Blood flow infection risk prediction device based on multi-scale attention feature enhancement

Abstract

The invention provides a bloodstream infection risk prediction device based on multi-scale attention feature enhancement. The bloodstream infection risk prediction device comprises a memory and a risk early warning device. The memory stores user physiological index data and a computer program, the processor reads the data and operates the program to realize risk prediction, and the early warning device executes graded early warning according to a prediction result. The prediction process comprises a data preprocessing (physiological index feature dimension expansion) module and a three-level series model module. A standardization processing module normalizes preprocessed data to eliminate dimension differences. The feature enhancement module is used for jointly mining multivariable interaction relations and cross-patient group similarity information in individual moments through multi-branch collaboration (linear transformation, a self-attention mechanism and a Mama layer), and effectively capturing multi-scale dependence of physiological indexes; and the linear prediction module outputs a risk value through a full connection layer with Softmax. The device solves the problems of single-dimension processing, insufficient multi-scale dependence capture and high missed and false alarm rate in the prior art, the prediction stability is improved, the early accurate early warning requirement of blood flow infection in an ICU high-risk scene is met, and clinical advanced intervention is assisted to improve the prognosis of a patient.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and machine learning technology, and in particular relates to a bloodstream infection risk prediction device based on multi-scale attention feature enhancement. Background Technology

[0002] Bloodstream infection (BSI) is a serious complication that frequently occurs in intensive care unit (ICU) patients. It has an insidious onset and rapid progression, often lacking typical clinical symptoms in its early stages. If intervention is missed, it can easily develop into sepsis and multiple organ failure, significantly increasing patient mortality, hospital stays, and healthcare costs. Therefore, achieving early and accurate risk prediction of BSI is crucial for helping clinicians intervene early, optimize treatment strategies, and improve patient outcomes, possessing significant clinical value and practical significance.

[0003] Existing technologies for predicting BSI risk mainly fall into two categories: one is based on traditional machine learning methods, which extract patients' static clinical features (such as single blood routine tests, inflammatory markers, liver and kidney function indicators, etc.) and use models such as gradient boosting trees, random forests, or logistic regression to build predictive models. These methods rely on manual feature selection and are difficult to capture the dynamic patterns of physiological indicators changing over time. The other category is based on deep learning methods, such as using recurrent neural networks (RNNs) and long short-term memory networks (LSTMs) to model the temporal changes of physiological indicators, or using the Transformer architecture to capture long-range dependencies. Some solutions also attempt to combine state-space models to optimize sequence processing efficiency. However, these methods tend to focus on single-dimensional feature processing—either only focusing on the longitudinal temporal changes of individual patients or only mining multivariate associations at a single moment, failing to combine intra-temporal multivariate interactions with cross-patient group similarity information.

[0004] Despite some progress in BSI prediction, existing technologies still face several critical technical challenges: First, current models lack the ability to integrate and mine multivariate synergistic effects and cross-patient group information within an individual's timeframe, making it difficult to comprehensively characterize the dynamic features of a patient's condition and group-related clues. Second, they are insufficient in capturing the multi-scale dependence of physiological indicators (such as short-term fluctuations and long-term trends) and have poor adaptability to irregular sampling data commonly found in clinical data, resulting in limited model robustness. Third, while some models have improved prediction accuracy, they suffer from weak interpretability and insufficient prediction stability, leading to high false negative or false positive rates in clinical applications, failing to meet the real-time and reliable BSI risk warning requirements in high-risk ICU scenarios. Summary of the Invention

[0005] To address the problems existing in the background technology, this invention provides a bloodstream infection risk prediction device based on multi-scale attention feature enhancement, comprising: a processor, a memory, and a risk warning device; the memory is used to store the user's physiological indicator data and a computer program; the processor is used to read the user's physiological indicator data and the computer program from the memory, and run the computer program to predict the user's bloodstream infection risk based on the user's physiological indicator data; the risk warning device provides a risk warning to the user based on the bloodstream infection risk prediction result; the bloodstream infection risk prediction based on the user's physiological indicator data includes: preprocessing the user's physiological indicator data, inputting the preprocessed physiological indicator data into a bloodstream infection risk prediction model for prediction, and obtaining a bloodstream infection risk prediction result; wherein, the bloodstream infection risk prediction model includes: a standardization processing module, a feature enhancement module, and a linear prediction module cascaded sequentially; the standardization processing module is used to normalize the user's input physiological indicator data, the feature enhancement module is used to enhance the input features, and the linear prediction module is used to predict the bloodstream infection risk based on the input features.

[0006] The present invention has at least the following beneficial effects

[0007] This invention reduces dimensional differences in physiological indicators by standardizing the data dimensions through a standardized processing module. It also leverages the multi-branch design of the feature enhancement module (combining linear transformation, self-attention mechanisms, and Mamba layers) to jointly mine multivariate interactions within an individual's timeframe and cross-patient group similarity information. This effectively captures the multi-scale dependence (short-term fluctuations and long-term trends) of physiological indicators and improves adaptability to irregular sampling data. Simultaneously, through the collaboration of the linear prediction module and the risk warning device, it broadens the usable threshold range of the model while ensuring prediction accuracy, reducing false negative and false positive rates, and enhancing prediction stability and interpretability. Ultimately, this achieves early and accurate warnings of bloodstream infections, assisting clinicians in early intervention to improve patient prognosis and better meeting the real-time diagnosis and treatment needs of high-risk ICU scenarios. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the device structure of the present invention;

[0009] Figure 2 This is a schematic diagram of the bloodstream infection risk prediction model of the present invention;

[0010] Figure 3 This is a schematic diagram comparing the AUC ROC performance curves of the present invention;

[0011] Figure 4 This is a schematic diagram comparing the AUPR performance curves of the present invention. Detailed Implementation

[0012] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0013] Please see Figure 1 and Figure 2 This invention provides a bloodstream infection risk prediction device based on multi-scale attention feature enhancement, comprising: a processor, a memory, and a risk warning device; the memory is used to store user physiological indicator data and a computer program; the processor is used to read the user's physiological indicator data and the computer program from the memory, and run the computer program to predict the user's bloodstream infection risk based on the user's physiological indicator data; the risk warning device provides a risk warning to the user based on the bloodstream infection risk prediction result; the bloodstream infection risk prediction based on the user's physiological indicator data includes: preprocessing the user's physiological indicator data, inputting the preprocessed physiological indicator data into a bloodstream infection risk prediction model for prediction, and obtaining a bloodstream infection risk prediction result; wherein, the bloodstream infection risk prediction model includes: a standardization processing module, a feature enhancement module, and a linear prediction module cascaded in sequence; the standardization processing module is used to normalize the user's input physiological indicator data, the feature enhancement module is used to enhance the input features, and the linear prediction module is used to predict the bloodstream infection risk based on the input features.

[0014] Preferably, the preprocessing of the user's physiological indicator data includes: assuming the user's physiological indicator data is represented as features. ,feature Adding a dimension yields features ;in, This indicates the number of times a user's most recent physiological indicator data has been recorded; This indicates the number of features in the physiological indicator data.

[0015] Preferably, the standardization processing module for normalizing user-input physiological indicator data includes: receiving preprocessed features. Normalize it and output the features. As input features for the feature enhancement module.

[0016] Preferably, the feature enhancement module performs feature enhancement processing on the input features by: applying a linear layer to the features. Performing linear transformations to obtain features Features are processed through linear layers. Performing linear transformations to obtain features Features are processed through an average pooling layer. Features are obtained by performing average pooling. ; Features and characteristics Features were obtained by performing self-attention mechanisms separately. and characteristics ; Features and characteristics Features are obtained by performing linear transformations through linear layers respectively. and characteristics ; Features and characteristics Features are obtained by adding features together. ; Features The features are obtained by forward propagation of the input into a feedforward neural network (FFN). ; Features The features are obtained by processing the input using the Silu activation function. ; Features Features are obtained by inference using the Mamba module. ; Features ,feature and characteristics Features are obtained by adding features together. .

[0017] In this embodiment, deep enhancement of physiological indicator features is achieved through multi-branch collaboration and multi-technology fusion: First, the standardized input features are subjected to two linear transformations to generate feature vectors adapted to different scales. Simultaneously, average pooling is performed on one set of features to filter noise and extract global statistical information. Then, self-attention mechanisms are applied to high-resolution features (after linear transformation) and low-resolution features (after pooling) respectively. The former captures the interaction correlation of physiological indicators of the same patient at different times, while the latter mines similarity information across patient groups. Subsequently, the two types of features are unified in dimension through linear transformation to meet the fusion conditions. After feature addition, the input is fed forward neural network (FFN) to enhance nonlinear expression ability. The Silu activation function enhances the model's adaptability to complex features. Then, the Mamba module is introduced to efficiently model long-range temporal dependencies to capture the dynamic trends of physiological indicators. Finally, the FFN output, Silu processing results, and Mamba inference results are added to form enhanced features containing multi-scale information, long-range dependencies, and group similarity cues. This process achieves multi-dimensional feature mining through "local interaction + global correlation + long-term time series", effectively improving feature discriminativeness and robustness, and laying the foundation for subsequent accurate prediction.

[0018] Preferably, the linear prediction module is used to predict the risk of bloodstream infection based on input features, including: the input features of the linear prediction module are features , will feature The risk of bloodstream infection for a user is obtained by regression prediction using a fully connected layer with Softmax input.

[0019] Preferably, the Mamba module first sets the features The features are obtained by performing linear transformations on two separate linear layers. and characteristics ; Features Features are obtained by processing through 3×3 causal convolution. ; Features The features are obtained by processing the input using the Silu activation function. ; Features Features are obtained by inputting data into the SSM module for inference. ; Features The features are obtained by processing the input using the Silu activation function. ; Features and characteristics Multiplication yields the features ; Features Features are obtained by performing a linear transformation on the input linear layer. .

[0020] Preferably, the warning method of the risk warning device includes at least one of audible and visual alarm, text prompt, or message push, and the warning intensity is set according to the level of bloodstream infection risk value.

[0021] In this embodiment, the risk warning device's core is to achieve accurate transmission and efficient response to bloodstream infection risks through multi-scenario adaptable warning methods and graded warning intensity design. The audible and visual alarms are primarily used in ICU settings and can be integrated into bedside monitors or centralized control consoles at nurse stations. Differentiated sound frequencies (e.g., high-frequency rapid beeping for high risk, low-frequency intermittent alerts for medium risk) and light colors (e.g., red for high risk, yellow for medium risk, green for low risk) allow medical staff to quickly perceive the risk level without viewing the device screen. Text prompts are simultaneously displayed on the patient's electronic medical record system interface. In addition to marking risk values ​​and levels on the mobile nursing terminal screen and monitor data panel, key abnormal physiological indicators are also briefly associated (such as "risk value 0.82 (high risk), associated with persistently elevated white blood cell count and abnormal procalcitonin"), providing early warning basis for medical staff; message push is designed for mobile work scenarios of medical staff, sending real-time notifications containing patient bed number, risk level and recommended intervention measures (such as "patient in bed 3 has high BSI risk, it is recommended to repeat blood gas analysis and blood routine within 1 hour") to the mobile APP of attending physicians and responsible nurses through the hospital's internal communication platform, avoiding the omission of early warnings due to the absence of personnel on site. Meanwhile, the warning intensity is strictly graded according to the bloodstream infection risk value: a risk value < 0.3 (low risk) triggers only a text prompt to reduce interference with routine diagnosis and treatment; a risk value ≤ 0.3 < 0.7 (medium risk) activates a text prompt plus a low-frequency audible and visual alarm to remind medical staff to strengthen indicator monitoring; a risk value ≥ 0.7 (high risk) simultaneously activates a high-frequency audible and visual alarm, a prominent text prompt, and an emergency message push to ensure that high-risk signals are given priority. This approach is suitable for the fast-paced diagnosis and treatment environment of the ICU and balances the sensitivity and practicality of the warning through a tiered mechanism, avoiding medical staff fatigue caused by excessive warnings.

[0022] Preferably, the user's physiological indicators include: heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, respiratory rate, blood oxygen saturation, arterial blood oxygen partial pressure, arterial carbon dioxide partial pressure, body temperature, peripheral perfusion index, lactate level, white blood cell count, procalcitonin, C-reactive protein, alanine aminotransferase, aspartate aminotransferase, serum creatinine, blood urea nitrogen, serum potassium, serum sodium, serum calcium, prothrombin time, activated partial thromboplastin time, D-dimer, platelet count, and hemoglobin, or multiple or all of these.

[0023] Experimental simulation:

[0024] Data sets and evaluation metrics

[0025] This invention utilizes the publicly available critical care medicine database MIMIC-IV, selecting initial admission records during adult ICU stays. Multivariate time series records containing continuous vital signs and laboratory indicators are retained, while samples with excessively high missing values ​​and insufficient stay are removed. The aim is to predict patients with early Bloodstream Infections (BSI). After patient-level deduplication, the training and test sets are randomly partitioned at a ratio of 7:3 to ensure that data from the same patient do not cross sets. All features are time-aligned, outlier truncation is performed, and missing value imputation is performed on the training set, and standardized using training set statistics. Only the processing parameters obtained during training are used in the testing phase. Model performance is primarily evaluated using AUROC and AUPR, supplemented by the most common binary classification metrics: F1 score, accuracy, and recall. AUPR more sensitively reflects ranking quality and false positive control under conditions of scarce positive classes, while AUROC provides a robust measure of overall discriminative ability.

[0026] Training settings

[0027] The model uses the Adam optimization algorithm with an initial learning rate of 1×10⁻³. Considering the patients' time series as a 10-hour sliding window, the batch size was set to 30, meaning 3 patients were selected each time, and a total of 500 iterations were trained. The hidden layer dimension of Mamba was set to 32. The model was trained and tested on an NVIDIA RTX4090 GPU and PyTorch platform.

[0028] Experimental results

[0029] To demonstrate the superiority of the proposed method, a systematic comparison was conducted on seven representative methods, including tree-based boosting models (LightGBM, GBDT), linear baseline (LR, Logistic Regression), sequence neural networks (RNN, LSTM), self-attention architectures (Transformer), and table modeling methods under the state-space model paradigm (MambaTab). The metrics used were the area under the receiver operating characteristic curve (AUC) and the area under the precision-recall curve (AUPR). To avoid evaluation bias caused by resampling and partitioning differences, all models were trained and evaluated on the same data partitioning, the same feature processing, and the same training-validation-test pipeline. The model output was the positive class probability at the time step level or the sample level, and AUC and AUPR were obtained through threshold-independent curve integration. Considering the prevalence of class imbalance in clinical data, AUC and AUPR curves are also reported (see [link to relevant documentation]). Figure 3 and Figure 4 Tables 1 and 2 provide the corresponding comparison methods and performance.

[0030] Table 1 Comparison Methods

[0031] Table 2 Comparison Results

[0032] As shown in Table 2, the comparison results based on MIMIC-IV show that LPFE achieved the best performance in the main indicators: AUROC=0.9615 and AUPR=0.7862, both significantly higher than the tree model (LightGBM AUROC=0.9309 / AUPR=0.7257, GBDT AUROC=0.7344 / AUPR=0.2716) and the traditional sequence model (RNN / LSTM both less than 0.64). Compared with the strong baseline Transformer and MambaTab, LPFE improved AUROC by about 2.9 and 0.7 percentage points, respectively, and AUPR by about 4.7 and 0.2 percentage points, respectively, showing better positive detection ability under imbalanced samples. Regarding threshold-related metrics, LPFE maintained a high recall (0.7012) and F1 score (0.7512) even at a precision of 0.9105, making it the most balanced among all reported methods. While LightGBM had the highest precision (0.9239), its recall was only 0.1495 and F1 score was 0.2592, indicating excessive conservatism and significant false negatives. MambaTab and Transformer were close to LPFE in recall, but their low precision and F1 scores suggested that while improving recall, they came with higher false positives. In summary, this invention effectively suppresses noise and captures multi-scale dependencies, significantly improving recall and overall discrimination while maintaining high precision, and offering a wider usable threshold range and stronger clinical deployability.

[0033] Figure 3 and Figure 4 The graphs for AUROC and AUPR are shown. In terms of overall discriminative ability, LPFE (this invention) and MambaTab significantly outperform other methods on the ROC curve, with AUCs of approximately 0.966 and 0.964, respectively, indicating that both maintain high true positive rates and suppress false positives across various thresholds. Transformer and LightGBM form the second tier, with AUCs of approximately 0.933 and 0.920, respectively. While traditional tree models show some performance, they lag significantly behind tabular / sequence deep models. GBDT further decreases to 0.733, while classic sequence or linear models (LR, RNN, LSTM) only reach around 0.63–0.64, rarely exceeding the random level, indicating insufficient ability to capture complex time series and multivariate dependencies.

[0034] Secondly, the differences are amplified in PR curves, which are more sensitive to scenarios with high class imbalance. LPFE has the highest AUPR of approximately 0.798, and its precision remains at a high level within the recall range of 0.3–0.8. Its curve decline is the most gradual, indicating a slower increase in false positives when expanding recall and a wider available threshold band. MambaTab follows closely with 0.783, but its precision drops sharply slightly earlier at the high recall end, suggesting it is more prone to false positives under extreme recall requirements. Transformer and LightGBM have AUPRs of 0.739 and 0.723, respectively. Although they achieve high precision in the low recall range, the precision declines more steeply as recall increases, indicating limited coverage of positive cases. GBDT has an AUPR of only 0.275, making it almost impossible to maintain effective precision within the available recall range. The AUPR of LR, RNN, and LSTM are all around 0.10 to 0.12, and the PR curves drop rapidly at low recall, making it difficult to meet the positive detection rate requirements of imbalanced detection tasks.

[0035] Combining the two figures, LPFE (the present invention) leads in both overall discrimination (AUC) and positive detection (AUPR), demonstrating its effective modeling of multi-scale dependencies and long-range time series, as well as its robust noise suppression capabilities. MambaTab performs similarly but is slightly less stable. Transformer / LightGBM is a usable baseline. GBDT and traditional deep / linear sequence models are not suitable for this task. Therefore, in practical deployments, if a significant improvement in recall is required while maintaining high precision, LPFE (the present invention) is more advantageous and transferable.

[0036] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0037] In summary, this invention reduces dimensional differences in physiological indicators by standardizing the processing module to unify the data dimensions, and leverages the multi-branch design of the feature enhancement module (combining linear transformation, self-attention mechanism, and Mamba layer) to jointly mine multivariate interactions within an individual timeframe and similarity information across patient groups. This effectively captures the multi-scale dependence (short-term fluctuations and long-term trends) of physiological indicators and improves adaptability to irregular sampling data. Simultaneously, through the collaboration of the linear prediction module and the risk warning device, the usable threshold range of the model is broadened while maintaining prediction accuracy, reducing false negative and false positive rates, enhancing prediction stability and interpretability. Ultimately, this achieves early and accurate warning of bloodstream infections, assisting clinicians in early intervention to improve patient prognosis, and better meeting the real-time diagnosis and treatment needs of high-risk ICU scenarios.

[0038] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A bloodstream infection risk prediction device based on multi-scale attention feature enhancement, characterized in that, include: Processor, memory, and risk warning device; The memory is used to store the user's physiological indicator data and computer programs; The processor is used to read the user's physiological index data and computer program from the memory, and run the computer program to predict the user's bloodstream infection risk based on the user's physiological index data. The risk warning device provides a risk warning to the user based on the bloodstream infection risk prediction result. The method of predicting the risk of bloodstream infection based on the user's physiological index data includes: preprocessing the user's physiological index data, inputting the preprocessed physiological index data into a bloodstream infection risk prediction model for prediction, and obtaining a bloodstream infection risk prediction result; wherein, the bloodstream infection risk prediction model includes: a standardization processing module, a feature enhancement module, and a linear prediction module cascaded in sequence; the standardization processing module is used to normalize the physiological index data input by the user, the feature enhancement module is used to enhance the input features, and the linear prediction module is used to predict the risk of bloodstream infection based on the input features.

2. The bloodstream infection risk prediction device based on multi-scale attention feature enhancement according to claim 1, characterized in that, The preprocessing of the user's physiological index data includes: assuming the user's physiological index data is represented as features. ,feature Adding a dimension yields features ;in, This indicates the number of times a user's most recent physiological indicator data has been recorded; This indicates the number of features in the physiological indicator data.

3. The bloodstream infection risk prediction device based on multi-scale attention feature enhancement according to claim 2, characterized in that, The standardization processing module is used to normalize the physiological indicator data input by the user, including: receiving preprocessed features. Normalize it and output the features. As input features for the feature enhancement module.

4. The bloodstream infection risk prediction device based on multi-scale attention feature enhancement according to claim 3, characterized in that, The feature enhancement module is used to perform feature enhancement processing on the input features, including: enhancing the features through a linear layer. Performing linear transformations to obtain features Features are processed through linear layers. Performing linear transformations to obtain features Features are processed through an average pooling layer. Features are obtained by performing average pooling. ; Features and characteristics Features were obtained by performing self-attention mechanisms separately. and characteristics ; Features and characteristics Features are obtained by performing linear transformations through linear layers respectively. and characteristics ; Features and characteristics Features are obtained by adding features together. ; Features The features are obtained by forward propagation of the input into a feedforward neural network (FFN). ; Features The features are obtained by processing the input using the Silu activation function. ; Features Features are obtained by inference using the Mamba module. ; Features ,feature and characteristics Features are obtained by adding features together. .

5. The bloodstream infection risk prediction device based on multi-scale attention feature enhancement according to claim 4, characterized in that, The linear prediction module is used to predict the risk of bloodstream infection based on input features, including: the input features of the linear prediction module are features. , will feature The risk of bloodstream infection for a user is obtained by regression prediction using a fully connected layer with Softmax input.

6. The bloodstream infection risk prediction device based on multi-scale attention feature enhancement according to claim 4, characterized in that, The Mamba module first identifies the features The features are obtained by performing linear transformations on two separate linear layers. and characteristics ; Features Features are obtained by processing through 3×3 causal convolution. ; Features The features are obtained by processing the input using the Silu activation function. ; Features Features are obtained by inputting data into the SSM module for inference. ; Features The features are obtained by processing the input using the Silu activation function. ; Features and characteristics Multiplication yields the features ; Features Features are obtained by performing a linear transformation on the input linear layer. .

7. The bloodstream infection risk prediction device based on multi-scale attention feature enhancement according to claim 3, characterized in that, The risk warning device provides warnings in at least one of the following ways: audible and visual alarms, text prompts, or message push notifications. The warning intensity is set according to the level of bloodstream infection risk.

8. The bloodstream infection risk prediction device based on multi-scale attention feature enhancement according to claim 1, characterized in that, The user's physiological indicators include: heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, respiratory rate, blood oxygen saturation, arterial blood oxygen partial pressure, arterial carbon dioxide partial pressure, body temperature, peripheral perfusion index, lactate level, white blood cell count, procalcitonin, C-reactive protein, alanine aminotransferase, aspartate aminotransferase, serum creatinine, blood urea nitrogen, serum potassium, serum sodium, serum calcium, prothrombin time, activated partial thromboplastin time, D-dimer, platelet count, and multiple or all of the following:

Images