Research status of prediction models for post-stroke neurological deterioration:a scoping review
10.3969/j.issn.1672-5921.2025.04.003
- VernacularTitle:卒中后神经功能恶化预测模型研究现状的范围综述
- Author:
Xiaohui SUN
1
;
Zhuoma PENGMAO
;
Xiaowei SONG
;
Ceshu GAO
;
Jian WU
Author Information
1. 102218 清华大学北京清华长庚医院神经内科
- Publication Type:Journal Article
- Keywords:
Stroke;
Neurological deterioration;
Prediction model;
Systematic review
- From:
Chinese Journal of Cerebrovascular Diseases
2025;22(4):235-251
- CountryChina
- Language:Chinese
-
Abstract:
Objective To evaluate the modeling characteristics and predictive performance of models for predicting post-stroke neurological deterioration(ND)published in existing literatures.Methods Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses(PRISMA)guidance for scoping reviews,a comprehensive search was conducted in PubMed,CINAHL,Cochrane Library,Embase,Web of Science,Scopus,CNKI,Wanfang Data,and VIP databases from inception to December 15,2024.The search strategy combined Medical Subject Headings(MeSH)and free-text terms,with key words including"Stroke""Ischemic Stroke""Neurological Deterioration""Nomograms""Risk Prediction""Predictive Models""卒中""脑梗死""脑出血""神经功能恶化"and"预测模型".Base on the data extraction checklist and critical appraisal,data extraction covered three domains:(1)basic characteristics,including author,publication year,country,study design(retrospective,prospective,registry-based),sample source(single-center,multicenter),stroke subtypes(acute ischemic stroke[AIS]-conservative therapy,AIS-intravenous thrombolysis[IVT],AIS-endovascular therapy[EVT],intracerebral hemorrhage[ICH]),ND time windows(acute[≤72 h],subacute[≤ 7 d],long-term[≤90 d]),and outcome types(single/composite endpoints);(2)model evaluation metrics,including missing data handling(complete-case analysis,multiple imputation),model development methodologies(multivariate Logistic regression,least absolute shrinkage and selection operator regression,machine learning),presentation formats(nomograms,web calculators,risk prediction tool),discrimination(area under the curve,C-index),calibration(Hosmer-Lemeshow test,calibration curve and slope),clinical utility(decision curve analysis[DCA],global metrics Brier score,R2,AIC),sample size(training set,internal validation set,external validation set),sample size requirements(events per variable[EPV]≥10 to mitigate overfitting),and validation(internal/external);(3)predictor features,including selection strategies(prior knowledge-driven,univariate analysis),quantity,and attributes(demographics,medical history,physical examination,treatment intervention information,imaging/laboratory indicators).Predictive models that meet exclusion criteria from prior literature were analyzed by their discrimination,calibration,clinical utility and global metrics.Forest plots were utilized to visualize discrimination(evaluated via difference in area under the curve)of the extracted models.The prediction model risk of bias assessment tool(PROBAST)was applied to assess bias risk and clinical applicability.Occurrence frequencies of the post-stroke neurological deterioration predictors were ranked and the top 6 high-frequency predictors were extracted.Results(1)Among 3 728 screened studies,25 were included based on the inclusion and exclusion criteria.(2)Basic characteristics:retrospective(72％[18/25])and single-center(64％[16/25])designs dominated.With most models targeted on AIS(92％[23/25]),and the rest(8％[2/25])on ICH.ND was primarily defined by neurological scale changes(60％[15/25];e.g.,National Institutes of Health stroke scale[NIHSS]score increase or Glasgow coma scale[GCS]score decrease),with time windows categorized as acute(36％[9/25]),subacute(48％[12/25]),or long-term(16％[4/25]).(3)Model evaluation:multivariate Logistic regression(96％[24/25])and nomograms(88％[22/25])were predominant.Only 24％(6/25)explicitly addressed missing data handling methods,and 52％(13/25)with EPV≥10.The median area under the curve was 0.865(range:0.650-0.981).44％(11/25)of the studies reported calibration curves,and 4％(1/25)reported calibration slopes.All studies utilized DCA to validate their clinical applicability,84％(21/25)of the studies conducted internal validation,while only 32％(8/25)conducted external validation.PROBAST evaluation revealed low overall bias risk in 8％(2/25;no error across participant,predictor,outcome,or analysis domains)and low clinical applicability risk in 44％(11/25;alignment with target populations,accessible predictors,and clinically relevant outcomes)of the studies.(4)Predictors:64％(16/25)of the predictor were screened predominantly through the prior knowledge-driven based strategy.The top 6 high-frequency predictors are NIHSS score(64％[16/25]),age(36％[9/25]),blood glucose/diabetes(36％[9/25]),blood pressure/hypertension(32％[8/25]),the Alberta stroke program early CT score(20％[5/25]),and neutrophil-to-lymphocyte ratio(20％[5/25]).AIS-ND predictors emphasized readily available metrics,such as NIHSS(65％[15/23]),age(35％[8/23]),while ICH-ND primarily relied on imaging markers(e.g.,baseline hematoma volume[2/2],location[1/2]).Conclusion Current post-stroke ND predictive models demonstrate satisfactory performance on discrimination and multimodal integration,but their practical application are hindered by insufficient calibration quantification,high bias risk,and limited clinical translatability.
