Prediction of Long-Term Poor Clinical Outcomes in Cerebral Venous Thrombosis Using Neural Networks Model: The BEAST Study

Research output: Contribution to journalArticlepeer-review

Abstract

Introduction
Risk prediction models are commonly performed with logistic regression analysis but are limited by skewed datasets. We utilised neural networks (NNs) model to identify independent predictors of poor outcomes in cerebral venous thrombosis (CVT) due to the limitations of logistic regression (LR) analysis with complex datasets.

Methods
We evaluated 1309 adult CVT patients from the prospective BEAST (Biorepository to Establish the Aetiology of Sinovenous Thrombosis) study. The area under the receiver operating characteristic (AUROC) curve confirmed the goodness-of-fit of prediction models. The normalised importance (NI) of the NNs determines the significance of independent predictors.

Results
The stepwise logistic regression model found thrombolysis (OR 32.1; 95% CI 3.6–287.0; P=0.002), craniotomy (OR 6.9; 95% CI 1.3–36.8; P=0.02), and cerebral haemorrhage (OR 4.5; 95% CI 1.3–15.4; P=0.01) as predictors of poor clinical outcome with the AUROC of 0.71. Conversely, the NNs model identified major independent predictors of long-term poor clinical outcomes as cerebral haemorrhage (NI 100%) and thrombolysis (NI 98%), as well as trivial predictors of age (NI 2.8%) and altered mental status (NI 3.5%). The accuracy of the NNs model was 95.1% and 94.1% for self-learned randomly selected training and testing samples with an AUROC of 0.82. Positive and negative predictive values for poor outcomes were 13.2% and 97.1% for the LR model, compared with the NNs model of 18.8% and 98.7%, respectively.

Conclusion
Cerebral haemorrhage and thrombolysis was a strong independent predictor, whereas age merely impacts the long-term poor clinical outcome in adult CVT. Integrating unorthodox neural networks risk prediction model can improve decision-making as it outperforms conventional logistic regression with complex datasets.
Original languageEnglish
Pages (from-to)2919-2930
Number of pages12
JournalInternational Journal of General Medicine
Volume17
DOIs
Publication statusPublished - 2 Jul 2024

Cite this