gms | German Medical Science

Objective: Both systemic and local inflammation in the brain parenchyma play a critical role during early brain injury (EBI) and delayed cerebral ischemia (DCI). Multiple biochemical markers are known to be associated with cell damage and post-SAH inflammation and hence contribute to the pathophysiology of aSAH. We investigated whether machine learning methods can identify those biomarkers to predict post-aSAH complications.

Methods: Blood concentration of eleven potential blood markers (Cytochrome B, D-loop, Cox-1, CCL5, CRP, leukocites, IL-6, IL-10, IL-23, sIL-17, and HMGB1) for 81 SAH patients were recorded on day 1, together with demographic and clinical data. Five potential complications – cerebral vasospasms (CVS), delayed ischemic neurologic deficit (DIND), shunt dependent hydrocephalus, convulsive epilepsy and delayed cerebral ischemia (DCI) – were considered. If one or more complications occurred, the day of the earliest occurrence was recorded. Missing values were imputed as medians or the most frequent value.

The data were randomly split into a training and a test set, in the 60:40 ratio. Two machine learning algorithms, Random Forest (RF) and Support Vector Machine (SVM), were trained to predict whether (classification) and when (regression) a patient will develop complications during his or her stay. All combinations of the blood markers, together with the Hunt & Hess score were systematically tested. For SVMs, the radial (Gaussian) kernel with the default width was used. For classification, the performance was quantified as sensitivity, specificity, and F1. For predicting the complications’ onset day (regression), ICC (3,k) was used.

Results: 63 patients (78%) developed post-aSAH complications. Most epilepsies (88%) occurred on days 1 or 2, and other complications mostly on days 4 or 5. Age (31-85, median 56) and sex did not correlate with complications. The SVM classifier achieved 100% sensitivity and 71% specificity (F1=0.96) on the test set. For Random Forest sensitivity was 100%, specificity 57% and F1=0.95. SVM regression reached an ICC (3,k) of 0.3 (p=0.19), while RF regression achieved ICC=0.34 (p=0.15).

Conclusion: Despite the small but high-dimensional training sets (48 observations for classification, 34 for regression), good performance was achieved for classification. Due to the small data set size we are cautious in regard to the potential of machine learning to identify serum predictive biomarkers for post-aSAH complications in intensive care.