A bioimpedance-based monitor for real-time detection and identification of secondary brain injury

Secondary brain injury impacts patient prognosis and can lead to long-term morbidity and mortality in cases of trauma. Continuous monitoring of secondary injury in acute clinical settings is primarily limited to intracranial pressure (ICP); however, ICP is unable to identify essential underlying etiologies of injury needed to guide treatment (e.g. immediate surgical intervention vs medical management). Here we show that a novel intracranial bioimpedance monitor (BIM) can detect onset of secondary injury, differentiate focal (e.g. hemorrhage) from global (e.g. edema) events, identify underlying etiology and provide localization of an intracranial mass effect. We found in an in vivo porcine model that the BIM detected changes in intracranial volume down to 0.38 mL, differentiated high impedance (e.g. ischemic) from low impedance (e.g. hemorrhagic) injuries (p < 0.001), separated focal from global events (p < 0.001) and provided coarse ‘imaging’ through localization of the mass effect. This work presents for the first time the full design, development, characterization and successful implementation of an intracranial bioimpedance monitor. This BIM technology could be further translated to clinical pathologies including but not limited to traumatic brain injury, intracerebral hemorrhage, stroke, hydrocephalus and post-surgical monitoring.


Supplementary Table of Contents
Supplementary Table 1 Figure 5: (Top) Phantom experiment in a saline tank of background conductivity 0.1 S/m tracking each sector as a Fogarty balloon inflates from 0 mL to 5 m in steps. Each sector represents a change in impedance-induced voltage from baseline (V=0). Experiment then repeated with the balloon in a different sector to validate spatial tracking and localization. (Bottom) Saline phantom with continuous inflation of a Fogarty balloon from 0 mL to 3 mL over 15 minutes shows higher resolution volume detect and satisfies design criteria (Supplementary  table 1 Figure 6: Change in impedance with volume inflation for detrended CT specific elements. As can be observed, there was moderate variability between electrodes, largely due to inherent experimental variables such as the hemostatic ability within subjects, or the varied proximity of the mass effect to the intracranial electrodes. Should it be close to the current sink (intracranial electrodes) its presence would extend beyond a single electrode channel. Vol Detect = 0.39mL+/-.  Figure 7: Change in impedance with balloon inflation without adjusting for baseline. While we do not anticipate this scenario clinically while undergoing continuous patient monitoring, should baseline-based detrending not be possible, the BIM still detects an ICV change in 8/9 pigs. Further analysis showed that a single five-minute collection window was sufficient to fit a curve to and detrend electrode drift beyond the change due to injury. Vol Detect = 0.29 mL +/-0.15 mL.    During initial analysis it became apparent that errant electrode traces indicated a need for validation to ensure quality data. A cohesive filtering logic was applied to all the collected raw impedance-induced voltage data unanimously. Primary objectives were to detect faulty electrodes and account for any physiological changes (e.g. a blood drip landing on an electrode). The logic tree for the applied filter can be seen in Supplementary Method 1 Figure 1, and is expanded upon further below.
Supplementary Method 1 Figure 1 -Logic tree for filter approach applied to raw impedance data.
First, all data were filtered using a median filter (N=50) to remove CT acquisition noise (Supplementary Method 1 Figure 2, left). The first level filter logic was an electrical filter which included a DC shift filter and phase filter. While voltage traces represented the dependent variable, phase factored in the stimulating current. An example of the DC filter followed by a phase identified trace example can be seen in Supplementary Method 1 Figures 2 and 3.
Supplementary Method 1 Figure 2 -All element baseline data for Pig 7 in raw form (left), with median filter applied (middle) and with DC filter applied (right). Example of Median filter removing the CT acquisition noise (farthest left red arrow), a large DC shift (middle red arrow) and a small DC shift not removed due to coarseness of threshold for DC detection (farthest right red arrow).It is important to remember that each of these eight traces were collected sequentially, so while displayed against time on the x-axis, each collection on each channel had the sequential channels collect before it's next sampling.

Supplementary Method 1 Figure 3 -Example bad trace in Pig 3 by phase filter
Due to the high noise environment of the CT bore and linked grounds from all systems with the AFE, the impedance traces would experience occasional DC shifts (Supplementary Figure 2, middle). To fairly represent the changes in impedance due to the injury, and not due to a digital signal jump, a filter was applied to subtract any sudden change larger than a specified voltage threshold. The threshold was determined heuristically and the filter applied across all data unanimously.
An example of a trace which the DC filter identified and corrected can be seen in Supplementary Method 1 Figure 2. In-spite of these shifts the impedance trends would persist, however the shifts falsely inflated our changing Z if not removed. Large shifts were successfully corrected using the DC filter. Small steps (e.g. Supplementary Method 1 Figure 2, right), if under the threshold, would remain. A tighter threshold would capture these, however, small noise changes may not always be correctable in clinic. Additionally, the DC filter code looked at changes between two time-steps (Zt2-Zt1, fs = 50 Hz). This will fail to catch any more gradual changes, however if a confounding gradual change was riding on an electrode this would not qualify as a DC shift, and if an errant electrode, should be caught in the following filter step. For these reasons we chose the more conservative coarser filter threshold, used single time-steps, and found our data trends to be robust in-spite of these small anomalies.
While filter one corrected for electrically-induced variations, filter two (i.e. Physiological Detect Filter) aimed to identify physiologic changes. In the second filter step all baseline traces were fit to a second order polynomial and subject to an R 2 based goodness-of-fit filter. The mean R 2 was higher for a second order polynomial than a first order one, suggesting this to be a more representative fit. Supplementary Method 1 Figure 4 shows an example of such a trace. Upon review of photos it was apparent that the element identified had CSF under the electrode, however, after a settling time, fluid compromised electrodes can still provide valuable information on changing impedance. Additionally, during baseline all traces tend to settle similarly unless acted upon otherwise (e.g. blood drip). Thus, if a baseline trace was identified as failing the polynomial fit, that trace would be replaced with an average of the neighboring elements. In further processing, should this electrode fail any injury event filters it would be discarded.
Supplementary Method 1 Figure 4 -Baseline data with a blood affected electrode (red arrow). This element was identified by the R 2 filter applied to baseline data.
Lastly, the Physiologic Filter was applied to all data and used to detect outliers. A second order polynomial was fit to each trace within each event (e.g. baseline, inflation, blood). An outlier was defined as any element whose slope was more than three standard deviations away from the mean of all the elements within that event (Supplementary Method 1 Figure 5). Once identified as an outlier, these elements were discarded.
Supplementary Method 1 Figure 5 -Pig 8 inflation data. All traces after median filter (left) and all traces after full filter logic (right). Note that element 5 has been identified as an outlier and discarded.
ICP detects both euthanasia and mannitol as "ICV" events when compared to baseline (p < 0.001, both), however does not differentiate them from each other.