Predicting In-hospital of Death of Patients with Acute Stroke in the ICU Using Stacking Model

doi:10.21203/rs.3.rs-4908107/v1

Objective: To establish the in-hospital death prediction model of acute stroke patients in ICU based on 8 kinds of machine learning algorithms (SVC, Logistics, RandomForest, XGboost, GBDT, LightGBM, Catboost, MLP).

Methods: The data of 1882 acute stroke patients in ICU of the Second Affiliated Hospital of Nanchang University from November 2006 to October 2022 were collected, Lasso regression was used to screen the features, multifactorial Logistics regression algorithm was utilized to mine the risk factors of acute stroke death in ICU, and eight machine learning algorithms were utilized to build ICU patient death prediction models, and selecting the four optimal algorithms as the Stacking model base learner, as well as selecting the optimal algorithms as the Stacking model meta-learners to construct ICU stroke death prediction models. The predictive performance of the model was evaluated using the area under the receiver operating characteristic curve (AUC) of the subjects, accuracy, sensitivity, and specificity, and the gain value of the model was evaluated using the decision curve.

Result: The multi-factorial logistics regression analysis showed that atrial fibrillation, pulmonary infection, coma, high creatinine, international normalized ratio(INR) of prothrombin time, serum sodium, neutrophil count and low platelet count were independent risk factors for in-hospital death in stroke patients (P<0.05). In the training set, validation set, and external validation set, the AUC values of the Stacking prediction model were 0.878, 0.871, and 0.809, respectively. The sensitivity values were 0.82, 0.85, and 0.87, respectively. The specificity values were 0.87, 0.84, and 0.68, respectively. The top four AUC values in the eight algorithms were MLP, XGBoost, GBDT, and CatBoost with correspondingly test set AUC values of 0.829, 0.786, 0.78, and 0.777. The decision curve showed that when the probability threshold predicted by the Stacking prediction model was greater than 0.1, the model had a positive net benefit.

Conclusion: The Stacking model has a better prediction effect on ICU in-hospital death in stroke patients and can be applied to early prediction of death in ICU stroke patients, providing a basis for early clinical intervention.

Biological sciences/Computational biology and bioinformatics

Biological sciences/Computational biology and bioinformatics/Machine learning

Health sciences/Biomarkers/Predictive markers

Health sciences/Biomarkers/Prognostic markers

Health sciences/Neurology/Neurological disorders

Health sciences/Neurology/Neurological disorders/Cerebrovascular disorders/Stroke

Biological sciences/Neuroscience/Cell death in the nervous system

Acute stroke

ICU

Death

Stacking model

Prediction

Ischemic stroke is one of the diseases with a significant global burden. Recent data statistics show that the mortality rate of stroke in China has been increasing year by year, and in 2019, the number of stroke-related deaths in China reached 2.19 million [1]. The condition of severe stroke patients is critical and complex, and the condition assessment after admission affects the prognosis of patients, but the assessment and decision of doctors are full of subjectivity and uncertainty [2]. For complex diseases such as stroke, machine learning can incorporate a large number of variables into the same stroke death prediction model [3], which is very suitable for assisting doctors to assess the condition of severe stroke patients and predict the mortality rate, so as to allocate and use medical resources more reasonably [4]. At present, relevant researches are applying standard algorithms such as random forest, artificial neural network and support vector machine to predict the prognosis of stroke, and certain results have been obtained [5–7]. At the same time, studies have proved that the Stacking model can further improve the forecasting performance of algorithms [8, 9], but it is not found that the Stacking model is applied to the prediction of stroke death. Therefore, this study selected four different types of optimal algorithms for the training of the stacking model from eight algorithms with the predictive ability index. It includes Logistics and Svc for linear algorithms, Randomforest for Bagging tree algorithm, Lightgbm, Xgboost, Gbdt and Catboost for boosting tree algorithm, and Mlp for neural networks.

In recent years, algorithm innovation has made it possible to build clinical prediction models based on big data to assist doctors in decision-making, and stroke prediction models based on national large databases have become increasingly mature [10–12]. However, when applying these predictive models to local clinical practice, they still need to be adjusted or updated according to different hospital environments [13, 14], which is undoubtedly a huge challenge for physicians, so predictive models developed specifically for local case big data have greater advantages [15]. The aim of this study was to develop a mortality prediction model based on the data of severe stroke patients admitted to the Second Affiliated Hospital of Nanchang University between November 2006 and October 2022, so that it could be directly applicable to the local population. In addition, external validation is a necessary prerequisite to be able to use the model in clinical practice, which reflects the feasibility of the prediction model for the prediction of new data. According to a systematic review, only 10% of the studies conducted external validation of the predictive model [16]. In this study, patients with severe stroke admitted from January 2023 to July 2023 were externally validated to further test and optimize the predictive performance of this prediction model.

2.1. Objects and Ethics of the study

In this study, a total of 1822 patients with acute stroke admitted to ICU from November 2006 to October 2022 were collected from the Second Affiliated Hospital of Nanchang University. These patients included both hemorrhagic and ischemic stroke patients. The following inclusion and exclusion criteria were used for the study: patients aged ≥ 18 years; primary diagnosis of stroke; and for patients with multiple ICU admissions, all ICU records were included in the statistical analysis. Stroke patients who were not admitted to our ICU and those who had in-hospital death or were discharged within 48 h of admission were excluded. The process of patient inclusion and exclusion is shown in Fig. 1. Diagnostic criteria: Patients diagnosed with acute stroke according to the International Classification of Diseases 10th Edition (ICD-10) diagnostic criteria.All experiments were conducted in accordance with relevant ethical guidelines and regulations. All experimental protocols for this study have been approved by the Biomedical Research Ethics Committee of the Second Affiliated Hospital of Nanchang University. In accordance with the ethics committee’s requirements, this study has been granted an exemption from informed consent review. All data were provided by the hospital’s clinical work platform and comply with the relevant regulations of the ethics committee.

2.2. Predictor variables and outcome indicators

All the data were collected from the clinical work platform of the Second Affiliated Hospital of Nanchang University. A total of 226 variables were derived based on literature reports, clinical experience and database characteristics, including patient demographic information (gender, age), previous medical history (diabetes, hypertension, etc.), personal history (smoking and drinking history), history of present disease (lucidity, lower limb weakness, coma, etc.), and laboratory tests (blood neutrophils, hemoglobin, creatinine, serum albumin, etc.). The variables with missing values > 30% were deleted, and death after 48 hours of hospitalization was used as the outcome variable. The features of data pre-processing were analyzed by single factor, and the features with significant differences (P < 0.05) were screened by Person correlation analysis method, and the features with p > 0.7 were eliminated. The features with VIF > 10 were eliminated by correlation collinearity analysis. LASSO regression feature screening was carried out, and the screened features were incorporated into the multi-factor Logistics algorithm to mine the death risk factors of stroke patients in ICU, and the prediction model was established in eight prediction model algorithms.

2.3. Research Methods

2.3.1. Predictive Modeling Methodology

This study uses Python 3.6.5 language as a development tool and utilizes Pandas, Numpy for data preprocessing. Sklearn was utilized for prediction model construction.

1) Stacking Integrated Learning Algorithm: the Stacking integrated learning approach is a model integration technique that combines information from multiple predictive models to generate a new model [17]. It mainly consists of two parts, selecting multiple base learners for training to get the Stacking base model, and inputting the prediction results of the base model as new features to the meta learner for training to get the Stacking meta model.

2) Data preprocessing: non-numeric variables are converted to numeric variables, discrete vacancies are filled with vertical numbers, and continuous variable vacancies are filled with mean values.

3) Model construction and evaluation: the training set and test set were randomly split in the ratio of 8:2. Whether to die after 48h after admission was used as the outcome variable. The Stacking algorithm was used to build the ICU acute stroke in-hospital death prediction model, and the model was trained using five-fold cross-validation, and the prediction effect of the model was judged by verifying the accuracy, sensitivity, specificity, and AUC value of the model in the test set. The DCA decision curve is used to evaluate the net benefit of the model at different thresholds. The significance of model features is quantified by SHAP.

2.3.2. Statistical Methods

SPSS22.0 statistical software was applied for statistical analysis, and the measurement information conforming to normal distribution was expressed as mean ± standard deviation, and t test was used for comparison between two groups. Measurement information with skewed distribution was expressed as median (interquartile spacing) [M(QR)], and non-parametric test was used for comparison between two groups. Count data were statistically described using frequencies, rates, and percentages, and comparisons between groups were statistically inferred using the Chi-squared test or Fisher test, and multifactorial analyses were performed using logistic regression analysis, with differences considered statistically significant at P < 0.05.

3.1. General information of the study population

A total of 1822 acute stroke patients were included in this study, and 439 patients died after 48h of admission, with a mortality rate of 24.09%. The mean age of the death group was 67 [56,75] years old, and the mean age of the improvement group was 65 [55,74] years old. Univariate analysis showed that age, sex, platelet count, hemoglobin, neutrophil ratio, c-reactive protein(CRP), creatinine, blood uric acid, serum sodium, serum chloride, INR, activated partial thromboplastin time(APTT), blood D-dimer, serum lactate quantification, total bilirubin, alanine aminotransferase(ALT), aspartate aminotransferase(AST), total cholesterol, low-density lipoproteins(LDL), hypertension, renal insufficiency, hyperlipidemia, pulmonary infection, atrial fibrillation, history of cerebrovascular disease, pulse, clarity, lower limb weakness, and coma were independent risk factors for the prediction of death in acute stroke patients in ICU, and the difference was statistically significant (P < 0.05)(Table 1).

Table 1

Comparison of Risk Factors Between Two Groups
Variable	Univariate Analysis
Variable	Improvement Group (n = 1383)	Death Group (n = 439)	t/X²/Z	P-value
Age (year, M(Q1,Q3))	65[55,74]	67[56,75]	-2.713	0.007
Sex n(%)			13.474	< 0.001
Male	900(65.076)	243(55.353)
Female	483(34.924)	196(44.647)
Platelet Count, mean(± SD)	206.052 ± 84.491	189.085 ± 82.299	3.116	0.002
Hemoglobin Quantitation, mean(± SD)	127.222 ± 22.802	122.863 ± 25.177	2.724	0.007
Neutrophil Ratio, mean(± SD)	75.840 ± 12.822	81.680 ± 11.822	-7.464	< 0.001
CRP, mean(± SD)	47.667 ± 62.895	71.444 ± 84.455	-4.326	< 0.001
Creatinine, mean(± SD)	89.429 ± 83.111	149.272 ± 168.431	-6.008	< 0.001
Uric Acid, mean(± SD)	290.053 ± 132.926	339.251 ± 177.820	-4.492	< 0.001
Serum Sodium, mean(± SD)	140.981 ± 6.338	144.337 ± 9.944	-5.569	< 0.001
Serum Chloride, mean(± SD)	104.675 ± 5.854	108.334 ± 9.419	-5.791	< 0.001
INR, mean(± SD)	1.049 ± 0.187	1.097 ± 0.403	-2.020	0.044
APTT, mean(± SD)	26.465 ± 6.275	28.785 ± 14.312	-2.760	0.006
D-Dimer Quantitation, mean(± SD)	4.373 ± 15.280	6.855 ± 17.251	-2.146	0.032
Serum Lactate Quantitation, mean(± SD)	2.072 ± 1.387	2.938 ± 2.248	-3.647	< 0.001
Total Bilirubin Quantitation, mean(± SD)	15.348 ± 10.531	16.975 ± 11.265	-2.308	0.021
ALT, mean(± SD)	30.703 ± 35.921	52.277 ± 161.437	-2.276	0.024
AST, mean(± SD)	48.404 ± 197.333	95.196 ± 289.770	-2.638	0.009
Total Cholesterol Quantitation, mean(± SD)	4.342 ± 1.138	4.141 ± 1.140	2.517	0.012
LDL Cholesterol Quantitation, mean(± SD)	2.589 ± 0.926	2.330 ± 0.839	3.843	< 0.001
History of Hypertension, n(%)	550(51.643)	132(42.857)	7.377	0.007
Renal Insufficiency, n(%)	954(89.577)	214(69.481)	75.969	< 0.001
Hyperlipidemia, n(%)	943(88.545)	287(93.182)	5.506	0.019
Pulmonary Infection, n(%)	622(58.404)	83(26.948)	94.625	< 0.001
History of Atrial Fibrillation, n(%)	933(87.606)	249(80.844)	9.120	0.003
History of Cerebrovascular Disease, n(%)	483(45.352)	75(24.351)	43.680	< 0.001
Pulse Rate, mean(± SD)	81.347 ± 17.642	93.641 ± 25.841	-7.376	< 0.001
Clear Consciousness, n(%)	940(88.263)	285(92.532)	4.528	0.033
Weakness in Lower Limbs, n(%)	1043(97.934)	307(99.675)	4.397	0.036
Coma, n(%)	1006(94.460)	241(78.247)	75.349	< 0.001

3.2. Feature screening and analysis of multifactorial logistic regression mortality risk factor

Features with significant differences in unifactorial analysis were screened out using Person correlation analysis for feature screening, and features with p > 0.7 were eliminated. And the features with VIF > 10 were eliminated using correlation shareability analysis to eliminate the feature covariance features. The screened features were included in the multifactorial Logistics regression model with whether death was the dependent variable, and the results showed that platelet count, percentage of neutrophils, serum sodium, INR, blood creatinine, coma, lung infection, and atrial fibrillation were independent risk factors for death of patients in ICUs of stroke patients (p < 0.05), which are shown in Table 2.

Table 2

Multivariate Logistic Regression Analysis
Risk Factor	β Value	Wald Value	OR Value	95% CI	P-value
Platelet Count	-0.002	6.832	0.998	0.996 ~ 0.999	0.009
Neutrophil Ratio	0.02	10.388	1.02	1.008 ~ 1.033	0.001
Serum Sodium	0.044	23.571	1.045	1.027 ~ 1.063	0
INR	0.542	3.846	1.719	1.032 ~ 3.056	0.05
Creatinine	0.004	45.105	1.004	1.003 ~ 1.005	0.0
Coma	0.855	13.809	1.495	1.495 ~ 3.689	0
Pulmonary Infection	0.652	18.413	1.426	1.426 ~ 2.588	0
Atrial Fibrillation	0.463	6.042	1.094	1.094 ~ 2.289	0.014

3.3. Analysis of Prediction Effectiveness Based on Machine Learning Models

In order to optimize the performance of Stacking prediction model, it needs to be based on the individual base model learning capability. First, experiments are designed to compare and analyze the results of individual prediction of each base learner on the training set. The statistically significant variables screened out by the one-way analysis are incorporated into the model, and the death prediction models based on SVC, Logistics, RandomForest, XGboost, GBDT, LightGBM, Catboost, and MLP are built respectively. The model AUC values in the test set were 0.761, 0.768, 0.759, 0.786, 0.78, 0.761, 0.777, and 0.829, respectively.The Receiver Operating Characteristic (ROC) curves of each model are shown in Fig. 2.

In order to construct the optimal Stacking prediction model, this study selected the models with the top 4 AUC values, XGboost, Catboost, GBDT, and MLP, as the base learner of the Stacking model, and selects the optimal model with the optimal AUC value, the MLP algorithm, as the meta-learners of the Stacking model, to construct the Stacking integration algorithm based ICU Prediction model. The model AUC values in the training set and test set are 0.878, 0.871, respectively.The optimal decision threshold for model probability is: 0.2333. The prediction model AUC graph is shown in Fig. 3.

To further validate the performance of the Stacking model in an independent dataset, the researcher prospectively extracted 88 ICU stroke patients from January 2023 to June 2023, of which 73 were improved and 15 died. The AUC values of the nine models are shown in Fig. 4, with the Stacking model demonstrating a prospective prediction accuracy of 0.72 (see Table 3).

Table 3

Predictive Performance of the Stacking Model
Dataset	Accuracy	Sensitivity	Specificity	NPV	PPV	AUC	Youden Index
Training set	0.86	0.82	0.87	0.94	0.65	0.878	0.69
Internal validation set	0.84	0.85	0.84	0.94	0.67	0.871	0.69
External validation set	0.72	0.87	0.68	0.96	0.36	0.809	0.55

The Dominance-based Classifier Accuracy (DCA) curve showed that all eight machine learning models showed good gain value, of which the Stacking prediction model had the highest gain value, and the model prediction threshold is greater than 0.1 model gain value was positive. The details are shown in Fig. 5.

3.4. Analysis of factors affecting hospital infections based on SHAP explanatory model

SHAP is an additive explanatory model inspired by Shapley value method. The overall visualization of the features was done by shap summary plot method. The SHAP predictive values showed that ICU stroke patients were at higher risk of death if they had higher blood creatinine, higher blood neutrophils, higher serum sodium, higher admission National Institutes of Health Stroke Scale(NIHSS) scores, higher total bilirubin, higher INR, older age, lower platelet counts, lower LDL, and hospitalization with pulmonary infection, coma, hypertension, and atrial fibrillation. (Fig. 6)

This study proposed a new idea of constructing the death risk prediction model of stroke patients based on the idea of Stacking integration. Four single models with good forecasting efficiency (XGboost, Catboost, GBDT, MLP) were selected and the Stacking models were integrated. Previous studies have shown that the four models have good performance in predicting stroke death [18]. In this study, the AUC values of these four models for predicting stroke death in the test set were 0.786, 0.777, 0.78 and 0.829, respectively. However, in order to further improve the performance of the algorithm and improve the accuracy and generalization ability of the model on this basis, this study combined the advantages of various model algorithms. Different single models with good forecasting efficiency are merged into Stacking models. The AUC values of the Stacking model in the training set and the test set were 0.878 and 0.871, respectively. To further validate the Stacking model’s performance in an independent dataset, data from 88 patients were prospectively collected from January 2023 to June 2023. The Stacking model demonstrated higher prospective prediction accuracy and AUC value compared to the other eight algorithms. This indicates that the Stacking prediction model has stronger accuracy and effectiveness in predicting mortality among ICU stroke patients.

In this study, the researchers found that the risk of poor prognosis in patients with acute stroke was significantly positively associated with elevated serum creatinine levels, increased blood neutrophil counts, increased serum sodium concentrations, increased total bilirubin levels, increased international normalized ratio (INR), older age at presentation, decreased platelet counts, decreased low-density lipoprotein levels, and higher NIHSS scores at admission. In addition, the risk of death was also increased if patients had pulmonary infections, coma, hypertension, or atrial fibrillation during hospitalization. Studies have shown a significant correlation between renal insufficiency and mortality in patients with ischemic stroke [19]. Aryandhito Widhi Nugroho and used multivariate logistic regression analysis to find that low EGFR based on creatinine level was closely related to in-hospital and discharge deaths [20]. In addition, studies have also revealed that traditional cardiovascular risk factors may lead to similar vascular damage in the kidney and cerebrovascular, so doctors can use serum creatinine to predict the risk of death in patients with stroke [21]. Lili Cui's research showed that patients with neutrophil counts exceeding 7.79×10^9/L havd a higher risk of death [22]. A meta-analysis by Lu Wang of 10,371 stroke patients showed that a higher neutrophil ratio was positively associated with bleeding conversion after stroke and the risk of death within three months [23]. By releasing reactive oxygen species, proteases, cytokines and chemokines, neutrophils are involved in destroying the blood-brain barrier, exacerbating ischemic injury and edema, which may aggravate cerebral infarction and even lead to cerebral hernia [24–25]. Amit Akirov found that abnormal serum sodium levels on admission were significantly associated with short and long-term mortality. Patients with hypernatremia are at higher risk of death than those with hyponatremia [26]. In the hypertonic state of hypernatremia, brain cell dehydration may cause brain atrophy, which may lead to vascular rupture, triggering intracerebral hemorrhage, subarachnoid hemorrhage, and permanent neurological impairment or severe fatal outcomes [27]. In an animal experiment, excess bilirubin was found to be significantly toxic to the nervous system and cause a large number of neuronal death [28]. Patients with acute stroke who have higher bilirubin levels may face more extensive cerebral infarction, more significant brain edema, and more severe reperfusion injury [29], which combined to significantly increase the risk of death in stroke patients.Xuewei Xie found that a high INR on admission was highly correlated with poor stroke outcomes in patients with acute ischemic stroke who did not receive anticoagulant therapy [30]. This correlation may be due to the hemorrhagic transformation of intracranial cerebral infarction lesions or hemorrhagic enlargement of cerebral hemorrhage lesions caused by elevated INR, which exacerbates the deterioration and even death of stroke patients. In an analysis of 1233 patients with acute ischemic stroke, Jason J. Sico and colleagues showed that a decreased platelet count was associated with higher mortality [31]. In stroke patients, thrombocytopenia may increase mortality by increasing the area of ischemic infarction [32], meanwhile, studies have shown that thrombocytopenia increases the risk of post-infectious intracerebral hemorrhage and death [33].In addition, the study by Gang Lv showed that the risk of poor outcome of stroke increased by 46.2% for every 1 mmol/L decrease in low-density lipoprotein cholesterol (LDL-C) level [34]. Bang found that LDL cholesterol levels were independently associated with abnormal permeability of the blood-brain barrier [35], suggesting that reduced lipids may compromise vascular integrity and cause blood to leak between unstable cerebrovascular endothelial cells [36], which increases the likelihood of bleeding transitions after stroke.

The NIHSS score at the time of admission not only has predictive value for the prognosis of cerebral infarction, but also has guiding significance for the prognosis of cerebral hemorrhage. C. M. Cheung used the NIHSS score at admission to predict 30-day mortality in patients with intracerebral hemorrhage, with sensitivity and specificity of 81% and 90%, respectively [37]. Studies on the relationship between the Glasgow Coma Scale (GCS) and the prediction of stroke death indicate that the GCS score is an independent predictor of the risk of death in stroke patients [38, 39]. Aspiration pneumonia is the most common type of pulmonary infection in stroke patients and one of the direct causes of death from stroke. Based on an analysis of 2,424,379 death certificates, Chia-Yu Chang pointed out that 5% of stroke patients died from aspiration pneumonia [40]. Lijuan Zhang found that stroke patients infected with the acute respiratory syndrome coronavirus were more likely to die [41]. Pulmonary infections can trigger a systemic inflammatory response, worsening brain edema and brain injury, and may lead to complications such as sepsis, which has a high mortality rate. Severe pulmonary infections interfere with normal pulmonary ventilation, resulting in inadequate oxygenation of the brain and the body, which leads to brain cell damage and death. In addition, hypertension and atrial fibrillation are considered to be important causes of stroke, and they also have important clinical significance in predicting mortality in patients with stroke. According to a meta-analysis of 61 prospective studies by Sarah Lewington, about 60% of stroke deaths worldwide are associated with high systolic blood pressure, and effective control of blood pressure has been identified as one of the most critical interventions for stroke prevention [42]. The study by Daniele Massera revealed an independent correlation between newly-onset atrial fibrillation and increased incidence and mortality of stroke in hospitalized patients. The patients involved had significantly increased mortality within 30 days and 1 year after admission [43]. These conditions may be related to factors such as cardiac embolism, fatal arrhythmias, myocardial ischemia, and myocardial infarction caused by atrial fibrillation. Besides, the prediction model constructed in this study takes the preliminary assessment, laboratory and diagnostic results of patients admitted to ICU as the inclusion indexes, so as to rapidly predict the early prognosis of patients in a short time, assist clinicians to evaluate the condition and make corresponding clinical diagnosis.

In summary, this study successfully constructed an in-hospital death prediction model for ICU acute stroke patients based on the Stacking integration algorithm. The model demonstrated strong prediction accuracy and effectiveness in the testing and validation of data from 1882 patients. By comparing the AUC values of the four base learners, the Stacking model outperforms a single model on both the training and test sets, highlighting the strong potential of the Stacking integration algorithm in the field of healthcare prediction. In addition, this study revealed biomarkers with significant predictive value for patients’ mortality risk, such as blood creatinine, blood neutrophils, serum sodium, total bilirubin, and INR, by using the SHAP feature interpretation method. The comprehensive analysis of these biomarkers and clinical characteristics provides clinicians with evidence-based recommendations for feasible interventions, and provides a strong scientific basis for reducing the risk of death in ICU acute stroke patients. The application of this predictive model helps clinicians to rapidly assess the risk of death at the time of patient admission and take targeted early interventions.

Nonetheless, there are some limitations to this study. For example, the data used in this study were all from the same hospital, and data from multiple hospitals can be added for comparison and validation in the future. In addition, more features can be added and new machine learning algorithms can be used to further improve the prediction effect during the process of feature screening and prediction model construction. In subsequent studies, the model can be gradually adjusted and optimized to cope with data updates and continuously improve the prediction performance to provide more effective clinical reference and support.

Acknowledgement

We would like to extend our heartfelt gratitude to all those who have contributed to the preparation of this manuscript. Special thanks go to Kai Wang, JianMo Liu, Fang Li, KeQi Lei, TingHao Guo, ZhiJuan Cheng, JiangLong Tu for their invaluable collaboration and co-authorship. Their persistent efforts and insightful feedback have been instrumental in bringing this study to fruition.

We also wish to acknowledge the support and guidance received from our colleagues and the staff of the medical and research department whose expertise and knowledge were a great source of inspiration. Furthermore, our appreciation extends to the patient and their family for their consent and cooperation, which was critical for this case study.

Our thanks are also due to the reviewers for their constructive comments, which significantly improved the quality of our paper. Lastly, we are grateful to the editorial team of the esteemed journal for considering our work for publication.
Contribution Details

1. Kai Wang collected and organized the data, conducted data analysis, and wrote the manuscript draft;

2. Jianmo Liu analyzed and statistically processed the data and developed a predictive model for stroke mortality;

3. Fang Li assisted in data collection and provided suggestions for writing the manuscript;

4. Keqi Lei assisted in data collection;

5. Tinghao Guo assisted in data collection;

6. Zhijuan Cheng proposed and designed experiments and helped revise the manuscript;

7. Jianglong Tu proposed and designed experiments and revised the manuscript;

8. All authors reviewed the manuscript.
Data Availability Statement

The data supporting the findings of this study are obtained from the Second Affiliated Hospital of Nanchang University. Due to privacy concerns and ongoing research, the data cannot be fully disclosed publicly. However, it can be obtained from the corresponding author upon reasonable request. Furthermore, the data must not be disclosed, as doing so may result in legal liability.

Declaration of Interest Statement

The authors declare that there are no conflicts of interest regarding the publication of this paper. All authors have no direct or indirect financial relationships, personal relationships, intellectual passions, or other circumstances that could be construed to unduly influence their work. Each contributor to this study has adhered to the highest ethical and academic standards to ensure the content provided is fair, objective, and governed solely by the evidence presented and the methodologies employed in the scientific research.

Ma, Q. et al. Temporal trend and attributable risk factors of stroke burden in China, 1990–2019: an analysis for the Global Burden of Disease Study 2019. Lancet Public. Health. 6 (12), e897–e906 (2021).
Knaus, W. A., Zimmerman, J. E., Wagner, D. P., Draper, E. A. & Lawrence, D. E. APACHE-acute physiology and chronic health evaluation: a physiologically based classification system. Crit. Care Med. 9 (8), 591–597 (1981).
Wang, W. et al. A systematic review of machine learning models for predicting outcomes of stroke with structured data. PLoS One. 15 (6), e0234722 (2020).
Ryan, L. et al. Mortality prediction model for the triage of COVID-19, pneumonia, and mechanically ventilated ICU patients: A retrospective study. Ann. Med. Surg. (Lond). 59, 207–216 (2020).
Monteiro, M. et al. Using Machine Learning to Improve the Prediction of Functional Outcome in Ischemic Stroke Patients. IEEE/ACM Trans. Comput. Biol. Bioinform. 15 (6), 1953–1959 (2018 Nov-Dec).
Liang, Y., Li, Q., Chen, P., Xu, L. & Li, J. Comparative Study of Back Propagation Artificial Neural Networks and Logistic Regression Model in Predicting Poor Prognosis after Acute Ischemic Stroke. Open. Med. (Wars). 14, 324–330 (2019).
Asadi, H., Dowling, R., Yan, B. & Mitchell, P. Machine learning for outcome prediction of acute ischemic stroke post intra-arterial therapy. PLoS One. 9 (2), e88225 (2014).
Huang, R. et al. Stroke mortality prediction based on ensemble learning and the combination of structured and textual data. Comput. Biol. Med. 155, 106176 (2023).
Wang, S. et al. Multidimensional Cell-Free DNA Fragmentomic Assay for Detection of Early-Stage Lung Cancer. Am. J. Respir Crit. Care Med. 207 (9), 1203–1213 (2023).
Seminog, O. O., Scarborough, P., Wright, F. L., Rayner, M. & Goldacre, M. J. Determinants of the decline in mortality from acute stroke in England: linked national database study of 795 869 adults. BMJ. 365, l1778 (2019).
Rajan, K. B., Aggarwal, N. T., Wilson, R. S., Everson-Rose, S. A. & Evans, D. A. Association of cognitive functioning, incident stroke, and mortality in older adults. Stroke. 45 (9), 2563–2567 (2014).
Lorking, N. et al. Seasonality of stroke: Winter admissions and mortality excess: A Thailand National Stroke population database study. Clin. Neurol. Neurosurg. 199, 106261 (2020).
Kappen, T. H. et al. Adaptation of clinical prediction models for application in local settings. Med. Decis. Mak. 32 (3), E1–10 (2012 May-Jun).
Janssen, K. J., Vergouwe, Y., Kalkman, C. J., Grobbee, D. E. & Moons, K. G. A simple method to adjust clinical prediction models to local circumstances. Can. J. Anaesth. 56 (3), 194–201 (2009).
Moons, K. G. et al. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): explanation and elaboration. Ann. Intern. Med. 162 (1), W1–73 (2015).
Wang, W. et al. A systematic review of machine learning models for predicting outcomes of stroke with structured data. PLoS One. 15 (6), e0234722 (2020).
Naimi, A. I. & Balzer, L. B. Stacked generalization: an introduction to super learning. Eur. J. Epidemiol. 33 (5), 459–464 (2018).
Jiang, C. et al. A simple and easily implemented risk model to predict 1-year ischemic stroke and systemic embolism in Chinese patients with atrial fibrillation. Chin. Med. J. (Engl). 134 (19), 2293–2298 (2021).
Hojs Fabjan, T., Penko, M., Hojs, R. & Cystatin, C. creatinine, estimated glomerular filtration, and long-term mortality in stroke patients. Ren. Fail. 36 (1), 81–86 (2014).
Widhi Nugroho, A. et al. The Association between Glomerular Filtration Rate Estimated on Admission and Acute Stroke Outcome: The Shiga Stroke Registry. J. Atheroscler Thromb. 25 (7), 570–579 (2018).
Tsagalis, G. et al. Renal dysfunction in acute stroke: an independent predictor of long-term all combined vascular events and overall mortality. Nephrol. Dial Transpl. 24 (1), 194–200 (2009).
Cui, L. L. et al. Early neutrophil count relates to infarct size and fatal outcome after large hemispheric infarction. CNS Neurosci. Ther. 26 (8), 829–836. 10.1111/cns.13381 (2020).
Wang, L. et al. Neutrophil to lymphocyte ratio predicts poor outcomes after acute ischemic stroke: A cohort study and systematic review. J. Neurol. Sci. 406, 116445 (2019).
Chu, H. X. et al. Immune cell infiltration in malignant middle cerebral artery infarction: comparison with transient cerebral ischemia. J. Cereb. Blood Flow. Metab. 34 (3), 450–459 (2014).
Jickling, G. C. et al. Targeting neutrophils in ischemic stroke: translational insights from experimental studies. J. Cereb. Blood Flow. Metab. 35 (6), 888–901 (2015).
Akirov, A., Diker-Cohen, T., Steinmetz, T., Amitai, O. & Shimon, I. Sodium levels on admission are associated with mortality risk in hospitalized patients. Eur. J. Intern. Med. 46, 25–29. 10.1016/j.ejim.2017.07.017 (2017). Epub 2017 Jul 8. PMID: 28697950.
Adrogué, H. J., Madias, N. E. & Hypernatremia N Engl. J. Med. ;342(20):1493–1499. (2000).
Silva, R. F., Rodrigues, C. M. & Brites, D. Rat cultured neuronal and glial cells respond differently to toxicity of unconjugated bilirubin. Pediatr. Res. 51 (4), 535–541 (2002).
Wang, Y. et al. Association of serum neuron-specific enolase and bilirubin levels with cerebral dysfunction and prognosis in large-artery atherosclerotic strokes. J. Cell. Biochem. 119 (12), 9685–9693 (2018).
Xie, X. et al. Prognostic Value of International Normalized Ratio in Ischemic Stroke Patients without Atrial Fibrillation or Anticoagulation Therapy. J. Atheroscler Thromb. 26 (4), 378–387 (2019).
Sico, J. J. et al. Thrombocytopenia and in-hospital mortality risk among ischemic stroke patients. J. Stroke Cerebrovasc. Dis. 22 (7), e99–e102 (2013).
Tohgi, H., Suzuki, H., Tamura, K. & Kimura, B. Platelet volume, aggregation, and adenosine triphosphate release in cerebral thrombosis. Stroke. 22 (1), 17–21 (1991).
Mendelow, A. D. Stroke: pathophysiology, diagnosis, and management[M] (Elsevier Health Sciences, 2015).
Lv, G. et al. Influences of blood lipids on the occurrence and prognosis of hemorrhagic transformation after acute cerebral infarction: a case-control study of 732 patients. Mil Med. Res. 6 (1), 2 (2019).
Bang, O. Y. et al. UCLA MRI Permeability Investigators. Patterns and predictors of blood-brain barrier permeability derangements in acute ischemic stroke. Stroke. 40 (2), 454–461 (2009).
Reed, D. M. The paradox of high risk of stroke in populations with low risk of coronary heart disease. Am. J. Epidemiol. 131 (4), 579–588 (1990).
Cheung, C. M. et al. Using the National Institutes of Health Stroke Scale (NIHSS) to predict the mortality and outcome of patients with intracerebral haemorrhage. Hong Kong Med. J. 14 (5), 367–370 (2008).
Handschu, R. et al. Mortality prediction in critical care for acute stroke: Severity of illness-score or coma-scale? J. Neurol. 252 (10), 1249–1254 (2005).
Sène Diouf, F. et al. Survie des accidents vasculaires cérébraux comateux à Dakar (Sénégal) [Survival of comatose stroke victims in a neurological department in Dakar]. Rev. Neurol. (Paris). 164 (5), 452–458 (2008). French.
Chang, C. Y. et al. Reporting of aspiration pneumonia or choking as a cause of death in patients who died with stroke. Stroke. 44 (4), 1182–1185 (2013).
Zhang, L. et al. Clinical Course and Mortality of Stroke Patients With Coronavirus Disease 2019 in Wuhan, China. Stroke. 51 (9), 2674–2682 (2020).
Lewington, S. et al. Age-specific relevance of usual blood pressure to vascular mortality: a meta-analysis of individual data for one million adults in 61 prospective studies. Lancet. 360 (9349), 1903–1913 (2002).
Massera, D., Wang, D., Vorchheimer, D. A., Negassa, A. & Garcia, M. J. Increased risk of stroke and mortality following new-onset atrial fibrillation during hospitalization. Europace. 19 (6), 929–936 (2017).

No competing interests reported.

Predicting In-hospital of Death of Patients with Acute Stroke in the ICU Using Stacking Model

Status:

Version 1

Abstract

Figures

1. Introduction

2. Objects and Methods

2.1. Objects and Ethics of the study

2.2. Predictor variables and outcome indicators

2.3. Research Methods

2.3.1. Predictive Modeling Methodology

2.3.2. Statistical Methods

3. Results

3.1. General information of the study population

3.2. Feature screening and analysis of multifactorial logistic regression mortality risk factor

3.3. Analysis of Prediction Effectiveness Based on Machine Learning Models

3.4. Analysis of factors affecting hospital infections based on SHAP explanatory model

4. Discussion

Declarations

References

Additional Declarations

Status:

Version 1