Deriving Automated Device Metadata From Intracranial Pressure Waveforms: A Transforming Research and Clinical Knowledge in Traumatic Brain Injury ICU Physiology Cohort Analysis.

Abstract

Importance: Treatment for intracranial pressure (ICP) has been increasingly informed by machine learning (ML)-derived ICP waveform characteristics. There are gaps, however, in understanding how ICP monitor type may bias waveform characteristics used for these predictive tools since differences between external ventricular drain (EVD) and intraparenchymal monitor (IPM)-derived waveforms have not been well accounted for.

Objectives: We sought to develop a proof-of-concept ML model differentiating ICP waveforms originating from an EVD or IPM.

Design, setting, and participants: We examined raw ICP waveform data from the ICU physiology cohort within the prospective Transforming Research and Clinical Knowledge in Traumatic Brain Injury multicenter study.

Main outcomes and measures: Nested patient-wise five-fold cross-validation and group analysis with bagged decision trees (BDT) and linear discriminant analysis were used for feature selection and fair evaluation. Nine patients were kept as unseen hold-outs for further evaluation.

Results: ICP waveform data totaling 14,110 hours were included from 82 patients (EVD, 47; IPM, 26; both, 9). Mean age, Glasgow Coma Scale (GCS) total, and GCS motor score upon admission, as well as the presence and amount of midline shift, were similar between groups. The model mean area under the receiver operating characteristic curve (AU-ROC) exceeded 0.874 across all folds. In additional rigorous cluster-based subgroup analysis, targeted at testing the resilience of models to cross-validation with smaller subsets constructed to develop models in one confounder set and test them in another subset, AU-ROC exceeded 0.811. In a similar analysis using propensity score-based rather than cluster-based subgroup analysis, the mean AU-ROC exceeded 0.827. Of 842 extracted ICP features, 62 were invariant within every analysis, representing the most accurate and robust differences between ICP monitor types. For the nine patient hold-outs, an AU-ROC of 0.826 was obtained using BDT.

Conclusions and relevance: The developed proof-of-concept ML model identified differences in EVD- and IPM-derived ICP signals, which can provide missing contextual data for large-scale retrospective datasets, prevent bias in computational models that ingest ICP data indiscriminately, and control for confounding using our model's output as a propensity score by to adjust for the monitoring method that was clinically indicated. Furthermore, the invariant features may be leveraged as ICP features for anomaly detection.