Machine learning-based prediction model for long-term mortality after ischemic stroke Authors: Hee-Soo Kim, Seung-Bo Lee, Changi Kim, Jiyeong Lee, Joon-myoung Kwon, Mi-Young Oh Published: 2026 ABSTRACT AbstractStroke, a leading cause of death, requires precise predictions of life expectancy. Traditional scores focused on short-term outcomes and required labor-intensive data collection. Recent machine learning (ML)-based prediction models showed superior performance. However, they were predominantly geared towards classification tasks, omitting crucial temporal aspects. Thus, an unmet need arises for more convenient, time-aware predictive methodologies for long-term mortality in managing acute stroke patients. The developmental cohort consisted of 3,411 patients from two tertiary hospitals, while the external test cohort included 502 patients from a secondary cardiovascular center. We developed various ML- and deep learning (DL)-based models to predict post-ischemic stroke mortality, utilizing clinical data. Among these, we selected the best-performing model and compared it with traditional risk scores. Furthermore, we assessed key features based on their importance, determining their optimal number for practical application. The total number of deceased patients in the developmental cohort was 136, and 118 in the external test cohort. The Gradient Boosting Cox Proportional Hazards model demonstrated the best performance with C-index of 0.785 in internal validation set, and C-index of 0.845 in external test set. It surpassed the conventional risk score, PREMISE score with C-index of 0.783(p = 0.017) in the external dataset. The key features included age, National Institutes of Health Stroke Scale score, hemoglobin, and so on. The study developed and validated a machine learning-based model tailored for post-stroke survival prediction, outperforming existing scores. It could enhance the individual stroke patient’s survival predictions and optimize medical resource allocation.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-53797-6.Keywords: Machine learning, Long-term survival, StrokeSubject terms: Neuroscience, Cardiology, Neurology Introduction Stroke has been the second leading cause of mortality, and the third leading cause of death and disability combined worldwide, with an estimated 5.5 million deaths annually1,2. From 1990 to 2019, the absolute number of strokes increased by 70%, the prevalence of stroke increased by 85%, and disability-adjusted life-years due to stroke increased by 32%3. Accordingly, the total medical costs for stroke care would be predicted to reach $1841 billion dollars by 20301. Given the immediate, life-threatening implications of stroke and increased demand of post-stroke care and cost, accurate prognostications of stroke are essential. It provides crucial insights to clinicians, enabling the customization of treatment strategies for the active treatment of high-risk groups in the early stages. Additionally, it facilitates the judicious allocation of limited medical resources, helping to prepare for long-term post-stroke care. Previous studies have developed risk stratification systems for predicting mortality after ischemic stroke. Mittal et al. introduced a scoring system predicting in-hospital mortality following stroke, with a strong emphasis on neurological deficits, particularly the initial neurological severity assessed using the National Institutes of Health Stroke Scale (NIHSS) and Glasgow Coma Scale4. Another predictive tool, the PREMISE score, estimates early mortality within 7 days after acute stroke, integrating clinical and imaging indicators such as, posterior circulation involvement and non-lacunar stroke types5. However, these models focus primarily on short-term mortality, overlooking the need for long-term care planning, which is crucial since nearly 50% of stroke survivors suffer moderate to severe impairment beyond the acute phase. Additionally, most models lack temporal dynamics, failing to account for how risk evolves over time. Saposnik et al. reported a model forecasting more extended survival up to one year following the ischemic stroke6. Despite being well-developed and validated, its clinical adoption was hindered by the labor-intensive manual data collection required. Moreover, most previous scores relied on traditional statistical models, such as the Cox proportional hazards model, which assume proportional hazards over time and may limit complex interactions between variables. This reveals an unmet need for a more streamlined prediction tool that operates independently of direct electronic medical records (EMR) extraction using the various data while efficiently predicting long-term stroke prognosis. Recent advancements in machine learning (ML) and deep learning (DL) have shown significant potential in improving predictions for ischemic stroke prognosis by effectively capturing the complex and non-linear relationships among clinical variables7–9. However, the previous ML- or DL- based prediction models also tended to focus on short-term mortality, akin to conventional risk scores. In addition, their emphasis on classification rather than the projection of survival duration may be insufficient for forecasting comprehensive survival trajectories for individual stroke patients. Despite their promising future, ML models face important challenges that must be acknowledged for a balanced perspective, particularly regarding interpretability and validation. ML techniques function as “black boxes,” making it difficult for clinicians to understand and trust model decisions. This lack of transparency can hinder clinical adoption, particularly in high-risk clinical scenarios such as stroke prognosis. Additionally, many ML models lack external validation, which limits their generalizability across diverse populations and healthcare systems. Moreover, practical integration into clinical workflows also remains challenging in the absence of streamlined variable inputs and prospective validation. These limitations underscore the need for models that not only enhance predictive accuracy but also address interpretability, usability, and real-world applicability. Accordingly, our study aimed to develop and validate a survival prediction model using advanced ML and DL techniques to improve the precision of personalized, longitudinal survival predictions compared with conventional risk scoring systems. Unlike previous models focused solely on short-term outcomes, our study captures extended survival trajectories and incorporates external validation to enhance generalizability and reliability. We prioritized selecting a minimal number of variables in our model to ensure efficient integration with EMR systems, while maintaining a high level of predictive accuracy for easy applicability to real-world practice. By streamlining the model’s design, we aim to create a practical, high -accuracy tool that supports broad clinical adoption and facilitates comprehensive care strategies across the pre- and post-stroke continuum, potentially enhancing patient outcome. Results Patient characteristics In this study, 136 patients in the developmental cohort and 118 patients in the external test cohort were deceased (Fig. 1). The patients in the developmental cohort were younger than those in the external test cohort. Comorbid conditions such as heart disease, peripheral arterial disease, diabetes, hyperlipidemia, atrial fibrillation, and cancer were more prevalent in the external test cohort. The NIHSS scores were higher in the external test cohort than that in the developmental cohort, indicating more severe stroke presentations. The hemoglobin and hematocrit levels were lower in the external test cohort. Additionally, the characteristics of stroke varied between the two cohorts. Strokes affecting the posterior circulation were more common in the external test cohort, whereas strokes due to small vessel occlusion were predominantly observed in the developmental cohort. Regarding treatment modalities, IV or IA treatment during the hyperacute phase of stroke were more frequently administered in the developmental cohort. In the developmental cohort, the longest period of follow-up extended to 1,962 days, whereas in the external test cohort, the maximum follow-up duration reached 2,129 days. We summarized the missing variables in the dataset in supplemental Table 1. The differences in baseline characteristics between included and excluded patients are shown in Supplemental Table 2 (Table 1). The flow chart of study population. Table 1Baseline characteristics of developmental and external test cohort.CharacteristicsDevelopment cohort(n = 3411)External test cohort(n = 502)P-valueAge (years)67.9912.7870.5313.52< 0.001aSex, male (%)1990 (58.30)284 (56.35)0.410BMI (kg/m2)23.9810.5524.053.210.021bMortality (%)136 (3.98)118 (23.41)< 0.001cNIHSS*3.664.326.879.66< 0.001aSystolic blood pressure (mmHg)145.6529.34143.6127.650.547bDiastolic blood pressure (mmHg)86.9517.8480.6716.23< 0.001aPoor premorbid state** (%)1089 (31.9)146 (27.80)0.112cPrevious stroke (%)519 (15.2)82 (16.27)0.535cPeripheral arterial disease (%)17 (0.50)22 (4.37)< 0.001cHeart disease (%)179 (5.24)118 (23.41)< 0.001cCancer (%)33 (0.97)20 (3.97)< 0.001cHypertension (%)2041 (59.78)127 (25.20)0.133cDiabetes (%)998 (29.23)319 (63.29)0.012cDyslipidemia (%)786 (23.02)175 (34.72)< 0.001cAtrial fibrillation (%)561 (16.43)155 (30.75)0.008cSmoking (%)1081 (31.66)103 (20.44)< 0.001cIntravenous thrombolytic treatment (%)298 (8.73)27 (5.36)0.010cIntra-arterial thrombolytic treatment (%)148 (4.34)35 (6.94)0.010cAntiplatelet medication (%)2696 (78.97)55 (10.91)< 0.001cAnticoagulant medication (%)522 (15.29)442 (87.70)< 0.001cPosterior circulation stroke (%)510 (14.94)189 (37.5)< 0.001cNon lacunar stroke (%)2873 (84.15)360 (71.43)< 0.001cWhite blood cell count (/ µl)7.862.988.323.38< 0.001aTotal cholesterol (mg/dL)171.6248.03164.0144.90< 0.001aBlood urea nitrogen (mg/dL)16.227.4418.9710.78< 0.001aCreatinine (mg/dL)1.181.591.050.690.002aHemoglobin (g/dL)14.212.4613.432.05< 0.001aPlatelet count (/µl)216.0272.74228.8767.13< 0.001aHematocrit (%)39.007.2939.685.570.014aTriglyceride (mg/dL)121.3980.61125.2770.57< 0.001bHigh density lipoprotein cholesterol (mg/dL)44.3112.5045.5410.770.013aLow density lipoprotein cholesterol (mg/dL)113.1041.3395.4232.69< 0.001aGlucose (mg/dL)147.0074.74151.1575.88< 0.001bPREMISE score3.611.954.602.38< 0.001aaStudent’s t-test, bMann–Whitney U test, cChi-square test.Data are presented as number (%) or mean ± SD.*NIHSS, National Institutes of Health Stroke Scale.**Poor premorbid state was defined by modified ranking scale, 3 to 5 before index stroke. Baseline characteristics of developmental and external test cohort. aStudent’s t-test, bMann–Whitney U test, cChi-square test. Data are presented as number (%) or mean ± SD. *NIHSS, National Institutes of Health Stroke Scale. **Poor premorbid state was defined by modified ranking scale, 3 to 5 before index stroke. Model performance The development of our predictive model utilized three ML and one DL techniques to forecast survival after ischemic stroke. The Elastic Cox model showed a C-index of 0.776 in the internal validation set with increased accuracy (C-index: 0.838) in the external test set. The RSF model recorded a C-index of 0.734 internally and external C-index of 0.834. The GradientCoxPH model exhibited a predictive power, with a C-index of 0.785 for the internal validation set and 0.845 for the external test set. The Deepsurv model demonstrated predictive abilities with a C-index of 0.758 internally and 0.734 externally. The GradientCoxPH model and elastic Cox model achieved the highest predictive accuracy and surpassed the conventional PREMISE score in terms of predictive performance. The statistical significance of these results was affirmed through comprehensive bootstrap analyses (p < 0.05), as shown in Table 2. Additionally, the model performance was not significantly different using variables with more than 20% missingness or without them (Supplemental Table 3.1). The model performance with or without excluded patients was not significantly different, as shown in Supplemental Table 3.2. Table 2Predictive performance of machine learning models, Cox regression, and PREMISE score.ModelC-indexp-value (vs. Traditional cox)p-value (vs. PREMISE score)Internal (95% CI)External (95% CI)InternalExternalInternalExternalTraditional cox0.730 (0.642, 0.806)0.830 (0.798, 0.857)0.9820.992PREMISE score0.743 (0.634, 0.843)0.783 (0.739, 0.820)0.8480.066––Elastic cox0.776 (0.692, 0.865)0.838 (0.810, 0.867)0.4200.5790.6340.030RSF0.734 (0.646, 0.814)0.834 (0.802, 0.862)0.9010.7230.8950.047GradientCoxPH0.785 (0.733, 0.847)0.845 (0.815, 0.876)0.2570.3900.4890.017Deepsurv0.758 (0.662, 0.842)0.734 (0.689, 0.773)0.6150.0000.8310.100 Predictive performance of machine learning models, Cox regression, and PREMISE score. Feature importance evaluation using SHAP SHAP analysis identified age at admission as the most influential predictor, and indicated the important role of initial stroke severity, measured by NIHSS score in assessing prognosis, shown in Fig. 2a, b. Other key variables affecting model predictions include hemoglobin, white blood cell count, antiplatelet medication, and so on. Top 20 feature of GradientCoxPH based survival prediction model. *NIHSS, National Institutes of Health Stroke Scale; White blood cell, White blood cell count; Platelet, platelet count; BMI, body mass index. (a) Internal validation set; (b) External test set. Optimal number of features with the maintenance of performance We used SHAP values to optimize feature selection in various predictive models to achieve high performance with a minimal number of features, shown in Fig. 3. We started with a baseline model using the single most influential feature and incrementally added up to 20 features. The Elastic Cox model’s performance improved from a C-index of 0.756 with one feature to 0.776 with 20 features internally, and from 0.767 to 0.827 externally. The RSF model demonstrated a similar trend, with a C-index increasing from 0.727 to 0.776 internally and 0.746 to 0.834 externally. The GradientCoxPH model increased in internal performance (from 0.749 to 0.785) with added features, and improved externally from 0.770 to 0.845. The Deepsurv model exhibited an increasing internal C-index from 0.744 to 0.758. The predictive performance did not show significant improvement when utilizing more than the top five features. Performance of models with a varying number of features. Elastic Cox, Elastic Net Cox regression; RSF, Random Survival Forests; GradientCoxPH, Gradient Boosting Cox Proportional Hazards model. Kaplan-Meier survival curve across feature sets Figure 4 shows the Kaplan–Meier survival analysis for patients classified into high- and low-risk categories based on GradientCoxPH model predictions. Survival probabilities were estimated across all features (A) and a selected set of top 5 features (B) over a period of 2000 days. The high-risk group showed significantly lower survival probabilities, with a clear early divergence from the low-risk group during the follow-up period (p < 0.001). Kaplan-Meier curves of GradientCoxPH based survival prediction model. (A) Survival prediction model using all features, and (B) survival prediction model using top 5 features. GradientCoxPH gradient boosting cox proportional hazards model. The Kaplan–Meier survival curves for high- and low- risk groups using all available features were shown in Fig. 4A. The high-risk group had a steeper decline in survival probability, while the low-risk group maintained a higher probability of survival throughout the observation period. The survival curve based on GradientCoxPH model using only Top 5 features was demonstrated in Fig. 4B. The high-risk group in this survival curve using the only top 5 features also showed faster decline in survival probability compared to the low-risk group. Furthermore, the consistent separation between the curves over the entire timeline demonstrated the discriminative strength of the prediction model. High-risk group for mortality identified by ML models To validate the clinical relevance of the ML-based risk groups, we compared the baseline characteristics between high- and low-risk groups (Supplemental Table 4). In both the development and external cohorts, the high-risk group was significantly older and presented with higher NIHSS scores, as well as a higher prevalence of comorbidities, including previous stroke and atrial fibrillation. Furthermore, high-risk patients exhibited poorer functional and physiological profiles, characterized by systemic inflammation (elevated white blood cell count), renal dysfunction (increased creatinine), and nutritional deficiencies (lower cholesterol and hemoglobin levels). These disparities were consistently observed across both datasets. Sensitivity analysis We conducted one-at-a-time (OAT) sensitivity analyses for key variables including age and initial NIHSS score. As shown in Fig. 5, the predicted risk remains relatively stable across age ranges until approximately 80 years, beyond which the risk increases more steeply. This suggests that higher age in the external cohort may partially contribute to the higher observed model performance, especially among the oldest patients. While the effect of NIHSS on predicted risk was relatively modest across its range, the age-related increase in sensitivity warrants careful interpretation. Nonetheless, we acknowledge that differential age distribution between cohorts may influence apparent model performance, and we have clarified this limitation in the discussion section. Sensitivity analysis according to age and NIHSS score. (a) Internal validation set. (b) External test set. Discussion We developed and validated a prediction model for survival after ischemic stroke. This model was evaluated using three ML and one DL-based approach. Among these, the GradientCoxPH model proved most effective and outperformed the conventional risk scoring system, PREMISE score in the external dataset. The key predictive features identified using SHAP values included age, initial NIHSS score upon admission, hemoglobin levels, and so on. Interestingly, the model’s predictive accuracy did not increase significantly with more than the top five features. Furthermore, the model effectively distinguished between high-risk and low-risk patient groups through distinct survival curves. The high-risk group was more likely to be older, physiologically vulnerable, and to have greater stroke severity along with a higher prevalence of comorbidities. The clustering of advanced age, severe neurological deficits, and multiple comorbidities identifies highly vulnerable patients with diminished physiological reserve. These findings validate the clinical utility of our model in effectively discriminating patients at the highest risk of mortality. Our ML-based model showed a higher C-index compared to traditional Cox regression, though these improvements lacked statistical significance. These limited improvements might be explained by the relatively small sample size or the dominance of well-established prognostic factors, such as age and initial NIHSS score, in determining post-stroke mortality outcomes. Additionally, the lack of substantial performance gains suggests that survival prediction in post-stroke patients may be largely driven by linear relationships rather than complex variable interactions, with age and initial NIHSS score being the predominant predictors. Nevertheless, ML approaches provided valuable insights beyond predictive performance metrics. Through feature importance analysis using SHAP, we were able to quantify the degree to which individual variables influenced mortality predictions, information not readily accessible through traditional statistical models. While the relatively small data size could not detect potentially meaningful performance gain with statistical value, the modest improvements observed in ML models compared to the clinical scoring system, PREMISE score may warrant further investigation to enhance current risk stratification tools in clinical practice. Future studies with larger, multi-center datasets are needed to establish whether ML-based model provide meaningful clinical utility in post-stroke mortality prediction and to better explore complex relationships that may emerge with increased statistical power. ML and DL models for survival prediction have predominantly focused on classification tasks of short-term mortality, often omitting crucial time-related information, potentially affecting the accuracy of predictions. For instance, Zhu et al. reported the various ML-based prediction model for survival using the dataset from 7,483 patients in Medical Information Mart from the Intensive Care (MIMIC)-IV database. However, the time dimension was not incorporated into this prediction model for classification of the short-term mortality, within 30 days after stroke9. Recently, ML and DL models for dealing with right-censored mortality data have been developed and exhibited excellent performance to predict the individual patient’s survival in various clinical situations8,10–12. In addition, Wang et al. reported that RSF based survival prediction model surpassed the performance of traditional Cox regression-based model in 4918 patients with hemorrhagic stroke13. However, there have been only a few studies that have validated these models using external datasets. The recent systematic meta-analyses pinpointed such caveat of the recent survival prediction model after ischemic stroke.14. The distinctiveness of the current study lies in the fact that the ML-based survival prediction model incorporated time information, which is crucial for estimating individual patient survival duration and identifying relevant clinical features for timely intervention with appropriate therapeutic targets. Furthermore, this prediction model was rigorously validated using external test data from another hospital, which differed in patient volume, baseline patient characteristics, and geographic location. The GradientCoxPH model proved effective, particularly in handling the critical aspect of time information after stroke in our study. This model adeptly managed multicollinearity, crucial when interpreting interrelated clinical predictors, and excelled in variable selection and regularization to avoid overfitting in our high-dimensional dataset. The integration of gradient boosting with Cox proportional hazards regression in the GradientCoxPH model allowed for a nuanced incorporation of time-dependent factors. This potential advantage allows the model to uncover complex temporal interactions within the data, which simpler linear models might overlook. By constructing a sophisticated ensemble of decision trees, the GradientCoxPH model effectively reduced bias and variance. These capabilities, especially the effective use of time information, underscored the GradientCoxPH model’s superiority in predicting long-term survival outcomes after a stroke, setting it apart from other ML and DL methodologies15. However, the model’s performance might vary across different datasets and clinical settings. Therefore, further studies involving diverse populations and various clinical scenarios are necessary to examine and validate the robustness of the model. The significant features identified using SHAP in our study were clinically relevant. Age and initial stroke severity measured by NIHSS score were the strongest clinical factors associated with outcomes after stroke, as supported by numerous studies16,17. Hemoglobin serves not only as a surrogate marker of general health but also as a modifiable factor for optimizing post-stroke recovery. Lower hemoglobin reduces systemic oxygen-carrying capacity, which can exacerbate ischemic injury and limit neuronal recovery. Inadequate hemoglobin levels also impair cerebral autoregulation, increase risk of hypo-perfusion, and reduce tolerance to secondary complications such as infections18,19. These findings strengthen the explainability and reliability of our model, demonstrating the multifactorial nature of stroke prognosis and the utility of SHAP values in elucidating the relative importance of clinical features. The baseline characteristics between the development and external test cohorts differed significantly, particularly in age, initial NIHSS score, and hemoglobin levels—key predictors identified by our model. The higher prevalence of these high-risk features in the external test cohort may have enhanced the model’s discriminative power. Sensitivity analyses confirmed that older age and higher NIHSS scores were consistently associated with higher predicted mortality risk, supporting the robustness of our model across varying population profiles and contextualizing the observed performance differences between cohorts. The next step for ML models is application in real-world clinical practice20,21. To facilitate this, we simplified the model by identifying the optimal number of variables. Remarkably, a model using only the top 5 clinical features performed comparably to one using all features. By identifying and retaining only the most influential clinical predictors, our model becomes practical for seamless integration into EMR systems, thus facilitating real-time predictions. This simplicity facilitates clinical adoption, as the model requires minimal manual data input while maintaining predictive accuracy. By enabling efficient use in real-world clinical settings, the model has the potential to support proactive decision-making in stroke care and resource allocation, ultimately improving patient outcomes22. Moreover, our model may have the potential to help predict not only stroke mortality but also incident stroke, and major cardiovascular events, such as acute coronary syndrome and recurrent stroke. By predicting stroke incidence, the model can support more proactive management strategies, potentially reducing both the occurrence of stroke events and post-stroke cardiovascular events through targeted preventive measures. This broader applicability in both pre- and post-stroke care contexts might demonstrate our model’s potential to reduce stroke-related morbidity and healthcare costs. This study has several limitations. First, the retrospective design carries an inherent risk of selection bias. Second, substantial disparities in mortality rates were observed between the development and external test cohorts (3.98% versus 23.41%, respectively), primarily due to distinct institutional characteristics. While the development cohort was from general tertiary hospitals, the external test cohort was from a specialized cardiovascular center managing high-risk patients involving complex cardiac and aortic interventions. The consistent performance of the model despite this substantial baseline heterogeneity and higher clinical severity, supports its robustness and potential generalizability across diverse clinical settings. Finally, this study did not collect information regarding the causes of death. Understanding these causes would provide clinicians with valuable insights into both acute and chronic care for stroke patients. Future studies should aim to clearly delineate the specific reasons for stroke in individual patients more clearly. Conclusions We have developed and validated a long-term survival prediction model for ischemic stroke, utilizing various clinical factors from the acute phase and incorporating the time dimension. This model is anticipated to aid in decision-making processes, potentially contributing to the reduction of mortality and morbidities following ischemic stroke. Methods Data source and study population The dataset used in this study was obtained from AI Hub (https://www.aihub.or.kr), a public platform operated by the National Information Society Agency (NIA) under the Ministry of Science and ICT of South Korea. AI Hub provides high-quality, expert-curated datasets and widely used in government-funded research and national AI development projects, ensuring their reliability and scientific integrity. The data collection, preprocessing, and distribution of AI Hub datasets were conducted by officially designated institutions, ensuring compliance with relevant ethical and technical standards. The specific dataset utilized in this study is titled “Medical Imaging Data for Neurological Disease Diagnosis” and is publicly accessible via AI Hub (https://www.aihub.or.kr/aihubdata/data/view.do?currMenu=115&dataSetSn=211&topMenu=100). This dataset includes annotated brain MRI scans and clinical information from hospitalized patients diagnosed with acute ischemic stroke, collected from geographically diverse, anonymized tertiary hospitals across South Korea between January 2011 and November 2019. Patients were included if they were diagnosed with ischemic stroke and underwent brain MRI within 7 days of symptom onset. All data were de-identified and curated under government-supervised protocols to ensure data quality and adherence to ethical standards. Imaging and clinical annotations were reviewed and labeled by board-certified neurologists to maintain high reliability and accuracy. The dataset from AI hub was used for developmental cohort including trainingand internal validation. In addition, we used geographically different external test cohort for validation set. The developmental cohort comprised 4,128 stroke patients admitted to two anonymous tertiary hospitals between January 2013 and June 2019. The patients in the developmental cohort were divided into training and internal validation datasets. Data from patients admitted to a secondary cardiovascular hospital (Bucheon Sejong hospital) between October, 2016 and May, 2022 were used as an external test cohort to evaluate the model. The exclusion criteria were as follows: (1) clinically abnormal data deemed erroneous due to data entry errors, BMI ≥ 400; Height = 999 ;Weight = 999; Hematocrit > 350;WBC = 0; Total cholesterol < = 20;Triglyceride = 0;HDL-cholesterol > 360;Prothrombin time = 0.01; Systolic blood pressure = 0 (2) poor-quality imaging, specifically diffusion-weighted (DWI) MRI inadequate for confirming the diagnosis of acute ischemic stroke (e.g., mismatch between diffusion-weighted images and the apparent diffusion coefficient map, or excessive motion artifacts that prevented reliable interpretation); (3) recurrent stroke during the enrollment period; and (4) missing data in categorical variables, smoking history or anti-platelet medication (5) unknown premorbid status before the index stroke, as illustrated in Fig. 1. As a result, 3,411 and 502 patients were included in the developmental and external test cohorts, respectively. Data collection The variables from the medical records at admission were collected including the following: The dataset consists of the following: Neuroimaging data:Initial brain DWI and apparent diffusion coefficient magnetic resonance imaging, which were used for confirmation of ischemic lesion, the location and characteristics of ischemic lesion, annotations, and imaging time. Neuroimaging data:Initial brain DWI and apparent diffusion coefficient magnetic resonance imaging, which were used for confirmation of ischemic lesion, the location and characteristics of ischemic lesion, annotations, and imaging time. Initial brain DWI and apparent diffusion coefficient magnetic resonance imaging, which were used for confirmation of ischemic lesion, the location and characteristics of ischemic lesion, annotations, and imaging time. Initial brain DWI and apparent diffusion coefficient magnetic resonance imaging, which were used for confirmation of ischemic lesion, the location and characteristics of ischemic lesion, annotations, and imaging time. Admission clinical data:Demographic: age, sex, height, weight, body mass indexMorbidities: hypertension, diabetes, hyperlipidemia, transient ischemic attack or stroke, peripheral artery disease, ischemic heart disease, atrial fibrillation, cancer and premorbid state.Initial stroke severity, indicated by National Institute of Health stroke scale (NIHSS) score.Hyperacute stroke treatment; intravenous (IV) or intra-arterial (IA) thrombolytic treatment.Initial vital sign and laboratory findings at admission: systolic and diastolic blood pressure, hemoglobin, hematocrit, white blood cell count, platelet count, blood urea nitrogen, creatinine, total cholesterol, high density lipoprotein (HDL) cholesterol, low density lipoprotein (LDL) cholesterol, glucose, prothrombin time. Admission clinical data:Demographic: age, sex, height, weight, body mass indexMorbidities: hypertension, diabetes, hyperlipidemia, transient ischemic attack or stroke, peripheral artery disease, ischemic heart disease, atrial fibrillation, cancer and premorbid state.Initial stroke severity, indicated by National Institute of Health stroke scale (NIHSS) score.Hyperacute stroke treatment; intravenous (IV) or intra-arterial (IA) thrombolytic treatment.Initial vital sign and laboratory findings at admission: systolic and diastolic blood pressure, hemoglobin, hematocrit, white blood cell count, platelet count, blood urea nitrogen, creatinine, total cholesterol, high density lipoprotein (HDL) cholesterol, low density lipoprotein (LDL) cholesterol, glucose, prothrombin time. Demographic: age, sex, height, weight, body mass index Demographic: age, sex, height, weight, body mass index Morbidities: hypertension, diabetes, hyperlipidemia, transient ischemic attack or stroke, peripheral artery disease, ischemic heart disease, atrial fibrillation, cancer and premorbid state. Morbidities: hypertension, diabetes, hyperlipidemia, transient ischemic attack or stroke, peripheral artery disease, ischemic heart disease, atrial fibrillation, cancer and premorbid state. Initial stroke severity, indicated by National Institute of Health stroke scale (NIHSS) score. Initial stroke severity, indicated by National Institute of Health stroke scale (NIHSS) score. Hyperacute stroke treatment; intravenous (IV) or intra-arterial (IA) thrombolytic treatment. Hyperacute stroke treatment; intravenous (IV) or intra-arterial (IA) thrombolytic treatment. Initial vital sign and laboratory findings at admission: systolic and diastolic blood pressure, hemoglobin, hematocrit, white blood cell count, platelet count, blood urea nitrogen, creatinine, total cholesterol, high density lipoprotein (HDL) cholesterol, low density lipoprotein (LDL) cholesterol, glucose, prothrombin time. Initial vital sign and laboratory findings at admission: systolic and diastolic blood pressure, hemoglobin, hematocrit, white blood cell count, platelet count, blood urea nitrogen, creatinine, total cholesterol, high density lipoprotein (HDL) cholesterol, low density lipoprotein (LDL) cholesterol, glucose, prothrombin time. Follow-up data3-month functional outcomes, and partially 1-year functional outcomes, and mortality information. Follow-up data3-month functional outcomes, and partially 1-year functional outcomes, and mortality information. 3-month functional outcomes, and partially 1-year functional outcomes, and mortality information. 3-month functional outcomes, and partially 1-year functional outcomes, and mortality information. 1) Missing values were imputed in order to mitigate the impact of missing data within the variables, imputation strategies were employed. The continuous variables were imputed using their respective mean values. Model development Survival prediction models were developed utilizing survival specific ML and DL models. The models implemented, include Elastic Net Cox regression (Elastic Cox), Random Survival Forests (RSF), Gradient Boosting Cox Proportional Hazards model (GradientCoxPH), and Deepsurv. Each model was selected for its ability to effectively analyze time-to-event data. The model that showed superior performance, as indicated by the highest concordance index (C-index), was selected as the final model for survival prediction. To ensure accuracy, all models underwent a 5-fold stratified cross-validation. This process was used to identify the best set of hyperparameters, determined by the highest average C-index. Elastic Cox: The Elastic Net Cox regression model combines the L1 and L2 penalties of the Lasso and Ridge regression methods, respectively. This approach allows selection of a subset of covariates, thus addressing potential multicollinearity and enhancing model interpretability. We used a grid search over a predefined range of penalizer and L1 ratio values23,24. Elastic Cox: The Elastic Net Cox regression model combines the L1 and L2 penalties of the Lasso and Ridge regression methods, respectively. This approach allows selection of a subset of covariates, thus addressing potential multicollinearity and enhancing model interpretability. We used a grid search over a predefined range of penalizer and L1 ratio values23,24. RSF: An ensemble tree method adapted for censored survival data, RSF extended the random forest algorithm to the survival setting to facilitate analysis of the importance of covariates in the presence of right-censored outcomes25. The tuning entailed a grid search across various hyperparameters, such as the number of estimators and tree depth. RSF: An ensemble tree method adapted for censored survival data, RSF extended the random forest algorithm to the survival setting to facilitate analysis of the importance of covariates in the presence of right-censored outcomes25. The tuning entailed a grid search across various hyperparameters, such as the number of estimators and tree depth. GradientCoxPH: This model integrated gradient boosting with Cox proportional hazards regression26. By iteratively fitting residual errors of the Cox model using boosting trees. For this model, comprehensive grid search was conducted over learning rates, max depths, the number of estimators, and subsample ratios. GradientCoxPH: This model integrated gradient boosting with Cox proportional hazards regression26. By iteratively fitting residual errors of the Cox model using boosting trees. For this model, comprehensive grid search was conducted over learning rates, max depths, the number of estimators, and subsample ratios. Deepsurv: A deep learning-based survival model that utilized a multi-layer perceptron architecture to model complex nonlinear relationships between covariates and the risk function27. The model incorporates two hidden layers, with each layer followed by a Rectified Linear Unit (ReLU) activation function28. The Deepsurv model utilized a learning rate of 0.001, which determines the step size at each iteration in the optimization process, and implements Cox loss as its loss function. An early stopping mechanism with a predefined patience parameter was used to ensure the model does not overfit to the training data. Deepsurv: A deep learning-based survival model that utilized a multi-layer perceptron architecture to model complex nonlinear relationships between covariates and the risk function27. The model incorporates two hidden layers, with each layer followed by a Rectified Linear Unit (ReLU) activation function28. The Deepsurv model utilized a learning rate of 0.001, which determines the step size at each iteration in the optimization process, and implements Cox loss as its loss function. An early stopping mechanism with a predefined patience parameter was used to ensure the model does not overfit to the training data. Statistical analysis Statistical analysis were performed in python 3.8 using the Scikit-Learn and Pytorch packages29. SHapley Additive exPlanations (SHAP) values version 0.40.0 were used30. Continuous variables were expressed as mean ± standard deviation or median with interquartile range, as appropriate. Categorical variables were summarized as frequencies and percentages. Comparisons between groups were conducted using the chi-squared test or Fisher’s exact test for categorical variables and the t-test or Mann-Whitney U test for continuous variables depending on the data distribution. Survival times were estimated using the Kaplan-Meier method, and differences between survival curves were assessed using the log-rank test31. To assess the predictive accuracy of the survival models, we used the concordance index (C-index), which measures the degree of concordance between predicted and observed survival times32. For hyperparameter tuning, we employed a grid search strategy with 5-fold cross-validation. Survival curves were visualized using the Kaplan-Meier method, categorizing patients into high- and low-risk groups based on the optimal threshold determined by the Youden index J33. Sensitivity analysis One-at-a-time (OAT) sensitivity analysis systematically varies one input variable at a time while holding all others constant at their baseline values to assess individual variable impact on model predictions. We applied this method to evaluate how changes in age and NIHSS score individually affected our model’s survival probability predictions, with each variable varied across its observed range while other predictors remained fixed at their mean values. Interpretability with SHapley Additive exPlanations (SHAP) value The interpretation of the model was enhanced through an extensive analysis of feature importance, utilizing SHAP values30. SHAP values provide a quantitative measure of the contribution each variable makes to the model’s predictive accuracy. We chose features by evaluating their SHAP values, focusing on those with the most significant impact on model performance. This approach ensured that only the most impactful features were included, optimizing the balance between model complexity and predictive power. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements None. Abbreviations Machine learning Deep learning National Institutes of Health Stroke Scale Elastic net cox regression Random survival forests Gradient boosting cox proportional hazards model Author contributions *H-S Kim and S-B Lee werecontributed equally as co-first authors. H-S Kim, S-B Lee was contributed to analysis, interpretation of data, drafted the work, and substantively revised it. C.Kim, JM. Kwon, JY. Leecontributed to data acquisition. M-Y Oh contributed to conception, design of the work, interpretation of data, drafted the work, and substantively revised it. Funding This research was supported by a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (grant number: HI22C1295) and was also supported by a grant of the Korea Health Technology R&D Projects through the Korea Heaath industry Development Institute(KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea[grant numbers: RS-2025-02213752]. Data availability This study was conducted using "Medical Imaging Data for Neurological Disease Diagnosis" supported by the Ministry of Science and ICT (MSIT) and the National Information Society Agency (NIA) of Korea. The datasets used and analyzed in the current study are available from the corresponding author on reasonable request. The dataset, accessible via AI Hub (https://www.aihub.or.kr), requires IRB approval and is available exclusively to South Korean citizens. Collaboration proposals will be evaluated for alignment with our criteria, and if suitable, access to the AI Hub datasets will be requested. Declarations Competing interests The authors declare no competing interests. Ethics approval and consent to participate This study was approved by the Institutional Review Board (IRB) of Bucheon Sejong Hospital (BSH 2021-09-005). The requirement for informed consent was waived by the Institutional Review Board of Bucheon Sejong Hospital due to the retrospective nature of this study. The study was reported in accordance with the TRIPOD guideline34. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Hee-Soo Kim and Seung-Bo Lee. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement This study was conducted using "Medical Imaging Data for Neurological Disease Diagnosis" supported by the Ministry of Science and ICT (MSIT) and the National Information Society Agency (NIA) of Korea. The datasets used and analyzed in the current study are available from the corresponding author on reasonable request. The dataset, accessible via AI Hub (https://www.aihub.or.kr), requires IRB approval and is available exclusively to South Korean citizens. Collaboration proposals will be evaluated for alignment with our criteria, and if suitable, access to the AI Hub datasets will be requested.