A deep learning-based postoperative debilitation risk dynamic prediction method

Abstract

The application discloses a kind of postoperative debilitation risk dynamic prediction methods based on deep learning, it is related to deep learning technical field, including, based on key feature subset, using key feature discriminant layer, trajectory evolution prediction layer and trajectory probability calculation layer, construct dynamic trajectory perception model, output debilitation risk comprehensive state quantity;Gradient back propagation and parameter contribution degree analysis are carried out to debilitation risk comprehensive state quantity, generate feature influence intensity and evaluate result stability, obtain core risk factor;Based on core risk factor, construct feature adjustment strategy set and carry out analog deduction, obtain training sample, feedback to dynamic trajectory perception model executes online learning update.The application carries out gradient back propagation and parameter contribution degree analysis to debilitation risk comprehensive state quantity under deep learning framework, improves the dynamic adaptation capability of postoperative debilitation risk prediction, prediction timeliness and clinical interpretability.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method for dynamic prediction of postoperative frailty risk based on deep learning. Background Technology

[0002] In recent years, the identification and management of postoperative frailty risk in the fields of surgical medicine and geriatrics have gradually shifted from experience-based assessment to data-driven, refined prediction methods represented by deep learning. With the advancement of hospital information technology, electronic medical records, laboratory tests, vital sign monitoring, and nursing records are gradually forming computable clinical information resources. Deep learning and machine learning technologies have been widely researched and applied in clinical risk prediction. Related methods increasingly emphasize the automatic learning ability of deep learning network structures to understand complex feature interactions, as well as their comprehensive representation of the uncertainty and evolutionary trends of risk output. This has enabled postoperative frailty risk prediction to move from single, static assessments to continuous tracking and dynamic updates.

[0003] While existing technologies can assess postoperative frailty risk to some extent, they still have two shortcomings: First, the lack of a unified and reproducible processing chain for multi-source heterogeneous continuous time-series data in terms of unified time granularity, scale difference elimination, and encoding mapping makes it difficult to form a consistent temporal feature expression among hospital information, wearable device data, and nursing records. Second, the lack of dynamic reasoning on the evolution of risk states and interpretable attribution of risk drivers makes it difficult for clinicians to transform prediction results into clear intervention basis. Furthermore, the model lacks online update capability when the condition and nursing behavior change, which easily leads to problems of risk estimation lag or mismatch. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based method for dynamic prediction of postoperative frailty risk to solve the problems of inconsistent fusion of multi-source continuous time-series data and the difficulty in achieving online updates of postoperative frailty risk prediction.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a deep learning-based method for dynamic prediction of postoperative frailty risk, comprising: collecting continuous time-series sample data and performing unified time granularity, scale normalization, and encoding mapping processing; generating a time-series feature sample set through cross-modal alignment and joint encoding; training the time-series feature sample set using machine learning algorithms to construct a nonlinear mapping function; obtaining a key feature subset through feature perturbation sensitivity screening; constructing a dynamic trajectory perception model based on the key feature subset using a key feature discrimination layer, a trajectory evolution prediction layer, and a trajectory probability calculation layer, and outputting a comprehensive state quantity of frailty risk; performing gradient backpropagation and parameter contribution analysis on the comprehensive state quantity of frailty risk, generating feature influence intensity and evaluating the stability of the results to obtain core risk factors; constructing a feature adjustment strategy set based on the core risk factors and performing simulation deduction to obtain training samples, which are then fed back to the dynamic trajectory perception model for online learning and updating.

[0008] As a preferred embodiment of the deep learning-based dynamic prediction method for postoperative frailty risk described in this invention, the continuous time-series sample data includes continuous time-series data of physiological state, continuous time-series data of behavior and activity state, and continuous time-series data of nursing and status records.

[0009] As a preferred embodiment of the deep learning-based dynamic prediction method for postoperative frailty risk described in this invention, the steps of collecting continuous time-series sample data and performing unified time granularity, scale normalization, and encoding mapping processing, and generating a time-series feature sample set through cross-modal alignment and joint encoding, are as follows:

[0010] Perform uniform time granularity, scale normalization and encoding mapping processing on continuous time series sample data to obtain standard time series sample data;

[0011] Cross-modal alignment is performed on standard time-series sample data to obtain cross-modal aligned data, and a time-series feature sample set is generated through joint encoding.

[0012] As a preferred embodiment of the deep learning-based dynamic prediction method for postoperative frailty risk described in this invention, the method employs a machine learning algorithm to train a set of time-series feature samples and construct a nonlinear mapping function. The specific steps are as follows:

[0013] Machine learning algorithms are used to train a set of time-series feature samples to obtain nonlinear temporal correlations;

[0014] Based on the temporal nonlinear correlation, machine learning algorithms are solidified into nonlinear mapping functions.

[0015] As a preferred embodiment of the deep learning-based dynamic prediction method for postoperative frailty risk described in this invention, the specific steps for obtaining the key feature subset are as follows:

[0016] Based on the nonlinear mapping function, feature perturbation is performed on the time-series feature sample set, and the changes in the mapping response are recorded;

[0017] Based on changes in the mapping response, a subset of key features is obtained by calculating the sensitivity to feature perturbations.

[0018] As a preferred embodiment of the deep learning-based dynamic prediction method for postoperative frailty risk described in this invention, the following steps are taken to construct a dynamic trajectory perception model based on a subset of key features, employing a key feature discrimination layer, a trajectory evolution prediction layer, and a trajectory probability calculation layer.

[0019] A key feature discrimination layer is used to perform feature discrimination processing on a subset of key features to obtain discriminative feature representations;

[0020] The trajectory evolution prediction layer uses a nonlinear mapping inference algorithm to perform mapping inference on the discriminative feature representation to generate a postoperative weakness risk state evolution prediction trajectory.

[0021] The trajectory probability calculation layer performs probability calculations based on the predicted trajectory of postoperative weakness risk state evolution to generate a risk probability distribution.

[0022] The key feature discrimination layer, trajectory evolution prediction layer, and trajectory probability calculation layer are passed and cascaded layer by layer to generate a dynamic trajectory perception model.

[0023] As a preferred embodiment of the deep learning-based dynamic prediction method for postoperative frailty risk described in this invention, the output comprehensive frailty risk state quantity refers to the comprehensive frailty risk state quantity obtained by fusing the risk probability distribution and the postoperative frailty risk state evolution prediction trajectory through collaborative reasoning based on a dynamic trajectory perception model.

[0024] As a preferred embodiment of the deep learning-based dynamic prediction method for postoperative frailty risk described in this invention, the specific steps for performing gradient backpropagation and parameter contribution analysis on the comprehensive state quantity of frailty risk to generate feature influence intensity are as follows.

[0025] Extract the model parameters from the dynamic trajectory perception model, combine them with the comprehensive state variables of the decay risk to perform gradient backpropagation calculation, and obtain gradient information;

[0026] Based on gradient information, parameter contribution analysis is performed on model parameters, and the results are mapped to a subset of key features to generate feature influence strength.

[0027] As a preferred embodiment of the deep learning-based dynamic prediction method for postoperative frailty risk described in this invention, the acquisition of core risk factors refers to calculating the stability score of the results based on the influence intensity of features and performing consistency discrimination screening to acquire core risk factors.

[0028] As a preferred embodiment of the deep learning-based dynamic prediction method for postoperative frailty risk described in this invention, the steps of constructing a feature adjustment strategy set based on core risk factors, performing simulations, obtaining training samples, and feeding them back to the dynamic trajectory perception model for online learning and updating are as follows:

[0029] A parameter perturbation generation algorithm is used to perform constrained perturbation combinations on the model parameters corresponding to the core risk factors to construct a set of feature adjustment strategies.

[0030] The feature adjustment strategy set is used in combination with the dynamic trajectory perception model to perform forward inference simulation and obtain training samples;

[0031] The training samples are fed back to the dynamic trajectory perception model to perform online learning updates and make adaptive adjustments for core risk factors.

[0032] The beneficial effects of this invention are as follows: by performing gradient backpropagation and parameter contribution analysis on the comprehensive state quantity of frailty risk under the deep learning framework, core risk factors are identified and stably screened. Furthermore, a set of feature adjustment strategies is constructed based on the core risk factors for simulation and online learning updates, thereby improving the dynamic adaptability, prediction timeliness, and clinical interpretability of postoperative frailty risk prediction in scenarios of continuous temporal changes. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a flowchart of a deep learning-based method for dynamically predicting postoperative frailty risk.

[0035] Figure 2 A flowchart for obtaining a set of time-series feature samples.

[0036] Figure 3 This is a flowchart for filtering a subset of key features.

[0037] Figure 4 A flowchart updated for online learning. Detailed Implementation

[0038] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0039] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0040] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0041] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for dynamically predicting postoperative frailty risk based on deep learning, including the following steps:

[0042] S1: Collect continuous time-series sample data and perform unified time granularity, scale normalization and encoding mapping processing. Generate a set of time-series feature samples through cross-modal alignment and joint encoding.

[0043] S1.1: Continuous time-series sample data includes continuous time-series data of physiological state, continuous time-series data of behavior and activity state, and continuous time-series data of nursing and status records.

[0044] Specifically, by continuously collecting physiological monitoring records generated during the patient's surgical period and postoperative recovery, the patient's vital signs and physiological indicators are continuously recorded, and the physiological states corresponding to each time point are organized in chronological order to form continuous time-series data of physiological states.

[0045] The patient's activity and behavior information was collected in chronological order at different times, and the activity status at each time point was arranged chronologically to form continuous time-series data of behavior and activity status that reflects changes in the patient's behavior and activity status.

[0046] By collecting nursing operation records and patient status records generated during clinical nursing processes and organizing them in chronological order, continuous time-series data of nursing and status records reflecting changes in the nursing process and patient status are generated.

[0047] S1.2: Perform uniform time granularity, scale normalization and encoding mapping processing on continuous time series sample data to obtain standard time series sample data.

[0048] Specifically, the process involves reading the recorded time information from continuous time-series data of physiological states, continuous time-series data of behavioral and activity states, and continuous time-series data of nursing and status records; arranging and organizing the continuous time-series sample data according to a unified time order. Using the same time position as the alignment benchmark, physiological state records, behavioral and activity state records, and nursing and status records at the same time position are aligned to obtain time-series sample data with a unified time granularity.

[0049] Based on the time series sample data after unifying the time granularity, the numerical distribution of continuous time series data of physiological state, continuous time series data of behavior and activity state, and continuous time series data of nursing and status records in the entire time series is read. The minimum and maximum values ​​of the corresponding values ​​in the time series are determined as the scaling benchmark. The time series sample data after unifying the time granularity is transformed to the same scale according to the scaling benchmark to obtain time series sample data with unified scale.

[0050] Using time-series sample data that has undergone scale normalization as input, encoding and mapping processing is performed on continuous time-series data of physiological states, continuous time-series data of behavioral and activity states, and continuous time-series data of nursing and status records, converting the numerical states and nursing events in the time series into corresponding encoded representations.

[0051] For example, at a certain point in time, the encoding representation can be a coding sequence 0.72, 0.35, 1; 0.72 corresponds to the numerical encoding of continuous time series data of physiological state, 0.35 corresponds to the numerical encoding of continuous time series data of behavior and activity state, and 1 corresponds to the nursing event identifier encoding that occurred in continuous time series data of nursing and status records, thereby obtaining standard time series sample data.

[0052] S1.3: Perform cross-modal alignment on standard time series sample data to obtain cross-modal aligned data, and generate a time series feature sample set through joint encoding.

[0053] Specifically, based on the unified time order in the standard time series sample data, the continuous time series data of physiological state, continuous time series data of behavior and activity state, and continuous time series data of nursing and status records are organized to correspond at each time position, so that data from different sources form a stable cross-modal correspondence at the same time position, and cross-modal aligned data is obtained.

[0054] Extract the coding results of continuous time-series data of physiological state, behavior and activity state, and nursing and status records at the same time location. Combine the multi-source coding results to form a joint feature representation that can simultaneously represent physiological state, behavior and activity state, and nursing status information. Arrange the joint feature representations corresponding to each time location in chronological order to generate a time-series feature sample set.

[0055] S2: Using machine learning algorithms, train the time series feature sample set, construct a nonlinear mapping function, and obtain a subset of key features through feature perturbation sensitivity screening.

[0056] S2.1: Use machine learning algorithms to train the time series feature sample set to obtain the nonlinear correlation relationship of time series.

[0057] Specifically, the time-series feature sample set is arranged in chronological order, and the joint feature representations corresponding to adjacent time positions are used as training sample pairs in turn.

[0058] The training sample pairs are input into machine learning algorithms (such as neural network algorithms, support vector machine algorithms, and ensemble learning algorithms). The joint feature representation of the previous time position in the training sample pair is used as the input data, and the joint feature representation of the next time position in the training sample pair is used as the output target. Through multiple rounds of parameter updates, the machine learning algorithm gradually fits the mapping relationship between the input data and the output target, and obtains the temporal nonlinear correlation relationship used to characterize the nonlinear dependence between features in the temporal feature sample set.

[0059] It should be noted that machine learning algorithms gradually form a parameterized mapping structure that can stably reflect the inherent laws of data by repeatedly training the mapping relationship between input data and corresponding output results. During the training process, the internal parameter values ​​are adjusted through multiple rounds of iteration, so that the output results gradually approach the target results, thereby realizing the automatic learning of complex nonlinear relationships, and after the training is completed, it is used to perform mapping inference on unknown data.

[0060] S2.2: Based on the temporal nonlinear correlation, the machine learning algorithm is solidified into a nonlinear mapping function.

[0061] Specifically, the temporal nonlinear correlations of the trained machine learning algorithms are fixed, and the fixed machine learning algorithms are saved as a deterministic mapping carrier between inputs and outputs.

[0062] To ensure that the parameter values ​​formed by the machine learning algorithm at the end of training do not change, when the machine learning algorithm receives the joint feature representation in the set of temporal feature samples, it outputs the corresponding mapping result according to the temporal nonlinear correlation relationship, and obtains the nonlinear mapping function.

[0063] It should be noted that the nonlinear mapping function characterizes the correspondence between the joint feature representation in the time-series feature sample set and the joint feature representation at subsequent time positions, and is obtained by machine learning algorithms trained on the time-series feature sample set. The nonlinear mapping function can organize the joint feature representation at a given time position into an input vector based on the established time-series nonlinear correlation, and substitute the input vector into the nonlinear mapping function. The nonlinear mapping function performs nonlinear transformation on the input vector layer by layer, and outputs a mapping result that reflects the trend of physiological state, behavioral activity state and nursing state over time.

[0064] S2.3: Perform feature perturbation on the time series feature sample set according to the nonlinear mapping function, and record the changes in the mapping response.

[0065] Specifically, the time series feature sample set is used as input, and the complete time series feature sample set is input into the nonlinear mapping function to obtain the corresponding benchmark mapping output result.

[0066] While keeping the time order of the time series feature sample set unchanged, for each feature in the time series feature sample set, all time position values ​​of the currently selected feature in the time series feature sample set are uniformly replaced with the average value of the entire time series feature sample set, forming a corresponding perturbed time series feature sample set. The perturbed time series feature sample set is then input into a nonlinear mapping function to obtain the mapping output result under the perturbed condition.

[0067] Compare the changes between the baseline mapping output and the perturbation mapping output, record the differences in mapping output caused by changes in feature values, and determine the recorded output differences as changes in mapping response.

[0068] S2.4: Based on the changes in the mapping response, a subset of key features is obtained by calculating the sensitivity to feature perturbation.

[0069] Specifically, based on the changes in the mapping response, the mapping response changes for each feature in the time-series feature sample set are summarized, and the feature perturbation sensitivity is calculated, expressed as follows:

[0070] ;

[0071] in, For characteristic perturbation sensitivity, This represents the total number of time positions in the time-series feature sample set. This is the time location index value. It is a nonlinear mapping function. For time position The joint feature representation, For the time position General The perturbation joint feature representation formed by uniformly replacing the value of each feature. This is the index for the feature value.

[0072] It should be noted that the output term of the nonlinear mapping function in the expression for calculating the sensitivity to feature perturbation belongs to the same mapping space before and after the perturbation and has been normalized by the ratio and time dimension. Therefore, the dimensions of each term are consistent, and the calculation result is a dimensionless quantity.

[0073] Sort the features according to their perturbation sensitivity values, and then determine the difference between the perturbation sensitivity values ​​of adjacent features in the sorted sequence. The sorted position with the largest difference is determined as the boundary position where the change in the perturbation sensitivity value occurs, and the time-series features in the time-series feature sample set before the boundary position are determined as the key feature subset.

[0074] S3: Based on a subset of key features, a dynamic trajectory perception model is constructed by employing a key feature discrimination layer, a trajectory evolution prediction layer, and a trajectory probability calculation layer, and outputs a comprehensive state quantity of attenuation risk.

[0075] S3.1: Use a key feature discrimination layer to perform feature discrimination processing on the key feature subset to obtain discriminative feature representations.

[0076] Specifically, using a subset of key features as the processing object, the values ​​of the subset of key features at each time position are read sequentially according to the time order in the time-series feature sample set, forming a key feature input sequence that corresponds one-to-one with the time order.

[0077] The key feature input sequence is input into the key feature discrimination layer at each time position. The feature perturbation sensitivity of each feature in the key feature subset at each time position is statistically analyzed to obtain intermediate discrimination results.

[0078] The intermediate discrimination results obtained at each time point are summarized and organized in chronological order, and the intermediate discrimination results from multiple time points are integrated into a continuous and consistent discrimination output. The output after chronological organization and integration of discrimination results is used as the discrimination feature representation corresponding to the key feature subset.

[0079] It should be noted that the discriminative feature representation refers to the temporal feature expression formed after the values ​​of the key feature subset at each time position are processed by the key feature discriminative layer, and is used to distinguish the differences between different risk states.

[0080] S3.2: The trajectory evolution prediction layer uses a nonlinear mapping inference algorithm to perform mapping inference on the discriminative feature representation to generate a postoperative weakness risk state evolution prediction trajectory.

[0081] Specifically, the discriminant feature representation output by the key feature discrimination layer is used as input. According to the time order in the discriminant feature representation, the discriminant feature representation corresponding to each time position is sequentially input into the trajectory evolution prediction layer. A nonlinear mapping inference algorithm is used to concatenate the discriminant feature representation of the current time position with the mapping output result of the previous time position to form a joint input vector, which is then input into the nonlinear mapping function to obtain the continuous inference result of the state change trend.

[0082] After completing the nonlinear mapping reasoning at a single time position, the continuous reasoning results are used as the reference input for the next time position, and the process is advanced step by step to realize the continuous mapping reasoning of the discriminant feature representation in the time dimension. The continuous reasoning results of the discriminant feature representation at all time positions are statistically analyzed to obtain a sequence of state evolution prediction results arranged in chronological order. The sequence of state evolution prediction results is then organized to form the postoperative weakness risk state evolution prediction trajectory.

[0083] It should be noted that the nonlinear mapping inference algorithm refers to the reasoning process that, after the machine learning algorithm has been trained and solidified into a nonlinear mapping function, uses the fixed mapping relationship to perform mapping calculations on the input features in chronological order. The nonlinear mapping inference algorithm takes the feature representation at a certain time position as input and uses the obtained mapping result as the inference output at the current time position, thereby realizing continuous mapping inference and state evolution inference of feature representation in the time dimension.

[0084] S3.3: The trajectory probability calculation layer performs probability calculations based on the predicted trajectory of postoperative weakness risk state evolution to generate a risk probability distribution.

[0085] Specifically, the state evolution prediction results corresponding to each time position are read sequentially according to the time sequence in the postoperative weakness risk state evolution prediction trajectory; the trajectory probability calculation layer calculates the occurrence probability value of the state evolution prediction result at each time position to obtain the risk probability result corresponding to each time position.

[0086] The expression for calculating the probability of occurrence is:

[0087] ;

[0088] in, The probability of occurrence is set to a value. For time position The state evolution prediction results are taken as follows: This represents the total number of distinguishable state results at the same time location. This is the index value of the state result. For the first The state evolution prediction result is assigned to each distinguishable state.

[0089] It should be noted that in the expression for calculating the probability of occurrence, all state evolution prediction results come from the same mapping space and participate in the calculation in the form of normalized ratios. The numerator and denominator have the same dimensions, and the probability of occurrence is a dimensionless quantity.

[0090] After completing the probability mapping at a single time location, the obtained risk probability results are continuously organized with the risk probability results at adjacent time locations to maintain the consistency of the temporal order in the postoperative weakness risk state evolution prediction trajectory, forming a sequence of risk probability results arranged in chronological order.

[0091] The risk probability result sequence is uniformly summarized to generate a risk probability distribution that characterizes how risk changes over time.

[0092] S3.4: The key feature discrimination layer, trajectory evolution prediction layer and trajectory probability calculation layer are passed and cascaded layer by layer to generate a dynamic trajectory perception model.

[0093] Specifically, the key feature subset is input into the key feature discrimination layer, which performs feature discrimination processing on the key feature subset according to the nonlinear mapping function and outputs the discriminative feature representation.

[0094] The discriminative feature representation is passed as input to the trajectory evolution prediction layer. The trajectory evolution prediction layer performs continuous mapping reasoning on the discriminative feature representation based on a nonlinear mapping reasoning algorithm to generate the postoperative weakness risk state evolution prediction trajectory.

[0095] The predicted trajectory of postoperative weakness risk state evolution is passed to the trajectory probability calculation layer. The trajectory probability calculation layer performs probability calculations at each time position based on the predicted trajectory of postoperative weakness risk state evolution and outputs the risk probability distribution.

[0096] By continuously connecting the key feature discrimination layer, trajectory evolution prediction layer, and trajectory probability calculation layer, a complete computational link from key feature subset to risk probability distribution is formed, generating a dynamic trajectory perception model.

[0097] S3.5: Based on the dynamic trajectory perception model, the risk probability distribution and the postoperative weakness risk state evolution prediction trajectory are fused through collaborative reasoning to obtain the comprehensive state quantity of weakness risk.

[0098] Specifically, the risk probability distribution and postoperative weakness risk state evolution prediction trajectory in the dynamic trajectory perception model are organized in a unified time sequence at each time position to form a combination of risk probability information and state evolution information that corresponds one-to-one at the time position.

[0099] Based on the correspondence between time and location, the risk probability information and state evolution information at the same time location are jointly input, and a nonlinear mapping function is used for mapping calculation to obtain a comprehensive reasoning result.

[0100] The number of inferences in the comprehensive inference results is counted, and the comprehensive inference results obtained at each time position are arranged and summarized in chronological order; by uniformly organizing the collaborative inference fusion results across the entire time range, a comprehensive state quantity of decay risk is generated.

[0101] S4: Perform gradient backpropagation and parameter contribution analysis on the comprehensive state variables of the decay risk, generate characteristic influence intensity and evaluate the stability of the results, and obtain the core risk factors.

[0102] S4.1: Extract the model parameters from the dynamic trajectory perception model, combine them with the weakening risk comprehensive state quantity to perform gradient backpropagation calculation, and obtain gradient information.

[0103] Specifically, following the cascaded order of the key feature discrimination layer, trajectory evolution prediction layer, and trajectory probability calculation layer, the model parameter values ​​involved in the calculation of the comprehensive state quantity of decay risk are read layer by layer, and the model parameter values ​​and their corresponding hierarchical relationships are organized together.

[0104] Using the weakening risk integrated state quantity as the target output of gradient backpropagation calculation, starting from the output position corresponding to the weakening risk integrated state quantity, the calculation is carried out layer by layer in reverse along the calculation path of the dynamic trajectory perception model.

[0105] The difference between the comprehensive state quantity of weakening risk and the model parameter values ​​of the trajectory probability calculation layer is extracted in the trajectory probability calculation layer; the difference between the comprehensive state quantity of weakening risk and the model parameter values ​​of the trajectory evolution prediction layer is extracted in the trajectory evolution prediction layer; and the difference between the comprehensive state quantity of weakening risk and the model parameter values ​​of the key feature discrimination layer is extracted at the key feature discrimination layer.

[0106] The differences between the key feature discrimination layer, the trajectory evolution prediction layer, and the trajectory probability calculation layer are used as gradient values, which are then compiled and summarized to obtain gradient information for subsequent parameter contribution analysis.

[0107] It should be noted that model parameters refer to a set of numerical values ​​determined during the training process of machine learning algorithms by repeatedly fitting the mapping relationship between input features and output results through a set of time-series feature samples. Model parameters are used to characterize the strength and combination of the role of each input feature in the mapping calculation in the nonlinear mapping function. After training, the values ​​are kept fixed so that when the nonlinear mapping function receives a set of time-series feature samples or a discriminant feature representation, it can perform deterministic mapping calculations according to the model parameters and output stable and consistent inference results.

[0108] S4.2: Perform parameter contribution analysis on the model parameters based on gradient information, and map them to a subset of key features to generate feature influence intensity.

[0109] Specifically, according to the hierarchical source of gradient information in the key feature discrimination layer, trajectory evolution prediction layer and trajectory probability calculation layer, the gradient values ​​corresponding to the model parameters are grouped and summarized, and sorted by gradient values ​​within the same level to form parameter contribution analysis results divided by level and gradient values.

[0110] The parameter contribution analysis results are mapped to the corresponding features in the key feature subset one by one to obtain the effect strength interpretation information (such as feature-level effect strength information, time-dimensional effect distribution information, and directional consistency information) for each feature in the key feature subset. The effect strength interpretation information for each feature in the key feature subset is then uniformly organized to generate the feature influence strength.

[0111] It should be noted that,

[0112] S4.3: Calculate the stability score of the results based on the influence strength of the features, and perform consistency discrimination screening to obtain the core risk factors.

[0113] Specifically, based on the intensity of the characteristic influence, the stability score is calculated according to the time sequence corresponding to the comprehensive state quantity of the decay risk, and the expression is:

[0114] ;

[0115] in, Score the stability of the results. For the number of inferences, This is the index value for the number of inferences. For the first The value of the comprehensive state quantity of weakening risk obtained from this reasoning is taken. This represents the average value of the comprehensive state quantity of the weakening risk. The mean of the squared deviations of the comprehensive state quantity of weakening risk relative to the mean value.

[0116] It should be noted that in the stability score expression of the calculation results, the value of the comprehensive state quantity of the weakening risk is in the same dimension space as the average value, and the squared deviation term is in the same dimension as the average value and participates in the calculation in the form of a ratio. Therefore, the stability score of the result is a dimensionless quantity, and the dimensions are kept consistent.

[0117] Using the outcome stability score as the criterion, the interpretation information of the effect intensity of the same feature at different time positions is compared one by one to determine whether the outcome stability score is consistent. Features with consistent outcome stability scores at all time positions are retained, while features with inconsistent outcome stability scores are eliminated to obtain the core risk factors.

[0118] S5: Construct a set of feature adjustment strategies based on core risk factors and conduct simulations to obtain training samples, which are then fed back to the dynamic trajectory perception model for online learning and updates.

[0119] S5.1: A parameter perturbation generation algorithm is used to perform constrained perturbation combinations on the model parameters corresponding to the core risk factors to construct a set of feature adjustment strategies.

[0120] Specifically, based on the parameter correspondence of the core risk factors in the dynamic trajectory perception model, the model parameters associated with the core risk factors are located one by one; a parameter perturbation generation algorithm is used to shift the values ​​of the located model parameters to obtain multiple different value states of the model parameters.

[0121] By combining and organizing different model parameter values ​​corresponding to the same core risk factor, multiple sets of parameter value combinations reflecting different adjustment scenarios of the core risk factor are obtained; and each set of parameter value combinations is organized into a set of characteristic adjustment strategies.

[0122] It should be noted that the parameter perturbation generation algorithm refers to an algorithmic process that makes small, limited changes to the values ​​of model parameters based on the already determined values ​​of model parameters. The parameter perturbation generation algorithm adjusts the values ​​of model parameters corresponding to the core risk factors in a controlled manner without destroying the overall structure of the original nonlinear mapping function, thereby forming multiple sets of different but comparable parameter combinations for subsequent forward inference simulation and training sample generation.

[0123] S5.2: Apply the feature adjustment strategy set to the dynamic trajectory perception model, perform forward inference simulation, and obtain training samples.

[0124] Specifically, a set of feature adjustment strategies from the feature adjustment strategy set is used as input. According to the feature adjustment strategy set, the model parameters corresponding to the core risk factors in the dynamic trajectory perception model are replaced to obtain the dynamic trajectory perception model after parameter replacement.

[0125] A subset of key features is used as input for forward inference and is fed into the key feature discrimination layer in chronological order according to the key feature input sequence. The key feature discrimination layer outputs a discrimination feature representation that corresponds one-to-one with the time position.

[0126] The discriminative feature representation is passed to the trajectory evolution prediction layer in chronological order. The trajectory evolution prediction layer performs a nonlinear mapping function on the discriminative feature representation at each time position and continuously organizes the mapping results of adjacent time positions in chronological order to generate the postoperative weakness risk state evolution prediction trajectory.

[0127] The predicted trajectory of postoperative weakness risk state evolution is passed to the trajectory probability calculation layer in chronological order and time position to obtain the risk probability distribution.

[0128] The risk probability distribution results corresponding to the same time location are matched with the postoperative frailty risk state evolution prediction trajectory results and used as joint inputs. The input is a nonlinear mapping function to obtain the comprehensive inference results at the same time location. The comprehensive inference results of all time locations are summarized and organized in chronological order to output the comprehensive frailty risk state quantity. The feature adjustment strategy, the parameter values ​​of the dynamic trajectory perception model after parameter replacement, the key feature subsets, and the comprehensive frailty risk state quantity are recorded in groups to generate training samples.

[0129] S5.3: Feed the training samples back to the dynamic trajectory perception model to perform online learning updates and make adaptive adjustments for core risk factors.

[0130] Specifically, according to the feature adjustment strategy and the comprehensive state of weakening risk recorded in the training samples, the training samples are sequentially input into the dynamic trajectory perception model in chronological order; based on the comprehensive state of weakening risk given in the training samples, the output results obtained by the dynamic trajectory perception model in the forward inference process are compared sample by sample, and the parameters are updated along the calculation path of the dynamic trajectory perception model to obtain the changing trend of the model parameters corresponding to the core risk factors in the training sample update.

[0131] The adjustment range of parameters associated with core risk factors is dynamically converged, so that the strength of the role of core risk factors in the dynamic trajectory perception model is adjusted according to the feedback results of training samples, thus completing online learning and updating based on the feedback of training samples.

[0132] In summary, this invention improves the dynamic adaptability, prediction timeliness, and clinical interpretability of postoperative frailty risk prediction in continuously changing scenarios by performing gradient backpropagation and parameter contribution analysis on the comprehensive state quantity of frailty risk within a deep learning framework, identifying and stabilizing core risk factors, and further constructing a set of feature adjustment strategies based on the core risk factors for simulation and driving online learning updates.

[0133] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for dynamic prediction of postoperative frailty risk based on deep learning, characterized in that: include, Collect continuous time-series sample data and perform uniform time granularity, scale normalization and encoding mapping processing. Generate a set of time-series feature samples through cross-modal alignment and joint encoding. Machine learning algorithms are used to train a set of time-series feature samples, construct a nonlinear mapping function, and obtain a subset of key features by filtering based on feature perturbation sensitivity. Based on a subset of key features, a dynamic trajectory perception model is constructed by employing a key feature discrimination layer, a trajectory evolution prediction layer, and a trajectory probability calculation layer, and outputs a comprehensive state quantity of attenuation risk. Gradient backpropagation and parameter contribution analysis are performed on the comprehensive state variables of the decay risk to generate the intensity of characteristic influences and evaluate the stability of the results, thereby obtaining the core risk factors. Based on core risk factors, a set of feature adjustment strategies is constructed and simulated to obtain training samples, which are then fed back to the dynamic trajectory perception model for online learning and updating.

2. The method for dynamic prediction of postoperative frailty risk based on deep learning as described in claim 1, characterized in that: The continuous time-series sample data includes continuous time-series data of physiological state, continuous time-series data of behavior and activity state, and continuous time-series data of nursing and status records.

3. The method for dynamic prediction of postoperative frailty risk based on deep learning as described in claim 2, characterized in that: The process involves collecting continuous time-series sample data and performing unified time granularity, scale normalization, and encoding mapping. Through cross-modal alignment and joint encoding, a time-series feature sample set is generated. The specific steps are as follows: Perform uniform time granularity, scale normalization and encoding mapping processing on continuous time series sample data to obtain standard time series sample data; Cross-modal alignment is performed on standard time-series sample data to obtain cross-modal aligned data, and a time-series feature sample set is generated through joint encoding.

4. The method for dynamic prediction of postoperative frailty risk based on deep learning as described in claim 3, characterized in that: The method employs machine learning algorithms to train a set of time-series feature samples and construct a nonlinear mapping function. The specific steps are as follows. Machine learning algorithms are used to train a set of time-series feature samples to obtain nonlinear temporal correlations; Based on the temporal nonlinear correlation, machine learning algorithms are solidified into nonlinear mapping functions.

5. The method for dynamic prediction of postoperative frailty risk based on deep learning as described in claim 4, characterized in that: The specific steps for obtaining the key feature subset are as follows: Based on the nonlinear mapping function, feature perturbation is applied to the time-series feature sample set, and the changes in the mapping response are recorded; Based on changes in the mapping response, a subset of key features is obtained by calculating the sensitivity to feature perturbations.

6. The method for dynamic prediction of postoperative frailty risk based on deep learning as described in claim 5, characterized in that: The dynamic trajectory perception model is constructed based on a subset of key features, employing a key feature discrimination layer, a trajectory evolution prediction layer, and a trajectory probability calculation layer. The specific steps are as follows: A key feature discrimination layer is used to perform feature discrimination processing on a subset of key features to obtain discriminative feature representations; The trajectory evolution prediction layer uses a nonlinear mapping inference algorithm to perform mapping inference on the discriminative feature representation to generate a postoperative weakness risk state evolution prediction trajectory. The trajectory probability calculation layer performs probability calculations based on the predicted trajectory of postoperative weakness risk state evolution to generate a risk probability distribution. The key feature discrimination layer, trajectory evolution prediction layer, and trajectory probability calculation layer are passed and cascaded layer by layer to generate a dynamic trajectory perception model.

7. The method for dynamic prediction of postoperative frailty risk based on deep learning as described in claim 6, characterized in that: The output comprehensive state quantity of frailty risk refers to the comprehensive state quantity of frailty risk obtained by fusing the risk probability distribution and the postoperative frailty risk state evolution prediction trajectory through collaborative reasoning based on the dynamic trajectory perception model.

8. The method for dynamic prediction of postoperative frailty risk based on deep learning as described in claim 7, characterized in that: The specific steps for performing gradient backpropagation and parameter contribution analysis on the comprehensive state variables of the attenuation risk to generate characteristic influence intensity are as follows. Extract the model parameters from the dynamic trajectory perception model, combine them with the comprehensive state variables of the decay risk to perform gradient backpropagation calculation, and obtain gradient information; Based on gradient information, parameter contribution analysis is performed on model parameters, and the results are mapped to a subset of key features to generate feature influence strength.

9. The method for dynamic prediction of postoperative frailty risk based on deep learning as described in claim 8, characterized in that: The process of obtaining core risk factors involves calculating the stability score of the results based on the strength of the influence of the features, and then performing consistency discrimination screening to obtain the core risk factors.

10. The method for dynamic prediction of postoperative frailty risk based on deep learning as described in claim 9, characterized in that: The specific steps are as follows: A set of feature adjustment strategies is constructed based on core risk factors, and simulations are performed to obtain training samples. These samples are then fed back to the dynamic trajectory perception model for online learning and updates. A parameter perturbation generation algorithm is used to perform constrained perturbation combinations on the model parameters corresponding to the core risk factors to construct a set of feature adjustment strategies. The feature adjustment strategy set is used in combination with the dynamic trajectory perception model to perform forward inference simulation and obtain training samples; The training samples are fed back to the dynamic trajectory perception model to perform online learning updates and make adaptive adjustments for core risk factors.

Images