Machine learning prediction of shunt-dependence in patients with subarachnoid hemorrhage based on time-series data of cerebrospinal fluid markers
Katharina Frank (Köln), Mihai Manu (Köln), Paiman Shalchian-Tehran (Köln), Makoto Nakamura (Köln)
Hydrocephalus is a common complication in patients with aneurysmal subarachnoid hemorrhage (SAH), resulting in a permanent shunt dependency in 8.9% to 48%1. We sought to validate a machine learning algorithm for predicting shunt-dependence in patients with subarachnoid hemorrhage (SAH) based on the initial post-ictal CT scan and simple CSF markers time-series data.
The CSF cell count, erythrocyte count, protein and lactic acid levels from 44 patients with SAH and ventricular drain, collected during the first 12 days after the event were so far characterized retrospectively in order to create a time series data. Using custom-written routines in Python, we first sought to extract meaningful features that capture the difference between the time series data of the shunt-dependent and shunt-independent patients.
The relevant CT features were extracted by passing the CT stack as input to a pre-trained deep network (ResNet50). Afterwards, in this new feature space, we trained and tested a random forest, multilayer perceptron and logistic regression classifier for predicting shunt dependency.
Of the 44 patients, 39% developed a shunt-dependent hydrocephalus (8/17 male, 9/17 female). The parametric and non-parametric statistical test revealed statistically significant differences in the normalized mean cell count and in the erythrocyte cell count trend over time (Mann-Kendall p=0.039 and Sen"s slope p=0.047, figure 1). Based on these engineered features, a random forest classificator achieved an accuracy of 89% in correctly identifying shunt-dependent/shunt-independent patients (precision 0.88 for shunt-independent and precision 1 for shunt-dependent patients).
A random forest classificator based on features extracted from simple time series CSF data, revealing the significance of slower degradation of erythrocytes in CSF, can achieve a high accuracy in predicting shunt-dependence. We expect an improvement in the classificatory task with addition of the imaging data.
We use cookies on our website. Cookies are small (text) files that are created and stored on your device (e.g., smartphone, notebook, tablet, PC). Some of these cookies are technically necessary to operate the website, other cookies are used to extend the functionality of the website or for marketing purposes. Apart from the technically necessary cookies, you are free to allow or not allow cookies when visiting our website.