave strategy network Sampling G candidate actions And calculating a group mean value and a standard deviation, wherein the formula is as follows: Wherein, the As a group mean value, Is the standard deviation of the two-dimensional image, For a numerical stability constant, a normalized dominance formula is defined as: The policy penalty function is: Wherein, the Representing the current policy Strategy with last round For the same action in the same state h Is used for the probability ratio of (a), For clipping coefficients.
  7. 7. The rare patient treatment regimen evaluation and recommendation method of claim 1, wherein step 6 uncertainty and interpretation comprises: Step 6-1, adopting a Monte Carlo Dropout method to carry out M times of forward propagation on the same input to obtain a plurality of groups of prediction results Average predicted value Confidence interval for final risk probability prediction Wherein And (3) with Respectively, are prediction results Outputting the interval as a predicted confidence, and marking the interval as a high uncertainty sample if the confidence interval span is greater than 0.1; step 6-2, for an image mode, adopting a gradient weighting type activation mapping Grad-CAM technology to visualize a concerned region of a model on an aortic image, generating a thermodynamic diagram through a gradient weighting characteristic diagram, and superposing the thermodynamic diagram to a gray image to display a key lesion region; for the structural EHR and the text mode, calculating the feature importance by using an attention weight analysis method, and generating a key variable contribution description; For the recommendation strategy, generating decision interpretation through anti-facts reasoning specifically comprises the following steps: the current optimal treatment scheme strategy is set as Alternative treatment strategies are Respectively calculating risk probability prediction corresponding to the two And (3) with Define the risk difference as If (if) This alternative is shown to reduce the risk of future aortic events in the patient.

Description

Rare patient treatment regimen evaluation and recommendation method Technical Field The invention relates to the technical field of artificial intelligence and medical intelligent diagnosis and treatment systems, in particular to a rare patient treatment scheme evaluation and recommendation method. Background In recent years, with the continuous development of medical informatization, genetic testing and image intelligent analysis technologies, the types of data available to clinicians in rare patient management are increasingly rich, and the data includes medical images (CT, MRI, echocardiography), structured electronic medical records (ElectronicHealthRecords, EHR), textual diagnostic records and genetic testing results. The data form a complete track of the disease progress of the patient in the time dimension, contain a large number of implicit rules related to risk evolution, treatment response and long-term prognosis, and are important carriers for promoting accurate medical treatment and intelligent decision. Existing studies have made progress in multi-modal clinical data analysis including medical image segmentation, aortic diameter measurement, genetic mutation association analysis, risk stratification, and prognostic assessment. However, these studies mostly stay at the level of specific tasks (e.g. lesion detection, risk scoring) and fail to achieve cross-modal, time-sequential, personalized system modeling. The aortic-related rare diseases represented by Marsdenia (MarfanSyndrome) have obvious individual variability and time dynamics in the evolution of the disease, wherein the common actions of aortic expansion speed, blood pressure fluctuation, genotype and medication compliance factors determine the operation time and treatment strategy of patients. Traditional manual evaluation methods mainly depend on doctor experience and discrete indexes, and are difficult to capture implicit dynamic relations of long-term dependence. With the rapid development of deep learning (DEEPLEARNING) technology, convolutional neural network (ConvolutionalNeuralNetwork, CNN), vision converter (VisionTransformer, viT), medical language model (BioBERT) and table Transformer (TabTransformer) model show excellent performance in medical data analysis, and at the same time, reinforcement learning (ReinforcementLearning, RL) method shows great potential in policy optimization and individualization recommendation. However, the existing methods still have serious limitations in rare disease treatment scheme recommendation, namely, on one hand, characteristic fusion is difficult due to time sequence misalignment and modal isomerism of multi-mode data, a model cannot model images, texts, structures and gene characteristics uniformly, and on the other hand, the existing reinforcement learning algorithm (such as near-end strategy optimization ProximalPolicyOptimization, PPO) depends on value network estimation, and a large number of interaction samples are needed to converge, so that the method is not suitable for rare disease data scenes with rare samples. Therefore, how to realize cross-modal fusion, time sequence risk prediction and controllable reinforcement learning strategy optimization under the condition of limited samples becomes a key technical problem to be solved urgently in the current intelligent medical field. ViT in the present invention comes out Dosovitskiy A. An image is worth 16x16 words: Transformers for image recognition at scale[J]. arXiv preprint arXiv:2010.11929, 2020. CNN comes out of 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. BioBERT out of 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. TabTransformer out of Huang X, Khetan A, Cvitkovic M, et al. Tabtransformer: Tabular data modeling using contextual embeddings[J]. arXiv preprint arXiv:2012.06678, 2020. ReinforcementLearning out Sutton R S, Barto A G. Reinforcement learning: An introduction[M]. Cambridge: MIT press, 1998,ProximalPolicyOptimization out Schulman J, Wolski F, Dhariwal P, et al. Proximal policy optimization algorithms[J]. arXiv preprint arXiv:1707.06347, 2017. GRPO out of Shao Z, Wang P, Zhu Q, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models[J]. arXiv preprint arXiv:2402.03300, 2024,. Monte Carlo Dropout comes out of Gal Y, Ghahramani Z. Dropout as a bayesian approximation: Representing model uncertainty in deep learning[C]//international conference on machine learning. PMLR, 2016: 1050-1059. The method of representing the gene modality using Embedding layers comes out 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. Disclosure of Invention T