A method and system for predicting mortality risk in critically ill sub-acute patients based on a pace-net model

Abstract

This invention proposes a method and system for predicting the mortality risk of critically ill subacute patients based on the PACE-Net model, collecting longitudinal clinical data from ICU patients. Tensors are input into the PACE-Net model, which includes parallel branches A and B: branch A uses a two-dimensional convolutional architecture to capture periodic physiological rhythms and short-term spatiotemporal patterns; branch B uses a Transformer-based architecture to model long-range temporal dependencies and global contextual relationships. After the two branches are encoded and concatenated, the predicted in-hospital mortality risk is output via a fully connected neural network and a sigmoid classifier. This invention addresses the shortcomings of traditional scoring tools, such as reliance on static data and inability to capture dynamic temporal series and nonlinear relationships, by using a dual-branch architecture to process complete longitudinal data, overcoming the deficiencies of traditional models; it employs a physiologically based time decay interpolation strategy to handle missing values ​​and enhances interpretability through SHAP analysis.

Description

Technical Field

[0001] This invention relates to the field of mortality risk prediction technology for critically ill subacute patients, and in particular to a method and system for predicting mortality risk in critically ill subacute patients based on the PACE-Net model. Background Technology

[0002] Currently, mortality risk prediction for intensive care unit (ICU) patients primarily relies on scoring tools such as the Acute Physiology and Chronic Health Evaluation (APACHE), the Simplified Acute Physiology Score (SAPS), and the Sequential Organ Failure Assessment (SOFA). These scoring systems are mainly based on discrete, static data points and lack sufficient sensitivity in identifying early deterioration of the condition. Clinicians and IT engineers struggle to effectively integrate continuous laboratory indicators and clinical monitoring data, and cannot capture dynamic physiological and pathological time-series information, resulting in significant delays and limitations in real-time prediction. The emergence of artificial intelligence offers hope for alleviating this problem.

[0003] Traditional clinical risk assessment methods have significant limitations. While they still have some value in clinical applications, they are subjective and suffer from poor consistency among raters. Established scoring systems such as the Simplified Acute Physiology Score (SAPSII), Acute Physiology and Chronic Health Assessment II (APACHEII), Modified Early Warning Score (MEWS), and Quick Sequential Organ Failure Assessment (qSOFA) all rely on a limited set of physiological variables and fixed thresholds. These models are typically used as static admission assessment tools and cannot fully reflect the dynamic temporal evolution of disease and the complex nonlinear interactions between clinical variables. These inherent limitations hinder the development of accurate, personalized risk prediction and often lead to decreased predictive accuracy when applied to external patient populations, challenging the model's generalization ability. Furthermore, these models are often "black boxes," with low interpretability, making translation from experimental to clinical applications difficult. In recent years, artificial intelligence (AI) technology has shown significant potential in medical big data analysis, complex pattern recognition, and risk prediction. By integrating multi-dimensional data such as real-time vital signs, laboratory parameters, and medication records, comprehensive predictive analysis can be conducted. This technology holds promise for early warning and prediction of mortality risk in ICU patients with subacute conditions, and for indirectly assessing the effectiveness of treatment interventions. Summary of the Invention

[0004] This invention aims to at least address the technical problems existing in the prior art, and innovatively proposes a method and system for predicting the risk of death in critically ill subacute patients based on the PACE-Net model.

[0005] To achieve the above-mentioned objectives of this invention, this invention provides a method for predicting the mortality risk of critically ill subacute patients based on the PACE-Net model, the method comprising: S1. Collect longitudinal clinical data of patients in the intensive care unit, including vital signs data, laboratory test data and clinical monitoring data; S2. Preprocess the longitudinal clinical data to obtain a two-dimensional tensor; S3. Input the two-dimensional tensor into the PACE-Net model for mortality risk prediction. The PACE-Net model includes parallel branches A and B. Branch A uses a two-dimensional convolutional architecture to process the two-dimensional tensor, capturing periodic physiological rhythms and short-term spatiotemporal patterns through stacked convolutional layers, batch normalization, and nonlinear activation functions. Branch B adopts a Transformer-based architecture combined with position embedding, and utilizes a multi-head self-attention mechanism and residual feedforward layer to model long-range temporal dependencies and global contextual relationships. S4. The encoded representations of branch A and branch B are concatenated, input into a fully connected neural network, and after Dropout regularization, the in-hospital mortality risk prediction result is output through a sigmoid classifier.

[0006] As an optional embodiment of the present invention, the preprocessing in step S2 may include: for laboratory test data with a missing value rate of less than n%, a time decay interpolation strategy based on physiological principles and incorporating a Kalman filter is used to reconstruct missing values.

[0007] As an optional embodiment of the present invention, the time decay interpolation strategy based on physiological principles may include: identifying key time windows using a time position-aware coding method that considers absolute time position and sampling interval, and reconstructing missing values ​​by integrating physiological prior knowledge and temporal continuity using a Kalman filter.

[0008] As an optional embodiment of the present invention, the position embedding in branch B may employ a masked multi-head attention mechanism to model the directionality of disease progression.

[0009] As an optional embodiment of the present invention, the method may further include: performing interpretability analysis on the prediction results of the PACE-Net model using SHAP analysis to quantify the marginal contribution of each clinical feature to the prediction of mortality risk; wherein the features affecting mortality risk include respiratory rate, heart rate, non-invasive systolic blood pressure, blood oxygen saturation, non-invasive diastolic blood pressure, and non-invasive mean arterial pressure.

[0010] As an optional embodiment of the present invention, the PACE-Net model may be trained and validated based on a multi-center database, which includes the internationally public databases MIMIC-III, MIMIC-IV, eICU, and at least one private hospital dataset.

[0011] On the other hand, the present invention also provides a system for predicting the risk of death in critically ill subacute patients based on the PACE-Net model, the system comprising: processor; Memory used to store processor-executable instructions; The processor is configured to implement the system for predicting the risk of death in critically ill subacute patients based on the PACE-Net model when executing the executable instructions.

[0012] The beneficial effects of this invention are as follows: This invention addresses the shortcomings of existing traditional rating tools (such as APACHE, SAPS, SOFA, etc.), which rely on discrete static data points, cannot capture dynamic temporal evolution and nonlinear interactions between variables, suffer from poor inter-rater consistency, insufficient generalization ability, and low interpretability. Through a dual-branch parallel architecture, branch A explicitly models periodic physiological structures such as circadian rhythms using two-dimensional convolution, while branch B utilizes a Transformer combined with positional embedding and occlusion multi-head attention mechanisms to capture long-term temporal dependence and global contextual associations. Furthermore, it is trained based on complete longitudinal disease data from patient admission to discharge, rather than a truncated observation window, effectively overcoming the limitations of traditional models in representing temporal dynamics and modeling nonlinear relationships. Simultaneously, it employs a Kalman filter... The device employs a time-decay interpolation strategy based on physiological principles to handle irregular missing values, ensuring the clinical rationality of the data and the stability of downstream models. Furthermore, SHAP interpretability analysis quantifies the marginal contribution of each clinical feature to the prediction, transforming the model from a "black box" to a transparent and acceptable one, thus overcoming a key obstacle in the translation from experimental to clinical practice. Validation on multi-center data shows that the method achieves an internal validation AUC of 0.960 and a recall of 0.860, and an external validation AUC of 0.937. It also improves the diagnostic AUC of young physicians from 0.702 to 0.876, significantly narrowing the gap with experienced physicians. This enables early, continuous, dynamic, and interpretable prediction of mortality risk in ICU subacute patients, providing a reliable basis for clinical decision support and treatment plan adjustments.

[0013] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0014] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of a method for predicting the risk of death in critically ill subacute patients based on the PACE-Net model according to the present invention; Figure 2 This is a schematic diagram of the two-dimensional convolutional architecture of the present invention; Figure 3 This is a schematic diagram of the structure of the PACE-Net model of this invention; Figure 4 This is a schematic diagram of the SHAP analysis results of the present invention; Figure 5 This is a schematic diagram of the SHAP analysis results of the present invention; Figure 6 This is the SHAP analysis statistical chart of the present invention. Detailed Implementation

[0015] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0016] like Figure 1 As shown, a method for predicting the risk of death in critically ill subacute patients based on the PACE-Net model is described, the method comprising: S1. Collect longitudinal clinical data of patients in the intensive care unit, including vital signs data, laboratory test data and clinical monitoring data; In step S1, it's important to note that the collected longitudinal clinical data covers the entire course of the patient's illness, from ICU admission to discharge, ensuring data continuity and completeness. Data sources include international multi-center public databases (MIMIC-III, MIMIC-IV, eICU) and real clinical data from top-tier hospitals in China, taking into account the heterogeneity of patient populations across different regions and healthcare systems, thus laying the foundation for the model's generalization ability. Vital signs data include key indicators such as respiratory rate, heart rate, non-invasive systolic blood pressure, non-invasive diastolic blood pressure, non-invasive mean arterial pressure, and blood oxygen saturation. Laboratory test data includes results from complete blood counts, biochemical indicators, and coagulation function. Clinical monitoring data includes ventilator usage parameters, vasoactive drug infusion status, and fluid balance. To address the issue of irregular missing values ​​in the data, laboratory test data with a missing value rate below 20% will be processed using a physiologically based time-decay interpolation strategy incorporating a Kalman filter to ensure the clinical validity and temporal continuity of the data, providing high-quality input data for subsequent model training.

[0017] S2. Preprocess the longitudinal clinical data to obtain a two-dimensional tensor. Preprocessing includes: for laboratory test data with a missing value rate below n%, a time-decay interpolation strategy based on physiological principles and incorporating a Kalman filter is used to reconstruct missing values. The time-decay interpolation strategy based on physiological principles includes: identifying key time windows using a time-position-aware coding method that considers absolute time position and sampling interval, and using a Kalman filter to integrate physiological prior knowledge and temporal continuity to reconstruct missing values.

[0018] In step S2, it should be noted that n is set to 20, meaning this imputation strategy is applied to laboratory test data with a missing rate of less than 20%. Time-location-aware encoding extracts key time nodes in the patient's disease progression (such as ICU admission time, clinical intervention implementation time, and turning points in disease deterioration or remission), combining them with the time interval weights of adjacent sampling points to accurately locate time windows that significantly impact prognosis. The Kalman filter constructs a state transition equation based on the intrinsic correlation between physiological parameters (such as the compensatory relationship between heart rate and blood pressure, and the synergistic changes in blood oxygen saturation and respiratory rate), integrating physiological prior knowledge into the dynamic estimation process of missing values. This ensures that the reconstructed data conforms to temporal continuity while adhering to clinical physiological logic. After imputation, all clinical feature sequences are resampled at a uniform time step (e.g., one data point every 30 minutes) and reshaped into a 224×F two-dimensional tensor. The 224 dimensions correspond to the integration of circadian rhythms and multi-day periodic time divisions, with F representing the number of clinical features. This provides structured input for the subsequent two-dimensional convolution module to extract periodic physiological patterns. In addition, preprocessing also includes standardizing the feature values ​​(such as using Z-score normalization) to eliminate the impact of differences in the units of different indicators on model training, thereby further improving data quality and model convergence efficiency.

[0019] S3. Input the two-dimensional tensor into the PACE-Net model for mortality risk prediction. The PACE-Net model includes parallel branches A and B. Branch A uses a two-dimensional convolutional architecture to process the two-dimensional tensor, capturing periodic physiological rhythms and short-term spatiotemporal patterns through stacked convolutional layers, batch normalization, and nonlinear activation functions. Branch B adopts a Transformer-based architecture combined with position embedding, and utilizes a multi-head self-attention mechanism and residual feedforward layer to model long-term temporal dependencies and global contextual associations; the position embedding in branch B adopts an occlusion multi-head attention mechanism to model the directionality of disease progression.

[0020] like Figure 2 and 3As shown, in step S3, it is necessary to explain in detail that the two-dimensional convolutional module of branch A consists of three stacked convolutional layers. The first convolutional layer uses a 3×3 convolutional kernel with 64 output channels, and extracts local periodic features through batch normalization and ReLU activation function. The second convolutional layer uses a 5×5 convolutional kernel with 128 output channels, further capturing the spatiotemporal correlation across time windows. The third convolutional layer uses a 1×1 convolutional kernel for feature dimensionality reduction, with an output dimension of 224×F' (F' is the number of features after dimensionality reduction). The Transformer architecture of branch B contains four multi-head self-attention layers, each with eight attention heads and a hidden layer dimension of 256. By using a masking multi-head attention mechanism, the model is restricted to focusing only on the current and previous time step data, avoiding the leakage of future information, thereby accurately modeling the unidirectional temporal characteristics of disease progression. Each self-attention layer is equipped with residual connections and layer normalization, and then the features are nonlinearly transformed through a feedforward neural network (FFN). The output features from the two branches are concatenated and input into a fusion module containing two fully connected layers. The first fully connected layer maps the concatenated features to 128 dimensions, and the second layer outputs a single risk probability value (range 0-1), representing the patient's current risk of death. During model training, a binary cross-entropy loss function is used, combined with the Adam optimizer for parameter updates. An early stopping strategy is also introduced to prevent overfitting and ensure the model's generalization ability on clinical data.

[0021] S4. The encoded representations of branch A and branch B are concatenated, input into a fully connected neural network, and after Dropout regularization, the in-hospital mortality risk prediction result is output through a sigmoid classifier.

[0022] In step S4, it is necessary to explain in detail that the two-dimensional feature map of branch A is first compressed into a one-dimensional vector (dimension F') through global average pooling, while the Transformer output sequence of branch B extracts global features (dimension 256) through mean pooling. After dimensional alignment, the two are concatenated into a fused feature vector of length F'+256. This fused vector is input to the fully connected neural network module, where the first fully connected layer has 128 neurons and uses the ReLU activation function for nonlinear transformation. To alleviate overfitting, Dropout regularization (probability set to 0.3) is introduced after the first fully connected layer to randomly discard some neurons to enhance the model's generalization ability. The output of the second fully connected layer is a single neuron, and the output value is mapped to the interval between 0 and 1 through the sigmoid activation function to obtain the patient's in-hospital mortality risk probability. In clinical applications, a risk probability threshold (e.g., 0.5) can be set. When the predicted value exceeds the threshold, it is considered a high risk of death, triggering a clinical warning. Simultaneously, combined with SHAP analysis results, the model can output key influencing features and their contribution, helping medical staff quickly identify the core driving factors of changes in the patient's condition and providing data support for developing personalized intervention plans. The risk probability output by the model will be dynamically adjusted according to the real-time updates of the patient's disease progress data, ensuring the timeliness and accuracy of risk assessment, and truly achieving continuous monitoring and early warning of the mortality risk of subacute critically ill patients.

[0023] As an optional embodiment of the present invention, the method may further include: performing interpretability analysis on the prediction results of the PACE-Net model using SHAP analysis to quantify the marginal contribution of each clinical feature to the prediction of mortality risk; wherein the features affecting mortality risk include respiratory rate, heart rate, non-invasive systolic blood pressure, blood oxygen saturation, non-invasive diastolic blood pressure, and non-invasive mean arterial pressure.

[0024] like Figure 4 , 5As shown in Figure 6, regarding SHAP analysis, it is important to explain in detail that this invention uses TreeExplainer as the SHAP interpreter because it has good compatibility with tree models and tree-based ensemble models, and is also compatible with the fully connected layer structure of the PACE-Net model fusion module. PACE-Net is a deep learning model that uses two-dimensional convolutional layers to periodically model physiological rhythms and utilizes the model architecture to highlight clinically relevant signals. Through specific data preprocessing methods, this model exhibits strong predictive performance. In internal validation, its area under the receiver operating characteristic (AUC) was 0.960 and its recall was 0.860. Subsequently, external validation confirmed its robustness, achieving an AUC of 0.937 in a multicenter study. When used as a clinical decision support system, PACE-Net improved the diagnostic accuracy of young physicians, increasing their AUC from 0.702 to 0.876, and significantly narrowing the performance gap between them and senior clinicians.

[0025] PACE-Net enables continuous and dynamic prediction of mortality risk, although its performance metrics gradually decline over time (e.g., precision decreases from 0.95 to 0.89, and recall from 0.86 to 0.74). This trend reflects the inherent difficulty of long-term prediction for critically ill patients. Notably, the model maintains early sensitivity in identifying high-risk individuals. In routine clinical practice, critically ill patients are typically reassessed every 24 hours, with additional assessments triggered only in cases of significant deterioration; PACE-Net, however, provides real-time, continuously updated risk assessments.

[0026] This model dynamically integrates complex longitudinal physiological data to generate interpretable risk scores for clinical decision support, whether used in routine ICU rounds or as a continuous monitoring tool. The system supports retrospective analysis of a patient's dynamic risk trajectory over the past 12 to 24 hours. When specific characteristics identified by the SHAP model (such as increased respiratory rate or decreased systolic blood pressure) lead to an increased risk probability, it prompts healthcare professionals to initiate a comprehensive assessment earlier, perform additional diagnostic tests (such as arterial blood gas analysis and chest X-ray), or consider enhanced treatment measures (such as using vasopressors or adjusting ventilator parameters). This invention overcomes the limitations of univariate analysis by utilizing artificial intelligence, developing a multivariate assessment system for accurately predicting the mortality risk of subacute intensive care unit patients. The six key characteristics identified by SHAP analysis are highly consistent with previous clinical studies, confirming the model's clinical effectiveness and demonstrating continuous prediction and early warning effects earlier and more intuitively than traditional static assessments, providing reference for adjusting treatment plans.

[0027] As an optional embodiment of the present invention, the PACE-Net model may be trained and validated based on a multi-center database, which includes the internationally public databases MIMIC-III, MIMIC-IV, eICU, and at least one private hospital dataset.

[0028] This invention utilizes data from internationally available databases MIMIC-III, MIMIC-IV, eICU, and data from the First Affiliated Hospital of Chongqing Medical University and Chongqing Traditional Chinese Medicine Hospital for training and validation, developing a novel artificial intelligence model called PACE-Net for predicting mortality rates in ICU patients, particularly those in the subacute phase. To overcome the limitations of traditional machine learning models, such as reliance on manual feature engineering, insufficient representation of temporal dynamics, and limited ability to model nonlinear relationships, this invention proposes PACE-Net (Periodic Attention Convolutional Encoder Network), a deep learning framework specifically designed for analyzing high-resolution time-series data from intensive care units. Unlike traditional methods that treat physiological signals as independent or uniformly sampled time series, PACE-Net is specifically designed to model two fundamental features of vital sign trajectories: periodic physiological rhythms (such as circadian rhythms) and the significance of feature-specific characteristics across multiple clinical variables. The PACE-Net architecture consists of two parallel and complementary encoding paths, each designed to capture different aspects of the input temporal dynamics.

[0029] Branch A employs a two-dimensional convolutional architecture to process the reconstructed input sequence, thereby capturing periodic and short-term spatiotemporal patterns. This architecture utilizes stacked convolutional layers to identify periodic physiological structures, such as circadian rhythms and other periodic trends, which often predict changes in the patient's physiological state.

[0030] Branch B employs a Transformer-based architecture combined with location embedding to model long-term temporal dependencies and global contextual relationships in patient clinical trajectories. This mechanism utilizes multi-head self-attention and residual feedforward layers to uncover important associations between different time points, thereby helping to elucidate key prognostic biomarkers and longitudinal trends indicating progressive decline in physiological function or the cumulative effects of clinical interventions.

[0031] The encoded representations of the two branches are then concatenated and fed into a fully connected neural network. This two-branch architecture can simultaneously leverage periodic regularity and feature-level distinguishability, thus providing a more physiologically plausible expression for downstream predictive tasks such as early detection of clinical deterioration.

[0032] The serialized medical data is organized into a T×F matrix, where T represents the time dimension and F represents the number of clinical features. In this invention, T is set to 48 discrete time steps, corresponding to a 12-hour observation window. The input sequence is reconstructed into a two-dimensional tensor of size 224×F to explicitly model circadian rhythms and other periodic dynamics, as shown in the figure. This structural transformation enables the extraction of periodic physiological rhythms and feature-level clinical significance from multivariate ICU time-series data within a unified framework using a convolutional encoder.

[0033] The reconstructed tensor (224F) is processed through a 2D convolutional module to model periodic physiological rhythms. This architecture learns structured spatiotemporal correlations, particularly those with periodicity over the 24-step time dimension. Representation learning is stabilized and optimized through batch normalization and non-linear activation functions. The feature maps are then flattened to obtain latent representations. PACE-Net can detect recurring physiological patterns, aberrant cycles, and deviations from standard rhythms, all important indicators of impending patient deterioration. The latent representations from the two branches are concatenated and then fed into a fully connected fusion layer. Dropout regularization is then applied, and finally, a sigmoid classifier is used to predict in-hospital mortality risk. This combined representation allows PACE-Net to simultaneously utilize periodic patterns discovered by the convolutional branch and physiologically relevant features learned by the attention branch.

[0034] By combining time-sensitive modeling with a Transformer architecture and utilizing longitudinal clinical data from admission to discharge, this model enables continuous mortality risk assessment. Unlike previous studies that were limited to short observation windows (e.g., 8-hour data), this invention employs complete patient trajectory data, allowing for more comprehensive longitudinal analysis. Training and validation at heterogeneous centers also improves the model's robustness and generalization ability in real-world clinical settings. SHAP analysis is used to elucidate the importance of key features. The study found that the most important features influencing mortality include respiratory rate, heart rate, non-invasive systolic blood pressure, oxygen saturation, non-invasive diastolic blood pressure, and non-invasive mean arterial pressure. SHAP values, by quantifying the marginal contribution of each clinical feature to prediction, comprehensively reveal the internal decision-making process of this deep learning model in predicting mortality risk in subacute intensive care unit patients, improving model interpretability, accelerating clinical application translation, and making the model's internal mechanisms more transparent and acceptable. This invention evaluates the model in a clinical setting and uses the DeLong test to statistically compare it with physician assessments. The model shows a statistically significant difference in improving the diagnostic capabilities of junior physicians.

[0035] This invention contributes to the field by integrating and expanding upon several key research directions. Based on a solid empirical foundation, it proposes a novel deep learning architecture called PACE-Net, designed to leverage the temporal and periodic structures in intensive care unit (ICU) data. The main innovation of PACE-Net lies in its use of two-dimensional convolutional layers to explicitly model the periodicity in physiological rhythms, combined with modules for dynamic weighting of clinically relevant signals. This architecture represents a significant advancement compared to traditional one-dimensional convolutional neural networks (1D-CNN) or long short-term memory (LSTM) backbone networks, which are less effective at capturing intra-cycle dynamics. Unlike previous methods that often truncate the observation window (e.g., only the first 48 hours), PACE-Net is trained and validated based on complete patient course data from admission to discharge, thus achieving truly longitudinal and comprehensive risk assessment.

[0036] A Physiologically Based Time-Decrease Interpolation Strategy: While vital signs of ICU patients are typically recorded at fixed intervals, laboratory test data is only obtained upon request by doctors. This results in a large number of irregular missing values, posing challenges for predictive modeling. To address this issue for laboratory test data with a missing value rate below 20%, the team employed a physiologically based time-decrease interpolation strategy incorporating a Kalman filter. This method integrates physiological prior knowledge and temporal continuity to reconstruct missing values ​​in a clinically plausible manner, thereby reducing noise and avoiding non-physiological fluctuations introduced by traditional interpolation methods. Consequently, the resulting time-series data better preserves the patient's intrinsic dynamic changes, thus improving the stability and predictive performance of downstream deep learning models.

[0037] A Progressive Evolutionary Framework Based on Location Encoding and Masked Multi-Head Attention: A progressive evolutionary framework based on location encoding and masked multi-head attention is proposed. First, the team employs a time-location-aware encoding method that considers absolute temporal location and sampling intervals to identify critical time windows. Subsequently, the team utilizes masked multi-head attention mechanisms to model the directionality of disease progression. PACE-Net was developed and internally validated on a large, multi-center public database, followed by external evaluation on independent private datasets. By combining time-sensitive modeling with a Transformer model and utilizing longitudinal clinical data from admission to discharge, the model enables continuous mortality risk assessment. Compared to previous studies limited to short observation windows (e.g., 8-hour data extraction), this invention utilizes complete patient trajectory data, enabling more comprehensive longitudinal assessments. Training and validation at heterogeneous centers further enhance the robustness and universality of the model in real-world clinical settings.

[0038] Multi-model prediction comparison and SHAP interpretability analysis: This invention evaluated ten prediction models, including logistic regression, XGBoost, FCN, LSTM-FCN, InceptionTime, ResCNN, XceptionTime, ROCKET, PACE-Net (CA only), and the full PACE-Net model. All models were trained on the training dataset and tested on internal and external validation sets. Performance metrics included area under the receiver operating characteristic curve (AUC), accuracy (ACC), precision, recall, and F1 score. Results from the external validation cohort confirmed the robustness and accuracy of the models. To elucidate the underlying mechanisms, SHAP (SHapley Additive ex Planations) analysis was used for visualization. This analysis revealed that six key features in the PACE-Net model's predictions were the most important determinants: respiratory rate, heart rate, non-invasive systolic blood pressure, pulse oxygen saturation, non-invasive diastolic blood pressure, and non-invasive mean arterial pressure. As an interpretive method for deep learning models, SHAP helps to understand the prediction mechanisms and quantifies the contribution of individual features to the model's output.

[0039] Example 2 A system for predicting the risk of death in critically ill subacute patients based on the PACE-Net model includes: a processor and a memory for storing processor-executable instructions; wherein the processor is configured to implement the PACE-Net model for predicting the risk of death in critically ill subacute patients when executing the executable instructions. It should be noted that the system includes: a processor, memory, and the computer device may further include one or more of multimedia components, input / output (I / O) interfaces, and communication components.

[0040] The processor controls the overall operation of the computer device to complete all or part of the steps in the method for predicting the risk of death in critically ill subacute patients based on the PACE-Net model.

[0041] Memory is used to store various types of data to support the operation of the computer device. This data may include, for example, instructions for any application or method used to operate on the computer device, as well as application-related data. Memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0042] The multimedia component may include a screen and an audio component, wherein the screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals; for example, the audio component may include a microphone for receiving external audio signals, the received audio signals may be further stored in memory or transmitted via a communication component; the audio component may also include at least one speaker for outputting audio signals.

[0043] I / O interfaces provide interfaces between the processor and other interface modules, such as keyboards, mice, buttons, etc.; these buttons can be virtual buttons or physical buttons.

[0044] The communication component is used for wired or wireless communication between the computer device and other devices; wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G or 5G, or one or more combinations thereof, and the corresponding communication component may include: Wi-Fi module, Bluetooth module, NFC module, mobile communication module.

[0045] As a preferred embodiment, the computer device may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described method for predicting the mortality risk of critically ill subacute patients based on the PACE-Net model.

[0046] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for predicting the risk of death in critically ill subacute patients based on the PACE-Net model, characterized in that, The method includes: S1. Collect longitudinal clinical data of patients in the intensive care unit, including vital signs data, laboratory test data and clinical monitoring data; S2. Preprocess the longitudinal clinical data to obtain a two-dimensional tensor; S3. Input the two-dimensional tensor into the PACE-Net model for mortality risk prediction. The PACE-Net model includes parallel branches A and B. Branch A uses a two-dimensional convolutional architecture to process the two-dimensional tensor, capturing periodic physiological rhythms and short-term spatiotemporal patterns through stacked convolutional layers, batch normalization, and nonlinear activation functions. Branch B adopts a Transformer-based architecture combined with position embedding, and utilizes a multi-head self-attention mechanism and residual feedforward layer to model long-range temporal dependencies and global contextual relationships. S4. The encoded representations of branch A and branch B are concatenated, input into a fully connected neural network, and after Dropout regularization, the in-hospital mortality risk prediction result is output through a sigmoid classifier.

2. The method for predicting the risk of death in critically ill subacute patients based on the PACE-Net model as described in claim 1, characterized in that, The preprocessing in step S2 includes: for laboratory test data with a missing value rate of less than n%, a time decay interpolation strategy based on physiological principles and incorporating a Kalman filter is used to reconstruct missing values.

3. The method for predicting the risk of death in critically ill subacute patients based on the PACE-Net model as described in claim 2, characterized in that, The time decay interpolation strategy based on physiological principles includes: using a time position-aware coding method that considers absolute time position and sampling interval to identify key time windows, and using a Kalman filter to integrate physiological prior knowledge and temporal continuity to reconstruct missing values.

4. The method for predicting the risk of death in critically ill subacute patients based on the PACE-Net model as described in claim 1, characterized in that, The location embedding in branch B employs a masked multi-head attention mechanism to model the directionality of disease progression.

5. The method for predicting the risk of death in critically ill subacute patients based on the PACE-Net model as described in claim 1, characterized in that, The method further includes: using SHAP analysis to perform interpretability analysis on the prediction results of the PACE-Net model, and quantifying the marginal contribution of each clinical feature to the prediction of mortality risk; wherein, the features affecting mortality risk include respiratory rate, heart rate, non-invasive systolic blood pressure, blood oxygen saturation, non-invasive diastolic blood pressure and non-invasive mean arterial pressure.

6. The method for predicting the risk of death in critically ill subacute patients based on the PACE-Net model as described in claim 1, characterized in that, The PACE-Net model is trained and validated based on a multi-center database, which includes the internationally public databases MIMIC-III, MIMIC-IV, eICU, and at least one private hospital dataset.

7. A system for predicting the risk of death in critically ill subacute patients based on the PACE-Net model, characterized in that, The system includes: processor; Memory used to store processor-executable instructions; The processor is configured to implement, when executing the executable instructions, the system for predicting the risk of death in critically ill subacute patients based on the PACE-Net model as described in any one of claims 1 to 6.

Images