The 2019 Global Burden of Disease Study findings reveal that stroke continues to be the second leading cause of death and ranks third in causing disability worldwide (1). Since 2015, stroke has emerged as the preeminent cause of death and disability in China, significantly impacting the nation’s public health and exerting substantial pressure on its healthcare infrastructure (2). Acute ischemic stroke (AIS), constituting over 80% of stroke cases (3), remains the predominant type. Statistics indicate that upwards of 50% of stroke survivors endure varying degrees of disability, with a recurrence rate of 17.7% (4), thereby placing a considerable strain on both families and society at large. Consequently, the early detection and prognosis of individuals at elevated risk of AIS, coupled with prompt intervention and treatment, hold critical importance in enhancing the quality of life for AIS patients.

Presently, studies on clinical prediction of ischemic stroke outcomes predominantly employ established scales or conventional models, such as the Essen stroke risk score, RRE-90 score, Logistic regression, and Cox regression (5–7). These tools, however, are limited by their inclusion of only a select few traditional risk factors and an inability to capture complex nonlinear interactions among various prognostic elements. In contrast, Machine Learning (ML)—a field exploring the simulation of human learning processes by computer algorithms (8)—holds distinct advantages. ML eschews the prerequisite assumptions about variable relationships and model structures inherent to traditional statistics and can adeptly manage datasets with incomplete entries. Additionally, ML models offer explanatory insights that surpass classical statistical methods (9). With the advent of big data analytics in healthcare research, machine learning is integrating increasingly vast data arrays to more accurately reflect human physiological complexity and the unpredictable aspects of disease traits. This progression heralds an innovative trajectory for disease diagnosis, prognosis, and risk evaluation, offering substantial practical application potential (10–12).

In this study, we develop and validate a predictive model for the early prognosis of patients experiencing their first AIS using relevant medical data. We employ various ML algorithms to assess the efficacy of the model and compare the predictive utility of each to determine the optimal mathematical model for forecasting early outcomes in AIS patients. The goal is to facilitate the identification of high-risk individuals with potentially poor prognoses at an early stage, thereby providing a scientific basis for selecting appropriate clinical treatment strategies.

In this study, we developed several machine learning models to predict the early prognosis of AIS patients using general clinical and laboratory data collected within 24 h of admission. Feature selection through the Lasso CV method identified the most critical predictors for early prognosis in AIS patients as WBC, HCY, D-Dimer, baseline NIHSS, FDP, and GLU. Among the nine machine learning models evaluated, the Random Forest model exhibited superior performance in the test set, achieving an AUC of 0.852, an accuracy rate of 0.818, a sensitivity of 0.654, a specificity of 0.945, and a recall rate of 0.900. These findings indicate that RF models utilizing general clinical and laboratory data from the initial 24 h of admission can effectively predict the early prognosis of AIS patients.

ML predictive models leverage algorithms to analyze data and forecast future occurrences or trends. These models have the capability to discern patterns in vast, complex datasets and integrate these patterns in a non-linear and highly interactive fashion (18). The application of machine learning in the medical domain, particularly in identifying risk factors and developing prognostic models for AIS patients, has seen significant growth in recent years, with notable contributions from researchers like Veerbeek et al. (19, 20), Xie et al. (12), Lin et al. (21), Ramos et al. (10), Su et al. (11), and Chen M et al. (22). These studies primarily derive their data from hospital records and publicly accessible databases. For instance, Lin et al. (21) assessed the capabilities of various ML models (SVM, RF, ANN, HANN) in predicting the 90-day mRS outcomes for stroke patients within the Taiwan Stroke Registry. Similarly, Su et al. (11) investigated the predictive performance of 4 ML models (SVM, RF, LGBM, DNN) in forecasting mRS scores and hospitalization outcomes at discharge among 2,780 AIS patients registered in the Chang Gung Healthcare System. Public databases offer vast data volume, broad coverage, and access ease, beneficial for swiftly compiling data required for predictive modeling. Nonetheless, ensuring the quality of data from these sources presents challenges, including risks of inconsistent patient assessment and treatment, incomplete records, and lost follow-up data (22). Predominantly, hospital data-driven studies incorporate imaging, demographic, and clinical information. For instance, Xie et al. (12) utilized imaging data and NIHSS scores to develop prediction models, while Ramos et al. (10) employed imaging and clinical baseline data for AIS prognosis. Common clinical assessments, like routine blood, coagulation, and biochemical tests, serve as conventional, accessible, and cost-effective diagnostics for AIS patients. Yet, the construction of an ML predictive model encompassing demographic, clinical, and 24-h admission blood test data for forecasting 3-month AIS patient outcomes lacks representation in the Chinese population. This study aims to fill this research void.

Previous research primarily focused on the prognostic impact of age (19, 20), atrial fibrillation (4, 23, 24), NIHSS scores at admission (25–27), D-dimer (28, 29), HCY (30–32), and GLU (33–35) on AIS outcomes. In the current study, we employed Lasso CV for feature selection, identifying WBC, HCY, D-dimer, NIHSS, FDP, and GLU as critical variables for developing prognostic models for AIS patients. Notably, WBC and FDP, indicators of the body’s inflammatory response and the fibrinolytic system’s activity, respectively, have been under-explored in AIS prognostication. While WBC has been associated with stroke prognosis and an increased risk of disability at discharge following every 1 × 10⁹/L increment at stroke onset (36), its independent predictive value remains unclear, potentially due to limited sample sizes in previous studies. Similarly, FDP and D-dimer serve as markers for hypercoagulability and hyper-fibrinolysis, pivotal for early diagnosis, monitoring treatment efficacy, and prognosticating thrombogenic conditions. Our findings underscore the importance of including D-dimer and FDP in early prognostic assessments for AIS patients, highlighting their significant role in predicting outcomes.

In this study, we utilized nine machine learning models for analysis, based on selected variables. Of these, the RF model demonstrated superior performance on the test set, achieving an AUC of 0.876 (95% CI, 0.754–0.991). This model provided a significantly better prediction capability. The RF algorithm, known for its robust noise resistance and low susceptibility to overfitting, has found extensive application in healthcare (37). Moreover, several studies employing the RF model for predicting stroke patient outcomes have affirmed its effective predictive power (10, 38, 39).

The current study is constrained by several limitations. Primarily, the modest sample size, despite being derived from two different centers, necessitates a larger dataset for robust validation. Also, this study lacked external validation. In the future, we will conduct a validation study with an external cohort. Moreover, while the study encompasses demographic information, past medical history, blood counts within 24 h of admission, and biochemical assays of AIS patients, it lacks imaging data pertaining to the location and extent of cerebral infarctions. The absence of these imaging details could potentially impact the predictive accuracy of the model. Finally, treatment decision-making is a complex process involving multiple considerations. Our model is just one of many aids whose results should be considered along with clinical judgment and other patient-specific factors, and is intended to provide additional information to support the decision-making process, not to replace professional medical judgment.

In conclusion, our study illustrates that utilizing demographic data, past medical histories, and 24-h laboratory information upon admission for ML modeling is a feasible approach to predicting the short-term prognosis of patients experiencing their first AIS. Notably, it reveals that the RF model could serve as an effective predictive tool. This finding is crucial for the early identification and prediction of individuals at elevated risk of developing AIS with an unfavorable prognosis. It underscores the importance of prompt and early intervention and treatment to enhance the quality of life for AIS patients.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving humans were approved by the Medical Ethics Committee of Bengbu Medical University. The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

QH: Writing – original draft, Methodology. G-LS: Data curation, Investigation, Writing – original draft. BS: Formal analysis, Methodology, Writing – review & editing. M-LL: Data curation, Investigation, Writing – review & editing. SZ: Methodology, Writing – original draft. MH: Methodology, Writing – review & editing. F-YH: Writing – original draft.