Advancing Clinical Decision-Making: Utilizing Machine Learning to Predict Postoperative Stroke Following Craniotomy

doi:10.21203/rs.3.rs-4190312/v1

Objection:

Postoperative stroke (PS) represents a significant and grave complication, which often remains challenging to detect until clear clinical symptoms emerge. The early identification of populations at high risk for perioperative stroke is essential for enabling timely intervention and enhancing postoperative outcomes. This study seeks to employ machine learning (ML) techniques to create a predictive model for PS following elective craniotomy.

Methods

This study encompassed a total of 1,349 cases that underwent elective craniotomy between January 2013 and August 2021. Perioperative data, encompassing demographics, etiology, laboratory results, comorbidities, and medications, were utilized to construct predictive models. Nine distinct machine learning models were developed for the prediction of postoperative stroke (PS) and assessed based on the area under the receiver-operating characteristic curve (AUC), along with sensitivity, specificity, and accuracy metrics.

Results

Among the 1,349 patients included in the study, 137 cases (10.2%) were diagnosed with postoperative stroke (PS), which was associated with a worse prognosis. Of the nine machine learning prediction models evaluated, the logistic regression (LR) model exhibited superior performance, as indicated by an area under the receiver-operating characteristic curve (AUC) value of 0.741 (0.64–0.85), and competitive performance metrics, including an accuracy of 0.668, sensitivity of 0.650, and specificity of 0.670. Notably, feature importance analysis identified "preoperative albumin," "ASA classification," and "preoperative hemoglobin" as the top three factors contributing to the prediction of PS.

Conclusion

Our study successfully developed a real-time and easily accessible parameter requiring LR-based PS prediction model for post-elective craniotomy patients.

intracranial surgery

perioperative stroke

machine learning

Stroke is globally acknowledged as the second leading cause of death and disability [1]. China, in particular, carries the highest burden of stroke, as indicated by its age-standardized prevalence rates [2]. Numerous studies have consistently shown that the Chinese population exhibits a slightly higher incidence of stroke and intracerebral hemorrhage compared to Caucasian populations [3].

Early postoperative stroke (PS), which is defined as a stroke occurring within 7 days following surgery, exhibits a relatively low incidence rate, thanks to advancements in surgical and anesthetic techniques. Nonetheless, the occurrence of PS significantly affects patient outcomes, manifesting as an 8-fold increase in mortality. The inconspicuous onset of stroke during intraoperative anesthesia, coupled with the administration of postoperative sedative drugs and associated complications, poses challenges in promptly identifying symptoms. As a result, the absence of early, clear, and reliable diagnostic methods impedes the prompt recognition and provision of effective treatment for patients experiencing early PS.

Anesthesiologists typically give precedence to factors such as preoperative laboratory tests, intraoperative fluid therapy, monitoring of intraoperative vital signs, and intraoperative drug therapy. However, these specific data points have not yet been formally integrated into diagnostic tools designed to aid in the early detection of postoperative stroke (PS). At present, making a comprehensive and accurate assessment based solely on physicians' electronic medical records and intraoperative data for prognostication is impractical. The adoption of machine learning (ML) technology presents an opportunity to augment physicians' diagnostic abilities in predicting PS by enabling detailed analyses of large datasets. Certain risk factors have been identified as contributing to a higher risk of perioperative stroke. Research utilizing machine learning has pinpointed hypertension, cancer, congestive heart failure, chronic pulmonary disease, and peripheral vascular disease as significant predictors that increase the risk of PS. Factors related to intraoperative anesthesia also significantly influence PS, yet previous studies have largely overlooked anesthesia's role. The incidence of postoperative stroke following craniotomy exceeds that observed in non-aortic surgeries due to the intricate nature of the surgical site. Additionally, it is crucial to acknowledge that the occurrence of postoperative stroke after craniotomy is not solely attributed to surgical factors but also influenced by the patients' perianaesthetic period condition. Consequently, the aim of this study was to identify independent risk factors for PS in patients undergoing elective intracranial surgery, to develop predictive models using ML algorithms, and to evaluate their effectiveness in predicting PS.

2.1 Ethics statements

This study, conducted at the Third Affiliated Hospital of Sun Yat-sen University, received approval from the Ethics Review Committee [No. (2019)02-609-01]. The study design involved extracting a restricted dataset from the perioperative database of the Department of Anesthesiology. Given the retrospective design, the requirement for informed patient consent was waived. The study was conducted according to STROBE guidelines.

2.2 Study population

In this retrospective study, the selection of the patient cohort and the study design were based on data retrieved from the electronic patient record (EPR) systems at the Third Affiliated Hospital of Sun Yat-sen University-Lingnan Hospital, situated in Guangzhou, Guangdong, China.

A comprehensive database platform was developed within the electronic patient record (EPR) systems of our hospital by integrating medical records from multiple systems, including the Hospital Information System (HIS), Laboratory Information System (LIS), Picture Archiving and Communication System (PACS), and the Docare Anesthesia System (2005–2020, Medical System Co., Ltd., Suzhou, China). This innovative database platform enabled the efficient retrieval of extensive data encompassing various facets of hospitalization, such as admission details, length of stay, and follow-up post-hospital visits. The platform provided access to a broad spectrum of information, including demographic details, daily documentation, laboratory test results, imaging findings, anesthesia records, and other pertinent clinical characteristics.

2.3 Primary outcome

The primary outcome of this study was the incidence of postoperative stroke (PS) following elective craniotomy. PS was defined as either cerebral ischemia or cerebral hemorrhage occurring within 7 days post-surgery. The diagnostic criteria for PS adhered to the guidelines provided by the American Stroke Association, which include: (1) Cerebral ischemia resulting from thrombosis, embolism, or systemic hypoperfusion; (2) Intracerebral hemorrhage, where hemorrhage within the brain parenchyma directly affects brain tissue, and hemorrhage within the subarachnoid space impacts the cerebrospinal fluid that surrounds the brain and spinal cord.

2.4 Data collection

The database platform provided data elements categorized into the following sections for analysis:

1. Demographics: This included age, sex, height, and weight of the patients.

2. Preoperative Comorbidities: Data on hypertension, coronary heart disease, myocardial infarction, diabetes mellitus, history of alcohol abuse, smoking, and previous surgeries were collected.

3. Etiology: The primary diseases necessitating intracranial surgery were identified, with a focus on brain tumors, cerebrovascular diseases, and other conditions.

4. Perioperative Laboratory Values: Laboratory results encompassing liver function, kidney function, electrolytes, and blood cell count were analyzed.

5. Preoperative Complications: Information was gathered on complications and metrics indicating patient severity, including complications from hypertension, documentation of treatment escalation such as length of stay in the ICU, use of continuous blood purification (CBP), and mechanical ventilation.

6. Intraoperative Events: Events indicative of hemodynamic instability, such as cardiac arrest, arrhythmia, lactic acidosis, acidosis, hypernatremia, hypokalemia, and hypotension, were recorded.

7. Intraoperative Fluid and Transfusion: Data on the total intraoperative fluid infusion and output, along with the total transfused blood products, were extracted. This included RBC transfusion, plasma transfusion, total blood product transfusion, and total fluid transfusion, categorized based on specific criteria.

8. Perioperative Medications: The administration of medications during the surgery and the subsequent seven-day postoperative period was closely monitored. This primarily included glucocorticoids, mannitol, and sedatives. After extraction of the original data, patients were divided into two groups with PS as the primary outcome measure. An analysis was then performed to examine the differences in the distribution of each characteristic between the two groups.

2.5 Variable selection

With a total of 50 variables, the potential for overfitting during the training process poses a risk to the overall performance of the model. Consequently, our initial approach involved the implementation of a univariate test to eliminate features that lacked statistical significance. Subsequently, we proceeded to remove variables that exhibited a missing data rate exceeding 30% within the training set, as well as those with a correlation coefficient greater than 0.7, which indicated poor predictive capability.

2.6 Model training

Before deploying machine learning (ML) models, all variables in the study were subjected to centralization and normalization processes. Continuous variables were transformed for normality using the Yeo-Johnson method and missing values were imputed with the mean, whereas categorical variables had their missing values filled with the mode. The dataset was then split into an 80% training set and a 20% test set. The training set was utilized for the development of predictive models, and the test set served to validate the performance of these models.

After preprocessing and feature selection, we constructed ML classifiers employing nine techniques, including logistic regression (LR) [4], random forest (RF) [5], adaptive boosting (Adaboosting) [6], XGboosting [7], bootstrap aggregating (Bagging) [8], gradient boosting (GB) [9], K-Nearest neighbor (KNN)[10], decision trees (DT)[11], and Gaussian NB [12]. The primary performance metric employed in the training/validation set was the area under the receiver operating characteristic curve (AUC). In addition to assessing model performance through AUC, we also evaluated the area under the exact recall curve, accuracy, sensitivity, specificity, positive predictive value, and negative predictive value for each test.

2.7 Statistical methods

The statistical analysis was performed using SPSS 22.0. The frequency or percentage was used to describe categorical variables, while the Pearson chi-square test was employed to compare groups. Mean ± standard deviation was used to describe continuous variable data, whereas median and quartile were used for skewed distribution data. T-test and nonparametric rank-sum test were utilized to compare normal and skewed distribution data, respectively.

The dataset comprised data from 3,297 patients who underwent elective craniotomy between January 2013 and August 2021. In the course of retrospective enrollment, patients were excluded from this study for the following reasons: preoperative stroke (n = 589), absence of electronic medical records (n = 273), undergoing a second surgery (n = 589), and undergoing non-intracranial surgery (n = 497).

This study conducted a retrospective analysis of the clinical data from 1,349 patients who underwent elective craniotomy at our hospital. Among these patients, 137 experienced early postoperative stroke (PS), yielding an incidence rate of 10.2% (Fig. 1).

3.1 Baseline characteristics

Table 1 presents the baseline characteristics of the study's patient population, offering summary statistics for the perioperative variables of the 1,349 patients included in the complete analysis cohort. The median age of the participants was 48.6 years, with males constituting 47% of the cohort. Compared to healthy controls, elderly patients with an ASA (American Society of Anesthesiologists) grade of ≥ 3, those requiring preoperative endotracheal intubation, and those showing preoperative reductions in fibrinogen (Fib) were observed to have a significantly higher prevalence of postoperative stroke (PS) (all P < 0.05). Furthermore, patients who experienced PS tended to have longer surgical durations and were associated with higher estimated blood loss.

Table 1

Baseline information of included patients
Category	Entire Cohort (n = 1349)	Stroke (n = 137)	Healthy Control (n = 1212)	p-value
Age,y	48.6 ± 14.4	52.9 ± 13.3	48.1 ± 14.4	0.414
Male sex	634(47%)	75(55%)	557(45.0%)	0.014
Height(cm)	163.3 ± 8.0	164.5 ± 7.6	163.1 ± 8.0	0.347
Weight(kg)	60.5 ± 11.3	61.8 ± 11.4	60.4 ± 11.3	0.360
BMI	22.7 ± 3.6	22.7 ± 3.8	22.7 ± 3.6	0.156
Diabetes	256(19.0%)	35(25.5%)	220(18.2%)	0.038
Hypertension	411(30.5%)	39(28.5%)	372(30.7%)	0.038
abnormal kidney function	29(2.1%)	6(4.4%)	23(1.9%)	0.375
abnormal liver function	52(3.9%)	5(3.6%)	47(3.9%)	0.288
abnormal Coagulation	38(2.8%)	4(2.9%)	34(2.8%)	0.280
Allergy history	31(2.3%)	5(3.6%)	26(2.1%)	0.921
ICU admission before surgery	7(0.5%)	0(0%)	7(0.6%)	0.320
Preoperative tracheal intubation	36(2.7%)	9(6.6%)	27(2.2%)	0.007
Preoperative WBC	6.88 ± 2.31	6.89 ± 2.3	6.81 ± 2.2	0.635
Preoperative HGB	132.12 ± 16.71	129.00 ± 15.01	132.5 ± 16.9	0.261
Preoperative NEUT%	0.60 ± 0.11	0.63 ± 0.10	0.59 ± 0.11	0.470
Preoperative Plt	250.04 ± 69.74	252.7 ± 83.5	249.7 ± 68.1	0.961
Preoperative APTT	36.94 ± 4.468	36.4 ± 4.3	37.0 ± 4.5	0.014
Preoperative PT	0.99 ± 0.08	0.98 ± 0.07	0.99 ± 0.09	0.718
Preoperative Fib	3.41 ± 1.02	3.6 ± 1.1	3.4 ± 1.0	0.422
Preoperative Na	141.30 ± 3.22	141.2 ± 3.1	141.3 ± 3.2	0.771
Preoperative K	3.90 ± 0.43	3.94 ± 0.40	3.90 ± 0.43	0.113
Preoperative Ca	2.39 ± 0.13	2.39 ± 0.13	2.39 ± 0.13	0.944
Preoperative AST	21.61 ± 14.66	24.51 ± 31.84	21.3 ± 11.3	0.953
Preoperative ALT	24.31 ± 22.45	28.1 ± 38.9	23.9 ± 19.8	0.411
Preoperative ALB	42.82 ± 4.25	41.9 ± 4.5	42.9 ± 4.2	0.310
Preoperative TBILL	9.86 ± 5.55	9.23 ± 9.5	9.9 ± 4.9	0.234
Preoperative Glu	5.62 ± 1.71	5.6 ± 1.6	5.6 ± 1.7	0.121
Preoperative A/G	1.65 ± 0.31	1.68 ± 0.32	1.65 ± 0.30	0.029
Preoperative BUN	5.10 ± 1.96	5.4 ± 2.5	5.1 ± 1.9	0.090
Preoperative eGFR	102.45 ± 19.70	104.79 ± 18.37	100.92 ± 19.49	0.057
Operation time (min)	173(105,267)	199.6(102,237)	197.6(105,283)	0.003
Intraoperative crystalloid volume (mL)	1300(600, 2200)	1300(700, 2000)	1375(600,2200)	0.256
Intraoperative colloid volume(mL)	500(500,1000)	500(500,1000)	500(500,1000)	0.220
Intraoperative urine(mL)	750(300,1300)	700(300,1300)	800(300,1400)	0.350
Intraoperative blood loss(mL)	100(20,200)	80(20,200)	100(30, 200)	0.025
Intraoperative amount(mL)	2000(1100,3200)	1950(1200,2900)	2000(1100,3225)	0.182
Intraoperative blood transfusion(mL)	0(0,0)	0(0,0)	0(0,0)	0.248
Intraoperative mannitol(mL)	0(0,62.5)	0(0,0)	0(0,100)	0.005
Intraoperative total output(mL)	715(212,1350)	700(230,1300)	750(210,1400)	0.350
postoperative hospital stay	9.0(5.0,15.0)	7.0(3.0,10.5)	10(6.0,17.0)	0
Total cost of hospitalization	61494(44732,87487)	56150(41623,79370)	64331(46626,90364)	0.047
Abbreviations:
BMI, body mass index; WBC, white blood cell; HGB, hemoglobin; NEUT%, percentage of neutrophils; Plt, platelet count; APTT, activated partial thromboplastin time; PT, prothrombin time; Fib, fibrinogen; AST, aspartate aminotransferase; ALT, alanine aminotransferase; ALB, albumin; TBILI, total bilirubin. Glu, glucose. A/G, albumin /globulin; BUN, blood urea nitrogen; eGFR, estimated glomerular filtration rate.

3.2 Feature selection.

Furthermore, we removed variables that exhibited sparse values in the Least Absolute Shrinkage and Selection Operator (LASSO) analysis based on the training set(Fig. 2A). Ultimately, we identified and retained 14 features that satisfied the criteria for inclusion in our machine learning (ML) models, with a significance level of P < 0.05(Fig. 2B). The 14 features are "gender," "preoperative ICU admission," "preoperative diabetes," "age," "ASA classification," "leukocyte count," "hemoglobin," "fibrinogen," "prothrombin time," "albumin," "alanine aminotransferase," "total bilirubin," "albumin/globulin ratio," and "serum potassium."

3.3 Model performance

To predict postoperative stoke (PS) following elective intracranial surgery, we utilized several essential machine learning (ML) algorithms, including logistic regression (LR), random forest (RF), and naive Bayes (NB). Based on the receiver operating characteristic (ROC) curves (Fig. 3) and the performance characteristics (Table 2) of the nine distinct ML models, the LR model demonstrated superior predictive performance, achieving the highest Area Under the Curve (AUC) of 0.741 (95% Confidence Interval [CI]: 0.64–0.85). Additionally, the LR model attained noteworthy sensitivity, specificity, and accuracy values of 0.650, 0.670, and 0.668, respectively.

Table 2

Features of each machine learning model.
Metrics	AUC	Cutoff	Sensitivity	Specificity	Accuracy
Logistic Regression	0.741 (0.638, 0.844)	0.111	0.650	0.670	0.668
AdaBoost	0.728 (0.607, 0.85)	0.195	0.800	0.644	0.659
Bagging	0.637 (0.511, 0.764)	0.120	0.600	0.598	0.598
Gradient Boost	0.645 (0.525, 0.764)	0.077	0.600	0.598	0.598
Random Forest	0.739 (0.637, 0.84)	0.111	0.700	0.701	0.701
XGBoost	0.578 (0.43, 0.726)	0.231	0.550	0.552	0.551
Gaussian NB	0.727 (0.616, 0.837)	0.030	0.650	0.649	0.650
KNN	0.628 (0.486, 0.769)	0.097	0.600	0.598	0.598
Decision Tree	0.537 (0.407, 0.667)	0.009	0.550	0.515	0.519
Abbreviations:
NB, naïve Bayesian; KNN, k-nearest neighbor.

3.4 Feature importance

The results depicted in Fig. 4 reveal that within the logistic regression (LR) model, the five features with the highest importance were "preoperative plasma albumin level," "ASA classification," "preoperative routine blood hemoglobin level," "plasma albumin/globulin ratio," and "total bilirubin level." These variables exhibited a significant impact on the model's outcome. In contrast, variables ranked lower demonstrated minimal significance and offered limited interpretability in the context of the study.

In our study, we enrolled 1,349 patients following elective intracranial surgery, observing a morbidity rate of postoperative stroke (PS) at 10.2%, with 6.3% attributed to hemorrhagic stroke and 4.7% to ischemic stroke. We developed nine machine learning (ML) prediction models, among which the logistic regression (LR) model demonstrated superior performance, achieving the highest Area Under the Curve (AUC) value of 0.741. Through feature importance analysis, it was identified that "preoperative plasma albumin level," "ASA classification," "preoperative routine blood hemoglobin level," "plasma albumin/globulin ratio," and "total bilirubin level" emerged as the top five significant features for predicting PS.

In this study, the efficacy of various machine learning (ML) models, including logistic regression (LR), Adaboost, Bagging, Gradient Boosting Decision Tree (GDB), Random Forest (RF), XGBoost, Gaussian Naive Bayes (NB), K-Nearest Neighbors (KNN), and Decision Tree (DT), was assessed using the test set. Our findings indicate that the LR model excelled in predicting postoperative seizures (PS), achieving an Area Under the Curve (AUC) of 0.741. Furthermore, the LR model exhibited an accuracy rate of 0.668, a sensitivity of 0.650, and a specificity of 0.670. Overall, the LR model showed a relatively high level of predictive accuracy, making it the preferred ML model. This preference was supported by its ability to highlight important features and provide some explanation for the variables' influence. For example, the LR model identified "preoperative plasma albumin level" as the most significant feature, suggesting that a lower preoperative plasma albumin level is correlated with an increased risk of PS. Similarly, the prominent rankings of "ASA classification", "preoperative routine hemoglobin level", "plasma albumin/globulin ratio" and "total bilirubin level" can be explained accordingly. Previous studies have consistently demonstrated a correlation between preoperative hemoglobin levels and the incidence and prognosis of PS. Bhupesh et al. found that women with lower preoperative plasma hemoglobin concentrations exhibited increased susceptibility to stroke [13]. Similarly, a Japanese trial revealed that low hemoglobin concentrations were linked to an increased risk of stroke in adults[14]. Moreover, a meta-analysis provided compelling evidence supporting the association between admission anemia and both ischemic and hemorrhagic stroke, as well as elevated mortality rates related to stroke [15]. Another contributing factor to the incidence of PS is the preoperative plasma albumin/globulin ratio. In a recent study, Beamer et al. observed that patients presenting with high clinical risk factors for stroke were significantly more likely to experience subsequent vascular events, which were associated with a lower blood albumin/globulin ratio. [16].

The indicators mentioned were also incorporated into our prediction model, which typically utilizes data routinely collected after admission, underscoring the model's practical feasibility and generalizability within a hospital setting. The predictive factors identified in our model align with the risk factors highlighted in several prior studies, reinforcing its validity. The logistic regression algorithm employed in this study offers distinct advantages over traditional statistical methods, including a transparent model structure and a probability derivation that is robust and open to scrutiny. The parameters within the model elucidate the impact of each feature on the outcome, providing high interpretability. Moreover, the model supports online learning, enabling easy parameter updates without the need for retraining the entire model. Consequently, this model not only integrates the predictive efficiency of key factors but also boasts commendable interpretability.

Currently, there is a paucity of research on PS in the field of neurosurgery. Nonetheless, among the sparse machine learning (ML) studies available, one conducted by Zhang et al. stands out. They discovered that the XGBoost model displayed the highest Area Under the Curve (AUC) of 0.78 for predicting PS in elderly patients. The study identified hypertension, cancer, congestive heart failure, chronic lung disease, and peripheral vascular disease as the top five predictors. [17]. It is important to highlight that the study did not place any restrictions on the types of surgical procedures evaluated, potentially encompassing high-risk interventions for stroke, such as cardiac macrovascular surgery. Additionally, another study identified advanced age, pre-existing valvular heart disease, previous stroke, emergency surgery, and postoperative hypotension as independent risk factors for postoperative stroke (PS). [18]. Patients necessitating emergency surgery exhibit a heightened vulnerability to postoperative complications, including postoperative seizures (PS). Consequently, these patients were excluded from the aforementioned studies to ensure a more focused and controlled analysis of risk factors and outcomes.

The incidence of postoperative stroke (PS) can be linked to a variety of pathologic and physiologic mechanisms. The administration of anesthesia and the execution of the surgical procedure provoke a systemic inflammatory response, leading to alterations in the concentrations of inflammatory mediators and the coagulation status of the blood. These alterations enhance the risk of perioperative thrombosis and intravascular plaque instability, subsequently elevating the probability of stroke in perioperative patients. [19, 20].

Our research represents the inaugural effort to construct a predictive model for postoperative stroke (PS) using a machine learning-based logistic regression (LR) algorithm, specifically designed for patients undergoing elective intracranial surgery. This model is distinguished by its explainability through feature importance analysis. Moreover, by statistically addressing the issue of unbalanced data, we have developed a practical predictive tool that enables clinicians to fine-tune clinical therapy. Nonetheless, our study is not without its limitations. Firstly, the presence of missing values in the original dataset could compromise the stability of our predictive ML model. Secondly, the study's sample size was relatively small, necessitating a larger, multicenter sample to further substantiate the predictive model's validity. Thirdly, this predictive model requires an independent dataset to assess its extrapolation and generalization capabilities accurately. Lastly, anesthesia-related incidents such as persistent intraoperative hypotension/hypertension during the perianesthetic period were not included due to insufficient data availability. Future endeavors will focus on gathering comprehensive external validation datasets to enhance the validity of this model.

This study has successfully created a real-time and readily accessible logistic regression (LR)-based prediction model for postoperative stroke(PS) in patients following elective craniotomy. The optimal application of this machine learning (ML) model facilitates prompt risk assessment and clinical decision-making, with the potential to enhance patient outcomes. Nonetheless, additional research and validation are essential to refine and broaden the model's applicability.

PS	perioperative stroke
ML	machine learning
ICU	intensive care unit
EPR system	electronic patient record system
HIS	hospital information system
PACS	picture archiving and communication systems
LIS	laboratory information management system
AI	artificial intelligence
APTT	activated partial thromboplastin time
PT	prothrombin time
ALT	alanine aminotransferase
AST	aspartate aminotransferase
BUN	blood urea nitrogen
GFR	glomerular filtration rate
WBC	white blood cell
HGB	hemoglobin
PLT	platelet count
NEUT	percentage of neutrophils
TBILI	total bilirubin
ALB	albumin
Fib	fibrinogen
K	serum potassium
Na	serum sodium
Ca	serum calcium
A/G	albumin /globulin
LR	logistics regression
GB	gradient boosting
KNN	k-nearest neighbor
DT	decision trees
RF	random forest
NB	naïve bayesian
BMI	body mass index

Funding

This study was partly supported by the Natural Science Foundation of Guangdong Province (Grant No. 2022A1515012603), Science and Technology Program of Guangzhou City (No.2024A04J4246 and 202201020429); Joint Funds of the National Natural Science Foundation of China (No.U22A20276), Science and Technology Planning Project of Guangdong Province-Regional Innovation Capacity and Support System Construction (No. 2023B110006). Provincial-enterprise Joint Funds of Guangdong Basic and Applied Basic Research Foundation (No.2021B1515230012) and the “Five and five” Project of the Third Affiliated Hospital of Sun Yat-Sen University (No.2023WW501)

Acknowledgement

The authors thank all the patients participated in the study.

Conflict of interest

The authors declare no competing interests.

Authors’ contributions

TY and HQ conducted the research design, TY and TS participated in the writing of the paper, TY and HQ analyzed and explained the data; YX collected and helped with analyzing the data of the study. All authors read and approved the final manuscript.

Ethics

Our study was approved by the Ethnic Committee of the Third Affiliated Hospital of Sun Yat-sen University [No. (2019)02-609-01].

Consent for publication

Not applicable.

Availability of data and materials

The datasets generated and analysed during the current study are not publicly available due to patient privacy data protection but are available from the corresponding author on reasonable request.

Global regional, national burden of neurological disorders. 1990–2016: a systematic analysis for the Global Burden of Disease Study 2016. Lancet Neurol. 2019;18(5):459–80.
Wang W, Jiang B, Sun H, Ru X, Sun D, Wang L, Wang L, Jiang Y, Li Y, Wang Y, et al. Prevalence, Incidence, and Mortality of Stroke in China: Results from a Nationwide Population-Based Survey of 480 687 Adults. Circulation. 2017;135(8):759–71.
Tsai C-F, Thomas B, Sudlow CLM. Epidemiology of stroke and its subtypes in Chinese vs white populations: a systematic review. Neurology. 2013;81(3):264–72.
Zhang Z, Ho KM, Hong Y. Machine learning for the prediction of volume responsiveness in patients with oliguric acute kidney injury in critical care. Crit Care (London England). 2019;23(1):112.
Heo J, Yoon JG, Park H, Kim YD, Nam HS, Heo JH. Machine Learning-Based Model for Prediction of Outcomes in Acute Stroke. Stroke. 2019;50(5):1263–5.
Wang J, Xue W, Shi X, Xu Y, Dong C. Adaboost-Based Machine Learning Improved the Modeling Robust and Estimation Accuracy of Pear Leaf Nitrogen Concentration by In-Field VIS-NIR Spectroscopy. Sensors (Basel) 2021, 21(18).
Hou N, Li M, He L, Xie B, Wang L, Zhang R, Yu Y, Sun X, Pan Z, Wang K. Predicting 30-days mortality for MIMIC-III patients with sepsis-3: a machine learning approach using XGboost. J Transl Med. 2020;18(1):462.
Lin E, Lin C-H, Lane H-Y. A bagging ensemble machine learning framework to predict overall cognitive function of schizophrenia patients with cognitive domains and tests. Asian J Psychiatr. 2022;69:103008.
Lee YW, Choi JW, Shin E-H. Machine learning model for predicting malaria using clinical information. Comput Biol Med. 2021;129:104151.
Liu W, Wang S, Ye Z, Xu P, Xia X, Guo M. Prediction of lung metastases in thyroid cancer using machine learning based on SEER database. Cancer Med. 2022;11(12):2503–15.
Elhazmi A, Al-Omari A, Sallam H, Mufti HN, Rabie AA, Alshahrani M, Mady A, Alghamdi A, Altalaq A, Azzam MH, et al. Machine learning decision tree algorithm role for predicting mortality in critically ill adult COVID-19 patients admitted to the ICU. J Infect Public Health. 2022;15(7):826–34.
Singh SK, Taylor RW, Pradhan B, Shirzadi A, Pham BT. Predicting sustainable arsenic mitigation using machine learning techniques. Ecotoxicol Environ Saf. 2022;232:113271.
Panwar B, Judd SE, Warnock DG, McClellan WM, Booth JN, Muntner P, Gutiérrez OM. Hemoglobin Concentration and Risk of Incident Stroke in Community-Living Adults. Stroke. 2016;47(8):2017–24.
Sato F, Nakamura Y, Kayaba K, Ishikawa S. Hemoglobin Concentration and the Incidence of Stroke in the General Japanese Population: The Jichi Medical School Cohort Study. J Epidemiol. 2022;32(3):125–30.
Barlas RS, Honney K, Loke YK, McCall SJ, Bettencourt-Silva JH, Clark AB, Bowles KM, Metcalf AK, Mamas MA, Potter JF et al. Impact of Hemoglobin Levels and Anemia on Mortality in Acute Stroke: Analysis of UK Regional Registry Data, Systematic Review, and Meta-Analysis. J Am Heart Assoc 2016, 5(8).
Beamer N, Coull BM, Sexton G, de Garmo P, Knox R, Seaman G. Fibrinogen and the albumin-globulin ratio in recurrent stroke. Stroke. 1993;24(8):1133–9.
Zhang X, Fei N, Zhang X, Wang Q, Fang Z. Machine Learning Prediction Models for Postoperative Stroke in Elderly Patients: Analyses of the MIMIC Database. Front Aging Neurosci. 2022;14:897611.
Vasivej T, Sathirapanya P, Kongkamol C. Incidence and Risk Factors of Perioperative Stroke in Noncardiac, and Nonaortic and Its Major Branches Surgery. J Stroke Cerebrovasc Dis. 2016;25(5):1172–6.
Schietroma M, Carlei F, Mownah A, Franchi L, Mazzotta C, Sozio A, Amicucci G. Changes in the blood coagulation, fibrinolysis, and cytokine profile during laparoscopic and open cholecystectomy. Surg Endosc. 2004;18(7):1090–6.
Rost NS, Wolf PA, Kase CS, Kelly-Hayes M, Silbershatz H, Massaro JM, D'Agostino RB, Franzblau C, Wilson PW. Plasma concentration of C-reactive protein and risk of ischemic stroke and transient ischemic attack: the Framingham study. Stroke. 2001;32(11):2575–9.

No competing interests reported.

Advancing Clinical Decision-Making: Utilizing Machine Learning to Predict Postoperative Stroke Following Craniotomy

Status:

Version 1

Abstract

Objection:

Methods

Results

Conclusion

Figures

1. Introduction

2. Materials and Methods

2.1 Ethics statements

2.2 Study population

2.3 Primary outcome

2.4 Data collection

2.5 Variable selection

2.6 Model training

2.7 Statistical methods

3. Results

3.1 Baseline characteristics

3.2 Feature selection.

3.3 Model performance

3.4 Feature importance

4. Discussion

5. Conclusion

Abbreviations

Declarations

References

Additional Declarations

Status:

Version 1