← Back
NeuroTrials.ai
Neurology Clinical Trial Database

PRERISK

PRERISK: A Personalized, Artificial Intelligence–Based and Statistically–Based Stroke Recurrence Predictor for Recurrent Stroke

Share Share Copied! Compare

Year of Publication: 2024

Authors: Giorgio Colangelo, DS, PhD; Marc Ribo, ..., PhD; Marta Rubiera

Journal: Stroke

Citation: Stroke. 2024;55:1200–1209. DOI: 10.1161/STROKEAHA.123.043691

Link: https://doi.org/10.1161/STROKEAHA.123.043691

PDF: https://www.ahajournals.org/doi/epub/10....EAHA.123.043691

Clinical Question

Can routinely collected clinical and socioeconomic data be used to build accurate statistical and machine-learning models (PRERISK) that predict early, late, and long-term stroke recurrence in individual patients after a first-ever stroke?

Bottom Line

In a large, population-based cohort, PRERISK machine-learning models (RF and AdaBoost) significantly outperformed Cox regression (P<0.05) for early, late, and long-term recurrence prediction (AUROC 0.76, 0.60, 0.71); a simplified model using key predictors had similar performance.

Major Points

  • Population-based dataset: 41,975 stroke admissions from 88 public health centers (Catalonia, 2014–2020); 36,118 first-ever IS/ICH patients selected; after excluding 408 who died within 7 days, analysis cohort was 36,114; 16.21% (5,932/36,114) had recurrence
  • Outcomes predicted at three windows: early (≤90 days), late (91–365 days), long-term (>365 days)
  • Model performance (ML AUROC): 0.76 (95% CI 0.74–0.77) early; 0.60 (0.58–0.61) late; 0.71 (0.69–0.72) long-term
  • Comparator performance (Cox AUROC): 0.73 (0.72–0.75) early; 0.59 (0.57–0.61) late; 0.67 (0.66–0.70) long-term
  • RF and AdaBoost showed statistically significant improvement over Cox (P<0.05) for all 3 windows; XGB showed no statistically significant improvement
  • Key predictors: time since previous stroke, Barthel Index, atrial fibrillation, dyslipidemia, age, diabetes, sex; simplified model with modifiable risk factors showed similar accuracy
  • Median follow-up 2.69 years (IQR 2.74); 953 (2.6%) had ≥2 recurrences; 90-day recurrence rate 3.8%; 12-month recurrence rate 7.2%

Design

Study Type: Population-based cohort analysis with statistical (Cox) and supervised machine-learning models

Randomization:

Enrollment Period: 2014–2020

Follow-up Duration: Median 2.69 years (IQR 2.74)

Centers: 88

Countries: Spain

Sample Size: 36114

Analysis: Supervised ML (Random Forest, AdaBoost, XGBoost) compared to Cox regression; performance assessed by AUROC/C-statistic; permutation importance for predictor contribution; 5-fold cross-validation

Inclusion Criteria

  • First-ever ischemic stroke (IS) or intracerebral hemorrhage (ICH) identified via ICD-9/10 codes
  • Admission within the Catalonia public healthcare system (2014–2020)
  • Survival ≥7 days after index stroke

Exclusion Criteria

  • Transient ischemic attack (TIA) not included in analysis cohort
  • Death within 7 days of index stroke
  • Recurrent stroke diagnoses within 24 hours of index event (considered fluctuation, not recurrence)
  • Patients with <365 days of tracking history

Baseline Characteristics

CharacteristicControlActive

Arms

FieldPopulation-based cohort
InterventionSupervised machine-learning models (Random Forest, AdaBoost, XGBoost) and Cox regression applied to routinely collected clinical + socioeconomic data to predict early (≤90 d), late (91–365 d), and long-term (>365 d) stroke recurrence.
DurationUp to long-term follow-up (>1 year)

Outcomes

OutcomeTypeControlInterventionHR / OR / RRP-value
Discrimination (AUROC) for prediction of stroke recurrence at early (≤90 d), late (91–365 d), and long-term (>365 d) windowsPrimaryCox AUROC: 0.73 (0.72–0.75); 0.59 (0.57–0.61); 0.67 (0.66–0.70)ML AUROC: 0.76 (0.74–0.77); 0.60 (0.58–0.61); 0.71 (0.69–0.72)RF and AdaBoost: P<0.05 vs Cox for all 3 windows; XGB: no statistically significant improvement over Cox
Recurrence rate and follow-upSecondary16.21% (5,932/36,114) had recurrence; 953 (2.6%) had ≥2 recurrences; 90-day recurrence rate 3.8%; 12-month recurrence rate 7.2%; median follow-up 2.69 years (IQR 2.74)
Predictor importance (modifiable risk factors)SecondaryFull ML models: modifiable risk factors accounted for 16.4%–36% of permutation importance (ADA 16%, RF 24%, XGB 36%); simplified MLʹ model: 16%–39% (ADA 16%, RF 23%, XGB 39%)
Retrospective model-building studyAdverseRetrospective AI-based recurrence prediction study - no AE data

Subgroup Analysis

A simplified model (MLʹ) using 7 key predictors (time since stroke, Barthel Index, AF, dyslipidemia, age, diabetes, sex) plus modifiable vascular risk factors achieved similar performance to the full ML models, maintaining statistically significant improvement over Cox.

Criticisms

  • Observational design using administrative/registry data susceptible to coding and selection bias
  • Late-window performance (AUROC 0.60) indicates modest discriminative ability
  • Large amount of missing data, especially for Barthel Index (58% missing), addressed by median imputation
  • Lack of data on treatment adherence and family history
  • Generalizability may be limited to similar healthcare systems (single-region: Catalonia, Spain)
  • Broad definition of stroke recurrence (24-hour window) may overestimate recurrences

Funding

Fundación Instituto Carlos III (grant PI20/01768); Ministerio de Asuntos Económicos y Transformación Digital (grant MIA.2021.M02.0005); European Commission under the Horizon Europe grant 101057263.

Based on: PRERISK (Stroke, 2024)

Authors: Giorgio Colangelo, DS, PhD; Marc Ribo, ..., PhD; Marta Rubiera

Citation: Stroke. 2024;55:1200–1209. DOI: 10.1161/STROKEAHA.123.043691

Content summarized and formatted by NeuroTrials.ai.