Artifical intelligence-powered delta-NIHSS-based model for predicting recurrence, disability and mortality after acute ischaemic strokes (DISCO): a modelling study Authors: Shiyao Cheng, Yuandan Wei, Huaguang Zheng, Qingrong Zhang, Xuanyan Yang, Zhen Lu, Zehua Huang, Zijun Wan, Yanfeng Shi, Jie Zhang, Jay Chol Choi, Chul-Hoo Kang, Hong Jun Kim, Joong-Goo Kim, Hee-Joon Bae, Juneyoung Lee, Jeong-Yoon Lee, Hannah Jung, Cang Guo, Xiaoling Liao, Jianping Wu, Xia Meng, Zixiao Li, Stéphanie Debette, Hao Li, Yongjun Wang, Si Cheng, Siyang Liu Published: 2025 ABSTRACT SummaryBackgroundStroke is a leading cause of mortality and disability globally, with high rates of adverse outcomes. We aimed to develop and test a clinically accessible and explainable predictive model for post-stroke composite outcomes that could identify patients at high-risk for targeted interventions.MethodsThis modelling study used artificial intelligence to develop and validate a prognostic predictive model for post-stroke outcomes at 3 months and over 5 years, using comprehensive data from the third China National Stroke Registry (CNSR-III; 2015–2018). The development cohort comprised 11,313 patients from CNSR-III. Validation cohort 1 (n = 2627 from CNSR-III; internal validation) were selected at random on the basis of hospital assignment. Validation cohort 2 (external validation) included 5158 patients with a history of ischaemic stroke, excluding those diagnosed with TIA, from the CHANCE-2 trial (2019–2021) done at 202 centres in China. Validation cohort 3 (external validation: disability and mortality) involved 3855 patients with a history of acute ischaemic stroke or TIA from a single centre in the Clinical Research Center for Stroke–Fifth Division (CRCS-5) cohort (2011–2024) in South Korea. We evaluated 309 hospitalisation variables, including baseline characteristics, medical history, hospitalisation data, biomarkers, geographical factors, National Institutes of Health Stroke Scale (NIHSS) scores, modified Rankin Scale (mRS), and stroke polygenic risk scores, using an extreme gradient boosting tree model. Feature importance was assessed via Shapley values. Primary outcomes were 3-month stroke recurrence, disability (mRS >2), and mortality. Secondary outcomes were assessed at six additional time points over 5 years. A nested cross-validation scheme was employed for feature selection and internal validation in 80% of patients (n = 11,313) from CNSR-III cohort. Testing was performed in the remaining 20% of patients (n = 2627) from CNSR-III (validation cohort 1). External validation was performed in validation cohorts 2 and 3.FindingsGlobal and domain-specific delta-NIHSS(admission-discharge) emerged as the strongest predictor of stroke recurrence, disability, and mortality. The Delta-NIHSS-based Predictor for Post-Stroke Composite Outcomes (DISCO) model, integrating 16 delta-NIHSS(admission-to-discharge) measures, age, gender, TOAST subtypes, history of TIA, stroke, diabetes and mRS at discharge, achieved the area under the receiver operating curve (AUCs) of 0.805 and 0.815 for recurrence and mortality and of 0.852 for disability at 3 months in validation cohort 2. The 1% patients with highest-risk exhibited a 19.27-fold relative risk (RR) for recurrence, 36.59-fold RR for disability, and 225.5-fold RR for mortality at 3 months. The DISCO model is available at https://www.discosysu.cn.InterpretationThe DISCO model, incorporating 23 clinical variables, demonstrates high accuracy, robustness, clinical accessibility, and explainability in predicting post-stroke outcomes. The predictive strength of delta-NIHSS(admission-discharge) provides mechanistic insights into stroke outcomes and could potentially inform future acute stroke treatment and rehabilitation strategies. Future studies, with external validation across diverse ancestries and mechanistic studies elucidating the predictive power of delta-NIHSS in stroke recovery, are warranted.Funding10.13039/501100012166National Key Research and Development Program of China, 10.13039/501100001809National Natural Science Foundation of China, Young Elite Scientists Sponsorship Program by the China Association for Science and Technology, and Young Talents Supporting Program of Capital Medical University. Research in contextEvidence before this studyStroke is a leading global cause of mortality and disability, with high rates of recurrence, disability, and mortality, particularly in underdeveloped regions. Existing risk scores for post-stroke outcomes (such as RRE, ASTRAL, IScore, ABCD) rely on regression models utilising clinical, imaging, or laboratory variables, but they are limited by multicollinearity, moderate predictive accuracy, and insufficient assessment of long-term outcomes. We searched MEDLINE, PubMed and Web of Science on September 5, 2025, using the terms: “stroke”, “artificial intelligence”, “machine learning”, “recurrence”, relapse”, “death”, “mortality” and “disability” (and assicuated MeSH terms), without language restrictions. Our research yielded 36 machine learning (ML)-based articles for stroke outcomes. However, most of these models exhibit suboptimal performance (only five models achieved AUC >0.8 for recurrence, one model AUC >0.9 for disability, and two models AUC >0.9 for mortality). Only three models have been externally validated, demonstrating a need for an accurate, robust, interpretable model based on easily obtainable clinical variables to predict post-stroke composite outcomes. We aimed to address this knowledge gap.Added value of this studyUsing 23 readily obtainable clinical variables selected from 309 hospitalisation variables in more than 15,000 patients with a history of stroke, we developed and externally validated the delta-NIHSS-Based Predictor for Post-Stroke Composite Outcomes (DISCO) model. The DISCO model predicts post-stroke recurrence and mortality with AUCs >0.8, and disability with AUC >0.9, outperforming existing models such as RRE (recurrence), ASTRAL (disability), and IScore (mortality) in head-to-head comparisons. Nested cross-validation using extreme gradient boosting tree (XGBoost) and Shapley additive explanations (SHAP) model confirmed the model's stability and interpretability compared with alternative ML and deep learning methods. Delta-NIHSS(admission-to-discharge) emerged as the most influential predictor. Model performance was tested in the development cohort (n = 5158 from CNSR-III), and externally validated in one Chinese cohort (n = 5158) and one South Korean cohort (n = 3855; only disability and mortality not stroke recrurrence), maintaining AUCs of >0.8 for recurrence and mortality, and >0.9 for disability. An online tool has been constructed for clinical use (https://www.discosysu.cn).Implications of all the available evidenceThe DISCO model does not require imaging or laboratory variables, has been validated across diverse healthcare settings, and is accessible via an online platform, indicating strong potential for future clinical application at discharge. Delta-NIHSSadmission-discharge reflects individualised biological and clinical responses to in-hospital treatment, providing a metric to capture underlying patient changes. More research, including prospective studies in diverse populations, is required. Research in context Evidence before this study Stroke is a leading global cause of mortality and disability, with high rates of recurrence, disability, and mortality, particularly in underdeveloped regions. Existing risk scores for post-stroke outcomes (such as RRE, ASTRAL, IScore, ABCD) rely on regression models utilising clinical, imaging, or laboratory variables, but they are limited by multicollinearity, moderate predictive accuracy, and insufficient assessment of long-term outcomes. We searched MEDLINE, PubMed and Web of Science on September 5, 2025, using the terms: “stroke”, “artificial intelligence”, “machine learning”, “recurrence”, relapse”, “death”, “mortality” and “disability” (and assicuated MeSH terms), without language restrictions. Our research yielded 36 machine learning (ML)-based articles for stroke outcomes. However, most of these models exhibit suboptimal performance (only five models achieved AUC >0.8 for recurrence, one model AUC >0.9 for disability, and two models AUC >0.9 for mortality). Only three models have been externally validated, demonstrating a need for an accurate, robust, interpretable model based on easily obtainable clinical variables to predict post-stroke composite outcomes. We aimed to address this knowledge gap. Added value of this study Using 23 readily obtainable clinical variables selected from 309 hospitalisation variables in more than 15,000 patients with a history of stroke, we developed and externally validated the delta-NIHSS-Based Predictor for Post-Stroke Composite Outcomes (DISCO) model. The DISCO model predicts post-stroke recurrence and mortality with AUCs >0.8, and disability with AUC >0.9, outperforming existing models such as RRE (recurrence), ASTRAL (disability), and IScore (mortality) in head-to-head comparisons. Nested cross-validation using extreme gradient boosting tree (XGBoost) and Shapley additive explanations (SHAP) model confirmed the model's stability and interpretability compared with alternative ML and deep learning methods. Delta-NIHSS(admission-to-discharge) emerged as the most influential predictor. Model performance was tested in the development cohort (n = 5158 from CNSR-III), and externally validated in one Chinese cohort (n = 5158) and one South Korean cohort (n = 3855; only disability and mortality not stroke recrurrence), maintaining AUCs of >0.8 for recurrence and mortality, and >0.9 for disability. An online tool has been constructed for clinical use (https://www.discosysu.cn). Implications of all the available evidence The DISCO model does not require imaging or laboratory variables, has been validated across diverse healthcare settings, and is accessible via an online platform, indicating strong potential for future clinical application at discharge. Delta-NIHSSadmission-discharge reflects individualised biological and clinical responses to in-hospital treatment, providing a metric to capture underlying patient changes. More research, including prospective studies in diverse populations, is required. Introduction Stroke is a leading cause of mortality and long-term disability worldwide, with high rates of recurrence, disability, and mortality, particularly in underdeveloped regions.1 In China, stroke recurrence rates are 6.33% at 3 months and 13.69% at 5 years; disability (mRS >2) occurs in 13.75% of patients at 3 months and 7.88% at 5 years; and mortality rates are 1.18% at 3 months and 9.05% at 5 years (Table S1). Effective prognostic prediction models for post-stroke outcomes are crucial for guiding clinical decision-making by identifying patients with high-risk who may benefit from targeted evaluation and intervention. Existing statistical tools, such as the Recurrence Risk Estimator (RRE) for early stroke recurrence,2 the Acute Stroke Registry and Analysis of Lausanne (ASTRAL) score for disability,3 the Ischaemic Stroke Predictive Risk Score (IScore) for stroke mortality,4 and the Age, Blood pressure, Clinical features, Duration of symptoms (ABCD) score for patients with a history of transient ischaemic attack (TIA),5 rely on clinical, imaging, or laboratory variables and are developed using multivariable regression models. However, regression-based models are susceptible to multicollinearity and fail to optimise the complex, nonlinear relationships between variables and outcomes, which limits their predictive accuracy and robustness. Moreover, these models depend on imaging and laboratory data that may not be readily available in resource-limited settings, thereby restricting their broader clinical applicability. Machine learning (ML) techniques have gained increasing attention in clinical prediction modelling due to their ability to address nonlinear relationships between variables and outcomes, capture complex variable interactions, and improve predictive accuracy.6,7 Several ML-based models, such as early warning systems for hypotension,8 have been successfully implemented in clinical practice with support from clinical trials. However, ensuring the transparency and interpretability of ML models is essential, as their lack thereof may hinder widespread clinical adoption.9,10 A systematic review of ML-based models for stroke recurrence, disability and mortality in PubMed and Web of Science revealed 15 models for stroke recurrence, 12 models for disability and 9 models for mortality (Expanded Methods in Supplementary Materials, Table S2). Most models exhibited suboptimal predictive performance, with only five achieving an AUC >0.8 for recurrence prediction11, 12, 13, 14), one achieving an AUC >0.9 for disability prediction,13 and two achieving an AUC >0.9 for mortality prediction.14,15 Notably, none of these models underwent external validation for recurrence, and only two for disability16,17 and one for mortality,15 raising concerns regarding their accuracy and generalisability. Furthermore, clinical accessibility remains a substantial challenge, as many of these models rely on limited feature sets and lack publicly available applications or source codes (Table S2).18,19 Additionally, existing ML-based models have not incorporated long-term stroke outcomes, further compromising their practical applicability for individualised patient care. In this study, we aimed to develop and externally validate a prognostic prediction model with high accuracy, robustness, clinical accessibility, and explainability for post-stroke composite outcomes, including stroke recurrence, disability and mortality at 3 months and over 5 years. Methods Study design and study population We employed explainable artificial intelligence (AI) techniques to evaluate the predictive importance of 309 clinical features, most of which are readily available during hospitalisation, among 11,813 patients with a history of ischaemic stroke in the CNSR-III, a nationwide multi-centre registry (40 independent hospitals) of ischaemic stroke and transient ischaemic attack (TIA) in China,.20 Through this approach, we identified previously unrecognised predictors of post-stroke outcomes. Based on these findings and a rigorous internal-external validation framework, we developed the DISCO model, which integrates 23 clinically accessible features. The model was validated using data from remaining 2627 patients with a history of ischaemic stroke in CNSR-III, and was externally validated in 5158 patients with a history of ischaemic stroke enrolled in the CHANCE-2 clinical trial.21 The CHANCE-2 trial21 was a randomised, double-blind, placebo-controlled trial conducted at 202 centres in China. Model performance for disability and mortality (but not stroke recurrence) was also externally validated in 3855 patients with a history of acute ischaemic stroke or TIA from one centre in the CRCS-5 cohort, a prospectively maintained multicentre registry in South Korea.22 Due to data availability, external validation for stroke recurrence was not feasible in CRCS-5. We assessed the model's risk stratification performance across development and validation cohorts and developed a user-friendly R Shiny web application (https://www.discosysu.cn/) to provide individualised risk predictions and explanations. A total of 15,166 patients were enrolled from CNSR-III (n = 11,313 training set; n = 2627 internal validation set), a nationwide registry of acute ischaemic stroke or transient ischaemic attack (TIA) in China between August 2015 and March 2018.23 Exclusion criteria included seven patients who experienced stroke recurrence during hospitalisation, 35 patients who died during hospitalisation, and 1184 who were diagnosed with TIA. The development cohort comprised 11,313 patients from CNSR-III. Validation cohort 1 included 2627 patients from CNSR-III, selected at random on the basis of hospital assignment. Validation cohort 2 included 5158 patients with a history of ischaemic stroke from the CHANCE-2 trial, excluding those diagnosed with TIA (details in Expanded Methods in Supplementary Materials).21 Validation cohort 3 involved 3855 patients with a history of acute ischaemic stroke or TIA from a single centre in the CRCS-5 cohort from South Korea.22 Polygenic risk scores (PRS) were calculated for 9402 patients with a history of stroke from CNSR-III with whole-genome sequencing (WGS) data, including 7640 in development cohort and 1762 in external cohort 1 (Fig. 1).20Fig. 1Workflow for predictive model development. Footnotes: In validation cohort 2, the CHANCE-2 trial initially enrolled 6412 patients with a history of stroke or TIA. After excluding patients with a history of TIA, 5158 patients with a history of stroke remained. TIA refers to transient ischaemic attack. Workflow for predictive model development. Footnotes: In validation cohort 2, the CHANCE-2 trial initially enrolled 6412 patients with a history of stroke or TIA. After excluding patients with a history of TIA, 5158 patients with a history of stroke remained. TIA refers to transient ischaemic attack. More detailed information on participant recruitment and informed consent procedures is provided in the CNSR-III, CHANCE-2, and CRCS-5 studies and Expanded Methods in Supplementary Materials.21, 22, 23 Ethics This study used data from three independent cohorts, each with separate ethical approval. The CNSR-III study was approved by the Ethics Committee of Beijing Tiantan Hospital (Institutional Review Board approval number KY2015-001–01); the CHANCE-2 trial was approved by the same committee (approval number KY2019-035–07); and the CRCS-5 cohort was approved by the Institutional Review Board of Jeju National University Hospital, South Korea (IRB No. JEJUNUH 2025-06-012). Ethics approval was also obtained from all participating centres. Written informed consent was obtained from all participants or their legally authorised representatives before enrolment. Candidate predictors Baseline data, along with biological and urine samples, were collected by physicians through face-to-face interview in CNSR-III. A total of 309 candidate predictors were collected, categorised into the following groups: basic information, disease history, hospitalisation information, biomarkers, geographical location, National Institutes of Health Stroke Scale (NIHSS)/modified Rankin Scale (mRS) score, and PRS. For the disease history variables, missing values were assumed to indicate the absence of the event, while missing values for biomarkers were retained as missing. When developing the model, missing values were filled by extreme gradient boosting tree (XGBoost)'s default direction method. Detailed descriptions and statistical summaries of the 309 predictors are provided in Table S3. Additionally, the 3-month recurrence risk was calculated using the RRE model within the CNSR-III cohort (Described in Table S4, Expanded Methods in Supplementary Materials). Due to the unavailability of genetic scores specific to stroke recurrence, we used the iPRS (metaPRS) panel for stroke susceptibility from the GIGASTROKE study of East Asians as a proxy24 (Expanded Methods in Supplementary Materials). Post-stroke outcomes Patients in the CNSR-III were followed up at 3 months, 6 months, and annually from 1 to 5 years, with follow-up completion rates exceeding 90% in 2 years and around 80% in 5 years (Table S5). Follow-up assessments were conducted within a 14-day window from each scheduled time point. Patients were interviewed face-to-face by physicians within the first year, followed by telephone interviews conducted by trained clinical research coordinators thereafter. Regardless of the mode of follow-up, patient evaluations adhered to the standardised operational manual of the follow-up case report form (CRF) (Details in Expanded Methods in Supplementary Materials). Primary outcomes The primary outcomes were stroke recurrence, disability, and all-cause mortality at 3 months. Stroke recurrence was defined as any new stroke event identified through readmission records or patient self-reports during follow-up. Disability was defined using a binary mRS classification, with mRS 0–2 considered non-events, and mRS 3–6 deemed as events. Each fatality was confirmed using the Population Mortality Registration and Management Information System within the Life Registration Information Management System of the Chinese Cent for Disease Control and Prevention. For assessing stroke recurrence, mortality adjudicated as due to stroke were classified as recurrence events, whereas mortality from non–stroke causes (e.g., cancer) were treated as non-events, consistent with a cause-specific endpoint definition. In the assessment of disability, patients who died during the observation period were excluded from the analysis. For external validation in the CHANCE-2 trial and CRCS-5 cohort, follow-ups were conducted at 3 months and 1 year using similar evaluation methods. Model performance against three of the primary outcomes (recurrence, disability and mortality) was possible in the CHANCE-2 cohort. Model performance against two of the primary outcomes (disability and mortality, but not stroke recurrence) was possible in the CRCS-5 cohort in South Korea. Secondary outcomes The secondary outcomes included stroke recurrence, disability, and death assessed at six time points within five years after the index stroke (6 months, 1, 2, 3, 4, and 5 years). In the development and validation cohort 1 (testing dataset), all six time points were evaluated. However, due to data limitations, only secondary outcomes in 1-year for stroke recurrence, disability, and death were available for external validation in validation cohort 2 and disability and death (no stroke recurrence) for validation cohort 3. Model construction, evaluation, and explanation The XGBoost algorithm25 was used for model construction due to its scalability and effectiveness in handling nonlinear, sparse, and class-imbalanced classification data. To enhance model robustness and prevent overfitting, tenfold cross-validation was employed, along with an inner tenfold cross-validation for hyperparameter tuning through a random grid search. The optimisation criterion was the area under the receiver operating curve (AUC) (Fig. 1). This nested cross-validation approach provides more robust and unbiased results, as demonstrated in previous studies.25, 26, 27 Using this methodology, models were initially developed using all 309 candidate features. Based on feature importance rankings and expert recommendations, 23 features were selected for the final model, incorporating 16 global and domain-specific delta-NIHSS(admission-discharge) variables and 7 clinical easy-assessed features (age, gender, TOAST subtypes, history of TIA, stroke and diabetes, and mRS at discharge). The details about decision rules, relative weighting and final results are provided in the Expanded Methods in Supplementary Materials. Delta-NIHSS(admission-discharge) is defined as total NIHSS change from admission to discharge. The model was termed DISCO (Delta-NIHSS-Based Predictor for Post-Stroke Composite Outcomes). To enhance interpretability, Shapley additive explanations (SHAP) values27 were used to assess feature importance, complementing XGBoost's importance scores to guide feature selection and provide individualised explanations for each feature's contribution. The DISCO model was trained using the development cohort, tested in validation cohort 1, and externally validated in validation cohorts 2 and 3. Model performance was evaluated using sensitivity, specificity, F1-score, and AUC, with 95% confidence intervals (CIs) derived from 200 bootstrap samples. Model calibration was evaluated using 200 bootstrap resamples. Apparent and test performance were compared to estimate optimism, which was used to bias-correct the calibration curve. Predicted probabilities were grouped into deciles, plotting mean predicted risk against observed event rates. Calibration was further assessed with the Brier score and 95% confidence intervals derived from bootstrap percentiles. To benchmark XGBoost's performance, the DISCO model's feature set was applied to other algorithms, including logistic regression, naive Bayes, random forest, and support vector machine. Head-to-head comparisons with the RRE model, ASTRAL and IScore in the CNSR-III cohort was also performed (Expanded Methods in Supplementary Materials). DeLong's test was used to compare the AUC between DISCO model and RRE model in the same dataset because it provides a non-parametric and unbiased estimate for correlated Receiver Operating Characteristic (ROC) curves.28 All statistical analyses were performed in R version 4.2.0. A detailed description of the model construction process and software specifications is provided in the Expanded Methods in Supplementary Materials. Development of web application To facilitate clinical application of the DISCO model, we developed an R Shiny web application, available at https://www.discosysu.cn/. This tool allows users to evaluate individual risk based on the 23 features in the DISCO model and provides personalised explanations of the predicted risk using SHAP values for enhanced interpretability. Sensitivity analysis To evaluate the robustness of the DISCO model across different clinical scenarios, sensitivity analyses were conducted under the following two conditions:(1)Scenario 1: Patients with a history of stroke admitted to the hospital over 48 h of stroke onset and with a hospital stay of equal or more than 14 days (n = 8262).(2)Scenario 2: Patients with a history of stroke who did not undergo intravenous thrombolysis with rt-PA (TPA) or mechanical thrombectomy (n = 12,681). (1)Scenario 1: Patients with a history of stroke admitted to the hospital over 48 h of stroke onset and with a hospital stay of equal or more than 14 days (n = 8262). Scenario 1: Patients with a history of stroke admitted to the hospital over 48 h of stroke onset and with a hospital stay of equal or more than 14 days (n = 8262). (2)Scenario 2: Patients with a history of stroke who did not undergo intravenous thrombolysis with rt-PA (TPA) or mechanical thrombectomy (n = 12,681). Scenario 2: Patients with a history of stroke who did not undergo intravenous thrombolysis with rt-PA (TPA) or mechanical thrombectomy (n = 12,681). Factors associated with delta-NIHSS Given the strong predictive power of delta-NIHSS(admission-discharge) for stroke recurrence, disability, and mortality, a post-hoc exploratory analysis was conducted to examine factors influencing delta-NIHSS(admission-discharge) in the full CNSR-III analysis, and two sensitivity scenarios, using multivariable logistic regression. Delta-NIHSS(admission-discharge) was dichotomised into two categories: delta-NIHSS(admission-discharge) ≥ 0 (non-events) and delta-NIHSS(admission-discharge) < 0 (events). Associations between baseline clinical variables and both global and domain-specific delta-NIHSS(admission-discharge) were assessed. Additionally, z-transformed biomarkers were analysed to explore potential biological mechanisms underlying delta-NIHSS(admission-discharge) variation. Biomarkers with significant missingness (>5000 missing values), and those exhibiting strong collinearity (assessed via variance inflation factor [VIF]) were excluded. A total of 38 biomarkers were included in the final regression analysis. We applied a Bonferroni correction based on 26 hospital characteristics and 38 biomarkers (i.e., Padjusted = 0.05/26 for hospital characteristics and Padjusted = 0.05/38 for biomarkers). To ensure robustness, sensitivity analyses were conducted within the two predefined subgroups. The detailed workflow is presented in Figure S1. Risk stratification Risk stratification for stroke recurrence, mRS, and mortality at 3 months was assessed using relative risks (RRs) across the development cohort and two validation cohorts. RRs were calculated using the formula shown below (Equation (1)), with patients in the 0–60% risk category serving as the reference group:(1)RR=Motality(orMorbidity)inthecomparedgroupMotality(orMorbidity)inthereferencegroup Role of the funding source The funders had no involvement in study design, data collection, data analyses, data interpretation, or the writing of the report. Results Study sample Among the 11,313 patients in the development cohort, 2627 patients in validation cohort 1, 5158 patients in validation cohort 2, and 3855 patients in validation cohort 3, the median ages were 62, 63, 65, and 73 years, respectively. The proportion of female participants was 31.10%, 32.74%, 32.80%, and 43%, respectively, and the median NIHSS scores at admission was 3, 3, 2, and 4 (Table 1). For mechanical thrombectomy, the proportions were 0.31%, 0.15%, and 0% respectively and they were 8.94%, 8.37%, and 0% for thrombolysis. According to the TOAST classification, most patients were diagnosed with large artery atherosclerosis (LAA), small vessel disease (SVS), or stroke of undetermined cause. At the 3-month follow-up, stroke recurrence rates were 6% (n = 716) for the development cohort, 6% (n = 147) for internal validation, and 7% (n = 356) for external validation. Equivalent disability rates were 14% (n = 1556), 17% (n = 448), and 3% (n = 147), and mortality rates were 1% (n = 134), 2% (n = 46), and 0.5% (n = 25). Rates of these three stroke outcomes during the subsequent 6-month follow-up, and annual follow-ups from years 1–5 are detailed in Table S1. Cumulative incidences curve for stroke recurrence at 3 months are presented in Figure S2.Table 1Baseline characteristics of the study cohorts.FeaturesDevelopment cohort (N = 11,313)Validation cohort 1 (N = 2627)Validation cohort 2 (N = 5158)Validation cohort 3 (N = 3855)Number of hospital/centres169402021Age, median (IQR), y62 (54–70)63 (56–71)65 (57–71)73 (61–80)Femalea, no. (%)3518 (31.10)860 (32.74)1692 (32.80)1659 (43)Delta NIHSSb (IQR)1 (0–3)1 (0–3)1 (0–1)1 (0–3)Admission NIHSS, median (IQR)3 (2–6)3 (2–6)2 (1–3)4 (2–9)Geographical location North China, no. (%)3107 (27.46)1096 (32.72)916 (17.76)0 Northeast China, no. (%)1329 (11.75)205 (7.80)673 (13.05)0 East China, no. (%)2754 (24.34)670 (25.50)2032 (39.40)0 South China, no. (%)479 (4.23)328 (12.49)60 (1.16)0 Central China, no. (%)2448 (21.64)239 (9.10)757 (14.68)0 Southwest China, no. (%)184 (1.63)88 (3.35)274 (5.31)0 Northwest China, no. (%)1012 (8.95)1 (0.038)375 (7.27)0 Korean, no. (%)0003855 (100)Disease history Transient ischaemic attack, no. (%)229 (2.02)56 (2.13)41 (0.79)87 (2.3) Diabetes mellitus, no. (%)2677 (23.66)597 (22.73)1260 (24.43)1184 (30.7) Hypertension, no. (%)7088 (62.65)1698 (64.64)3153 (61.13)2512 (65.2) Lipid metabolism disorder, no. (%)827 (7.31)226 (8.60)614 (11.90)1432 (37.10) Coronary heart disease, no. (%)1208 (10.68)251 (9.55)421 (8.16)381 (9.9)TOASTcsubtypes Large artery atherosclerosis, no. (%)2924 (25.84)669 (25.47)1446 (28.03)1162 (30.1) Cardioembolism, no. (%)671 (5.93)186 (7.08)/792 (20.5) Small vessel stroke, no. (%)2557 (22.60)531 (20.21)1708 (33.11)888 (23) Other cause, no. (%)140 (1.24)34 (1.29)206 (3.99)201 (5.2) Undetermined cause, no. (%)5021 (44.38)1207 (45.95)1533 (29.72)812 (21.1)Hospitalisation treatment Anticoagulation, no. (%)1185 (10.47)253 (9.63)52 (1.01)577 (15.0) Lipid lowering, no. (%)10,851 (95.92)2525 (96.12)1964 (38.08)3201 (83.0) Antiplatelet, no. (%)10,896 (96.31)2554 (97.22)5158 (100)3074 (79.7) Antihypertensive treatment, no. (%)5164 (45.65)1319 (50.21)2491 (48.29)1737 (45.1) Hypoglycemic treatment, no. (%)2910 (25.72)648 (24.67)1375 (26.66)942 (24.4) Thrombolytic treatment, no. (%)1011 (8.94)220 (8.37)0 (0)517 (13.4) Mechanical thrombectomy, no. (%)35 (0.31)4 (0.15)0 (0)458 (11.9)aParticipants provided self-reported information on gender identity during data collection (male, female, or other as specified). For the present analysis, gender categories were used as reported.bThe term refers to the change in NIHSS score from hospital admission to discharge, calculated as NIHSS score at admission—NIHSS score at discharge.cThe TOAST classification is a commonly used system for categorising the aetiology of ischaemic strokes. It includes the following subtypes: Large Artery Atherosclerosis, Cardioembolism, Small Vessel Disease, Other Cause and Undetermined Cause. Table 1Baseline characteristics of the study cohorts.FeaturesDevelopment cohort (N = 11,313)Validation cohort 1 (N = 2627)Validation cohort 2 (N = 5158)Validation cohort 3 (N = 3855)Number of hospital/centres169402021Age, median (IQR), y62 (54–70)63 (56–71)65 (57–71)73 (61–80)Femalea, no. (%)3518 (31.10)860 (32.74)1692 (32.80)1659 (43)Delta NIHSSb (IQR)1 (0–3)1 (0–3)1 (0–1)1 (0–3)Admission NIHSS, median (IQR)3 (2–6)3 (2–6)2 (1–3)4 (2–9)Geographical location North China, no. (%)3107 (27.46)1096 (32.72)916 (17.76)0 Northeast China, no. (%)1329 (11.75)205 (7.80)673 (13.05)0 East China, no. (%)2754 (24.34)670 (25.50)2032 (39.40)0 South China, no. (%)479 (4.23)328 (12.49)60 (1.16)0 Central China, no. (%)2448 (21.64)239 (9.10)757 (14.68)0 Southwest China, no. (%)184 (1.63)88 (3.35)274 (5.31)0 Northwest China, no. (%)1012 (8.95)1 (0.038)375 (7.27)0 Korean, no. (%)0003855 (100)Disease history Transient ischaemic attack, no. (%)229 (2.02)56 (2.13)41 (0.79)87 (2.3) Diabetes mellitus, no. (%)2677 (23.66)597 (22.73)1260 (24.43)1184 (30.7) Hypertension, no. (%)7088 (62.65)1698 (64.64)3153 (61.13)2512 (65.2) Lipid metabolism disorder, no. (%)827 (7.31)226 (8.60)614 (11.90)1432 (37.10) Coronary heart disease, no. (%)1208 (10.68)251 (9.55)421 (8.16)381 (9.9)TOASTcsubtypes Large artery atherosclerosis, no. (%)2924 (25.84)669 (25.47)1446 (28.03)1162 (30.1) Cardioembolism, no. (%)671 (5.93)186 (7.08)/792 (20.5) Small vessel stroke, no. (%)2557 (22.60)531 (20.21)1708 (33.11)888 (23) Other cause, no. (%)140 (1.24)34 (1.29)206 (3.99)201 (5.2) Undetermined cause, no. (%)5021 (44.38)1207 (45.95)1533 (29.72)812 (21.1)Hospitalisation treatment Anticoagulation, no. (%)1185 (10.47)253 (9.63)52 (1.01)577 (15.0) Lipid lowering, no. (%)10,851 (95.92)2525 (96.12)1964 (38.08)3201 (83.0) Antiplatelet, no. (%)10,896 (96.31)2554 (97.22)5158 (100)3074 (79.7) Antihypertensive treatment, no. (%)5164 (45.65)1319 (50.21)2491 (48.29)1737 (45.1) Hypoglycemic treatment, no. (%)2910 (25.72)648 (24.67)1375 (26.66)942 (24.4) Thrombolytic treatment, no. (%)1011 (8.94)220 (8.37)0 (0)517 (13.4) Mechanical thrombectomy, no. (%)35 (0.31)4 (0.15)0 (0)458 (11.9)aParticipants provided self-reported information on gender identity during data collection (male, female, or other as specified). For the present analysis, gender categories were used as reported.bThe term refers to the change in NIHSS score from hospital admission to discharge, calculated as NIHSS score at admission—NIHSS score at discharge.cThe TOAST classification is a commonly used system for categorising the aetiology of ischaemic strokes. It includes the following subtypes: Large Artery Atherosclerosis, Cardioembolism, Small Vessel Disease, Other Cause and Undetermined Cause. Baseline characteristics of the study cohorts. Participants provided self-reported information on gender identity during data collection (male, female, or other as specified). For the present analysis, gender categories were used as reported. The term refers to the change in NIHSS score from hospital admission to discharge, calculated as NIHSS score at admission—NIHSS score at discharge. The TOAST classification is a commonly used system for categorising the aetiology of ischaemic strokes. It includes the following subtypes: Large Artery Atherosclerosis, Cardioembolism, Small Vessel Disease, Other Cause and Undetermined Cause. Feature selection The feature importance of all 309 candidate predictors, assessed using SHAP values and XGBoost importance scores, is presented in Table S6, with the top 20 features visualised in Figure S3. Notably, the global delta-NIHSS(admission-discharge) emerged as the strongest predictor of stroke recurrence at 3 months (Figure S3 and Table S6), accounting for more than 30% of the model's predictive power. Additional key predictors included mRS at discharge, molecular biomarkers such as choline and fasting blood glucose, domain-specific delta-NIHSS(admission-discharge), history of disease, and age (Figure S3 and Table S6). In contrast, the PRS for stroke susceptibility, despite being normally distributed, contributed to less than 0.5% to the prediction of 3-month stroke recurrence (Figure S4, Table S6), suggesting that distinct genetic factors and biological mechanisms drive stroke occurrence and recurrence risk. Based on a combination of feature importance rankings and clinical feasibility, as recommended by stroke specialists, we selected 23 features for the final model (Table S7). These included the global delta-NIHSS(admission-discharge), 15 domain-specific delta-NIHSS(admission-discharge) features, and 7 core clinical variables: mRS at discharge, age, gender, TOAST classification, history of stroke, diabetes and TIA. This final 23-feature model was designated as the DISCO model (Fig. 1). Model performance In internal cross-validation, the DISCO model for predicting 3-month stroke recurrence achieved an AUC of 0.805 (95% CI: 0.748–0.863) and an F1 score of 0.975. In the two validation cohorts for stroke recurrence, the AUCs were 0.731 (0.683–0.78) in validation cohort 1 (testing dataset) and 0.803 (0.776–0.83) in validation cohort 2, respectively (Fig. 2A). The DISCO model outperformed the RRE model for 3-month recurrence risk, which yielded an AUC of 0.556 (95% CI: 0.53–0.581) in CNSR-III (Fig. 2A). Additional models using all 309 features, revascularisation features, global and domain-specific delta-NIHSS(admission-discharge) features, biomarker features, and the top 20 features were also evaluated. The 23-feature DISCO model demonstrated performance comparable to the full 309-feature model (AUC: 0.815; 95% CI: 0.762–0.868) for 3-month stroke recurrence (Figure S5). Other models exhibited lower AUCs than the DISCO model.Fig. 2ROC curves and feature importance of the DISCO model. Footnotes: (A) The ROC curve illustrating the predictive performance of the DISCO model for 3-month stroke recurrence, with the RRE model estimates derived from 7261 patients in the CNSR-III cohort via the algorithm on the RRE website (https://www.nmr.mgh.harvard.edu/RRE/). (B) Feature importance rankings. The ‘Others’ represents 14 features utilised in DISCO model. (C) The ROC curve for predicting 3-month disability (mRS >2), with the ASTRAL score estimates derived from 11,544 patients in CNSR-III cohort. (D) The ROC curve for the DISCO model in predicting 3-month mortality, with IScore estimated derived from 11,544 patients in CNSR-III cohort visa the IScore website (https://www.sorcan.ca/iscore/index.html). ROC: Receiver Operating Characteristic; AUC: the area under the receiver operating curve. ROC curves and feature importance of the DISCO model. Footnotes: (A) The ROC curve illustrating the predictive performance of the DISCO model for 3-month stroke recurrence, with the RRE model estimates derived from 7261 patients in the CNSR-III cohort via the algorithm on the RRE website (https://www.nmr.mgh.harvard.edu/RRE/). (B) Feature importance rankings. The ‘Others’ represents 14 features utilised in DISCO model. (C) The ROC curve for predicting 3-month disability (mRS >2), with the ASTRAL score estimates derived from 11,544 patients in CNSR-III cohort. (D) The ROC curve for the DISCO model in predicting 3-month mortality, with IScore estimated derived from 11,544 patients in CNSR-III cohort visa the IScore website (https://www.sorcan.ca/iscore/index.html). ROC: Receiver Operating Characteristic; AUC: the area under the receiver operating curve. Delta-NIHSS(admission-discharge) was the most important predictor in the DISCO model, contributing 34.87% to the model's prediction power (Fig. 2B). Owing to k-fold cross-validation, the apparent and bias-corrected curves were nearly identical and the brier scores are 0.025 (95% CI: 0.02–0.032) and 0.032 (95% CI: 0.028–0.036) in validation cohorts 1 and 2, respectively (Figures S6 and S7). Decision Curve Analysis indicated a positive net benefit for the DISCO model. The optimal cut-off value in validation cohort 1, based on the maximum Youden index, was 0.154 (sensitivity = 0.756, specificity = 0.619, Youden index = 0.375), higher than that of the RRE model (Figure S8). In validation cohort 2, the sensitivity, specificity and Youden index are 0.868, 0.640 and 0.508 (Figure S9). Additionally, the DISCO model demonstrated the highest AUC for predicting 3-month stroke recurrence compared to other five statistical or machine learning algorithms including Cox regression, random forest, support vector machine, naive bayes, and the TabPFN, which is a deep learning model developed recently for tabular data (Fig. 3).29 AUCs for internal and external validation, along with parameters for secondary endpoint events, are detailed in Tables S8 and S9.Fig. 3ROC curves comparing methods using 23 DISCO model features for predicting main outcomes in 3 moths in validation cohort 2. Footnotes: The nested cross-validated XGBoost model achieved the highest AUC compared to the six other models, including traditional statistical models, machine learning models, and deep learning models for tabular data. All models were developed using the same features as the DISCO model, trained in the development cohort, and evaluated in external cohort 2. Due to the lack of the record of endpoint for disability (mRS), the comparison for disability lacks COX regression and limits in 6 methods. Fig. 3A, B and C represts the prediction of stroke recurrence, binary mRS and mortality in 3 months. ROC represents Receiver Operating Characteristic; AUC represents the area under the receiver operating curve. ROC curves comparing methods using 23 DISCO model features for predicting main outcomes in 3 moths in validation cohort 2. Footnotes: The nested cross-validated XGBoost model achieved the highest AUC compared to the six other models, including traditional statistical models, machine learning models, and deep learning models for tabular data. All models were developed using the same features as the DISCO model, trained in the development cohort, and evaluated in external cohort 2. Due to the lack of the record of endpoint for disability (mRS), the comparison for disability lacks COX regression and limits in 6 methods. Fig. 3A, B and C represts the prediction of stroke recurrence, binary mRS and mortality in 3 months. ROC represents Receiver Operating Characteristic; AUC represents the area under the receiver operating curve. For predicting 3-month post-discharge disability events, the DISCO model achieved an internal validation AUC of 0.892 (95% CI: 0.872–0.929) in development cohort and 0.904 (95% CI: 0.888–0.92) in validation cohort 1 (testing dataset). As for external validation, the DISCO model had AUCs of 0.852 (95% CI: 0.813–0.891) in validation cohort 2 and 0.953 (0.944–0.96) in validation cohort 3 (Fig. 2C, Tables S8 and S9). The DISCO model outperformed the ASTRAL model, which achieved an AUC of 0.794 (95% CI: 0.782–0.807) for predicting 3-month disability. In terms of 3-month mortality prediction, the DISCO model achieved an AUC of 0.943 (95% CI: 0.858–1) and 0.872 (0.826–0.917) in development cohort and validation cohort 1 (testing dataset) for internal validation. In the external validation cohorts, AUCs were 0.813 (95% CI: 0.719–0.908) in external cohort 2, and 0.817 (0.789–0.843) in external cohort 3 (Fig. 2D, Tables S8 and S9). The AUCs for 1-month and 1-year mortality using the IScore model were 0.815 (95% CI: 0.776–0.855) and 0.82 (95% CI: 0.782–0.859), respectively. AUCs decreased with longer follow-up durations (Tables S8, S9). In the development and two validation cohorts, AUCs for 1-year stroke recurrence decreased from 0.805 (three month stroke recurrence) to 0.72 (1 year stroke recurrence). Sensitivity analyses did not show significant AUC reductions for 3-month stroke recurrence in scenarios 1 and 2 in external cohort 1 (Figure S11). Adding PRS didn't improve AUC for predicting 3-month stroke recurrence in external cohort 1 (Figure S12). Influencing factors of global and domain-specific delta-NIHSS(admission-discharge) To better understand the significant contribution of delta-NIHSS(admission-discharge) in predicting stroke composite outcomes, we investigated the associations between hospitalisation factors and both global and domain-specific delta-NIHSS(admission-discharge). In a retrospective analysis, the distribution of NIHSS scores at admission and discharge differed between patients who did not experience stroke recurrence at 3 months (Fig. 4A) and those who did (Fig. 4B). Patients without recurrence tended to have a higher proportion of NIHSS scores ≤3 at discharge compared to admission, whereas patients with recurrence were more likely to have NIHSS scores >3 at discharge. In addition, based on NIHSS changes from admission to discharge across different baseline NIHSS scores, patients with worsening delta-NIHSS(admission-discharge) during hospitalisation tend to have poorer prognoses (Table S10). The distribution of both global and domain-specific delta-NIHSS(admission-discharge) was zero-inflated (Figure S13), prompting us to dichotomise patients into events (delta−NIHSS<0) and non-events (delta−NIHSS≥0).Fig. 4Patterns and potential influencing factors of NIHSS changes in CNSR-III. Footnotes: (A) NIHSS changes from admission to discharge in patients without recurrent stroke. (B) NIHSS changes from admission to discharge in patients with recurrent stroke. (C) Results of multivariable regression analysis showing the associations between baseline clinical variables and dichotomised delta-NIHSS (admission-discharge). We applied a Bonferroni correction based on 26 hospital characteristics. NIHSS represents National Institutes of Health Stroke Scale; TOAST1 represnts Large Artery Atherosclerosis stroke; TOAST2 represents Cardioembolism stroke; TOAST3 represnts Small Vessel Stroke; TOAST5 represnts Undetermined Cause stroke. Delta-NIHSS Delta-NIHSS is the National Institutes of Health Stroke Scale score between administration and discharge.; Delta NIH1A represents the change in the level of consciousness (responsiveness) score between admission and discharge; Delta NIH1B: Denotes the change in the level of consciousness—questions (month and age orientation) score between admission and discharge; Delta NIH1C: Refers to the change in the level of consciousness—commands (eye opening and grip release) score between admission and discharge; Delta NIH2 represents the change in best gaze (horizontal eye movement) score between admission and discharge; Delta NIH3 represents the change in visual field loss score between admission and discharge; Delta NIH4R represents the change in facial palsy (facial muscle movement) score between admission and discharge; Delta NIH5A represents the change in left arm motor function score between admission and discharge; Delta NIH5B represents the change in right arm motor function score between admission and discharge; Delta NIH6A represents the change in left leg motor function score between admission and discharge; Delta NIH6B represents the change in right leg motor function score between admission and discharge; Delta NIH7 represents the change in limb ataxia (finger–nose or heel–shin coordination) score between admission and discharge; Delta NIH8 represents the change in sensory (pinprick or light touch response) score between admission and discharge; Delta NIH9 represents the change in best language (comprehension and expression) score between admission and discharge; Delta NIH10 represents the change in dysarthria (articulation and speech clarity) score between admission and discharge; Delta NIH11 represents the change in extinction and inattention (neglect) score between admission and discha. Patterns and potential influencing factors of NIHSS changes in CNSR-III. Footnotes: (A) NIHSS changes from admission to discharge in patients without recurrent stroke. (B) NIHSS changes from admission to discharge in patients with recurrent stroke. (C) Results of multivariable regression analysis showing the associations between baseline clinical variables and dichotomised delta-NIHSS (admission-discharge). We applied a Bonferroni correction based on 26 hospital characteristics. NIHSS represents National Institutes of Health Stroke Scale; TOAST1 represnts Large Artery Atherosclerosis stroke; TOAST2 represents Cardioembolism stroke; TOAST3 represnts Small Vessel Stroke; TOAST5 represnts Undetermined Cause stroke. Delta-NIHSS Delta-NIHSS is the National Institutes of Health Stroke Scale score between administration and discharge.; Delta NIH1A represents the change in the level of consciousness (responsiveness) score between admission and discharge; Delta NIH1B: Denotes the change in the level of consciousness—questions (month and age orientation) score between admission and discharge; Delta NIH1C: Refers to the change in the level of consciousness—commands (eye opening and grip release) score between admission and discharge; Delta NIH2 represents the change in best gaze (horizontal eye movement) score between admission and discharge; Delta NIH3 represents the change in visual field loss score between admission and discharge; Delta NIH4R represents the change in facial palsy (facial muscle movement) score between admission and discharge; Delta NIH5A represents the change in left arm motor function score between admission and discharge; Delta NIH5B represents the change in right arm motor function score between admission and discharge; Delta NIH6A represents the change in left leg motor function score between admission and discharge; Delta NIH6B represents the change in right leg motor function score between admission and discharge; Delta NIH7 represents the change in limb ataxia (finger–nose or heel–shin coordination) score between admission and discharge; Delta NIH8 represents the change in sensory (pinprick or light touch response) score between admission and discharge; Delta NIH9 represents the change in best language (comprehension and expression) score between admission and discharge; Delta NIH10 represents the change in dysarthria (articulation and speech clarity) score between admission and discharge; Delta NIH11 represents the change in extinction and inattention (neglect) score between admission and discha. When regressing dichotomised global and domain-specific delta-NIHSS(admission-discharge) on 28 hospitalisation characteristics, with Bonferroni correction applied, we found that age, geographic location, NIHSS at admission, anticoagulation treatment, TPA treatment, and a history of atrial fibrillation were significantly associated with multiple delta-NIHSS(admission-discharge) domains (Fig. 4C, Table S11). Additional factors, including lipid-lowering treatments, antiplatelet and hypoglycemic treatments, and mechanical thrombectomy, were significantly associated with specific delta-NIHSS(admission-discharge) domains. By replacing the overall and domain-specific delta-NIHSS in DISCO model as NIHSS in admission or discharge, the AUCs of DISCO model decreased in both the external cohort 1 and 2 (Figure S14). In scenario 1 which restricted patients to those who arrived at the hospitals within 48 h and stayed no more than 14 days, mechanical thrombectomy became more strongly and positively associated with multiple delta-NIHSS(admission-discharge), while rt-PA remained negatively associated with several delta-NIHSS(admission-discharge) (Figure S15, Table S12). In scenario 2, after excluding patients who received TPA or mechanical thrombectomy, age, geographical location, NIHSS at admission, anticoagulation treatment, and a history of atrial fibrillation remained the most significant factors associated with dichotomised delta-NIHSS(admission-discharge) (Figure S16, Table S13). At the molecular level, when regressing dichotomised delta-NIHSS(admission-discharge) against 38 biomarkers assayed at the time of patient admission, IL6 demonstrated the most widespread Bonferroni-corrected significance across domain-specific delta-NIHSS(admission-discharge) in all three scenarios (Figures S17–S19, Tables S14–S16). Risk stratification and clinical application The DISCO model demonstrates strong capacity in identifying patients with high-risk. For stroke recurrence, disability, and mortality at 3 months, the relative risks (RRs) for patients in the highest risk group (99%–100%) compared to the reference group (0%–60%) were 19.27, 36.59, and 225.5, respectively (Fig. 5). The model's stratification performance was consistent in validation cohort 1 (Figure S20), where the RRs for the highest-risk group compared to the reference group patients were 38.21, 42.17, and 201.27, and in validation cohort 2 (Figure S21), where the RRs were 36.64, 92.24, and 104.71.Fig. 5Risk stratification performance of the DISCO model for 3-months stroke outcomes. Risk stratification performance of the DISCO model for 3-months stroke outcomes. To illustrate personalised risk assessment, two patients from validation cohort 1 were randomly selected for analysis (Figure S22). For Patient A, the delta-NIHSS4 (poor facial palsy, −2) had the highest SHAP value, suggesting the need for focused rehabilitation on facial palsy post-discharge. For Patient B, delta-NIHSS5A (poor left motor arm, −2) had the highest SHAP value, indicating a need for post-discharge motor rehabilitation. These results illustrate the utility of personalised feature importance explanations to inform tailored clinical management strategies. To facilitate real-world applications, we developed a web-based application for risk estimation and personalised result explanations, which is accessible at https://www.discosysu.cn. Discussion In this study, we employed a robust machine learning framework, incorporating explainable artificial intelligence techniques, to systematically evaluate the importance of 309 hospitalisation features in predicting three critical post-stroke outcomes: recurrence, disability, and mortality at 3 months, as well as across six additional time points over 5 years. For the first time, we identified clinically accessible global and domain-specific delta-NIHSS(admission-discharge) as key predictors of post-stroke recurrence, disability and mortality at 3 months. Based on these findings, we developed and externally validated the DISCO model, which demonstrated dramatically improved discriminative ability compared to existing models, achieving AUCs more than 0.8 for 3-month stroke recurrence and mortality, 0.9 for disability in 3 validation cohorts. The NIHSS is a simple, widely used clinical measure that offers valuable insights into stroke severity and recovery.30 It is used as an endpoint in randomised controlled trials (RCTs)30,31 and associated with 90-day disability.14,32 Previous study has indicated the importance of NIHSS at admission or discharge in stroke outcome prediction.33, 34, 35, 36 However, NIHSS at admission or discharge alone did not outperform delta-NIHSS(admission–discharge), highlighting the prognostic value of dynamic NIHSS changes. Delta-NIHSS(admission-to-discharge) reflected patient's individualised biological and clinical responses to treatment during hospitalisation. Besides, the definitions of stroke recurrence, disability and NIHSS are based on symptomatic observation, which may explain why delta-NIHSS(admission-discharge) is a strong predictor of post-stroke composite outcomes. The predictive contribution of delta-NIHSS to 3-month composite outcomes underscores the reassessment at discharge and the importance of recovery during hospitalisation for secondary ischaemic stroke prevention and improving long-term quality of patients. Previous observational and genetic studies have suggested that specific NIHSS domains correspond to distinct biological mechanisms and testing modalities,30,37,38 such as NIHSS 1A, which aligns with the Montreal Cognitive Assessment Scale (MOCA).30,39 Our findings of global and domain-specific delta-NIHSS(admission-discharge) as novel and essential predictive variables for 3-month stroke composite outcomes, are consistent with the biological understanding of NIHSS. These findings provide new insights into the mechanisms of post-stroke recurrence and disability. Since domain-specific delta-NIHSS(admission-discharge) reflects deficits in a specific functional domain, its strong predictive power for 3-month stroke recurrence suggests that recurrence may represent the deterioration or progression of the initial ischaemic stroke rather than the onset of a new stroke. The DISCO model demonstrated robust identification of patients at high-risk with substantially elevated relative risks in both internal and external validation, supporting its utility for future risk stratification and personalised management. By relying solely on standard clinical variables, the model incurs no additional costs, ensuring feasibility and equity of implementation across diverse health systems. To facilitate translation into practice, we developed a web-based tool that allows clinicians to generate individualised risk estimates for recurrence, disability, and mortality at 3 months (https://www.discosysu.cn/). Similar AI-assisted personalised risk information and expert recommendations has been proven to be an effective non-pharmacological intervention in multicentre RCTs, such as those focused on parental check-ups for term planned delivery40 and primary diabetes care.41 Broad implementation of the DISCO model has the potential to reduce long-term disability, alleviate caregiver and societal burden, and advance global efforts to decrease stroke-related disability-adjusted life years (DALYs). In clinical practice, the DISCO model offers several key applications. First, it can enable personalised discharge planning, which helps determine whether these patients are suitable for rehabitation medical centre or home-based care. Second, it can support risk stratification in clinical research, guiding personalised follow-up, optimising rehabilitation strategies, and improving the allocation of limited healthcare resources toward those most at risk. Third, it can provide individualised recovery expectations, allowing clinicians, nurses, and caregivers to set realistic goals and tailor patient-centred care. The study has several limitations. First, although the DISCO model underwent a validation process in a cohort of patients from South Korea (n = 3855), its performance in individuals of other ancestry beyond the East Asia cohort remains to be established. Further research in different regions would provide greater insights on generalisabiilty. It is also notable that model performance for only disability and mortality (but not stroke recurrence) were evaluated in this cohort; external validation for stroke recurrence was not feasible due to data availability. Given the model's emphasis on clinical accessibility, the inclusion of 23 easily assessable features facilitates its validation in independent, multi-ancestry cohorts, which would help evaluate and extend its applicability. Second, the PRS used in this study were developed for ischaemic stroke occurrence rather than recurrence, as no PRS models specifically designed for stroke outcomes currently exist. Therefore, we cannot rule out the potential contribution of genetic factors to stroke outcome prediction. Further investigations are warranted when genetic studies specifically targeting secondary stroke prevention become available. Third, the mechanistic understanding of global and domain-specific delta-NIHSS(admission-discharge) remains limited. While our observational analysis identified several factors significantly associated with delta-NIHSS(admission-discharge), future research is needed to explore their causal effects. Such investigations may provide insights into more precise treatment and intervention strategies for acute stroke management and post-stroke neurorehabilitation. Fourth, the external cohort 1 has lower AUC in primary and secondary outcomes. This is the result of the combination of reduced sample size and event count. But, consistent or higher AUCs in cohorts 2 and 3 support overall robustness. In conclusion, the DISCO model, an AI-driven, delta-NIHSS-based predictor of composite ischaemic stroke outcomes, demonstrates high accuracy, robustness, clinical accessibility, and explainability. By highlighting the significance of delta-NIHSS(admission-discharge) representing biological and clinical responses during hospitalisation, it provides a valuable tool for personalised risk prediction, supporting personalised neurorehabilitation interventions for patients with a history of ischaemic stroke. Contributors ShC, YDW, SL, and SiC conceived and designed the study. QZ, HL, SL, and SiC contributed to the conceptualisation of the research framework. ShC and YW performed the formal analyses. YJW, HL, SiC, and SL obtained the funding for this study. ShC and QZ conducted the investigation of previous study. Methodology was developed by YDW, ShC, and HL. Project administration was managed by SL and SiC. Resources were provided by HZ, YS, JZ, CG, XL,HZ, JW, XM, ZL, HL, YJW, and SiC. YDW developed the code used in the analyses. Validation of the findings was undertaken by ShC, SiC, YDW, SL, CHK, HJK, JGK, HJB, JL, JCC, JYL and HJ. Visualisation was performed by SiC, XY, and ZH. ShC drafted the original manuscript. SL, SiC, HL, ZL, SD and ZW critically reviewed and edited the manuscript. ShC, YW, SiC and SL assessed and verified the underlying data. All authors read and approved the final version of the manuscript and took final responsibility for the decision to submit it for publication. Data sharing statement The DISCO model can be assessed through online software (https://www.discosysu.cn), and the application code is available at (https://github.com/liusylab/Disco_model_external_validation). Raw data underlying the results of this study are available from the corresponding authors upon reasonable request, in accordance with ethical guidelines. Declaration of interests We declare no competing interests.