A comparative study on the influencing factors and risk prediction models for stroke- associated pneumonia in patients with acute ischemic stroke and atrial fibrillation

doi:10.21203/rs.3.rs-5318884/v1

Background

To analyze the factors influencing stroke-associated pneumonia (SAP) in patients with acute ischemic stroke (AIS) and atrial fibrillation (AF), and to explore an optimal model for risk prediction.

Methods

Data were sourced from the Shandong Provincial Center for Disease Control and Prevention, encompassing all patients diagnosed with AIS and AF from 2020 to 2023. First, univariate analysis and LASSO (Least absolute shrinkage and selection operator) regression analysis methods were used to screen predictors. Secondly, the patients with AIS and AF were randomly divided into a training set, validation set, and test set in a ratio of 7:2:1, which were utilized for model training, model parameter adjustment, and model performance evaluation, respectively. The training set was balanced by synthetic minority oversampling technique (SMOTE), logistic regression, random forest (RF), and support vector machine (SVM),extreme gradient boosting (XGboost) models were constructed. Finally, we compared the models based on accuracy, sensitivity, specificity, AUC (area under the curve), and Youden index. We clarified the optimal prediction model and influencing factors ,the nomogram for risk prediction was constructed for SAP in patients with AIS and AF.

Results

Among the 4496 patients with AIS and AF, SAP was identified in 10.16% of cases. In the test set, the AUC for logistic regression, RF, SVM, and XGboost models were 0.866, 0.817, 0.816, and 0.838, respectively. The most predictive factors included coronary heart disease [OR = 1.05 (1.03, 1.07), p < 0.001], hypertension [OR = 1.05 (1.04, 1.07), p < 0.001], consciousness disorder [OR = 1.19 (1.16–1.23), p < 0.001], cognitive impairment [OR = 1.10 (1.08–1.13), p < 0.001], limb movement disorder [OR = 1.07 (1.04–1.09), p < 0.001], dysphagia [OR = 1.13 (1.08–1.19), p < 0.001], nasal feeding [OR = 0.95 (0.92–0.98), p = 0.003], and oxygen intake [OR = 0.65 (0.62–0.67), p < 0.001]. the nomogram average absolute error of calibration curve was 0.014.

Conclusions

Coronary artery disease, hypertension, consciousness disorder, cognitive impairment, limb movement disorder, and dysphagia were identified as independent risk factors for SAP in patients with AIS and AF. In contrast, nasal feeding and oxygen intake served as independent protective factors. The logistic regression model demonstrated the best predictive performance for SAP in patients with AIS and AF compared to RF, SVM, and XGboost models. The risk prediction model established by nomogram can better predict the risk of SAP.

acute ischemic stroke

atrial fibrillation

stroke-associated pneumonia

machine learning

nomogram

Stroke is the second leading cause of death worldwide and the primary cause of disability-adjusted life years (DALYs) in China.^[1] Acute ischemic stroke (AIS) is a critical nervous system disorder caused by thrombosis and embolism that block cerebral arteries,^[3] accounting for 85% of stroke types. AIS is associated with the development of neuroinflammation and may also arise from cerebral hemodynamic abnormalities. Atrial fibrillation (AF) is a significant contributor to the development of stroke.^[4] Approximately 24% of patients with AIS also have AF, and individuals with AF face a risk of AIS that is 4–5 times higher than those without AF,^[5] which increase the overall risk of AIS for these patients. Additionally, the inflammatory response plays a crucial role in both AIS and AF, as well as their related complications.^[6]

Due to the suppression of immune function following stroke, individuals are at an increased risk of developing lung infections. Stroke-associated pneumonia (SAP) is a common and serious complication that contribute to morbidity and mortality in stroke patients.^[7] The global incidence rate of SAP ranges frame 7–38%,^[8] with the acute mortality rate reaching as high as 30% -40%.^[12] This significantly impacts patients prognosis, prolongs hospitalization time, and increases the medical burden.^[13] In addition, research has shown that pneumonia occurs significantly more frequently in patients with AF compared to those without AF, with rates of 9.8% versus 5.3%, respectively.^[16]

As a core technology of artificial intelligence, machine learning (ML) refers to the ability to recognize patterns and learn from data.^[17] In recent years, it has been widely applied in various areas, including disease prediction, disease prognosis assessment, disease-assisted diagnosis, and health management.^[18] ML models, including Extreme Gradient Boosting (XGboost), Support Vector Machines (SVM), Random Forest (RF), Logistic Regression (LR), and Deep Neural Networks (DNN), can capture complex non-linear relationships and identify unknown correlations in big data, providing deeper insights. RF is a non-parametric method based on the bagging principle, which adopts the outputs of integrating algorithms of multiple decision trees. This approach enhances the overall model performance by strengthening multiple weak classifiers, making it more effective than a single decision tree.^[19] SVM introduces the concept of kernel functions and employs the principle of structural risk minimization. This allows for nonlinear decision-making in the original space by identifying a linear hyperplane within a high-dimensional space.^[20] The XGboost algorithm is a large-scale machine learning algorithm that represents an efficient and extensible variant of gradient enhancement.^[21]

The purpose of this study is to compare the predictive performance of various ML models for SAP in patients with AIS and AF by using predictive factors effectively. This approach aims to facilitate the early identification of high-risk groups in clinical settings, support individualized treatment, improve patient outcomes, and reduce social burden.

2.1 Data sources

This study was designed as a case-control study. All data were obtained from the network management system of chronic disease monitoring information managed by Shandong Provincial Center for Disease Control and Prevention. After thorough communication and selection, 24 hospitals across 8 cities in Shandong Province were chosen by the center (Table 1). These medical institutions reported relevant cases, and the center performed a comprehensive review of the data. This study utilized data collected by the Shandong Provincial Center for Disease Control and Prevention from 2020 to 2023. The subjects were patients diagnosed with AIS and AF.

Table 1

Information on 24 hospitals across 8 cities in Shandong Province
City	Hospitals
Jinan	Shandong Provincial Qianfoshan Hospital, Jinan Shizhong District People's Hospital, Shanghe People's Hospital
Qingdao	Qingdao Central Medical Hospital, Qingdao Chengyang People's Hospital, Qingdao West Coast New Area District Hospital
Yantai	Yantai Yuhuangding Hospital, Yantai Penglai People's Hospital, Haiyang People's Hospital
Weifang	Weifang People's Hospital, Weifang No.2 People's Hospital, Qingzhou People's Hospital
Jining	Jining No.3 People's Hospital, Yanzhou District People's Hospital, Jining City, Wenshang County People's Hospital
Dezhou	Dezhou People's Hospital, Qilu Hospital of Shandong University Dezhou Hospital, Qihe County People's Hospital
Liaocheng	Liaocheng People's Hospital, The People's Hospital of Chiping District, Linqing People's Hospital
Linyi	Linyi People's Hospital, First People's Hospital of Tancheng, Yishui County People's Hospital

2.2 Study samples

In this study, a total of 10967 patients (n = 10967) diagnosed with AIS and AF were identified. Among those, 6471 were excluded according to the following criteria: 2341 were under 18 years of age, 2028 patients had a hospital stay of less than 24 hours, 356 had an infection within one week prior to the onset of AIS, 337 had a tumor, and 1139 had missing values for variable exceeding 20%. Ultimately, 4496 patients with AIS and AF were enrolled in this study, including 457 cases of SAP during hospitalization, and 4039 cases did not develop SAP.

2.3 Diagnosis of AIS, AF, and SAP

Inclusion criteria: (1) Meet the diagnostic criteria for AIS as outlined in the Chinese Guidelines for the Diagnosis and Treatment of AIS; (2) Diagnosed with AF either during the current screening electrocardiogram (12-lead electrocardiogram or single-lead electrocardiogram with AF rhythm ≥ 30s); or through prior medical diagnoses, electrocardiograms, or medical records; (3) Over 18 of age.

Exclusion criteria: (1) Automatically discharged from hospital within 24 hours of admission or died during hospitalization; (2) Lung infection prior to or shortly before the onset of stroke; (3) With infections affecting other tissues and organs; (4) With cancer, severe liver or kidney dysfunction, severe hematological disorders, or autoimmune diseases; (5) With incomplete medical records.

Diagnostic criteria for SAP Refer to the "Chinese Expert Consensus on the Diagnosis and Treatment of SAP" and meet at least one of the following criteria: (1) Fever ≥ 38°C without identifiable alternative causes; (2) Decreased (≤ 4×10⁹/L) or increased (≥ 10×10⁹/L) peripheral blood leukocyte counts; (3) Elderly persons aged ≥ 70 years with a sudden change in consciousness. Additionally, patients must fulfill at least 2 of the following secondary criteria: (1) Newly developed cough, increased respiratory rate, or even difficulty breathing; (2) New sputum production or changes in sputum within 24 hours; (3) Presence of rales, bronchial breath sounds, or crackling sounds in the lungs; (4) Impaired gas exchange. Chest imaging should demonstrate at least one of the following findings: (1) New or progressive infiltrating shadows; (2) New or progressive solid shading; (3) New or progressive ground-glass shadows. (Note: for patients without prior cardiopulmonary conditions, a single chest imaging examination showing any one of the above manifestations may be sufficient).

2.4 Collection of medical history

Demographic data: age, sex, smoking, drinking, medical history (whether it was the first cerebral infarction, coronary heart disease, hypertension, diabetes, duration of AF), medication history (anticoagulants, antiplatelet medications, antihypertensive medications, lipid-lowering medications, heart rate controlling medications), and hospital day.

Clinical data: body temperature, heart rate, respiration rate, diastolic blood pressure (DBP), systolic blood pressure (SBP), admission date, consciousness disorders, cognitive disorders, limb movement disorders, dysphagia, oxygen intake, nasal feeding requirements.

Laboratory tests: red blood cell count (RBC), white blood cell count (WBC), hemoglobin (HB), hematocrit (HCT), platelet count (PLT), plateletcrit (PCT), neutrophil count (NEUT), lymphocyte count (LYM), monocyte count (MONO), eosinophil (EO), basophil (BA), neutrophil ratio (NEUT%), C-reactive protein (CPR), international normalized ratio (INR), prothrombin time (PT), activated partial prothrombin time (APTT), D-dimer (DD), fibrinogen (FIB), albumin (ALB), direct bilirubin (DBIL), creatinine (Cr), total cholesterol (TC), homocysteine (HCY), low-density lipoprotein (LDL), high-density lipoprotein (HDL).

Inflammatory markers calculated using formulae: PLR = platelet count/lymphocyte count, NLR = neutrophil count/lymphocyte count, MLR = monocyte count/lymphocyte count, NRAP = neutrophil percentage/albumin, SIRI=(neutrophil count×monocyte count)/lymphocyte count, SII = platelets count×(neutrophil count/lymphocyte count),CAR = C-reactive protein/albumin.

2.5 Data analysis

Before establishing the model, whether SAP occurred in patients with AIS and AF was evaluated. These patients were divided into two groups: the SAP group and the non-SAP group. Univariate analysis was used to describe the demographic information, clinical data, and laboratory indicators of AIS and AF patients. The statistically significant variables identified from univariate analyses were then included in the LASSO (Least absolute shrinkage and selection operator) regression to determine the predictors for inclusion in the model. When building the model, AIS and AF patients were randomly divided into three sets: the training set, verification set, and test set, according to the ratio of 7:2:1. Due to the significant imbalance between the positive and negative samples, a synthetic minority oversampling technique (SMOTE) was utilized to balance the datasets within training set.

The predictors were included in the training set after SMOTE balance, which was used to construct logistic regression, RF, SVM, and XGboost models. The verification set was employed to adjust model parameters and optimize model performance. In the logistic regression, a stepwise regression approach was utilized for multifactor logistic regression analysis. The RF model utilized bootstrap sampling along with five-fold cross-validation. For the SVM model, a grid search method combined with ten-fold cross-validation was applied. The XGBoost model incorporated hyperparameter optimization along with ten-fold cross-validation to enhance its performance.

The test set was used to evaluate the performance of four models. We compared the receiver-operating-characteristic (ROC) curves among the models and computed various metrics, including accuracy, sensitivity, specificity, Youden index, and AUC (area under the receiver operating characteristic curve) value. This analysis assisted to identify the best predictive model, forest map and nomogram to clarify independent influences on the occurrence of SAP in patients with AF and AIS. Calibration curve to evaluate the performance of a nomogram.

2.6 Statistical methods

This study described the characteristics of various datasets and performed different satistical tests. For continuous data, we utilized means and standard deviations, or medians and quartiles, to describe the variables. The Kruskal-Wallis rank sum test was applied to compare differences among groups. For categorical data, we described the data using rates and absolute numbers, employing the chi-square test to assess differences between groups. A two-sided P-value < 0.05 was considered statistically significant. The predictive ability of models was determined based on the AUC value, with the best cutoff point defined as the one that maximized the Youden index. ROC analyses and calibration curves were used to evaluate the predictive capabilities of the models. Calibration curves are used to evaluate the performance of the nomogram. The more diagonally the curve fits the calibration chart, the better the agreement between the observed results and the predicted probabilities. Statistical analyses were performed using IBM SPSS version 26.0, and R version 4.3.2.

3.1 Characteristics of the study population

A total of 4496 patients with AIS and AF were included in this study. The average age was 73.56 ± 9.95 years (mean ± standard), with 2679 (59.59%) patients being male. Among the total sample, 457 (10.16%) patients developed SAP. The baseline characteristics of the SAP patients were shown in Table 2. Patients in the SAP group had higher age, higher prevalence of coronary heart disease, hypertension, and diabetes. They also experienced longer hospital stays, and exhibited more frequent consciousness disorders, cognitive impairments, limb movement disorders, dysphagia, nasal feeding requirements, and oxygen intake. Additionally, the SAP group showed significantly elevated levels of body temperature, heart rate, respiration rate, systolic blood pressure, diastolic blood pressure, neutrophil count, neutrophil ratio, NLR, NPAR, SIRI, and SII. In contrast, the levels of hematocrit were significantly lower in the SAP group compared to the non-SAP group (P < 0.05).

Table 2

Baseline characteristics of patients with AIS and AF in the two groups
	SAP (n = 4039)	Non-SAP (n = 457)	t/x²/Z	p
Demographics
Age(years,x̄±s) Sex [n%]	73.4 ± 10	75.1 ± 9.7	-3.51 1.72	< 0.001 0.181
Female	2420 (59.9)	259 (56.7)
Male	1619 (40.1)	198 (43.3)
Smoking, n (%)	670 (16.6)	76 (16.6)	0.001	0.982
Drinking, n (%)	730 (18.1)	91 (19.9)	0.930	0.335
Cerebral infarction, n (%)	1151 (28.5)	150 (32.8)	3.76	0.053
Coronary heart disease, n (%)	814 (20.2)	160 (35)	53.4	< 0.001
Diabetes, n (%)	460 (11.4)	98 (21.4)	38.19	< 0.001
Hypertension, n (%)	1235 (30.6)	214 (46.8)	49.63	< 0.001
Duration of AF (years, x̄±s)	7.38 ± 4.17	7.76 ± 4.86	-1.80	0.072
Anticoagulants, n (%)	180 (4.5)	27 (5.9)	1.65	0.158
Antiplatelet drugs, n (%)	207 (5.1)	18 (3.9)	0.98	0.309
Antihypertensive drugs, n (%)	140 (3.5)	20 (4.4)	0.74	0.349
Lipid lowering drugs, n (%)	151 (3.7)	11 (2.4)	1.73	0.184
Antiarrhythmic drugs, n (%)	34 (0.8)	7 (1.5)	1.47	0.185
Clinical data
Hospital day (days, x̄±s)	9.09 ± 5.30	11.07 ± 6.44	-6.80	< 0.001
Consciousness disorder, n (%)	263 (6.5)	209 (45.7)	672.15	< 0.001
Cognitive disorder, n (%)	748 (18.5)	318 (69.6)	591.85	< 0.001
limb movement disorders, n (%)	1013 (25.1)	335 (73.3)	454.79	< 0.001
Dysphagia, n (%)	67 (1.5)	69 (15.1)	252.79	< 0.001
Nasal feeding, n (%)	177 (4.4)	105 (23)	241.44	< 0.001
Oxygen uptake, n (%)	64 (1.6)	121 (26.5)	644.77	< 0.001
Temperature (℃,x̄±s)	36.37 ± 0.31	36.46 ± 0.43	-4.17	< 0.001
Respiration rate (Times/minute,x̄±s)	18.83 ± 1.83	19.24 ± 2.71	-3.09	< 0.002
Heart rate (Times/minute,x̄±s)	88.60 ± 17.87	91.64 ± 20.23	-3.12	< 0.002
SBP (mmHg, x̄±s)	141.02 ± 20.04	146.35 ± 24.43	-4.50	< 0.001
DBP (mmHg, x̄±s)	84.16 ± 12.20	86.20 ± 15.17	-2.78	< 0.006
Admission date (h, x̄±s)	7.85 ± 6.27	7.96 ± 6.47	-0.35	0.725
Laboratory tests
RBC (10^12/L)	4.51 ± 1.22	4.47 ± 1.18	0.73	0.465
WBC(10^9/L)	7.41 ± 3.56	7.64 ± 3.69	-1.33	0.184
NEUT (10^9/L)	6.06 (4.41,7.43)	6.55 (4.65,8.00)	-3.97	< 0.001
LYM (10^9/L)	1.98 ± 1.48	2.13 ± 1.77	-1.74	0.082
MONO (10^9/L)	0.54 (0.38,0.70)	0.54 (0.36,0.71)	-0.29	0.769
EO (10^9/L)	0.23 ± 0.29	0.22 ± 0.29	0.59	0.554
BA (10^9/L)	0.45 ± 0.33	0.48 ± 0.34	-1.72	0.085
HB (g/L)	130.75 ± 18.79	129.68 ± 19.92	1.149	0.251
NEUT% (%)	64.75 ± 8.05	65.58 ± 6.55	-2.48	0.013
PLT (10^9/L)	194 (157,232)	194 (158,233)	-0.06	0.953
PCT (%)	0.21 ± 0.06	0.21 ± 0.07	-1.23	0.218
HCT (L/L)	0.40 ± 0.70	0.39 ± 0.07	2.22	0.026
CPR (mg/L)	5.58 ± 3.25	5.62 ± 3.29	-0.26	0.796
INR	1.14 ± 0.19	1.14 ± 0.17	0.75	0.452
PT (s)	12.91 ± 1.48	12.97 ± 1.45	-0.74	0.457
APTT (s)	31.70 ± 4.92	31.96 ± 4.92	-1.09	0.276
DD (mg/L)	0.36 ± 0.21	0.36 ± 0.22	0.33	0.735
FIB (g/L)	3.24 ± 0.81	3.28 ± 0.83	-1.08	0.280
ALB (g/L)	37.42 ± 4.86	37.05 ± 5.25	1.47	0.143
DBIL (µmol/L)	4.78 ± 2.23	4.79 ± 2.23	-0.44	0.965
Cr (mg/dL)	76.74 ± 25.68	75.84 ± 27.37	0.71	0.480
TC (mmol/L)	4.02 ± 1.09	3.97 ± 1.04	0.88	0.377
LDL (mmol/L)	2.42 ± 0.82	2.39 ± 0.81	0.66	0.509
HDL (mmol/L)	1.21 ± 0.27	1.21 ± 0.28	0.34	0.731
HCY (µmol/L)	13.20 ± 3.28	13.18 ± 3.49	0.12	0.906
PLR	125.00 (89.22,171.81)	129.41 (79.22,181.66)	-0.68	0.497
NLR	4.13 ± 2.52	4.49 ± 2.91	-2.55	0.011
MLR	0.36 ± 0.23	0.38 ± 0.25	-1.23	0.220
NPAR	1.73 (1.55,1.95)	1.80 (1.62,2.00)	-3.76	< 0.001
SIRI	2.30 ± 1.97	2.56 ± 2.28	-2.36	0.019
SII	822.53 ± 587.33	899.73 ± 678.67	-2.34	0.020
CAP	0.15 ± 0.09	0.16 ± 0.10	-0.88	0.401

3.2 Predictors screened by LASSO regression analysis

Using SAP as the dependent variable, we included 23 statistically significant variables (P < 0.05) identified in prior univariate analysis as independent variables. LASSO regression was employed to screen for predictors, and ten-fold cross-validation was conducted to determine the optimal λ value, which was found to be 0.00234. Ultimately, eight variables were identified as having the best performance with the least number of variables: coronary heart disease, hypertension, consciousness disorder, cognitive impairment, limb movement disorder, dysphagia, nasal feeding requirement, and oxygen intake. The process of selecting coefficients for the LASSO regression variables was shown in Figs. 1A and 1B. These findings highlighted critical risk factors associated with the development of SAP in patients with AIS and AF, aiding in clinical decision-making and targeted interventions.

3.3 Model construction, optimization, and evaluation

AIS and AF patients (n = 4496) were randomly divided into three sets: the training set (n = 3147), verification set (n = 905), and test set (n = 440), according to the ratio of 7:2:1. In the training set, there were 312 positive samples (SAP) and 2835 negative samples (non-SAP). We utilized SMOTE to balance the datasets within training set, construct logistic regression, RF, SVM, and XGboost models

Optimize the model in validation set, AUC value in ROC was suitable to evaluate the predictive performance of unbalanced samples. The AUC (95% CI) for logistic regression, RF, SVM, and XGboost models were 0.891 (0.855–0.927), 0.868 (0.826–0.906), 0.847 (0.821–0.864), and 0.877 (0.836–0.918), respectively; evaluate the model in test set, the AUC (95% CI) for logistic regression, RF, and SVM, XGboost models were 0.866 (0.816–0.917), 0.818 (0.728–0.877), 0.817 (0.780–0.820), and 0.838 (0.780–0.896), respectively(Fig. 2A,2B). Logistic regression had the best predictive performance (AUC = 0.866), highest accuracy (90.13%), sensitivity (79.49%)and Youden index (0.63) in models(Table 3). Among these, the LR model demonstrated the optimal performance in both the validation set and the test set.

Note

AUC is the area under the ROC curve, and the closer the AUC value is to 1, the better the model performs ,Conversely, the closer the AUC value is to 0, the worse the model performance.

Table 3

Prediction performance of four models for SAP in patients with AIS and AF.
	Accuracy (%)	Sensitivity (%)	Specificity (%)	AUC (95%CI)	Youden index
Validation set Logistic regression RF SVM XGboost Test set Logistic regression RF SVM XGboost	91.36 87.49 87.04 87.26 90.13 84.98 84.30 84.61	88.42 78.09 82.10 86.32 79.49 70.00 78.03 66.01	77.97 88.42 86.39 79.33 83.04 86.11 82.03 88.64	0.891 (0.855–0.927) 0.868 (0.826–0.906) 0.847 (0.821–0.864) 0.877 (0.836–0.918) 0.866 (0.816–0.917) 0.818 (0.728–0.877) 0.817 (0.780–0.820) 0.838 (0.780–0.896)	0.66 0.67 0.68 0.66 0.63 0.56 0.60 0.55

3.4 Optimal model and independent influences

The findings from the logistic regression model indicated that coronary heart disease, hypertension, consciousness disorder, cognitive impairment, limb movement disorders, dysphagia, nasal feeding, and oxygen intake were independent factors influencing the occurrence of SAP in patients with AIS and AF (P < 0.05). Specifically, coronary heart disease, hypertension, consciousness disorder, cognitive impairment, limb movement disorders, and dysphagia were identified as the risk factors for SAP in patients with AIS and AF (OR > 1). In contrast, nasal feeding (OR = 0.95), and oxygen intake (OR = 0.65) were recognized as protective factors (OR < 1) (Table 4,Fig. 3).

To facilitate the clinical service, we converted the complex mathematical model into a nomogram (Fig. 4). It was necessary to sum the scores of variables included in the model. And then a vertical line at the total score was drawn and making it intersect with the one lines representing the predicted SAP. The calibration curve showed that the logistic regression model predicted the risk of SAP in patients with AIS and AF with good consistency with the actual risk of SAP, and the mean absolute error in the validation set and the test set were 0.012 and 0.014, respectively(Fig. 5A,5B). It shows that the nomogram had good distinguishing ability.

Table 4

Logistic regression analysis results
	\(\:\beta\:\)	Standard error	WaldX²	OR(95%CI)	P
Coronary heart disease No				1.00 (Reference)
Yes	0.6681	0.1378	4.848	1.05 (1.03,1.07)	< 0.001
Hypertension No Yes Consciousness disorder No Yes Cognitive impairment No Yes Limb movement disorders No Yes Dysphagia No Yes Nasal feeding No Yes Oxygen intake No Yes	0.7941 1.0186 1.1442 0.9721 0.6324 -0.3337 -2.8299	0.1324 0.1504 0.1551 0.1482 0.2260 0.1833 0.1981	5.996 6.775 7.377 6.561 2.798 -1.831 -14.288	1.00 (Reference) 1.05 (1.04,1.07) 1.00 (Reference) 1.19 (1.16,1.23) 1.00 (Reference) 1.10 (1.08,1.13) 1.00 (Reference) 1.07 (1.04,1.09) 1.00 (Reference) 1.13 (1.08,1.19) 1.00 (Reference) 0.95 (0.92,0.98) 1.00 (Reference) 0.65 (0.62,0.67)	< 0.001 < 0.001 < 0.001 < 0.001 < 0.001 0.003 < 0.001

AIS pathophysiological mechanisms are closely related to immunity, it can disrupt the balance between immunity and the central nervous system by activating the autonomic nervous system and the stress axis, which leads to secondary immune deficiency and increases the risk of infections and SAP.^[23] After AIS, the inflammatory response acts as a defense mechanism against infection, promoting tissue regeneration and removal of necrotic cells. However, an excessive inflammatory response can lead to secondary injury. Pneumonia is the most common type of infection following AIS and significantly impacts the recovery of neurological function.^[24]

Immunoinflammatory indicators play a crucial role in the inflammatory response. The neutrophil-lymphocyte ratio (NLR) and platelet-lymphocyte ratio (PLR) are commonly used markers in clinical practice to assess the degree of immune-inflammatory response.^[25] The systemic immunoinflammatory index (SII), which integrates peripheral blood neutrophil, platelet, and lymphocyte counts, serves as a novel biomarker for malignancies and inflammatory diseases, frequently utilized to evaluate systemic inflammatory status. In this study, SII and SIRI were introduced as markers to assess the overall immune and inflammatory status of SAP in patients with AIS and AF. Previous research has demonstrated that SIRI, SII, and NLR were more predictive of pneumonia than traditional inflammatory factors.^[26] However, our research findings were inconsistent with the previous researches. This discrepancy may stem from the incomplete understanding of the pathogenesis of AF, where atrial electrical remodeling serves as a crucial pathophysiological mechanism.^[27]

Patients with AF typically exhibit weakened immune function and are prone to inducing inflammation. At the same time, inflammation may promote the development of AF, thus creating a vicious circle between the two.^[28] Numerous studies have confirmed the various immune-inflammatory indexes, such as CPR, interleukins, white blood cell count, are significantly higher in patients with AF compared to those without AF.^[29] This elevation in inflammatory markers may be attributed to the presence of underlying conditions in these patients, leading to the increased levels of inflammatory factors even before the onset of acute cerebral infarction. Consequently, immunoinflammatory indicators may not accurately reflect changes in the inflammatory state of patients with AIS with AF during the onset of SAP.

Hypertension, diabetes, consciousness disorder, dysphagia, cognitive impairment, and limb movement disorders identified in this study along with previous findings and are acknowledged to increase the risk of SAP.^[30] These comorbidities can contribute to a more complex clinical picture, making patients more vulnerable to respiratory complications.

Nasal feeding remains a controversial topic. It is typically used in patients with dysphagia to provide essential nutritional support, helping to mitigate the risk of malnutrition and related complications.^[32] In clinical practice, enteral nutrition is generally considered as the first choice, and feeding via nasogastric tube is more common in Asian countries.^[33] Studies have confirmed that nasal feeding in patients with massive cerebral infarction could effectively correct metabolic disorders, promote neurological recovery, and reduce the occurrence of related infections.^[34] The findings of our study align with previous study, indicating that nasal feeding may act as a protective factor against SAP in patients with AIS and AF, thereby reducing the incidence of this complication. However, prolonged use of nasogastric feeding may increase the risk of nasal infection, thereby raising the likelihood of SAP in turn.^[35] Therefore, it is crucial to maintain cleanliness of the nasal passages and to regularly change the nasal tubes to minimize the chance of infection when implementing nasal feeding therapy.

In this study, oxygen intake was identified as a risk factor for SAP in patients with AIS and AF. Since its introduction in 1855, supplemental oxygen has been widely used in acute care, and physicians consider it a harmless and potentially beneficial treatments, even in the absence of hypoxemia.^[36] Our research results also demonstrated the same conclusion. However, a systematic review and meta-analysis of more than 16000 patients with acute illnesses have indicated that supplemental oxygen levels exceeding the range of SpO₂ (94%-96%) may lead to vasoconstriction in the pulmonary, cardiovascular, and neurological systems, as well as inflammatory responses, and oxidative stress, potentially resulting in life-threatening conditions.^[39] Therefore, when administering oxygen to improve clinical symptoms in patients with acute cerebral infarction, it is crucial for clinicians to conduct thorough assessment before providing oxygen, in order to minimize the risk of pulmonary inflammation.

This study possesses several limitations. Firstly, due to the varying levels of hospital comprehensiveness and differing judgmental criteria among clinicians, CHA2DS2-VASc scores, NIHSS scores, and MRS scores were not included, which limited the ability to assess the severity of the patients' conditions. Secondly, while the target population comprised patients with AIS and AF, it remains unverified whether the model can be generalized to the border population, highlighting the need to consider its applicability across different demographics. Lastly, this study was validated and tested in 24 hospitals across 8 cities in Shandong Province, and further validation is required to determine its relevance in other regions.

In this study, we used a multi-center, large sample, case-control design to compare multiple ML models. Our aim was to determine the independent influencing factors and risk factors for SAP in patients with AIS and AF, as well as to determine the effective model. The research findings are illustrated through a nomogram, which serves an objective clinical risk assessment tool for developing individualized treatment plans. The primary objective is to establish a decision-making framework of high-risk patients, allows for the adjustment of therapeutic strategies, enhances patient prognosis, and reduces the overall burden associated with stroke. Additionally, this study will also contribute to inproving SAP screening in patients with AIS and AF while assisting in the identification of key intervention targets.

SAP	Stroke-associated pneumonia
AIS	Acute ischemic stroke
AF	Atrial fibrillation
LASSO	Least absolute shrinkage and selection operator
SMOTE	Synthetic minority oversampling technique
RF	Random forest
SVM	Support vector machine
XGboost	Extreme Gradient Boosting
DBP	Diastolic blood pressure
SBP	Systolic blood pressure
RBC	Red blood cell count
WBC	White blood cell count
HB	Hemoglobin
HCT	Hematocrit
PLT	Platelet count
PCT	Plateletcrit
NEUT	Neutrophil count
LYM	Lymphocyte count
MONO	Monocyte count
EO	Eosinophil
BA	Basophil
NEUT%	Neutrophil ratio
CPR	C-reactive protein
INR	International normalized ratio
PT	Prothrombin time
APTT	Activated partial prothrombin time
DD	D-dimer
FIB	Fibrinogen
ALB	Albumin
DBIL	Direct bilirubin
Cr	Creatinine
TC	Total cholesterol
HCY	Homocysteine
LDL	Low-density lipoprotein
HDL	High-density lipoprotein
ROC	Receiver-operating-characteristic
AUC	Area under the receiver operating characteristic curve
CI	Confidence interval

Author Contribution

All authors contributed to the conception and design of the study.

GBD 2019 Stroke Collaborators. Global, regional, and national burden of stroke and its risk factors, 1990–2019: a systematic analysis for the Global Burden of Disease Study 2019 [J]. Lancet Neurol. 2021;20(10):795–820. 10.1016/S1474-4422(21)00252-0.
Zhao Y, Hua X, Ren X, et al. Increasing burden of stroke in China: A systematic review and meta-analysis of prevalence, incidence, mortality, and case fatality. Int J Stroke. 2023;18(3):259–67. 10.1177/17474930221135983.
Cerami C, Perani D. Imaging neuroinflammation in ischemic stroke and in the atherosclerotic vascular disease. Curr Vasc Pharmacol. 2015;13(2):218–22. 10.2174/15701611113116660168.
Kelley RE, Kelley BP. Heart-Brain Relationship in Stroke. Biomedicines. 2021;9(12):1835. 10.3390/biomedicines9121835. Published 2021 Dec 4.
Kimura T, Kashimura S, Nishiyama T, et al. Asymptomatic Cerebral Infarction During Catheter Ablation for Atrial Fibrillation: Comparing Uninterrupted Rivaroxaban and Warfarin (ASCERTAIN). JACC Clin Electrophysiol. 2018;4(12):1598–609. 10.1016/j.jacep.2018.08.003.
Xu M, Wang J, Zhan C, et al. Association of follow-up neutrophil-to-lymphocyte ratio and systemic inflammation response index with stroke-associated pneumonia and functional outcomes in cerebral hemorrhage patients: a case-controlled study. Int J Surg. 2024;110(7):4014–22. 10.1097/JS9.0000000000001329. Published 2024 Jul 1.
Gittins M, Lobo Chaves MA, Vail A, Smith CJ. Does stroke-associated pneumonia play an important role on risk of in-hospital mortality associated with severe stroke? A four-way decomposition analysis of a national cohort of stroke patients. Int J Stroke. 2023;18(9):1092–101. 10.1177/17474930231177881.
Ji R, Wang D, Shen H, et al. Interrelationship among common medical complications after acute stroke: pneumonia plays an important role. Stroke. 2013;44(12):3436–44. 10.1161/STROKEAHA.113.001931.
Smith CJ, Bray BD, Hoffman A et al. Can a novel clinical risk score improve pneumonia prediction in acute stroke care? A UK multicenter cohort study. J Am Heart Assoc. 2015;4(1):e001307. Published 2015 Jan 13. 10.1161/JAHA.114.001307
Teh WH, Smith CJ, Barlas RS, et al. Impact of stroke-associated pneumonia on mortality, length of hospitalization, and functional outcome. Acta Neurol Scand. 2018;138(4):293–300. 10.1111/ane.12956.
Yu YJ, Weng WC, Su FC, et al. Association between pneumonia in acute stroke stage and 3-year mortality in patients with acute first-ever ischemic stroke. J Clin Neurosci. 2016;33:124–8. 10.1016/j.jocn.2016.02.039.
Bai Q, Liu J, Wang G. Ferroptosis, a Regulated Neuronal Cell Death Type After Intracerebral Hemorrhage. Front Cell Neurosci. 2020;14:591874. 10.3389/fncel.2020.591874. Published 2020 Nov 16.
Teh WH, Smith CJ, Barlas RS, et al. Impact of stroke-associated pneumonia on mortality, length of hospitalization, and functional outcome. Acta Neurol Scand. 2018;138(4):293–300. 10.1111/ane.12956.
Yu YJ, Weng WC, Su FC, et al. Association between pneumonia in acute stroke stage and 3-year mortality in patients with acute first-ever ischemic stroke. J Clin Neurosci. 2016;33:124–8. 10.1016/j.jocn.2016.02.039.
Koennecke HC, Belz W, Berfelde D, et al. Factors influencing in-hospital mortality and morbidity in patients treated on a stroke unit. Neurology. 2011;77(10):965–72. 10.1212/WNL.0b013e31822dc795.
Keller K, Hobohm L, Wenzel P, Münzel T, Espinola-Klein C, 29., Ostad MA et al. Impact of atrial fibrillation/flutter on the inhospital mortality of ischemic stroke patients. Heart Rhythm. 2020; 17(3):383 – 90. 10.1016/j.hrthm.2019. 10.001.
Ting Sim JZ, Fong QW, Huang W, Tan CH. Machine learning in medicine: what clinicians should know. Singap Med J. 2023;64(2):91–7. 10.11622/smedj.2021054.
Shurrab S, Guerra-Manzanares A, Magid A, Piechowski-Jozwiak B, Atashzar SF, Shamout FE. Multimodal Machine Learning for Stroke Prognosis and Diagnosis: A Systematic Review. IEEE J Biomed Health Inf Published online August. 2024;22. 10.1109/JBHI.2024.3448238.
Hu J, Szymczak S. A review on longitudinal data analysis with random forest. Brief Bioinform. 2023;24(2):bbad002. 10.1093/bib/bbad002.
Si Q, Yang Z, Ye J. Symmetric LINEX loss twin support vector machine for robust classification and its fast iterative algorithm. Neural Netw. 2023;168:143–60. 10.1016/j.neunet.2023.08.055.
Li Y, Zou Z, Gao Z, et al. Prediction of lung cancer risk in Chinese population with genetic-environment factor using extreme gradient boosting. Cancer Med. 2022;11(23):4469–78. 10.1002/cam4.4800.
Meisel C, Schwab JM, Prass K, Meisel A, Dirnagl U. Central nervous system injury-induced immune deficiency syndrome. Nat Rev Neurosci. 2005;6(10):775–86. 10.1038/nrn1765.
Hotter B, Hoffmann S, Ulm L, et al. Inflammatory and stress markers predicting pneumonia, outcome, and etiology in patients with stroke: Biomarkers for predicting pneumonia, functional outcome, and death after stroke. Neurol Neuroimmunol Neuroinflamm. 2020;7(3):e692. 10.1212/NXI.0000000000000692. Published 2020 Feb 25.
Liu DD, Chu SF, Chen C, Yang PF, Chen NH, He X. Research progress in stroke-induced immunodepression syndrome (SIDS) and stroke-associated pneumonia (SAP). Neurochem Int. 2018;114:42–54. 10.1016/j.neuint.2018.01.002.
Chang TI, Wu X, Boström KI, Tran HA, Friedlander AH. Red Cell Distribution Width, Unlike Neutrophil Lymphocyte Ratio Is Unable to Accurately Gauge Enhanced Systemic Inflammation Associated With Panoramic Imaged Carotid Plaque. Mil Med. 2021;186(1–2):39–43. 10.1093/milmed/usaa302.
Wang RH, Wen WX, Jiang ZP, Du ZP, Ma ZH, Lu AL, Li HP, Yuan F, Wu SB, Guo JW, Cai YF, Huang Y, Wang LX, Lu HJ. The clinical value of neutrophil-to-lymphocyte ratio (NLR), systemic immune-inflammation index (SII), platelet-to-lymphocyte ratio (PLR) and systemic inflammation response index (SIRI) for predicting the occurrence and severity of pneumonia in patients with intracerebral hemorrhage. Front Immunol. 2023;14:1115031. 10.3389/fimmu.2023.1115031. PMID: 36860868; PMCID: PMC9969881.
Korantzopoulos P, Letsas KP, Tse G, Fragakis N, Goudis CA, Liu T. Inflammation and atrial fibrillation: A comprehensive review. J Arrhythm. 2018;34(4):394–401. 10.1002/joa3.12077. Published 2018 Jun 4.
Ihara K, Sasano T. Role of Inflammation in the Pathogenesis of Atrial Fibrillation. Front Physiol. 2022;13:862164. 10.3389/fphys.2022.862164. Published 2022 Apr 14.
Nso N, Bookani KR, Metzl M, Radparvar F. Role of inflammation in atrial fibrillation: A comprehensive review of current knowledge. J Arrhythm. 2020;37(1):1–10. 10.1002/joa3.12473. Published 2020 Dec 23.
Chang MC, Choo YJ, Seo KC, Yang S. The Relationship Between Dysphagia and Pneumonia in Acute Stroke Patients: A Systematic Review and Meta-Analysis. Front Neurol. 2022;13:834240. 10.3389/fneur.2022.834240. Published 2022 Mar 17.
Li C, Qiu C, Shi Y, Yang T, Shao X, Zheng D. Experiences and perceptions of stroke patients living with dysphagia: A qualitative meta-synthesis. J Clin Nurs. 2022;31(7–8):820–31. 10.1111/jocn.15995.
Galovic M, Stauber AJ, Leisi N, Krammer W, Brugger F, Vehoff J, Balcerak P, Müller A, Rosenfeld J, Polymeris A, et al. Development and validation of a prognostic model of swallowing recovery and enteral tube feeding after ischemic stroke. Jama Neurol. 2019;76:561–70. 10.1001/jamaneurol.2018.4858.
Minelli C, Luvizutto GJ, Cacho RO, et al. Brazilian practice guidelines for stroke rehabilitation: Part II. Diretrizes brasileiras para reabilitação no acidente vascular cerebral: parte II. Arq Neuropsiquiatr. 2022;80(7):741–58. 10.1055/s-0042-1757692.
Patel UK, Kodumuri N, Dave M, et al. Stroke-Associated Pneumonia: A Retrospective Study of Risk Factors and Outcomes. Neurologist. 2020;25(3):39–48. 10.1097/NRL.0000000000000269.
Park JH, Kim E, Cho H, Park DW, Choi J, Jang SH. Brain activation in response to visual sexual stimuli in male patients with right middle cerebral artery infarction: The first case-control functional magnetic resonance imaging study. Med (Baltim). 2021;100(21):e25823. 10.1097/MD.0000000000025823.
Shultz SM, Hartmann PM. George E Holtzapple (1862–1946) and oxygen therapy for lobar pneumonia: the first reported case (1887) and a review of the contemporary literature to 1899. J Med Biogr. 2005;13(4):201–6. 10.1177/096777200501300405.
Kelly CA, Lynes D, O'Brien MR, Shaw B. A wolf in sheep's clothing? Patients' and healthcare professionals' perceptions of oxygen therapy: An interpretative phenomenological analysis. Clin Respir J. 2018;12(2):616–32. 10.1111/crj.12571.
Kelly CA, Maden M. How do health-care professionals perceive oxygen therapy? A critical interpretive synthesis of the literature. Chron Respir Dis. 2015;12(1):11–23. 10.1177/1479972314562408.
Chu DK, Kim LH, Young PJ, et al. Mortality and morbidity in acutely ill adults treated with liberal versus conservative oxygen therapy (IOTA): a systematic review and meta-analysis. Lancet. 2018;391(10131):1693–705. 10.1016/S0140-6736(18)30479-3.
O'Driscoll BR, Howard LS, Earis J, Mak V, British Thoracic Society Emergency Oxygen Guideline Group; BTS Emergency Oxygen Guideline Development Group. BTS guideline for oxygen use in adults in healthcare and emergency settings. Thorax. 2017;72(Suppl 1):ii1–90. 10.1136/thoraxjnl-2016-209729.
Hafner S, Beloncle F, Koch A, Radermacher P, Asfar P. Hyperoxia in intensive care, emergency, and peri-operative medicine: Dr. Jekyll or Mr. Hyde? A 2015 upd ate. Ann Intensive Care. 2015;5(1):42. 10.1186/s13613-015-0084-6.

No competing interests reported.

A comparative study on the influencing factors and risk prediction models for stroke- associated pneumonia in patients with acute ischemic stroke and atrial fibrillation

Status:

Version 1

Abstract

Background

Methods

Results

Conclusions

Figures

1. INTRODUCTION

2. Method

2.1 Data sources

2.2 Study samples

2.3 Diagnosis of AIS, AF, and SAP

2.4 Collection of medical history

2.5 Data analysis

2.6 Statistical methods

3. Results

3.1 Characteristics of the study population

3.2 Predictors screened by LASSO regression analysis

3.3 Model construction, optimization, and evaluation

3.4 Optimal model and independent influences

4. DISCUSSION

5. CONCLUSIONS

Abbreviations

Declarations

Author Contribution

References

Additional Declarations

Status:

Version 1