ICU (Intensive Care Unit) duration prediction method based on adaptive meta-learning probabilistic reasoning
Through the adaptive meta-learning probabilistic reasoning framework and frequency domain sequence decomposition mechanism, the accuracy and stability problems of early ICU hospitalization length prediction are solved, individualized ICU length prediction is achieved, the adaptability and generalization ability of the model are improved, and clinical decision-making is supported.
Patent Information
- Application Number
- CN202510808621.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-17
- Publication Date
- 2025-09-19
AI Technical Summary
Existing ICU length of stay prediction methods are difficult to achieve efficient, stable, and accurate individualized regression prediction in the early stages of ICU admission, especially under conditions of small samples, heterogeneous, and highly dynamic data, where the accuracy is low and the stability is insufficient.
An adaptive meta-learning probabilistic reasoning framework is adopted, combined with the frequency domain sequence decomposition mechanism and the adaptive meta-learning task construction strategy. The clinical variable sequence is decomposed into seasonal and trend components through the frequency domain decomposition module. Cluster matching is performed using the variational autoencoder, and a probabilistic reasoning module is constructed to predict ICU length of stay.
It achieves efficient, stable and accurate prediction of patients' length of stay in the early stage of ICU admission, improves the adaptability and generalization ability of the model in heterogeneous populations, and supports clinical decision-making such as the timing of department transfer and bed rotation arrangements.
Smart Images

Figure CN120674078A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of smart medical technology, and specifically relates to an ICU duration prediction method based on adaptive meta-learning probabilistic reasoning. Background Art
[0002] In the intensive care unit (ICU), accurately predicting a patient's length of stay is a key factor influencing multiple decision-making aspects, including medical resource allocation, transfer planning, cost management, and risk control. The course of illness varies greatly among ICU patients, with hospital stays ranging from one or two days to several months. Accurately predicting a patient's ICU length of stay can help hospitals optimize bed transfers, develop transfer plans, predict nursing intensity, and coordinate cross-departmental resource allocation. Therefore, in the context of limited bed capacity and medical resources, predicting length of stay directly impacts the efficiency and responsiveness of the entire hospital system, and thus has significant practical significance.
[0003] Current mainstream length of stay prediction methods still rely on static scoring systems or traditional machine learning models, which fail to meet the urgent clinical need for efficient, accurate, and personalized predictions. In recent years, numerous studies have explored the use of neural networks to model length of stay for inpatients (including those in the ICU and general wards). However, these studies generally simplify length of stay prediction to a binary classification task, such as identifying patients at risk of a "long stay" (e.g., >7 days) to facilitate early intervention and transfer planning. This classification modeling approach has significant limitations: First, the artificial discretization of continuous time leads to information loss, reducing prediction accuracy and clinical applicability; second, the lack of uniformity in classification criteria leads to different definitions of "long stay" boundaries across hospitals and departments, limiting the generalization and portability of the model. Therefore, compared to classification, direct modeling of length of stay using regression approaches is more aligned with actual medical needs and more capable of meeting the requirements of downstream tasks such as high-precision resource forecasting, cost management, and path optimization.
[0004] In clinical practice, the early stages of ICU admission (e.g., the first 48 hours) present a golden window for implementing critical interventions and adjusting decisions. However, leveraging this early data to achieve accurate, individualized regression predictions of length of stay is a highly challenging task. The limited amount of data available within these 48 hours is highly heterogeneous and highly dimensional (including time-series data such as vital signs and laboratory parameters). Furthermore, the patient population varies greatly (in age, disease type, and underlying conditions), significantly increasing the difficulty of modeling. Furthermore, due to the high patient mobility, rapid disease progression, and ongoing adjustments to treatment strategies within the ICU, the data distribution constantly shifts over time. Existing methods often lack rapid adaptability, resulting in low prediction accuracy and insufficient stability under these conditions of small-sample, heterogeneous, and highly dynamic data. Summary of the Invention
[0005] In view of the above, the purpose of the present invention is to provide an ICU length of stay prediction method based on adaptive meta-learning probabilistic reasoning. By introducing a meta-learning probabilistic reasoning framework and designing a frequency domain sequence decomposition mechanism and an adaptive meta-learning task construction strategy, a model system is established that can efficiently, stably, and accurately predict the length of hospital stay of patients in the early stage of ICU admission, thereby achieving accurate prediction of ICU length of stay.
[0006] To achieve the above-mentioned object of the invention, an embodiment provides a method for predicting ICU duration based on adaptive meta-learning probabilistic reasoning, comprising the following steps: Collect clinical variable sequence data of ICU patients and future ICU duration as sample data and corresponding labels, and divide them into support set and query set; A model for predicting ICU duration was constructed, comprising a frequency-domain decomposition module, an adaptive meta-learning task construction module, and a probabilistic reasoning module. The frequency-domain decomposition module was used to decompose each clinical variable sequence into seasonal and trend components. The adaptive meta-learning task construction module was used to learn global representations of the seasonal and trend components, respectively, and based on the global representations, the seasonal and trend components were divided into different tasks using a cluster matching method. The probabilistic reasoning module was used to learn specific task parameters for different tasks using a support set, and to generalize the task-specific features of the query set samples from the task-specific parameters to infer the ICU duration of the query set samples. After constructing the loss function to optimize the parameters of the ICU duration prediction model, the ICU duration prediction model is used to predict the ICU duration of each sample to be tested.
[0007] Preferably, the clinical variables include diastolic blood pressure, inspired oxygen fraction, blood glucose, heart rate, mean blood pressure, blood oxygen saturation, respiratory rate, systolic blood pressure, body temperature and pH value.
[0008] Preferably, in the frequency domain decomposition module, a fast Fourier transform is used to decompose each input clinical time series into a seasonal component and a trend component, so as to independently model the dynamic characteristics of the two components, which can be expressed as follows:
[0009] in, stands for Fast Fourier Transform, is the inverse fast Fourier transform, is the threshold for dividing the frequency components, Indicates higher than The high frequency part of and After transformation, seasonal components are obtained , Indicates lower than The low-frequency part of and After transformation, trend components are obtained .
[0010] Preferably, the seasonal component is further subjected to denoising processing, and the denoised seasonal component is input into the adaptive meta-learning task construction module, wherein the denoising processing includes: filtering the Top-K frequencies corresponding to the amplitude to reduce the noise in the seasonal component.
[0011] Preferably, in the adaptive meta-learning task building module, the encoder in the variational autoencoder is used to learn the global representation of the seasonal component and the trend component respectively; Kmeans clustering is used based on the global representation of each component. According to the two dimensions of seasonal component and trend component, the support set samples are divided into different clusters. Each cluster corresponds to a task, which is expressed as: and ,in, Indicates the task index, and The number of tasks representing seasonal components and trend components respectively, and Respectively represent n The seasonal components and their labels corresponding to the tasks, represents the seasonal task set corresponding to the seasonal component dimension, and Respectively represent n The trend components and labels corresponding to each task, Represents the trend task set corresponding to the trend component dimension.
[0012] Preferably, in the probabilistic reasoning module, two neural network encoders are used to encode the seasonal component and trend component of all samples in the support set and query set respectively from the two dimensions of seasonal component and trend component, to obtain the potential representation of each component, and the input features of each task are constructed based on the potential representation of each component of all samples in the support set; A two-stream parameter generation network is used to predict the specific task parameters of each type of task corresponding to the support set samples based on the input features of two dimensions; For each query set sample, its global representation is matched with the cluster center of each task cluster to determine the matching task and its specific task parameters. Based on the specific task parameters, the two-dimensional specific task features of the query set sample are learned, and the decoder is used to predict the ICU duration of the query set sample based on the two-dimensional specific task features.
[0013] Preferably, the input features for each task are constructed based on the potential representation of each component of all samples in the support set, including: For each dimension, the potential representation of each sample in each task in the support set is concatenated with its label to obtain the input features of each sample, and then the input features of all samples in each task are concatenated to form the input features of each task.
[0014] Preferably, a dual-stream parameter generation network is used to predict the specific task parameters of each type of task corresponding to the support set samples based on the input features of the two dimensions, including: The dual-stream parameter generation network uses two data stream branches, each of which uses a feedforward neural network; The input features corresponding to each task in the two dimensions are respectively input into the two data flow branches to predict the mean and variance of the posterior distribution of the specific task parameters corresponding to each type of task corresponding to the support set samples, assuming that the posterior distribution of the specific task parameters is a Gaussian distribution.
[0015] Preferably, the probability distribution of ICU time in two dimensions of the query set samples is learned based on specific task parameters, including: The potential representation of the corresponding component of each dimension of the query set sample is spliced with the matching specific task parameters and then input into the neural network for fusion to obtain the probability distribution of ICU time for each dimension.
[0016] Preferably, the constructed loss function includes the ELBO loss function for the variational autoencoder, and also includes the task loss of the query set in different tasks, wherein the task loss adopts Monte Carlo sampling of specific task parameters, and obtains the approximate predicted ICU length of the query set sample through multiple sampling.
[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. How to achieve personalized and accurate regression prediction of hospital stay in the early stages of ICU admission. Hospital stay is one of the core indicators of critical care resource management. This invention uses regression modeling to directly output the patient's expected continuous hospital stay, which can support clinical decision-making in a more granular manner, such as determining the timing of department transfer, arranging bed rotation, cost estimation, and path optimization.
[0018] 2. How to extract key trends and fluctuations from dynamic time series data. Early ICU admission data exhibits significant fluctuations and non-stationarity, making it difficult to capture its underlying structure through direct modeling. This paper proposes a frequency domain sequence decomposition mechanism to split the input clinical variable sequence data into seasonal and trend components, which are then independently modeled. This allows the model to capture the impact of both short-term fluctuations and long-term changes, thereby improving the expressiveness and accuracy of duration predictions.
[0019] 3. How to improve the adaptability and generalization ability of the model in the heterogeneous population of ICU. The conditions of ICU patients are complex and highly individualized. There are huge differences in the treatment paths and recovery speeds of different patients. Traditional models are difficult to generalize to new patients. This invention introduces the meta-learning probabilistic reasoning framework into the ICU duration prediction model for the first time: the duration prediction of each new ICU patient is modeled as a new task. The model captures shared patterns by transferring knowledge from similar patients in historical tasks, thereby quickly inferring the probability distribution of the new patient's ICU duration, improving rapid adaptation and stability in a dynamically changing environment.
[0020] 4. How to rationally divide meta-learning tasks to enhance meta-learning effectiveness. Unlike most meta-learning methods that rely on artificial task division (such as grouping by disease type or age), this paper designs an adaptive task construction mechanism based on clustering. It uses the global representation generated by the variational autoencoder to perform unsupervised clustering of support set samples and adaptively construct seasonal and trend task clusters. This avoids the introduction of bias due to reliance on prior knowledge and improves the robustness and accuracy of the model in the dynamic environment of the ICU. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0022] Figure 1 is a flowchart of an ICU duration prediction method based on adaptive meta-learning probabilistic reasoning provided in an embodiment; Figure 2 This is a schematic diagram of the structure and flow of the ICU duration prediction model provided in the embodiment; Figure 3 This is a time-series frequency domain decomposition effect diagram provided by the embodiment; Figure 4 Schematic diagram of clustering task construction provided by the embodiment; Figure 5 Schematic diagram of a probabilistic reasoning module provided in an embodiment. DETAILED DESCRIPTION
[0023] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.
[0024] The inventive concept of the present invention is to solve the three core problems currently faced in ICU hospital stay prediction: insufficient model adaptability due to large changes in the distribution of ICU indicator data over time; weak model generalization ability due to the small number of ICU samples; and low model accuracy due to large individual differences among ICU patients. The embodiment of the present invention provides an ICU duration prediction scheme based on adaptive meta-learning probabilistic reasoning, which introduces a meta-learning probabilistic reasoning framework and designs a frequency domain sequence decomposition mechanism and an adaptive meta-learning task construction strategy to establish a model system that can perform efficient, stable, and accurate predictions on patient hospital stays in the early stages of ICU admission, thereby achieving accurate prediction of ICU duration.
[0025] like Figure 1 As shown, the embodiment provides a method for predicting ICU duration based on adaptive meta-learning probabilistic reasoning, comprising the following steps: S1, collects clinical variable sequence data of ICU patients and future ICU duration as sample data and corresponding labels, and divides them into support set and query set.
[0026] In this example, clinical variable sequence data for the first 48 hours of ICU admission is first sampled every two hours, and the remaining hospital stay after 48 hours is recorded as the corresponding label. The 10 continuous clinical variables routinely monitored in the ICU are extracted: diastolic blood pressure, inspired oxygen fraction, blood glucose, heart rate, mean blood pressure, blood oxygen saturation, respiratory rate, systolic blood pressure, body temperature, and pH value, to construct time series data corresponding to each variable.
[0027] Then, each set of sample data (including clinical variable sequences and corresponding labels) is divided into a support set and queryset ,in represents the sequence and corresponding label contained in the mth sample in the support set, The mth sample in the query set contains the sequence and corresponding label, where N and M are the number of samples. The support set is used to generate task-specific parameters to guide the model to quickly adapt to the current task. The purpose of the query set is to evaluate the model's prediction ability and calculate the loss based on the parameters of the support set, driving parameter updates. The support set and query set do not overlap with each other to ensure the model's generalization ability. Meta-learning is used to learn shared knowledge from the support set and model the conditional probability distribution of the query set. , thereby predicting the labels of the query set .
[0028] S2, builds an ICU duration prediction model that includes a frequency domain decomposition module, an adaptive meta-learning task construction module, and a probabilistic reasoning module.
[0029] like Figure 2As shown, the constructed ICU duration prediction model includes a frequency domain decomposition module, an adaptive meta-learning task construction module, and a probabilistic reasoning module. The performance is improved by the meta-learning probabilistic reasoning framework, the frequency domain sequence decomposition mechanism, and the adaptive meta-learning task construction strategy in these three modules. Meta-learning achieves rapid prediction of new tasks by capturing shared knowledge between historical tasks. When a new patient is admitted to the ICU, the model constructed by the present invention can quickly predict the length of hospital stay of the new patient by transferring knowledge of similar patients in history. Through frequency domain decomposition, the local physiological fluctuations in the time series of clinical indicators are effectively decoupled from the long-term change trends, which helps the model capture short-term and long-term features respectively and improve the expression ability. The adaptive meta-learning task construction strategy breaks through the limitations of traditional task construction based on prior rules, realizes the automatic division of meta-tasks according to the structure of the data itself, enhances the rationality of migration between tasks, and improves the model's generalization ability for unknown tasks. Finally, the model predicts the label of the sample by modeling the conditional probability distribution of the sample label based on historical tasks.
[0030] In the embodiment, the frequency domain decomposition module is used to decompose each clinical variable sequence into a seasonal component and a trend component. Specifically, a fast Fourier transform is used to decompose each input clinical time series into a seasonal component and a trend component, so as to independently model the dynamic characteristics of the two components, which can be expressed as follows:
[0031] in, stands for Fast Fourier Transform, is the inverse fast Fourier transform, is the threshold for dividing the frequency components, Indicates higher than The high frequency part of and After transformation, seasonal components are obtained , Indicates lower than The low-frequency part of and After transformation, trend components are obtained . Figure 3 This is the effect diagram of time series frequency domain decomposition. It can be seen intuitively that complex mixed time series are decomposed into stable components, which helps to capture short-term fluctuations and long-term progress respectively, enhance representation capabilities, and improve prediction accuracy.
[0032] Due to seasonal ingredients Usually contains a lot of noise. In the regression prediction task, the presence of noise often reduces the generalization ability of the model. Therefore, the present invention also performs denoising on the seasonal component, which may include: filtering the Top-K frequencies corresponding to the amplitude to reduce the noise in the seasonal component, thereby obtaining the final seasonal component , seasonal components after denoising Input into the adaptive meta-learning task building module.
[0033] In the embodiment, after the sequence of all samples is decoupled, the seasonal components and trend components are adaptively constructed into meta-learning tasks through the cluster matching method. The specific adaptive meta-learning task construction module is used to learn the global representation of the seasonal components and trend components respectively, and divide the seasonal components and trend components into different tasks based on the global representation through the cluster matching method. Specifically, first, the seasonal components are respectively and trend components Input-independent standard variational autoencoder ( and ), used to learn and extract the underlying global representation and The feature extraction of the variational autoencoder helps to build a stable task embedding space. The variational autoencoder is optimized using the standard ELBO (Evidence Lower Bound) loss function. The optimization goal is to find the best approximate posterior distribution q(Z|X) by maximizing the ELBO.
[0034] Then, Kmeans clustering is used to obtain the global representation of each component. and , according to the two dimensions of seasonal components and trend components, the support set samples are divided into different clusters, each cluster corresponds to a task, expressed as: and ,in, Indicates the task index, and The number of tasks representing seasonal components and trend components respectively, and Respectively represent n The seasonal components and their labels corresponding to the tasks, represents the seasonal task set corresponding to the seasonal component dimension, and Respectively represent n The trend components and labels corresponding to each task, Represents the trend task set corresponding to the trend component dimension.
[0035] The sample of the support set is decomposed into This means that a sample in the query set has a seasonal component. possible task assignments, for the trend component there are possible task assignments.
[0036] In the embodiment, the query set samples are also matched with each type of task. Specifically, for each query set sample, the task and its specific task parameters corresponding to the query set sample are determined based on its global representation and the cluster center of each task cluster. 、 The process is: calculate the cosine similarity between each sample in the query set and each cluster center in the support set, and assign the sample in the query set to the category with the highest similarity. The specific formula is: ,in, and denote the category and global representation vectors used to cluster the mth sample in the query set, respectively. represents the cluster center of each category in the support set, Represents the cosine similarity calculation function.
[0037] In the embodiment, the probabilistic reasoning module is used to use the support set to learn the specific task parameters of different tasks, and use the specific task parameters to generalize the specific task features of the query set samples and infer the ICU time of the query set samples. Specifically, Figure 5 As shown: First, from the two dimensions of seasonal components and trend components, two neural network encoders are used Encode the seasonal components and trend components of all samples in the support set and query set respectively to obtain the potential representation of each component; Then, the input features of each task under each dimension are constructed based on the potential representation of each component of all samples in the support set. and Specifically, for each dimension, the potential representation of each sample in each task in the support set is concatenated with its label to obtain the input features of each sample, and then the input features of all samples in each task are concatenated to form the input features of each task. and .
[0038] Next, the dual-stream parameter generation network is based on the input features of two dimensions and Predict the specific task parameters for each type of task corresponding to the support set samples 、 Among them, the dual-stream parameter generation network uses two data stream branches, each of which uses a feedforward neural network and , and The network structure is a single linear layer, a pooling layer, and a single linear layer in sequence; the input features corresponding to each task in the two dimensions and Input into two data flow branches respectively, the posterior distribution of the specific task parameters is Gaussian distribution, and the mean and variance of the Gaussian distribution corresponding to each type of task are predicted based on the input features ( and )and( and ) constitutes the posterior distribution of each class of specific task parameters and , the posterior distribution of each task is used as a specific task parameter 、 .
[0039] Finally, based on the Bayesian probability formula, the conditional probability distribution of the query set sample labels is The formula is:
[0040] in, Indicates that the parameters in a specific task 、 The prior probability distribution of the query set sample labels under the condition.
[0041] Since the integral is not analytic, the present invention uses the specific task parameters and Perform Monte Carlo sampling by multiple sampling , get the approximate predicted ICU time :
[0042] in, and is the number of sampling times, is the probability density function, Indicates that, given a specific task parameter Next, for the query set sample and corresponding labels probability.
[0043] Specifically, after determining the task and specific task parameters corresponding to each query set sample based on the above matching, the probability distribution of the ICU duration of the query set samples in two dimensions is learned based on the specific task parameters, including: the potential representation of the corresponding component of each dimension of the query set sample and , matching the specific task parameters 、 After splicing, input into the neural network Fusion is performed to obtain specific task features for each dimension and ,Right now .
[0044] Finally, a decoder using a nonlinear neural network The ICU duration of the query set samples is fitted based on the specific task features of the two dimensions. ,Right now: .
[0045] S3, after constructing the loss function to optimize the parameters of the ICU duration prediction model, the ICU duration prediction model is used to predict the ICU duration of each sample to be tested.
[0046] In the embodiment, the model uses the loss of query sets in different tasks to train parameters through back propagation and ,The loss of the query set is calculated using Monte Carlo sampling to obtain the predicted log-likelihood.
[0047]
[0048] in, is the number of samples of query set task k, and Represents the sampling index, and the final loss function is ,express and Respectively and The sampling times index in the middle, K represents the total number of tasks.
[0049] The embodiment also implemented an ICU length of stay prediction performance evaluation module, testing it with patient samples that had never participated in training, and comparing the model's prediction results with the actual results. During the testing phase, a support set was randomly extracted from the training set, and the extracted support set and new samples were input into the trained model together. The specific task parameters on the support set captured by the model were used to predict the ICU length of stay of the new samples. After obtaining the prediction results for different new samples, various evaluation indicators can be used to objectively evaluate the model's prediction accuracy. Among them, the mean absolute error (MAE) and the root mean square error (RMSE) are two commonly used evaluation indicators used to measure the degree of difference between the model's predicted value and the actual observed value. These indicators can help us understand the overall prediction accuracy of the model, thereby providing a reference for further optimization and improvement.
[0050] After training, each sample to be tested is input into the ICU duration prediction model, and the ICU duration is predicted after being processed by the frequency domain decomposition module, the adaptive meta-learning task construction module, and the probabilistic reasoning module.
[0051] This method is used to predict ICU length of stay by regression using time-series clinical variables within the first 48 hours of a patient's ICU admission. This method effectively addresses the low precision and poor generalization of current methods, significantly improving the accuracy, generalization, and clinical adaptability of early ICU length of stay predictions. It has important clinical and engineering application value, specifically including: 1. Support bed scheduling and resource management. Accurately predicting ICU patient length of stay is essential for improving bed turnover, reducing congestion, and optimizing department transfers. The model outputs continuous predictions that can be directly used to support management decisions such as bed scheduling and nursing staff pre-allocation.
[0052] 2. Assisting with precise stratification and early intervention. This invention can identify patients at high risk of long-term stay at an early stage and, in conjunction with other models (such as mortality and multi-morbidity risk models), build a more comprehensive clinical decision support system, thereby improving the efficiency of medical intervention and prognosis.
[0053] 3. Rapid adaptation and small sample learning capabilities. By modeling the duration prediction for each new patient as an independent task through meta-learning and rapidly transferring learning capabilities from historical tasks, the model can still make effective predictions quickly even with a small sample size and a variety of tasks.
[0054] 4. Applicable to multiple institutions and multiple disease settings, with strong portability. Because the present invention uses clustering to automatically construct task clusters, the model does not rely on hospital type, department distribution, or manual labeling, and has cross-institutional portability.
[0055] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for predicting ICU duration based on adaptive meta-learning probabilistic reasoning, characterized by: The following steps are involved: Collect clinical variable sequence data of ICU patients and future ICU duration as sample data and corresponding labels, and divide them into support set and query set; Build an ICU length of stay prediction model that includes a frequency domain decomposition module, an adaptive meta-learning task construction module, and a probabilistic reasoning module. The frequency domain decomposition module is used to decompose each clinical variable sequence into seasonal components and trend components. The adaptive meta-learning task building module is used to learn the global representation of seasonal components and trend components respectively, and then divide the seasonal components and trend components into different tasks through cluster matching method based on the global representation; The probabilistic reasoning module is used to learn the specific task parameters of different tasks using the support set, and to generalize the specific task features of the query set samples using the specific task parameters and infer the ICU duration of the query set samples; After constructing the loss function to optimize the parameters of the ICU duration prediction model, the ICU duration prediction model is used to predict the ICU duration of each sample to be tested.
2. The ICU duration prediction method based on adaptive meta-learning probabilistic reasoning according to claim 1 is characterized in that The clinical variables included diastolic blood pressure, fraction of inspired oxygen, blood glucose, heart rate, mean blood pressure, blood oxygen saturation, respiratory rate, systolic blood pressure, body temperature and pH.
3. The ICU duration prediction method based on adaptive meta-learning probabilistic reasoning according to claim 1 is characterized in that In the frequency domain decomposition module, fast Fourier transform is used to decompose each input clinical time series into seasonal components and trend components to independently model the dynamic characteristics of the two components, which can be expressed as follows: ; in, stands for Fast Fourier Transform, is the inverse fast Fourier transform, is the threshold for dividing the frequency components, Indicates higher than The high frequency part of and After transformation, seasonal components are obtained , Indicates lower than The low-frequency part of and After transformation, trend components are obtained .
4. The ICU duration prediction method based on adaptive meta-learning probabilistic reasoning according to claim 1 or 3, characterized in that: It also includes denoising the seasonal components, and the denoised seasonal components are input into the adaptive meta-learning task construction module, where the denoising process includes filtering the Top-K frequencies corresponding to the amplitude to reduce the noise in the seasonal components.
5. The ICU duration prediction method based on adaptive meta-learning probabilistic reasoning according to claim 1 is characterized in that In the adaptive meta-learning task building module, a variational autoencoder is used to learn the global representation of seasonal components and trend components respectively; Kmeans clustering is used based on the global representation of each component. According to the two dimensions of seasonal component and trend component, the support set samples are divided into different clusters. Each cluster corresponds to a task, which is expressed as: and ,in, Indicates the task index, and The number of tasks representing seasonal components and trend components respectively, and Respectively represent n The seasonal components and their labels corresponding to the tasks, represents the seasonal task set corresponding to the seasonal component dimension, and Respectively represent n The trend components and labels corresponding to each task, Represents the trend task set corresponding to the trend component dimension.
6. The ICU duration prediction method based on adaptive meta-learning probabilistic reasoning according to claim 1 is characterized in that In the probabilistic reasoning module, two neural network encoders are used to encode the seasonal component and trend component of all samples in the support set and query set respectively from the two dimensions of seasonal component and trend component, to obtain the latent representation of each component. The input features of each task are constructed based on the latent representation of each component of all samples in the support set. A two-stream parameter generation network is used to predict the specific task parameters of each type of task corresponding to the support set samples based on the input features of two dimensions; For each query set sample, its global representation is matched with the cluster center of each task cluster to determine the matching task and its specific task parameters. The two-dimensional specific task features of the query set sample are learned based on the specific task parameters. The decoder is used to infer the two-dimensional specific task features and predict the ICU time of the query set sample.
7. The ICU duration prediction method based on adaptive meta-learning probabilistic reasoning according to claim 6 is characterized in that: The input features for each task are constructed based on the latent representation of each component of all samples in the support set, including: For each dimension, the potential representation of each sample in each task in the support set is concatenated with its label to obtain the input features of each sample, and then the input features of all samples in each task are concatenated to form the input features of each task.
8. The ICU duration prediction method based on adaptive meta-learning probabilistic reasoning according to claim 6 is characterized in that: A dual-stream parameter generation network is used to predict the specific task parameters of each type of task corresponding to the support set samples based on the input features of two dimensions, including: The dual-stream parameter generation network uses two data stream branches, each of which uses a feedforward neural network; The input features corresponding to each task in the two dimensions are respectively input into the two data flow branches to predict the mean and variance of the posterior distribution of the specific task parameters corresponding to each type of task corresponding to the support set samples, assuming that the posterior distribution of the specific task parameters is a Gaussian distribution.
9. The ICU duration prediction method based on adaptive meta-learning probabilistic reasoning according to claim 6 is characterized in that The probability distribution of ICU time in two dimensions of the query set samples is learned based on specific task parameters, including: The potential representation of the corresponding component of each dimension of the query set sample is spliced with the matching specific task parameters and then input into the neural network for fusion to obtain the probability distribution of ICU time for each dimension.
10. The ICU duration prediction method based on adaptive meta-learning probabilistic reasoning according to claim 1 is characterized in that The constructed loss functions include the ELBO loss function for the variational autoencoder, as well as the task loss of the query set in different tasks. The task loss uses Monte Carlo sampling of specific task parameters to obtain an approximate predicted ICU length of time for the query set samples through multiple sampling.