Critical elderly patient death risk prediction method and device based on dual-view heterogeneous fusion and storage medium thereof

Abstract

The application provides a critical and elderly patient death risk prediction method and device based on dual-view heterogeneous fusion. A static data encoder is used to process static data of a subject, capture baseline information of the subject, and obtain static data features. A time series data encoder is used to process time series data of the subject, model dynamic trend visual perception and explicit statistical relationship of the time series data, and obtain image view features and statistical view features. An adaptive gating fusion module is used to perform heterogeneous feature fusion on the image view features and the statistical view features, and time series data features are obtained. A static feature encoding and fusion module is used to encode the static data features by using a multilayer perceptron to obtain static encoded features. The static encoded features and the time series data features are input into a cross-modal bidirectional cross-attention layer, a multi-head attention mechanism is used to perform double-feature interaction fusion and splicing, and fusion features are obtained. The fusion features are input into a prediction module, and a death risk probability is output by the prediction module.

Description

Technical Field

[0001] This application relates to the application of artificial intelligence in clinical medicine, and in particular to a method, device and storage medium for predicting the mortality risk of critically ill elderly patients based on dual-perspective heterogeneous fusion. Background Technology

[0002] In recent years, artificial intelligence technology has developed rapidly, and machine learning and deep learning have been widely used in the field of clinical medicine. Although traditional critical care scoring systems (such as APACHE, SOFA, SAPS, etc.) have played an important role in clinical practice, their linear or logistic regression models based on finite variables are difficult to fully explore the nonlinear relationships contained in high-dimensional data.

[0003] In machine learning-based prognostic assessment studies, Dong Wang et al. used a random forest model to predict mortality rates in ICU patients with sepsis, demonstrating high accuracy and reliability. Xiujuan Zhao et al. used a random forest model to predict functional recovery and short-term prognosis in elderly stroke patients, showing significantly better performance than the traditional logistic regression model. Li et al. used an XGBoost model to predict in-hospital mortality rates in patients with heart failure, achieving excellent results, and enhanced the model's interpretability through SHAP analysis. Xiaoli Liu et al. constructed a disease severity assessment model for elderly critically ill patients based on the XGBoost model, achieving good predictive performance.

[0004] However, the above studies share a common limitation: most focus on the first measurement or a single static indicator within 24 hours of the patient's admission to the ICU for modeling. This assessment method based on a single data collection point has the following shortcomings: (1) it cannot reflect the dynamic trend of the patient's condition over time, ignoring the rich information on the changes in the condition contained in the time series data; (2) the measurement value at a single time point is easily affected by instantaneous interference (such as stress response, drug effects, etc.), resulting in insufficient measurement stability; (3) it cannot capture the temporal dependence and synergistic change patterns between different physiological indicators. In fact, the vital signs continuously monitored in the ICU and the results of multiple laboratory tests constitute a wealth of high-dimensional time series data, and the dynamic change information contained therein (such as the fluctuation trend of blood pressure, the evolution trajectory of oxygenation index, etc.) is of great value for prognostic assessment, but existing studies have not explored this aspect sufficiently.

[0005] Furthermore, current research on machine learning-based risk stratification specifically for elderly patients (≥80 years old) is relatively limited. The pathophysiological characteristics, pharmacokinetic features, and prognostic factors of elderly patients differ significantly from those of general adult patients, and directly applying risk assessment models to the general population may lead to predictive bias. Therefore, there is an urgent need to develop a mortality risk prediction model that can effectively utilize multidimensional time-series data, capture dynamic information about disease progression, and is specifically designed for elderly critically ill patients. Summary of the Invention

[0006] In view of the above problems, this application aims to propose a method for predicting the mortality risk of critically ill elderly patients based on dual-perspective heterogeneous fusion.

[0007] The method for predicting mortality risk in critically ill elderly patients based on dual-perspective heterogeneous fusion in this application includes: Static data encoders are used to process the static data of subjects, capture the baseline information of subjects, and obtain static data features; The temporal data encoder is used to process the subject's temporal data, model the dynamic trend visual perception and explicit statistical relationship of the temporal data, and obtain image perspective features and statistical perspective features. The image perspective features and statistical perspective features are fused heterogeneously through an adaptive gating fusion module to obtain temporal data features. In the static feature encoding and fusion module, static data features are encoded using a multilayer perceptron to obtain static encoded features; the static encoded features and time-series data features are input into a bidirectional cross-attention layer, and the dual features are interactively fused and concatenated through a multi-head attention mechanism to obtain fused features; The fused features are input into the prediction module, which outputs the probability of death risk. The prediction module consists of a fully connected layer and an output layer. The static data of the subjects included: age, gender, BMI, type of hospital admission, and length of hospital stay before ICU. The subjects' time-series data included: heart rate, systolic blood pressure, mean arterial pressure, respiratory rate, body temperature, blood oxygen saturation, central venous pressure, albumin, anion gap, bicarbonate, urea nitrogen, calcium, chloride, creatinine, glucose, sodium, potassium, fibrinogen, international normalized ratio, prothrombin time, partial thromboplastin time, hematocrit, hemoglobin, platelets, white blood cell count, alanine aminotransferase, aspartate aminotransferase, bilirubin, arterial blood oxygen partial pressure, arterial blood carbon dioxide partial pressure, inhaled oxygen concentration, oxygenation index, pH, base excess, lactate, troponin, magnesium, brain natriuretic peptide, lymphocytes, neutrophils, alkaline phosphatase, vasoactive drugs, and ventilation support. These data were collected multiple times within a 24-hour time window to form the time series.

[0008] Correspondingly, this application also proposes a mortality risk prediction device for critically ill elderly patients based on dual-view heterogeneous fusion, which includes: a static data encoder, a time-series data encoder, a static feature encoding and fusion module, and a prediction module. A static data encoder is used to process the static data of the subjects to capture the subjects' baseline information and obtain static data features; The temporal data encoder is used to process the subject's temporal data to model the dynamic trends and explicit statistical relationships of the temporal data and obtain image perspective features and statistical perspective features. The temporal data encoder includes an adaptive gating fusion module, which performs heterogeneous feature fusion of image perspective features and statistical perspective features to obtain temporal data features. The static feature encoding and fusion module includes a multilayer perceptron and a bidirectional cross-attention layer. Static data features are encoded by the multilayer perceptron to obtain static encoded features. The static encoded features and time-series data features are input into the bidirectional cross-attention layer, and the dual features are interacted, fused, and concatenated through a multi-head attention mechanism to obtain fused features. The prediction module consists of a fully connected layer and an output layer. After the fused features are input into the prediction module, the prediction module outputs the probability of death risk. The static data of the subjects included: age, gender, BMI, type of hospital admission, and length of hospital stay before ICU. The subjects' time-series data included: heart rate, systolic blood pressure, mean arterial pressure, respiratory rate, body temperature, blood oxygen saturation, central venous pressure, albumin, anion gap, bicarbonate, urea nitrogen, calcium, chloride, creatinine, glucose, sodium, potassium, fibrinogen, international normalized ratio, prothrombin time, partial thromboplastin time, hematocrit, hemoglobin, platelets, white blood cell count, alanine aminotransferase, aspartate aminotransferase, bilirubin, arterial blood oxygen partial pressure, arterial blood carbon dioxide partial pressure, inhaled oxygen concentration, oxygenation index, pH, base excess, lactate, troponin, magnesium, brain natriuretic peptide, lymphocytes, neutrophils, alkaline phosphatase, vasoactive drugs, and ventilation support. These data were collected multiple times within a 24-hour time window to form the time series.

[0009] Preferably, the static data encoder includes two fully connected layers.

[0010] Preferably, the time-series data encoder includes an image viewpoint branch and a statistical feature viewpoint branch; In the image perspective branch, the time series data is converted into a multi-channel GAF image using Gram angle field, and the dynamic trend of the time series data is modeled through a multi-scale, multi-channel Vision Transformer network. In the statistical feature perspective branch, the statistical features of the time series data are extracted: mean, standard deviation, maximum, minimum, median, skewness, and kurtosis, and machine learning algorithms are used to model explicit statistical relationships.

[0011] Preferably, the mathematical definition of the adaptive gating fusion module is as follows: Given the image viewpoint feature F_img and the statistical viewpoint feature F_stat, firstly, the statistical viewpoint feature F_stat is broadcast-expanded to form the expanded statistical viewpoint feature F_stat_expanded, and then the gate signal G is calculated: G = σ(W_g · [F_img ∥ F_stat_expanded] + b_g); Where σ is the Sigmoid activation function, W_g and b_g are learnable parameters, and [·∥·] represents the vector concatenation operation; each element of the gate signal G controls the fusion weight of the corresponding feature dimension; The fusion feature F_fused is: F_fused = G ⊙ F_img + (1 - G) ⊙ F_stat_expanded; Here, ⊙ represents element-wise multiplication.

[0012] Preferably, the machine learning algorithm is XGBoost, Random Forest, Logistic Regression, or Support Vector Machine.

[0013] This application also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the above-described method for predicting the mortality risk of critically ill elderly patients based on dual-view heterogeneous fusion.

[0014] This application develops a dual-branch coding neural network architecture for dynamic and static data. Its core components include: a static encoder, which uses two fully connected layers to process static features and capture the patient's baseline characteristics; and a temporal encoder, which employs a dual-view heterogeneous fusion temporal coding network to innovatively jointly model multidimensional temporal data from two dimensions: an "image perspective" and an "explicit statistical feature perspective." The image perspective focuses on the dynamic trend changes of the temporal data, while the statistical perspective focuses on the explicit statistical relationships of the temporal data, forming a complementary relationship. In the image perspective branch, temporal data is converted into multi-channel images using Gram angle field (GAF), and a multi-scale Vision Transformer is introduced for GAF image feature extraction: the self-attention mechanism of the Vision Transformer is used to capture global dependencies in GAF images, and temporal patterns are extracted at different scales through a multi-scale patch embedding module, effectively overcoming the limitations of the local receptive field of traditional CNNs; in the statistical feature perspective branch, seven types of statistical features are extracted from the temporal data, and four machine learning algorithms are used to model explicit statistical relationships; in the fusion stage, an adaptive gating fusion mechanism is designed to fuse static features, and the contribution of image features and statistical features is dynamically weighed through a learnable gating network.

[0015] Finally, static and temporal data features are input into a bidirectional cross-attention layer: through a multi-head attention mechanism, static and temporal features query and compute each other, performing dynamic and static data fusion representation learning. Finally, the fused features are input into the prediction layer to output the probability of death risk.

[0016] Dynamic assessment of patient condition based on multi-dimensional time-series data: By utilizing complete time-series observation data of patients within 24 hours of admission to the ICU, the dynamic evolution trajectory of various physiological indicators is modeled, breaking through the limitations of traditional single-measurement methods.

[0017] The present invention has the following beneficial effects: (1) Significantly improved prediction accuracy: In the test of 15,517 patients in the MIMIC database, the dual-view heterogeneous fusion model achieved an AUROC of 0.730, which is 2% higher than the baseline GAF+CNN (AUROC=0.710).

[0018] (2) Fully capture information on the dynamic evolution of the disease: Using complete 24-hour time series observation data, GAF+ multi-scale ViT is used to model the time series dynamic trend, breaking through the limitation of traditional methods that rely only on single measurement values.

[0019] (3) Comprehensive integration of multi-source heterogeneous data: Simultaneously utilize dynamic time series data (43 indicators and 24-hour evolution trajectory) and static clinical characteristics (age, gender, BMI, admission type, etc.). Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method and device for predicting mortality risk in critically ill elderly patients based on dual-view heterogeneous fusion according to the present invention. Figure 2 Flowchart for screening research subjects; Figure 3 Here is an example of an image after Gram's angle field transformation; Figure 4 Differences in the distribution of age, BMI, CCI score, APSIII, OASIS, and SAPS in the death and non-death groups; Figure 5 The results of the contribution analysis of the top 20 features based on SHAP; Figure 6 Interpretation of SHAP individuals (predicted as positive); Figure 7 Explanation of SHAP individuals (predicted as negative). Detailed Implementation

[0021] I. Research Methods 1.1 Research Design This study, based on the MIMIC database, included critically ill patients aged 80 years and older who had spent at least 24 hours in the ICU. Demographic indicators, laboratory tests, and other physiological and biochemical indicators (dynamic time-series data) were extracted from patients within 24 hours of ICU admission, along with static characteristics such as age, sex, BMI, and admission type. The dataset was divided into training and test sets in an 8:2 ratio.

[0022] 1.2 Data Sources The data for this study were obtained from the MIMIC-III and MIMIC-IV databases. The MIMIC (Medical InformationMart for Intensive Care) database, developed and maintained by the Computational Physiology Laboratory at MIT, is a large, publicly available database for critical care medicine. The data source is Beth Israel Deaconess Medical Center (BIDMC) in Boston, covering clinical records from 2001 to 2019, including over 100,000 hospitalizations for more than 80,000 patients.

[0023] Inclusion criteria: (1) Hospitalization between 2001 and 2019; (2) Patients aged ≥80 years; (3) Patients with a first ICU hospitalization record. Exclusion criteria: (1) ICU stay of less than 24 hours; (2) Lack of in-hospital outcome data; (3) Patients with missing vital sign observation records on the first day of ICU admission.

[0024] A total of 15,517 patients were ultimately included. The variables included in the study were divided into two categories: (1) Dynamic time-series variables (43 items): including heart rate, systolic blood pressure, mean arterial pressure, respiratory rate, body temperature, blood oxygen saturation, central venous pressure, albumin, anion gap, bicarbonate, urea nitrogen, calcium, chloride, creatinine, glucose, sodium, potassium, fibrinogen, international normalized ratio, prothrombin time, partial thromboplastin time, hematocrit, hemoglobin, platelets, white blood cell count, alanine aminotransferase, aspartate aminotransferase, bilirubin, arterial blood oxygen partial pressure, arterial blood carbon dioxide partial pressure, inhaled oxygen concentration, oxygenation index, pH, base excess, lactate, troponin, magnesium, brain natriuretic peptide, lymphocytes, neutrophils, alkaline phosphatase, vasoactive drugs, and ventilation support. Multiple data collections were conducted within a 24-hour time window to form a time-series trajectory. (2) Static variables (5 items): age, gender, BMI, admission type (emergency / elective / observation / other), and length of hospital stay before ICU. The outcome measure was in-hospital mortality.

[0025] The flowchart for the selection of research subjects in this application is shown below. Figure 2 .

[0026] 1.3 Data Preprocessing For missing values, a multiple imputation strategy is employed: variables with approximately normal distributions are imputed using the mean, and variables with skewed distributions are imputed using the median. Categorical variables are encoded using One-Hot encoding. Continuous variables are standardized to eliminate the influence of unit weight. The dataset is randomly divided into training and test sets in an 8:2 ratio to ensure no information leakage; the test set is used only for final model evaluation.

[0027] 1.4 Multimodal Model Framework This application develops a dynamic and static data dual-branch coding neural network architecture. The core components include: a static encoder, which uses a two-layer fully connected layer to process static features and capture the patient's baseline feature information; and a temporal encoder, which uses a dual-view heterogeneous fusion temporal coding network to innovatively jointly model multidimensional temporal data from two dimensions: "image perspective" and "explicit statistical feature perspective". The dual-view heterogeneous fusion temporal coding network includes five core modules: (1) temporal data image transformation; (2) image perspective feature extraction based on multi-scale multi-channel Vision Transformer; (3) statistical perspective feature extraction based on statistical features and machine learning; (4) adaptive gating fusion mechanism; and (5) static feature encoding and fusion.

[0028] 1.4.1 Visualization of Time Series Data (GAF) To capture high-order feature relationships in time series data from a visual perspective, Gram angle field (GAF) is used to convert one-dimensional time series data into two-dimensional images. Given a time series of length N, the conversion steps are as follows: ① Polar coordinate mapping: The MinMaxScaler function in the sci-kit learn Python library is used to normalize the time series to the [-1,1] interval; ② Construct polar coordinate mapping: Formula (1); ③ Image generation: An N×N dimensional image matrix is ​​generated through trigonometric function operations. Formula (2); The final result is a 43-channel GAF image (24×24) corresponding to 43 variables. This paper selects eight features to illustrate examples of the image after Gram-based corner field transformation, such as… Figure 3 As shown, A: Heart rate; B: Systolic blood pressure; C: Diastolic blood pressure; D: Respiratory rate; E: Body temperature; F: Blood oxygen saturation; G: Urine output; H: Central venous pressure. The color changes from blue to red to indicate that the variable value increases.

[0029] 1.4.2 Image Perspective: Multi-scale, Multi-channel Vision Transformer To overcome the limitations of traditional CNNs, which are restricted to local receptive fields, this study introduces a multi-scale VisionTransformer (ViT) for feature extraction from GAF images. The design of multi-scale ViT is inspired by the human visual system's ability to perceive information at different scales. It captures local details and global trends in temporal data by segmenting images and extracting features at different patch scales.

[0030] (1) Multi-scale Patch Embedding Module: The input GAF image X ∈ R^{C×H×W} (where C=43 is the number of channels, H=W=24 is the spatial dimension) is embedded with patches at different scales through three parallel branches. First, each branch uses a 1×1 convolution to reduce the number of channels from 43 to 3, which serves as the unified feature basis for subsequent multi-scale segmentation. Then: Small-scale branch (patch_size=4×4): Divide the image into 6×6=36 patches. Each patch (48-dimensional after flattening) is mapped to an embedding vector through linear projection, with an embedding dimension d_model=256. This branch focuses on capturing local details and short-term variations.

[0031] The mesoscale branch (patch_size=6×6) divides the image into 4×4=16 patches. This branch balances the capture of local and global information.

[0032] Large-scale branch (patch_size=8×8): Divides the image into 3×3=9 patches. This branch focuses on capturing global trends and long-range dependencies.

[0033] The embedding vectors of each branch are concatenated along the sequence dimension to obtain a token sequence with a total length of 36+16+9=61, and learnable positional encoding is added. (2) Transformer encoder: The concatenated multi-scale token sequence is fed into the standard Transformer encoder, which consists of L=6 layers of multi-head self-attention (MHSA) and feed-forward network (FFN). The self-attention mechanism enables the model to focus on the dependency between any two positions in the GAF image, effectively capturing the global contextual information of the temporal data.

[0034] (3) Feature output: Global average pooling is performed on the token sequence output by the Transformer encoder to obtain a compact feature representation of the image view F_img ∈ R^{256}.

[0035] The detailed parameter configurations for the multi-scale Vision Transformer model are shown in Table 1: Table 1. Multi-scale, multi-channel Vision Transformer network parameter configuration 1.4.3 Statistical Feature Perspective: Machine Learning Models To model explicit statistical relationships in time series data, this study extracted seven statistical features from the time series data: (1) mean; (2) standard deviation; (3) maximum; (4) minimum; (5) median; (6) skewness; and (7) kurtosis. For 43 dynamic variables, each statistical feature was calculated within a 24-hour window, resulting in a 43×7=301-dimensional statistical feature vector.

[0036] This study selected four machine learning algorithms to model statistical features: (1) XGBoost (eXtreme Gradient Boosting): A gradient boosting algorithm based on the Boosting framework. It optimizes the loss function by iteratively adding decision trees and introduces regularization terms to control model complexity.

[0037] (2) Random Forest (RF): An ensemble learning algorithm based on the Bagging strategy. It effectively reduces variance by constructing multiple decision trees and aggregating their prediction results, and has good tolerance for noisy data.

[0038] (3) Logistic Regression (LR): A classic generalized linear model that maps linear output to probability values ​​through the Sigmoid function, and has good interpretability.

[0039] (4) Support Vector Machine (SVM): The kernel function maps the data to a high-dimensional space to find the optimal classification hyperplane. This study uses the RBF kernel function.

[0040] All four algorithms optimized their hyperparameters using 5-fold cross-validation. The output probability of the best-performing model (Random Forest) was ultimately selected as the statistical feature F_stat.

[0041] 1.4.4 Adaptive Gating Fusion Mechanism Traditional multimodal fusion methods (such as simple stitching and fixed-weighted fusion) fail to fully utilize the dynamic interaction between features from different perspectives. This study designs an adaptive gating fusion module that uses a learnable gating mechanism to dynamically weigh the contributions of image perspective and statistical perspective based on the input data.

[0042] The mathematical definition of the adaptive gating fusion module is as follows: Given image viewpoint features F_img ∈ R^{256} and statistical viewpoint features F_stat ∈ R^{1}, we first broadcast-expand F_stat to 256 dimensions to form extended statistical viewpoint features F_stat_expanded, and then calculate the gated signal: G = σ(W_g · [F_img ∥ F_stat_expanded] + b_g); Where σ is the Sigmoid activation function, W_g ∈ R^{512×256} and b_g ∈ R^{256} are learnable parameters, and [·∥·] denotes vector concatenation operation. Each element of the gate signal G ∈ (0,1)^{256} controls the fusion weights of the corresponding feature dimensions.

[0043] The fusion features are: F_fused = G ⊙ F_img + (1 - G) ⊙ F_stat_expanded; Where ⊙ represents element-wise multiplication. The advantages of this mechanism are: (1) the gating signal is dynamically generated based on the joint information of dual-view features, rather than fixed weights; (2) element-wise gating allows each feature dimension to be adjusted independently, which is more flexible; (3) the complementary fusion strategy ensures the complete transmission of information.

[0044] 1.4.5 Static Feature Encoding and Fusion In addition to dynamic time-series data, patients' demographic indicators and baseline clinical characteristics (such as age, sex, BMI, hospital type, and pre-ICU length of stay) have significant independent predictive value for mortality risk. To enable the model to utilize complementary information from both static and dynamic time-series features, this study designed a static feature encoding and fusion module.

[0045] Specifically, firstly, five static features (including the expanded dimensions of the categorical variables after One-Hot encoding, totaling nine dimensions) are input into a two-layer fully connected layer: F_static = ReLU(W2 · ReLU(W1 · x_static + b1) + b2); Where W1 ∈ R^{32×9} and W2 ∈ R^{64×32} are weight matrices, and the encoded static feature vector F_static ∈ R^{64}. Then, the dual-view fused feature F_fused ∈ R^{256} output by the adaptive gated fusion module is combined with the static encoded feature F_static input through a cross-modal bidirectional cross-attention mechanism to perform dual-feature interaction fusion. F_final = [F_fused ∥ F_static]; in, , [·∥·] represents the vector concatenation operation.

[0046] Finally, F_final is input into the prediction layer, which consists of a fully connected layer (320→128, ReLU activation, Dropout=0.3) and a Softmax output layer (128→2), and outputs the probability of death risk.

[0047] This hierarchical fusion strategy of "first gating and fusing dual-perspective temporal features, then encoding and fusing static features" has the following advantages: (1) The dynamic trends and statistical relationships within the temporal data interact fully through the gating mechanism; (2) Static features are independently encoded as complementary information and participate in prediction, avoiding the dilution of static features in high-dimensional gating computation; (3) The overall architecture has flexible scalability and can easily add more types of static features.

[0048] 1.4.6 Model Training and Optimization The multi-scale, multi-channel ViT model uses the Adam optimizer with an initial learning rate of 3e-4 and employs a cosine annealing learning rate decay strategy. The loss function is a weighted cross-entropy loss (weight ratio of 1:3 to mitigate class imbalance). The batch size is 32, the maximum training epochs are 100, and an early stopping strategy (patience=10) is used. The training parameters of the GAF+CNN+CBAM baseline model remain unchanged.

[0049] The training of the entire model adopts a two-stage strategy: the first stage trains the image perspective branch (multi-scale ViT) and the statistical feature perspective branch (4 ML models) independently; the second stage fixes the parameters of the pre-trained branches and jointly trains the adaptive gating fusion module, the static feature MLP encoder and the prediction layer.

[0050] 1.5 Model Evaluation Methods This study evaluated the model's discriminative ability using the area under the receiver operating characteristic curve (AUROC). The true positive rate and false positive rate were calculated at different thresholds, and the AUC value was calculated. The closer the AUROC value is to 1, the stronger the model's discriminative ability. (2) Classification accuracy was evaluated using a comprehensive assessment of the model's classification performance, including accuracy, sensitivity / recall, specificity, precision, and F1 score.

[0051] II. Research Results 2.1 Baseline Characteristics The study included 15,517 patients, with 2,522 (16.25%) dying and 12,995 (83.75%) not dying. Baseline characteristics analysis showed that the age of patients in the dying group was significantly higher than that in the non-dying group (median: 86 vs 85 years, P<0.001), but there was no statistically significant difference in gender distribution (males accounted for 47.98% vs 46.69%, P=0.243). The dying group had a lower BMI (24.97 vs 25.78, P<0.001) and a significantly lower proportion undergoing elective surgery than the non-dying group (1.35% vs 4.43%, P<0.001).

[0052] Regarding disease severity, patients in the death group had higher APSIII scores (62 vs 44, P<0.001), OASIS scores (41 vs 34, P<0.001), and SAPS scores (49 vs 40, P<0.001). The distribution of admission types showed that the proportion of emergency admissions in the death group was significantly higher than that in the non-death group (76.84% vs 67.23%), while the proportion of elective admissions was lower (1.55% vs 5.87%), with statistically significant differences between groups (P<0.001). Nonparametric tests were used for all variable comparisons between groups. Baseline characteristics are shown in Table 2, with categorical variables expressed as quantity (proportion). Continuous variables were expressed as median and quartiles. Differences in the distribution of age, BMI, CCI score, APSIII, OASIS, and SAPS between the death and non-death groups are shown in Table 2. Figure 4 As shown.

[0053] Table 2 Baseline Characteristics 2.2 Model Performance Comparison This study trained image-viewpoint models (multi-scale ViT and baseline GAF+convolutional neural networks) and statistical feature-viewpoint models (four ML algorithms). Table 3 shows the performance comparison of each model on the test set.

[0054] Table 3 Performance Comparison of Each Model on the Test Set 3.3 Interpretability Analysis Since machine learning models are highly interpretable and random forest models have the highest AUROC, this study uses random forest models for interpretive analysis. Figure 5 The results of the contribution analysis of the top 20 features based on SHAP are presented. The horizontal axis represents the marginal contribution of the features to the model output (Shapley Value Contribution), and the vertical axis represents the feature names, arranged in descending order of mean SHApley value. The English variable names in the figure correspond to the Chinese names as follows: pao2fio2ratio: oxygenation index; ventilation: ventilation support; vasopressor: vasoactive drug; admission_type_EMERGENCY: admission type_emergency; troponin_standard_deviation: standard deviation of troponin; spo2_variance: variance of blood oxygen saturation; gender_F: gender_female; lactate_variance: variance of lactate; albumin_variance: variance of albumin; pt_standard_deviation: standard deviation of prothrombin time; temperature_variance: variance of body temperature; age: age; creatinine_variance: variance of creatinine; pao2: standard deviation of arterial blood oxygen partial pressure; hematocrit_median: median of hematocrit; inr_minimum: minimum of international normalized ratio. The color gradient (from blue to red) represents the high and low values ​​of the feature, where blue represents low values ​​and red represents high values. The black vertical dashed line is the baseline reference line (usually zero contribution or global mean).

[0055] Oxygenation index was the most important characteristic, with high values ​​(red area) significantly enhancing its positive contribution, indicating that larger ratio fluctuations were closely associated with increased disease risk. Ventilation support-related parameters all showed strong correlations with outcomes, with high values ​​(red area) corresponding to stronger positive contributions, suggesting that ventilation support needs might indicate more severe respiratory failure. The use of vasoactive drugs also showed a strong positive contribution, reflecting a direct link between the intensity of vasoactive drug use and disease severity. Low values ​​for emergency admission type (blue area) showed a negative contribution, suggesting lower potential risk in non-emergency admissions. High troponin levels amplified the negative impact of cardiovascular injury on outcomes through a positive contribution; blood oxygen saturation was consistent with oxygenation index, revealing that larger ratio fluctuations were closely associated with increased disease risk. Age's contribution was close to the baseline, indicating a limited independent effect, but it may indirectly affect outcomes through interactions with other variables.

[0056] This study proposes a mortality risk prediction model for critically ill elderly patients based on dual-perspective heterogeneous fusion. By modeling multidimensional time-series data from both image and explicit statistical feature perspectives, and fusing static clinical features, a dynamic-static dual-branch coding network is constructed, achieving efficient integration of multi-source heterogeneous information. Experimental results show that the model achieves optimal performance across all four evaluation dimensions.

[0057] The core innovation of this study lies in the design of a dual-perspective heterogeneous fusion framework. Unlike existing studies that often rely on single measurements within 24 hours of ICU admission, this study utilizes complete 24-hour time-series observation data to model the dynamics of the patient's condition from two complementary representation spaces: an image perspective and a statistical perspective. The image perspective captures the global dependencies of the time-series data through multi-scale ViT, while the statistical perspective models explicit statistical relationships through machine learning models. This design more comprehensively characterizes the information about the evolution of the patient's condition contained in the multi-dimensional time-series data.

[0058] This study proposes a mortality risk prediction model for critically ill elderly patients based on dual-view heterogeneous fusion. This model innovatively integrates dynamic image features from multi-scale Vision Transformer, statistical relational features from machine learning, and static clinical features. An adaptive gating fusion mechanism is employed to achieve efficient integration of heterogeneous features, overcoming the limitation of relying on single-measurement data. It fully utilizes the dynamic evolution information of the patient's condition contained in 24-hour multidimensional time-series data, achieving optimal performance across all four assessment dimensions. SHAP analysis revealed key risk factors such as vasoactive drug use, oxygenation index, ventilation support, dynamic lactate variability, and age, providing important references for clinical intervention. This model offers an intelligent decision support tool with both predictive efficacy and clinical interpretability for precise risk stratification of critically ill elderly patients.

Claims

1. A method for predicting mortality risk in critically ill elderly patients based on dual-perspective heterogeneous fusion, characterized in that: Static data encoders are used to process the static data of subjects, capture the baseline information of subjects, and obtain static data features; The temporal data encoder is used to process the subject's temporal data, model the dynamic trend visual perception and explicit statistical relationship of the temporal data, and obtain image perspective features and statistical perspective features. The image perspective features and statistical perspective features are fused heterogeneously through an adaptive gating fusion module to obtain temporal data features. In the static feature encoding and fusion module, static data features are encoded using a multilayer perceptron to obtain static encoded features; Static coding features and time-series data features are input into a bidirectional cross-attention layer. Through a multi-head attention mechanism, the two features are interactively fused and concatenated to obtain fused features. The fused features are input into the prediction module, which then outputs the probability of death risk. The prediction module consists of a fully connected layer and an output layer; The static data of the subjects included: age, gender, BMI, type of hospital admission, and length of hospital stay before ICU. The subjects' time-series data included: heart rate, systolic blood pressure, mean arterial pressure, respiratory rate, body temperature, blood oxygen saturation, central venous pressure, albumin, anion gap, bicarbonate, urea nitrogen, calcium, chloride, creatinine, glucose, sodium, potassium, fibrinogen, international normalized ratio, prothrombin time, partial thromboplastin time, hematocrit, hemoglobin, platelets, white blood cell count, alanine aminotransferase, aspartate aminotransferase, bilirubin, arterial blood oxygen partial pressure, arterial blood carbon dioxide partial pressure, inhaled oxygen concentration, oxygenation index, pH, base excess, lactate, troponin, magnesium, brain natriuretic peptide, lymphocytes, neutrophils, alkaline phosphatase, vasoactive drugs, and ventilation support. These data were collected multiple times within a 24-hour time window to form the time series.

2. The method for predicting mortality risk in critically ill elderly patients based on dual-view heterogeneous fusion according to claim 1, characterized in that: The static data encoder includes two fully connected layers.

3. The method for predicting mortality risk in critically ill elderly patients based on dual-view heterogeneous fusion according to claim 1, characterized in that: The temporal data encoder includes an image perspective branch and a statistical feature perspective branch; In the image perspective branch, the time series data is converted into a multi-channel GAF image using Gram angle field, and the dynamic trend of the time series data is modeled through a multi-scale, multi-channel Vision Transformer network. In the statistical feature perspective branch, the statistical features of the time series data are extracted: mean, standard deviation, maximum, minimum, median, skewness, and kurtosis, and machine learning algorithms are used to model explicit statistical relationships.

4. The method for predicting mortality risk in critically ill elderly patients based on dual-view heterogeneous fusion according to claim 1, characterized in that: In the adaptive gating fusion module Given image viewpoint features F_img and statistical viewpoint features F_stat, the statistical viewpoint features F_stat are first broadcast-expanded to form expanded statistical viewpoint features F_stat_expanded. Then, the two are concatenated along the feature dimension, input into a fully connected layer, and activated by a Sigmoid function to generate a gated signal G. G = σ(W_g · [F_img ∥ F_stat_expanded] + b_g); Where σ is the Sigmoid activation function, W_g and b_g are learnable parameters, and [·∥·] represents the vector concatenation operation; each element of the gate signal G takes a value between (0,1) and is used to control the fusion ratio of the two features in each dimension; The fusion feature F_fused is: F_fused = G ⊙ F_img + (1 - G) ⊙ F_stat_expanded; Here, ⊙ represents element-wise multiplication, a mechanism that enables the model to adaptively assign fusion weights to different feature dimensions.

5. The method for predicting mortality risk in critically ill elderly patients based on dual-view heterogeneous fusion according to claim 3, characterized in that: The machine learning algorithm is XGBoost, Random Forest, Logistic Regression, or Support Vector Machine.

6. A device for predicting mortality risk in critically ill elderly patients based on dual-view heterogeneous fusion, comprising: Static data encoder, time-series data encoder, static feature encoding and fusion module, prediction module; A static data encoder is used to process the static data of the subjects to capture the subjects' baseline information and obtain static data features; The temporal data encoder is used to process the subject's temporal data to model the dynamic trends and explicit statistical relationships of the temporal data and obtain image perspective features and statistical perspective features. The temporal data encoder includes an adaptive gating fusion module, which performs heterogeneous feature fusion of image perspective features and statistical perspective features to obtain temporal data features. The static feature encoding and fusion module includes a multilayer perceptron and a cross-modal bidirectional cross-attention layer. Static data features are encoded by the multilayer perceptron to obtain static encoded features. The static encoded features and time-series data features are input into the bidirectional cross-attention layer, and the dual features are interacted, fused, and concatenated through a multi-head attention mechanism to obtain fused features. The prediction module consists of a fully connected layer and an output layer. After the fused features are input into the prediction module, the prediction module outputs the probability of death risk. The static data of the subjects included: age, gender, BMI, type of hospital admission, and length of hospital stay before ICU. The subjects' time-series data included: heart rate, systolic blood pressure, mean arterial pressure, respiratory rate, body temperature, blood oxygen saturation, central venous pressure, albumin, anion gap, bicarbonate, urea nitrogen, calcium, chloride, creatinine, glucose, sodium, potassium, fibrinogen, international normalized ratio, prothrombin time, partial thromboplastin time, hematocrit, hemoglobin, platelets, white blood cell count, alanine aminotransferase, aspartate aminotransferase, bilirubin, arterial blood oxygen partial pressure, arterial blood carbon dioxide partial pressure, inhaled oxygen concentration, oxygenation index, pH, base excess, lactate, troponin, magnesium, brain natriuretic peptide, lymphocytes, neutrophils, alkaline phosphatase, vasoactive drugs, and ventilation support. These data were collected multiple times within a 24-hour time window to form the time series.

7. The mortality risk prediction device for critically ill elderly patients based on dual-view heterogeneous fusion according to claim 6, characterized in that: The static data encoder includes two layers of multilayer perceptrons.

8. The mortality risk prediction device for critically ill elderly patients based on dual-view heterogeneous fusion according to claim 6, characterized in that: The temporal data encoder includes an image perspective branch and a statistical feature perspective branch; The image perspective branch includes an image transformation unit and a multi-scale, multi-channel Vision Transformer network unit. In the image transformation unit, the temporal data is converted into a multi-channel GAF image using the Gram angle field. In the multi-scale, multi-channel Vision Transformer network unit, the multi-channel GAF image is used to model the dynamic trend of the temporal data through the multi-scale Vision Transformer network. The statistical features perspective branch includes a machine learning algorithm unit that extracts statistical features from time series data: mean, standard deviation, maximum, minimum, median, skewness, and kurtosis, and models explicit statistical relationships through the machine learning algorithm unit.

9. The mortality risk prediction device for critically ill elderly patients based on dual-view heterogeneous fusion according to claim 6, characterized in that: In the adaptive gating fusion module Given the image viewpoint feature F_img and the statistical viewpoint feature F_stat, the statistical viewpoint feature F_stat is first broadcast-expanded to form the expanded statistical viewpoint feature F_stat_expanded, and then the gate signal G is calculated: G = σ(W_g · [F_img ∥ F_stat_expanded] + b_g); Where σ is the Sigmoid activation function, W_g and b_g are learnable parameters, and [·∥·] represents the vector concatenation operation; each element of the gate signal G controls the fusion weight of the corresponding feature dimension; The fusion feature F_fused is: F_fused = G ⊙ F_img + (1 - G) ⊙ F_stat_expanded; Here, ⊙ represents element-wise multiplication.

10. A computer-readable storage medium having a computer program stored thereon, said computer program, when executed, implementing the method for predicting mortality risk in critically ill elderly patients based on dual-view heterogeneous fusion as described in any one of claims 1-5.

Images