Deep learning prediction system for cancer recurrence probability based on multi-modal temporal features

Abstract

The application discloses a cancer recurrence probability deep learning prediction system based on multi-modal time sequence characteristics, and relates to the technical field of medical data processing; the application collects clinical static data, dynamic treatment data and time sequence monitoring data of patients, constructs a multi-modal time sequence data set, carries out missing value filling and abnormal value elimination based on a sliding Z-score on the data, forms an effective data set, uses a random forest model to screen key clinical static features, constructs a radiotherapy cumulative toxicity model to quantize residual toxicity, combines a dose change rate tri-state feature to generate dynamic features, converts the features into a four-dimensional time sequence tensor through a dynamic sliding window, adopts random search to optimize LSTM hyperparameters, trains a high-precision prediction model to output a recurrence probability, and finally triggers differentiated clinical intervention based on risk grading, forms a "prediction-decision" closed-loop system, and significantly improves the initiative and accuracy of cancer recurrence diagnosis and treatment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical data processing, and in particular, relates to a cancer recurrence probability deep learning prediction system based on multi-modal time sequence features. BACKGROUND

[0002] The application belongs to the field of medical information technology and artificial intelligence in the medical industry, and specifically relates to a system for accurately predicting cancer recurrence probability using multi-modal time sequence data through deep learning algorithms. The system is directly applied to the prognosis evaluation link in the clinical treatment process of cancer, providing scientific basis for developing personalized treatment plans to improve the survival rate and quality of life of cancer patients.

[0003] Currently, cancer recurrence prediction mainly relies on a single type of data and traditional statistical analysis methods. In clinical practice, it is common to assess the recurrence risk based on limited information such as the patient's clinical stage and pathological features. For example, only through indicators such as tumor size, grade, and lymph node metastasis to determine the likelihood of cancer recurrence. However, this approach ignores the dynamic changes of multiple factors during cancer development and the complex interactions between factors. From the data source, most existing methods only consider the patient's static clinical information, such as the patient's basic demographic information, initial diagnosis of the tumor, etc., and fail to fully utilize the multi-modal time sequence data such as drug treatment information, radiotherapy information, follow-up information, and tumor event information during the treatment process. These data contain dynamic changes during cancer treatment and are of great significance for accurately predicting cancer recurrence probability. In terms of analysis methods, traditional statistical models such as logistic regression and decision trees are difficult to handle high-dimensional, complex multi-modal data and non-linear relationships. These models usually assume a linear relationship between data, which cannot capture complex patterns and potential information in the data, resulting in limited accuracy and reliability of the prediction results.

[0004] The existing technology has the problems of insufficient utilization of dynamic treatment data, lack of multi-modal time sequence feature fusion, interference of abnormal values and redundant features, and lack of deep learning application. SUMMARY

[0005] (I) Technical problems solved

[0006] To solve the problems in the related art, the present application provides a deep learning prediction method for cancer recurrence probability based on multi-modal time sequence features to overcome the above technical problems existing in the prior art.

[0007] (II) Technical solutions

[0008] To solve the above technical problems, the present application is realized by the following technical solutions:

[0009] S1, collect clinical static data, dynamic treatment data and time series monitoring data of the patient to obtain multi-modal time series data of the patient;

[0010] S2, perform missing data processing and outlier processing on the multi-modal time series data of the patient to obtain effective multi-modal time series data of the patient;

[0011] S3, screening selected clinical static features from the effective multi-modal time series data of the patient based on a random forest model to obtain selected static feature data of the patient;

[0012] S4, calculating cumulative toxicity index and dose rate of change features of drug treatment in the effective multi-modal time series data of the patient to obtain dynamic feature data of the patient;

[0013] Performing time series formatting operation on the selected static feature data of the patient and the dynamic feature data of the patient to obtain a four-dimensional time series feature tensor;

[0014] S5, constructing an LSTM model; training the LSTM model based on the labeled historical four-dimensional time series feature tensor data combined with a random search algorithm to obtain an LSTM cancer recurrence prediction model;

[0015] Inputting the four-dimensional time series feature tensor into the LSTM cancer recurrence prediction model to obtain a predicted cancer recurrence probability of the patient;

[0016] S6, setting a recurrence probability treatment rule; based on the recurrence probability treatment rule and the predicted recurrence probability of the patient, treating the cancer patient;

[0017] The application realizes high-precision prediction of cancer recurrence probability by fusing multi-modal time series data, combining intelligent screening of key features and innovative radiotherapy cumulative toxicity model, constructing an LSTM deep learning model with optimized four-dimensional time series tensor input, triggering differentiated clinical intervention based on risk classification, forming a "prediction-decision" closed loop, significantly improving the initiative and accuracy of cancer recurrence diagnosis and treatment, and promoting the cancer recurrence diagnosis and treatment from passive response to active prevention and control.

[0018] Preferably, the S1 comprises the following steps:

[0019] S11, collecting the age, gender, tumor stage and pathological type of the patient from the hospital information system to obtain the clinical static data of the patient;

[0020] S12, extracting the drug treatment sequence and radiotherapy parameters of the patient from the electronic medical record to obtain the dynamic treatment data of the patient; the drug treatment sequence includes drug name, single dose and administration time point; the radiotherapy parameters include each radiotherapy dose, irradiation site and cumulative dose change curve;

[0021] S13, obtaining the detection value and detection timestamp of the tumor marker of the patient and the key indicator change of the imaging examination result from the follow-up system to obtain the time series monitoring data of the patient;

[0022] The patient clinical static data, the patient dynamic treatment data and the patient time series monitoring data jointly constitute the patient multi-modal time series data;

[0023] The present application integrates the hospital information system, the electronic medical record and the follow-up system, and constructs the multi-modal time series data set containing dynamic and static characteristics and accurate timestamps, so as to provide a comprehensive and coherent patient treatment process for subsequent analysis, effectively avoid the data fragmentation problem, and support accurate recurrence prediction.

[0024] Preferably, the S2 comprises the following steps:

[0025] S21, for continuous variables in the patient multi-modal time series data, time series linear interpolation is used for filling, and the filled patient multi-modal time series data is obtained;

[0026] S22, for the category variable in the filled patient multi-modal time series data, the mode filling is used, and the filled patient multi-modal time series data is obtained;

[0027] S23, setting an invalid data threshold; calculating the sliding Zscore of the tumor marker detection value in the filled patient multi-modal time series data to obtain a sliding Zscore set; regarding the value of the sliding Zscore in the sliding Zscore set as the invalid sliding Zscore when the sliding Zscore is greater than the invalid data threshold, and obtaining an invalid sliding Zscore set;

[0028] The tumor marker detection value in the filled patient multi-modal time series data corresponding to the invalid sliding Zscore set is removed, and the effective patient multi-modal time series data is obtained;

[0029] The present application solves the data missing problem by time series linear interpolation of continuous variables and mode filling of category variables, and further accurately identifies and removes abnormal tumor marker detection values based on the sliding Z-score dynamic threshold, so as to significantly improve the integrity and reliability of the multi-modal data and lay a high-quality data foundation for subsequent modeling.

[0030] Preferably, the S3 comprises the following steps:

[0031] S31, setting tumor clinical characteristics to obtain a tumor clinical characteristic set; the tumor clinical characteristic set comprises all clinical characteristics of the tumor;

[0032] Based on the tumor clinical characteristic set, historical tumor clinical characteristic data is collected;

[0033] S32, initialize the random forest model, set the feature importance evaluation standard of the random forest model as the Gini impurity reduction;

[0034] The random forest model is trained using historical tumor clinical feature data combined with a cross-validation algorithm to obtain a trained random forest classifier.

[0035] S33, calculate the purity improvement sum of each tumor clinical feature in the effective patient multi-modal time series data at the decision tree node split by using the trained random forest classifier, and obtain an improvement sum set;

[0036] S34, sort all features in the historical tumor clinical feature data in descending order according to the size of the improvement sum in the improvement sum set, and perform normalization processing, record the importance score of the historical tumor clinical feature, and obtain a tumor clinical feature importance ranking;

[0037] Set the correlation threshold and the number of key features; select the key features with the highest importance score according to the tumor clinical feature importance ranking and the number of key features, and obtain a key tumor clinical feature set;

[0038] Verify the clinical significance of the key features in the key tumor clinical feature set, exclude redundant features with a correlation greater than the correlation threshold, and obtain a selected static feature set;

[0039] Based on the selected static feature set, the selected static features of the patient are extracted from the effective patient multi-modal time series data to obtain patient selected static feature data.

[0040] The present application quantifies the importance of each clinical feature and sorts them by using random forest model combined with cross-validation training; selects key features, verifies clinical significance and excludes redundant variables, and finally extracts high-value static features; significantly improves the representativeness of the features and reduces the complexity of the model, providing a core index set for accurate recurrence prediction.

[0041] Preferably, the S4 comprises the following steps:

[0042] S41, collect the medication time data of the patient from the effective patient multi-modal time series data to obtain a standardized drug administration time sequence;

[0043] Set the drug attenuation parameter, construct a radiotherapy cumulative toxicity model; calculate the residual toxicity of each drug administration at the current time point by using the radiotherapy cumulative toxicity model, and obtain the real-time toxicity weight of each drug administration;

[0044] S42, based on the real-time toxicity weight of each drug administration, accumulate the toxicity weight of all effective drug administrations at the same time point, correct the toxicity superposition of the combined drug administration scene, generate a time-toxicity curve, and obtain a dynamic toxicity cumulative feature curve;

[0045] Based on the dynamic toxicity cumulative characteristic curve, the actual liver and kidney function detection values of the patient are compared, the drug attenuation parameters are adjusted, the curve is made to conform to the clinical toxicity performance, and the toxicity grade is divided, and the calibrated clinical usable toxicity curve is obtained;

[0046] S43, extract the patient radiotherapy dose time sequence record from the effective patient multi-modal time sequence data, arrange the dose data in time sequence, form a time sequence; check the data continuity, and obtain a standardized dose time sequence;

[0047] The difference between each dose in the standardized dose time sequence is calculated by using the change amount of adjacent periods; the dose increase is set as a positive value, and the dose decrease is set as a negative value, to obtain a dose change amount sequence;

[0048] S44, set a change amount threshold; based on the dose change amount sequence and the change amount threshold, convert the change amount into a treatment intensity adjustment mark, and generate a three-state feature value, to obtain a dose change rate feature vector;

[0049] S45, align all feature data according to time points by using the patient static feature data, the calibrated clinical usable toxicity curve and the dose change rate feature vector, to obtain an original time sequence feature matrix;

[0050] The original feature matrix is subjected to window slicing and filling operation by using a dynamic sliding window method, to obtain a four-dimensional time sequence feature tensor;

[0051] The present application constructs a radiotherapy cumulative toxicity model by innovation, generates a clinical calibration curve by toxicity superposition correction; extracts a dose change three-state feature to represent treatment intensity adjustment; finally, a dynamic sliding window is used to convert multi-source features into a four-dimensional time sequence tensor, which not only accurately quantifies the treatment dynamic effect, but also solves the variable length time sequence data input problem, and provides high-value structured features for the LSTM model.

[0052] Preferably, the S5 comprises the following steps:

[0053] S51, construct an LSTM model, and set the hyperparameters of the LSTM model training;

[0054] S52, collect four-dimensional feature tensor data of historical cancer patients and label data of whether historical cancer patients relapse, to obtain labeled historical four-dimensional time sequence feature tensor data;

[0055] S53, train the LSTM model by using the labeled historical four-dimensional time sequence feature tensor data, find the optimal hyperparameters of the LSTM model by using a random search algorithm during the training process, obtain an optimal solution; use the optimal solution as the hyperparameters of the LSTM model, and obtain an LSTM cancer recurrence prediction model;

[0056] S54, input the four-dimensional time sequence feature tensor to the LSTM cancer recurrence prediction model to obtain a predicted recurrence probability of the patient;

[0057] The application trains a high-precision cancer recurrence prediction model by constructing an LSTM model, using labeled four-dimensional time sequence feature tensor data, and efficiently optimizing hyperparameters by combining a random search algorithm.

[0058] Preferably, the step of obtaining the optimal solution in S53 includes the following steps:

[0059] S531, set the parameter search space as L , set the maximum number of random searches as M , set the number of cross-validation folds as E , and initialize the optimal hyperparameters;

[0060] S532, under the optimal hyperparameters, train and evaluate the LSTM model using labeled historical four-dimensional time sequence feature tensor data to obtain the best prediction accuracy of the LSTM model d ;

[0061] S533, define a candidate hyperparameter set; randomly select hyperparameters from the candidate hyperparameter set as f , divide the training data in the labeled historical four-dimensional time sequence feature tensor data into K folds, for each fold k , ; train the LSTM model from scratch using all training data in the labeled historical four-dimensional time sequence feature tensor data except the k th fold, evaluate the model using the training data in the k th fold of the labeled historical four-dimensional time sequence feature tensor data to obtain the prediction accuracy, calculate the average prediction accuracy of all folds to obtain the average prediction accuracy g ;

[0062] If the average prediction accuracy g is greater than the best prediction accuracy d , update the best prediction accuracy to g , and update the optimal hyperparameters f ; otherwise, keep the original best prediction accuracy and optimal hyperparameters;

[0063] S534, repeat S533, and when the maximum number of random searches M is reached, stop iteration to obtain the optimal solution;

[0064] The application efficiently explores the LSTM hyperparameter combination by combining the random search algorithm with K-fold cross-validation: each time a hyperparameter is randomly selected, K-1-fold data is trained, one-fold validation is performed, and the average accuracy g is calculated; when g exceeds the historical best d, the optimal hyperparameter is updated; this mechanism locks the best configuration with limited computing cost, significantly improves the model generalization ability and prediction stability, and avoids the risk of overfitting.

[0065] Preferably, the S6 comprises the following steps:

[0066] S61, setting a recurrence probability treatment rule;

[0067] S62, based on the set recurrence probability treatment rule, predicting the recurrence probability of the patient, and treating the cancer patient;

[0068] The application converts the prediction probability into differentiated clinical intervention measures by setting the recurrence probability treatment rule, realizes seamless conversion from the prediction result to the treatment decision, improves the diagnosis and treatment accuracy, and avoids excessive medical treatment or delayed treatment.

[0069] The cancer recurrence probability deep learning prediction system based on multi-modal time sequence characteristics is used to realize the cancer recurrence probability deep learning prediction method based on multi-modal time sequence characteristics, and comprises a data acquisition module, a data processing module, a static feature screening module, a dynamic feature calculation and formatting module, a model construction and prediction module, and a treatment decision module.

[0070] The data acquisition module collects the original data of the patient from a multi-source medical system, including clinical static data, dynamic treatment data and time sequence monitoring data; these data jointly constitute the multi-modal time sequence data of the patient, providing basic input for subsequent analysis;

[0071] The data processing module is used for pre-processing the collected multi-modal time sequence data to solve the data quality problem; the missing values of continuous variables are filled by using time series linear interpolation, the missing values of categorical variables are filled by using mode, and then the sliding Z-score of tumor marker detection values is calculated to identify and eliminate abnormal values; finally, the effective patient multi-modal time sequence data is output, ensuring the integrity and reliability of the data;

[0072] The feature screening module screens the most valuable static features from the effective data by using a random forest model; the random forest model is initialized, the model is trained using historical tumor clinical feature data, and the feature importance score is calculated based on the Gini impurity reduction; then the features are sorted and normalized, the key features with the highest importance score are selected, the clinical significance is verified and redundant features are excluded, and finally the patient selected static feature data is output, the feature set is optimized to improve the prediction accuracy;

[0073] The feature calculation and formatting module calculates dynamic treatment features and converts all features into a format that can be input into the model; a radiotherapy cumulative toxicity model is constructed to calculate a cumulative toxicity index, generate a time-toxicity curve and calibrate the toxicity level according to clinical data; the dose change rate feature is extracted from the radiotherapy dose time record, the change amount sequence is calculated by the adjacent dose difference, and the three-state feature value is converted; the static features, toxicity curves and dose change features are aligned according to the time points, and a four-dimensional time sequence feature tensor is generated using a dynamic sliding window method, which is convenient for deep learning model processing;

[0074] The model construction and prediction module is used for constructing and optimizing an LSTM deep learning model to predict the probability of cancer recurrence; by collecting labeled historical four-dimensional time sequence feature tensor data, an LSTM model is constructed and the hyperparameters are optimized using a random search algorithm; after training, an LSTM cancer recurrence prediction model is obtained, the four-dimensional time sequence feature tensor of the current patient is input, and the probability of cancer recurrence of the patient is output, realizing high-precision time sequence prediction;

[0075] The treatment decision module formulates a clinical treatment strategy based on the predicted recurrence probability; sets the recurrence probability treatment rule; and carries out personalized treatment intervention for cancer patients according to the predicted probability, so as to ensure that the prediction result is converted into actual clinical action.

[0076] (Three) beneficial effects

[0077] The present application has the following beneficial effects:

[0078] The present application realizes high-precision prediction of the probability of cancer recurrence by fusing multi-modal time sequence data, combining intelligent screening of key features and an innovative radiotherapy cumulative toxicity model, constructing an optimized LSTM deep learning model with four-dimensional time sequence tensor input, triggering differentiated clinical intervention based on risk classification, forming a "prediction-decision" closed loop, and significantly improving the initiative and accuracy of cancer recurrence diagnosis and treatment, and promoting the cancer recurrence diagnosis and treatment from passive response to active prevention.

[0079] The present application constructs a multi-modal data set with a unified time axis by structuring the collection of clinical static data, dynamic treatment data and time sequence monitoring data; adopts sliding Z-score to eliminate abnormal values of tumor markers, and combines random forest feature screening to compress redundant static features, which significantly improves the data quality and feature representativeness.

[0080] The present application introduces a drug attenuation parameter to create a radiotherapy cumulative toxicity index model, quantifies the residual toxicity at different time points, generates a time-toxicity curve by toxicity superposition correction, and calibrates the model based on liver and kidney function test values; the three-state feature of the dose change rate is extracted, the treatment intensity adjustment is converted into a calculable time sequence marker, and the key treatment decision signal is provided for the model.

[0081] The application solves the problem of variable-length time series data input by converting the original feature matrix into a four-dimensional tensor through a dynamic sliding window; the window overlap design retains the continuity of the treatment stage, and the front-end zero padding strategy is compatible with short-period patient data; combined with random search hyperparameter optimization, the optimal configuration of LSTM is efficiently locked with the cross-validation average accuracy as the index, and the reliability of recurrence probability prediction is improved.

[0082] The application sets a hierarchical treatment rule based on the predicted probability, avoids over-treatment, and can screen out recurrence clues earlier and treat patients with high recurrence probability in time.

[0083] Of course, implementing any product of the application does not necessarily need to achieve all the advantages described above at the same time. BRIEF DESCRIPTION OF DRAWINGS

[0084] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed for the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0085] Figure 1 The flowchart of the cancer recurrence probability deep learning prediction method based on multi-modal time series features of the application;

[0086] Figure 2 The flowchart of processing patient data in the cancer recurrence probability deep learning prediction method based on multi-modal time series features of the application;

[0087] Figure 3 The flowchart of collecting historical cancer recurrence patient time series feature data in the cancer recurrence probability deep learning prediction method based on multi-modal time series features of the application;

[0088] Figure 4 The module diagram of the cancer recurrence probability deep learning prediction system based on multi-modal time series features of the application. DETAILED DESCRIPTION

[0089] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only some embodiments of the application, not all embodiments. Based on the embodiments of the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0090] In the description of the present application, it should be understood that the terms "opening", "upper", "lower", "top", "middle", "inner" and the like indicate the orientation or positional relationship, only for the convenience of describing the application and simplifying the description, and do not indicate or imply that the components or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore cannot be understood as a limitation of the application.

[0091] Embodiment one:

[0092] Please refer to Figure 1 、 Figure 2 、 Figure 3 The application discloses a cancer recurrence probability deep learning prediction method based on multi-modal time sequence characteristics, comprising the following steps:

[0093] S1, collecting the clinical static data, dynamic treatment data and time sequence monitoring data of the patient to obtain the multi-modal time sequence data of the patient;

[0094] The S1 comprises the following steps:

[0095] S11, collecting the age, gender, tumor stage and pathological type of the patient from the hospital information system (HIS) to obtain the clinical static data of the patient;

[0096] S12, extracting the drug treatment sequence and radiotherapy parameters of the patient from the electronic medical record (EMR) to obtain the dynamic treatment data of the patient; the drug treatment sequence includes drug name, single dose and administration time point (accurate to day); the radiotherapy parameters include each radiotherapy dose, irradiation site and cumulative dose change curve;

[0097] S13, obtaining the multiple detection values and detection time stamps of tumor markers (such as CEA, CA125) and the key index changes of imaging examination results (CT / MRI) of the patient from the follow-up system to obtain the time sequence monitoring data of the patient;

[0098] The patient clinical static data, patient dynamic treatment data and patient time sequence monitoring data jointly constitute the multi-modal time sequence data of the patient;

[0099] S2, performing missing value processing and outlier processing on the multi-modal time sequence data of the patient to obtain effective multi-modal time sequence data of the patient;

[0100] The S2 comprises the following steps:

[0101] S21, for continuous variables (such as drug concentration) in the multi-modal time sequence data of the patient, linear interpolation filling is adopted to obtain the filled multi-modal time sequence data of the patient;

[0102] S22, for the filled patient multi-modal time series data, the mode filling is used for the category type variable (such as radiotherapy type) in the filled patient multi-modal time series data, and the filled patient multi-modal time series data is obtained;

[0103] S23, setting an invalid data threshold; calculating the sliding Zscore of the tumor marker detection value in the filled patient multi-modal time series data, and obtaining the sliding Zscore set; the value of the sliding Zscore in the sliding Zscore set is greater than the invalid data threshold, and the invalid sliding Zscore is obtained;

[0104] The tumor marker detection value in the filled patient multi-modal time series data corresponding to the invalid sliding Zscore set is removed, and the effective patient multi-modal time series data is obtained;

[0105] S3, based on the random forest model, the selected clinical static features are selected from the effective patient multi-modal time series data, and the patient selected static feature data is obtained;

[0106] The S3 includes the following steps:

[0107] S31, setting a tumor clinical feature, obtaining a tumor clinical feature set; the tumor clinical feature set includes all clinical features of the tumor; such as TNM stage, Ki67 index;

[0108] Based on the tumor clinical feature set, the historical tumor clinical feature data is collected;

[0109] S32, initializing a random forest model containing 100 decision trees, and setting the feature importance evaluation standard of the random forest model as the Gini impurity reduction;

[0110] The historical tumor clinical feature data is used to train the random forest model combined with the cross-validation algorithm, and the trained random forest classifier is obtained;

[0111] S33, the purity improvement sum of each tumor clinical feature in the effective patient multi-modal time series data at the decision tree node splitting is calculated by the trained random forest classifier, and the improvement sum set is obtained;

[0112] S34, according to the size of the improvement sum in the improvement sum set, all features in the historical tumor clinical feature data are sorted in descending order, and normalized, and the importance score of the historical tumor clinical feature is recorded, and the tumor clinical feature importance ranking is obtained;

[0113] The correlation threshold and the number of key features are set to 10; according to the tumor clinical feature importance ranking and the number of key features, the top 10 features with the highest importance score are selected, and the key tumor clinical feature set is obtained;

[0114] Verify the clinical significance of key features in the key tumor clinical feature set (such as the Ki-67 index reflecting cell proliferation activity), exclude redundant features with correlations greater than the correlation threshold, and obtain a selected static feature set (including core indicators such as EGFR mutation status);

[0115] Based on the selected static feature set, the selected static features of the patient are extracted from the effective patient multimodal time series data to obtain the patient selected static feature data;

[0116] S4. Calculate the cumulative toxicity index and dose change rate characteristics of drug treatment in the multimodal time series data of valid patients to obtain patient dynamic characteristic data;

[0117] Perform time series formatting operations on the patient's selected static feature data and the patient's dynamic feature data to obtain a four-dimensional time series feature tensor;

[0118] The S4 comprises the following steps:

[0119] S41. Collecting the patient's medication time data from the valid patient multimodal time series data to obtain a standardized medication time series;

[0120] Setting drug decay parameters Lambda , construct a radiotherapy cumulative toxicity model; the radiotherapy cumulative toxicity model formula is as follows,

[0121] ;

[0122] in, Tox cum ( t ) indicates the current time point t The residual toxicity n The total number of radiotherapy treatments, D i Indicates the i The dose of radiotherapy, t i Indicates the i The timing of the dose of radiotherapy e represents the base of natural logarithms;

[0123] The radiotherapy cumulative toxicity model is used to calculate the residual toxicity of each administration at the current time point, and the real-time toxicity weight of each administration is obtained. If the drug is recently administered, the toxicity remains above 90%; if it was administered 30 days ago, the toxicity decreases to about 35%; if it was administered 60 days ago, the toxicity decreases to about 12%;

[0124] S42. Based on the real-time toxicity weight of each drug administration, the toxicity weights of all effective drug administrations at the same time point are accumulated, and toxicity superposition correction is performed on the combined drug administration scenario to generate a time-toxicity curve to obtain a dynamic toxicity accumulation characteristic curve;

[0125] Based on the dynamic toxicity cumulative characteristic curve, the actual liver and kidney function detection value of the patient is compared, the drug attenuation parameter is adjusted, the curve is made to conform to the clinical toxicity performance, and the toxicity grade is divided, and the calibrated clinical usable toxicity curve is obtained; such as mild 0%~20%, moderate 20%~50%, severe >50%;

[0126] Taking a breast cancer patient as an example, after 8 cycles of chemotherapy, the cumulative dose of paclitaxel of the breast cancer patient is 960mg / m², and the calculated cumulative toxicity value is 58.3% (severe), and the actual performance is grade 3 neutropenia (verified consistency);

[0127] S43, extract the patient radiotherapy dose time record (such as single dose value recorded by day / week) from the effective patient multi-modal time series data, arrange the dose data in chronological order to form a time series; check the data continuity (ensure no date discontinuity) to obtain a standardized dose time series;

[0128] The difference between each dose in the standardized dose time series is calculated by using the change amount of adjacent periods; set the dose increase as positive and the dose decrease as negative to obtain a dose change amount sequence;

[0129] S44, set a change amount threshold; based on the dose change amount sequence and the change amount threshold, convert the change amount into a treatment intensity adjustment mark (such as an increase threshold >1Gy for intensive treatment, i.e. a dose reduction adjustment), and generate a three-state feature value (record the increase threshold >1Gy as +1, indicating significant dose increase; record the change amount between +1 and +1 as 0, indicating stable dose; record the decrease threshold <-1Gy as -1, indicating significant dose reduction), to obtain a dose change rate feature vector;

[0130] S45, align all feature data according to the time point by using the patient static feature data, the calibrated clinical usable toxicity curve and the dose change rate feature vector, to obtain an original time series feature matrix;

[0131] The original feature matrix is subjected to window slicing and filling operation by using a dynamic sliding window method to obtain a four-dimensional time series feature tensor; the specific steps are as follows:

[0132] Construct an original time series feature matrix (rows represent time points and columns represent features); set the window length to cover 10 consecutive time points (about 3-6 months of treatment), set the sliding step to move 2 time points each time (new and old windows overlap by 50%), and set the edge processing to zero fill when the front end is insufficient and keep the new data at the end;

[0133] Based on the time sequence feature matrix, complete window slicing is performed, for example, if a patient has 24 months of data, 7 windows can be cut out; zero padding is performed at the front end, for example, if a patient has only 5 months of data, the window needs to be padded with 5 zero values; end retention rules are executed, without forced padding, and actual data (such as the last window) is retained, to obtain a window sequence set;

[0134] For all window sequences of the window sequence set of the patients, stacking is performed according to dimensions, to obtain a stacked feature time sequence tensor; for example, dimension 1 is the number of patient samples (450 cases), dimension 2 is the number of time windows (for example, 7 per patient), dimension 3 is the time step in the window (fixed at 10 steps), and dimension 4 is the feature dimension (18 dimensions);

[0135] Zero padding is performed on the missing values in the stacked feature time sequence tensor, to obtain a four-dimensional time sequence feature tensor;

[0136] S5, constructing an LSTM model; based on the labeled historical four-dimensional time sequence feature tensor data and a random search algorithm, the LSTM model is trained, to obtain an LSTM cancer recurrence prediction model;

[0137] The four-dimensional time sequence feature tensor is input into the LSTM cancer recurrence prediction model, to obtain a predicted cancer recurrence probability of the patient;

[0138] The S5 includes the following steps:

[0139] S51, constructing an LSTM model, and setting hyperparameters for training of the LSTM model;

[0140] S52, collecting four-dimensional feature tensor data of historical cancer patients and label data of whether historical cancer patients have recurrence, to obtain labeled historical four-dimensional time sequence feature tensor data;

[0141] S53, training the LSTM model using the labeled historical four-dimensional time sequence feature tensor data, and using a random search algorithm to find optimal hyperparameters of the LSTM model during the training process, to obtain an optimal solution; the optimal solution is used as the hyperparameters of the LSTM model, to obtain an LSTM cancer recurrence prediction model;

[0142] The S53 includes the following steps:

[0143] S531, setting a parameter search space as L ( L a candidate set of learning rates, that is, a set of predefined hyperparameters), setting a maximum number of random searches as M , setting a cross-validation fold number as E , and initializing optimal hyperparameters;

[0144] S532. Under the optimal hyperparameters, use the labeled historical four-dimensional time series feature tensor data to train and evaluate the LSTM model to obtain the best prediction accuracy of the LSTM model;

[0145] S533, define the candidate hyperparameter set (the candidate learning rate set is all candidates in L); randomly select hyperparameters from the candidate hyperparameter set as f , the training data in the labeled historical four-dimensional time series feature tensor data is divided into K Fold, for each fold k , ; Use the training data in all the labeled historical four-dimensional time series feature tensor data except the w-th fold to train the LSTM model from scratch, and use the k The training data evaluation model in the folded labeled historical four-dimensional time series feature tensor data is used to obtain the prediction accuracy. The mean of the prediction accuracy of all folds is calculated to obtain the average prediction accuracy. g , the calculation formula is as follows,

[0146] ;

[0147] in, g represents the average prediction accuracy, k Indicates the k A discount, K represents the number of cross-validation folds, T k Indicates the k The prediction accuracy of each fold;

[0148] If the average prediction accuracy g >Best prediction accuracy d , then the updated best prediction accuracy is g , update the optimal hyperparameters f Otherwise, the original best prediction accuracy and optimal hyperparameters are maintained;

[0149] S534, repeat S533, and when the maximum number of random searches M is reached, stop the iteration and obtain the optimal solution;

[0150] S54, inputting the four-dimensional time series feature tensor into the LSTM cancer recurrence prediction model to obtain the predicted patient recurrence probability;

[0151] S6. setting a recurrence probability treatment rule; treating the cancer patient based on the set recurrence probability treatment rule and the predicted recurrence probability of the patient;

[0152] The S6 comprises the following steps:

[0153] S61, set recurrence probability treatment rules; such as low risk (<0.3), recommend routine 3-month follow-up; medium risk (0.3 and 0.7), additional circulating tumor DNA detection; high risk (>0.7), start second-line treatment regimen;

[0154] S62, based on the recurrence probability level, predict the recurrence probability of the patient, and treat the cancer patient.

[0155] Example two:

[0156] Please refer to Figure 4 , a cancer recurrence probability deep learning prediction system based on multi-modal time series features, for implementing the above-mentioned cancer recurrence probability deep learning prediction method based on multi-modal time series features, including a data acquisition module, a data processing module, a static feature screening module, a dynamic feature calculation and formatting module, a model construction and prediction module, and a treatment decision module;

[0157] The data acquisition module collects the original data of the patient from the multi-source medical system, including clinical static data, dynamic treatment data, and time series monitoring data; these data together constitute the multi-modal time series data of the patient, providing basic input for subsequent analysis;

[0158] The data processing module is used for pre-processing the collected multi-modal time series data to solve the data quality problem; the missing values of continuous variables are filled by using time series linear interpolation, the missing values of categorical variables are filled by using mode, and then the sliding Z-score of tumor marker detection values is calculated to identify and eliminate outliers; finally, the valid patient multi-modal time series data is output to ensure the integrity and reliability of the data;

[0159] The feature screening module uses a random forest model to screen the most valuable static features from the valid data; by initializing the random forest model, the model is trained using historical tumor clinical feature data, and the feature importance score is calculated based on the Gini impurity reduction; then the features are sorted and normalized, the key features with the highest importance score are selected, their clinical significance is verified and redundant features are excluded, and finally the patient selected static feature data is output, the feature set is optimized to improve the prediction accuracy;

[0160] The feature calculation and formatting module calculates dynamic treatment features and converts all features into a format that can be input into the model; a radiotherapy cumulative toxicity model is constructed to calculate the cumulative toxicity index, a time-toxicity curve is generated and the toxicity level is calibrated according to the clinical data; the dose rate feature is extracted from the radiotherapy dose time series record, the change amount sequence is calculated by the adjacent dose difference, and is converted into a three-state feature value; the static features, toxicity curve and dose change features are aligned by time point, and a four-dimensional time series feature tensor is generated using a dynamic sliding window method to facilitate deep learning model processing;

[0161] The model construction and prediction module is used to construct and optimize an LSTM deep learning model to predict the probability of cancer recurrence; by collecting labeled historical four-dimensional time series feature tensor data, an LSTM model is constructed and the hyperparameters are optimized using a random search algorithm; after training, an LSTM cancer recurrence prediction model is obtained, the four-dimensional time series feature tensor of the current patient is input, and the probability of predicting the recurrence of the patient's cancer is output, realizing high-precision time series prediction;

[0162] The treatment decision module formulates a clinical treatment strategy based on the predicted probability of recurrence; sets the recurrence probability treatment rules; and performs personalized treatment intervention on cancer patients according to the predicted probability, ensuring that the prediction results are converted into actual clinical actions.

[0163] In the description of the present specification, the description of the terms "one embodiment", "example", "specific example" and the like means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are contained in at least one embodiment or example of the invention. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0164] The preferred embodiments of the above disclosed invention are only used to help explain the invention. The preferred embodiments do not describe all the details and limit the invention to only the specific embodiments described. Obviously, many modifications and changes can be made according to the content of the present specification. The present specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the invention, so that those skilled in the art can well understand and utilize the invention.

Claims

1. A deep learning prediction method for cancer recurrence probability based on multi-modal temporal features, characterized in that, The method comprises the following steps: S1, collecting clinical static data, dynamic treatment data and time series monitoring data of a patient to obtain multi-modal time series data of the patient; S2, performing missing value processing and outlier processing on the multi-modal time series data of the patient to obtain effective multi-modal time series data of the patient; S3, screening selected clinical static features from the effective multi-modal time series data of the patient based on a random forest model to obtain selected static feature data of the patient; S4, calculating a cumulative toxicity index and a dose change rate feature of drug treatment in the effective multi-modal time series data of the patient to obtain dynamic feature data of the patient; performing time series formatting operation on the selected static feature data of the patient and the dynamic feature data of the patient to obtain a four-dimensional time series feature tensor; The S4 comprises the following steps: S41, collecting medication time data of the patient from the effective multi-modal time series data of the patient to obtain a standardized drug administration time sequence; setting a drug attenuation parameter, constructing a radiotherapy cumulative toxicity model, calculating the residual toxicity of each administration at the current time point through the radiotherapy cumulative toxicity model, and obtaining real-time toxicity weights of each administration; S42, based on the real-time toxicity weights of each administration, accumulating the toxicity weights of all effective administrations at the same time point, correcting the toxicity superposition of the combined drug use scene, generating a time-toxicity curve, and obtaining a dynamic toxicity accumulation feature curve; based on the dynamic toxicity accumulation feature curve, comparing the actual liver and kidney function detection values of the patient, adjusting the drug attenuation parameter, making the curve consistent with the clinical toxicity performance, and dividing the toxicity grade to obtain a calibrated clinically available toxicity curve; S43, extracting the radiotherapy dose time series record of the patient from the effective multi-modal time series data of the patient, arranging the dose data in the order of treatment time to form a time sequence, and verifying the data continuity to obtain a standardized dose time sequence; calculating the difference between each dose in the standardized dose time sequence by using the change amount of adjacent time periods; setting the dose increase as a positive value and the dose decrease as a negative value to obtain a dose change amount sequence; S44, setting a change amount threshold; based on the dose change amount sequence and the change amount threshold, converting the change amount into a treatment intensity adjustment flag, and generating a three-state feature value to obtain a dose change rate feature vector; S45, aligning all feature data according to the time point by using the selected static feature data of the patient, the calibrated clinically available toxicity curve and the dose change rate feature vector to obtain an original time series feature matrix; performing window slicing and filling operation on the original feature matrix by using a dynamic sliding window method to obtain a four-dimensional time series feature tensor; S5, constructing an LSTM model; training the LSTM model based on the labeled historical four-dimensional time series feature tensor data combined with a random search algorithm to obtain an LSTM cancer recurrence prediction model; inputting the four-dimensional time series feature tensor into the LSTM cancer recurrence prediction model to obtain a predicted cancer recurrence probability of the patient; S6, setting a recurrence probability treatment rule; treating the cancer patient based on the recurrence probability treatment rule and the predicted cancer recurrence probability of the patient. 2.The deep learning method for predicting cancer recurrence probability based on multi-modal time-series features according to claim 1, wherein, The S1 comprises the following steps: S11, collecting the age, gender, tumor stage and pathological type of the patient from a hospital information system to obtain the clinical static data of the patient; S12, extracting a patient's medication sequence and radiotherapy parameters from an electronic medical record to obtain patient dynamic treatment data; S13, obtaining a patient's tumor marker detection value and detection timestamp and key index changes of imaging examination results from a follow-up system to obtain patient time series monitoring data; The patient clinical static data, the patient dynamic treatment data, and the patient time series monitoring data jointly constitute patient multi-modal time series data. 3.The deep learning method for predicting cancer recurrence probability based on multi-modal time-series features according to claim 1, wherein, The S2 includes the following steps: S21, using time series linear interpolation to fill in continuous variables in the patient multi-modal time series data to obtain filled patient multi-modal time series data; S22, using mode filling for categorical variables in the filled patient multi-modal time series data to obtain filled patient multi-modal time series data; S23, setting an invalid data threshold; calculating the sliding Zscore of the tumor marker detection value in the filled patient multi-modal time series data to obtain a sliding Zscore set; regarding the values of the sliding Zscore in the sliding Zscore set that are greater than the invalid data threshold as invalid sliding Zscore to obtain an invalid sliding Zscore set; Removing the tumor marker detection value in the filled patient multi-modal time series data corresponding to the invalid sliding Zscore set to obtain effective patient multi-modal time series data. 4.The deep learning method for predicting cancer recurrence probability based on multi-modal time-series features according to claim 1, wherein, The S3 includes the following steps: S31, setting tumor clinical features to obtain a tumor clinical feature set; collecting historical tumor clinical feature data based on the tumor clinical feature set; S32, initializing a random forest model and setting the feature importance evaluation criterion of the random forest model as Gini impurity reduction; Training the random forest model using the historical tumor clinical feature data combined with a cross-validation algorithm to obtain a trained random forest classifier; S33, calculating the purity improvement sum of each tumor clinical feature in the effective patient multi-modal time series data at the decision tree node split using the trained random forest classifier to obtain an improvement sum set; S34, sorting all features in the historical tumor clinical feature data in descending order according to the size of the improvement sum in the improvement sum set and performing normalization processing, recording the importance score of the historical tumor clinical feature to obtain a tumor clinical feature importance ranking; Setting a correlation threshold and a number of key features; selecting the key features with the highest importance score according to the tumor clinical feature importance ranking and the number of key features to obtain a key tumor clinical feature set; Verifying the clinical significance of the key features in the key tumor clinical feature set and excluding redundant features with a correlation greater than the correlation threshold to obtain a selected static feature set; Extracting the selected static features of the patient from the effective patient multi-modal time series data based on the selected static feature set to obtain patient selected static feature data. 5.The deep learning-based cancer recurrence probability prediction method based on multi-modal timing features according to claim 1, characterized in that, The S5 includes the following steps: S51, constructing an LSTM model and setting the hyperparameters of the LSTM model training; S52, collecting four-dimensional feature tensor data of historical cancer patients and label data of whether the historical cancer patients relapse to obtain labeled historical four-dimensional time series feature tensor data; S53, training the LSTM model using the labeled historical four-dimensional time series feature tensor data, using a random search algorithm to find the optimal hyperparameters of the LSTM model during the training process to obtain an optimal solution; using the optimal solution as the hyperparameters of the LSTM model to obtain an LSTM cancer recurrence prediction model; S54, inputting the four-dimensional time series feature tensor into the LSTM cancer recurrence prediction model to obtain a predicted patient recurrence probability. 6.The deep learning method for predicting cancer recurrence probability based on multi-modal time-series features according to claim 5, wherein, The S53 includes the following steps: S531、set the parameter search space as L , set the maximum random search number as M , set the cross-validation number as E , initialize the optimal hyperparameters; S532, under the optimal hyperparameters, using the labeled historical four-dimensional time series feature tensor data, training and evaluating the LSTM model to obtain the best prediction accuracy of the LSTM model d ; S533, defining a candidate hyperparameter set; randomly selecting a hyperparameter from the candidate hyperparameter set as f , dividing the training data in the labeled historical four-dimensional time-series feature tensor data into K folds, for each fold k , ; training the LSTM model from scratch using the training data in all the labeled historical four-dimensional time-series feature tensor data except the training data in the k fold, evaluating the model using the training data in the labeled historical four-dimensional time-series feature tensor data in the k fold to obtain a prediction accuracy, calculating the mean of the prediction accuracies of all the folds to obtain an average prediction accuracy g ; If the average prediction accuracy g is greater than the best prediction accuracy d , then update the best prediction accuracy as g , and update the optimal hyperparameters as f ; otherwise, keep the original best prediction accuracy and optimal hyperparameters. S534, repeating S533, and when the maximum random search number M is reached, stopping iteration to obtain an optimal solution; S54, inputting the four-dimensional time series feature tensor into the LSTM cancer recurrence prediction model to obtain a predicted patient recurrence probability. 7.The deep learning-based cancer recurrence probability prediction method based on multi-modal timing features according to claim 1, characterized in that, The S6 includes the following steps: S61, setting a recurrence probability treatment rule; S62, based on the set recurrence probability treatment rule and the predicted patient recurrence probability, treating the cancer patient.

8. A deep learning prediction system for cancer recurrence probability based on multi-modal temporal features, characterized in that, The system for implementing the cancer recurrence probability deep learning prediction method based on multi-modal time series features according to any one of claims 1-7, the system comprising a data acquisition module, a data processing module, a static feature screening module, a dynamic feature calculation and formatting module, a model construction and prediction module, and a treatment decision module.

9. A storage medium, characterized by A program is stored thereon, and when the program is executed by a processor, the cancer recurrence probability deep learning prediction method based on multi-modal time series features according to any one of claims 1-7 is implemented.

Images