Although delayed
cerebral ischemia (DCI) is a well-known complication after
subarachnoid hemorrhage (SAH), there are no reliable
biomarkers to predict DCI development. Matricellular
proteins (MCPs) have been reported relevant to DCI and expected to become
biomarkers. As machine learning (ML) enables the classification of various input data and the result prediction, the aim of this study was to construct early prediction models of DCI development with clinical variables and MCPs using ML analyses. Early-stage clinical data of 95 SAH patients in a prospective cohort were analyzed and applied to a ML algorithm, random forest, to construct three prediction models: (1) a model with only clinical variables on admission, (2) a model with only plasma levels of MCP (
periostin,
osteopontin, and
galectin-3) at post-onset days 1-3, and (3) a model with both clinical variables on admission and MCP values at days 1-3. The prediction accuracy of the development of DCI, angiographic vasospasm, or
cerebral infarction and the importance of each feature were computed. The prediction accuracy of DCI development was 93.9% in model 1, 87.2% in model 2, and 95.1% in model 3, but that of angiographic vasospasm or
cerebral infarction was lower. The three most important features in model 3 for DCI were
periostin,
osteopontin, and
galectin-3, followed by
aneurysm location. All of the early-stage prediction models of DCI development constructed by ML worked with high accuracy and sensitivity. One-time early-stage measurement of plasma MCPs served for reliable prediction of DCI development, suggesting their potential utility as
biomarkers.