Correlating the passive response of eye and brain to head impact using MEMS IMUs on 3D-printed human head phantom

Six 6-degree-of-freedom (DoF) LSM6DS3 IMUs by STMicroelectronics^TM (2016), each containing a 3-DoF accelerometer and a 3-DoF gyroscope, were used to capture data in this study. Five of the IMUs were mounted on ribbon-style flexible PCBs made of 25.4-µm-thick DuPont^TM Pyralux, coated with silicone conformal coating, and connected to interface PCBs as seen in figure 1. As an improvement compared to the design in Meng et al (2017), the polyimide substrate material was trimmed into a half-moon shape designed to fit the sensor to the surface of the eyeball in the lower eyelid, and better couple the motions of the eye and sensor at the position of the IMU. The sensor's silicone conformal coating created a moisture barrier to protect the circuits and made the sensor less likely to scratch the eye. The last IMU was mounted on a SparkFun^TM breakout board as a reference sensor for data calibration and relative acceleration computation.

A data acquisition (DAQ) system containing three Teensy^TM 3.6 micro-controllers (MCUs) (PJRC 2019) was created to control and communicate with the six IMUs via I²C buses, acquire sensing data, and log the data into a micro SD card for further processing (figure 2). The MCUs were plugged on a motherboard and synchronized by sharing the same trigger interrupt. Each MCU hosts two IMUs. Through the MCUs, the gyroscopes were programmed to operate with a sensing range of ±2000° s⁻¹ and a sensitivity of 70 mdps/LSB. The sampling rates and bandwidths of the gyroscopes were set as 1.66 kHz and 400 Hz respectively to obtain the best performance of the IMUs. The output data rate (ODR) was set as 6664 Hz. The bandwidth is determined by the ODR selection suggested by the data sheet to set an anti-aliasing filter for high performance. All the sensing data, including temperature readings from the six IMUs, were buffered in the 256 KB RAM of each MCU before being logged to its micro SD card. The overall sampling rate was 1 kHz which was sufficient for this application (Meng et al 2017).

Zoom In Zoom Out Reset image size

To mimic a human head, a scaled plastic phantom human skull was 3D-printed and employed as the main test unit, as shown in figure 3. The 3D model was based on scanning data of a real person's head (MacGyver 2013). The phantom skull had multiple openings that needed to be closed to create the gelatin brain. Oatey^TM plumber's putty and 100% silicone were used to prepare the skull for creation of the phantom brain and eyes by filling these holes. Furthermore, the back of the eye sockets were also filled with putty to prevent leaking of the gelatin surrounding the phantom eyeballs. Once all the holes were filled, Dremel^TM and drill power tools were used along with multiple screws to drill four small holes into the skull around the eye sockets to position the phantom eyeballs in the anatomically correct place within the eye sockets (figure 3(a)). Finally, three slits and one hole were created in the top piece of the skull using a Dremel^TM. The IMU sensors were inserted through these slits and the hole was for pouring of the gelatin. The three slits for the IMUs were placed on a diagonal with respect to one another.

Zoom In Zoom Out Reset image size

The eyeball and eyelid replicas were made in the same way as introduced in Meng et al (2017). In preparing the phantom eyes, a suspension platform was mechanically attached on the 3D-printed phantom skull using the four small holes mentioned above, and held the two phantom eyeballs in the correct anatomical position inside the sockets while the gelatin surrounding the eyeballs solidified (figure 3(a)). Both 2% and 4% gelatin solutions were used for the right and the left eyes, respectively. A piece of PDMS eyelid was then glued to the skull around each eye socket, had a slit cut in the center, and had an IMU placed within the slit where it was touching both the eyelid and the eyeball (figure 3(b)).

Among the six sensors, one was tightly fastened on the forehead as the reference which moves with the skull (SK) as shown in figures 3(c) and (d), two were inserted through the horizontal slits of the eyelids and held tightly on the eyeballs by the eyelids (figure 3(b)), one in the left eye (LE) and one in the right eye (RE). The other three were inserted in the gelatin brain in the skull cavity through three slits on the skull top (figure 3(c)). These slits formed a diagonal line so that motion/deformation of be brain tissue in different directions can be monitored. There was one IMU in the anterior left of the gelatin phantom brain (AB), one in the middle of the brain (MB), and one in the posterior right of the brain (PB). The three sensors were planted vertically into the brain, facing front. They were on the same depth of approximate 40 mm from the top of the brain. The sensor interface boards were taped to the skull with electrical tape, and the top of the skull was stuck on to the bottom using silicon and plumber's putty. The gelatin solution with mass ratio of 5% was created and poured into the assembled phantom skull through the pour hole, covered the sensors, and was put in the 4 °C refrigerator overnight to allow for it to solidify.

A vestibulo-ocular response (VOR) chair, provided by the Vision Sciences Department at the University of Alabama at Birmingham (UAB), was used to actuate rotational motions of the skull phantom, so as to collect data about the movement of the IMUs within the brain and on the eyes (figure 3(d)). The phantom skull was secured on a metal plate above the seat of the VOR chair. A data collection task of 'hit' motion were performed with this chair. A hit-motion consists of a very quick jerk in a random direction around the y-axis of the chair (figure 3(d)), either clockwise or counterclockwise viewing from above. The peak velocity during the hit-motion task was 160° s⁻¹ and the peak acceleration was 1025° s⁻².

The hit-motion experiments provided a good understanding of the time delays associated with the start of rotation of the test IMUs compared to the starting time point of the reference IMU. Angular accelerations recorded by the gyroscopes were integrated to get time-varying rotation angles. Figure 4 shows a plot of rotational data around the y-axes (angles of θ) collected by all the six IMUs during a typical counterclockwise hit-motion at the start. The y-axes were studied because the IMUs were all aligned in a way that the majority of rotations happened around their y-axes, which is aligned with the y-axis of the VOR chair, as shown in figure 3. Three thresholds of 0.5°, 1°, and 1.5° were applied to study the time delays. Though varying in different rotational thresholds, the time delays compared to the reference forehead IMU, from short to long, all tend to follow the order of posterior brain, left eye, middle brain, right eye, and anterior brain, as shown in figure 5. This tendency was repeatable in the ten hit-motions in the experiment of four clockwise rotations and six counterclockwise rotations. The averages and standard deviations (STD) of all the processed data are recorded in table B1 in appendix B with the averages visualized in figure 5.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The results demonstrate that the rotational motions of both eyes were good reflections of the rotations of the brain since the time delays of both eyes from the skull rotation fell into the range of the time delays of the different parts of the brain. Specifically, on average, both eye delays were longer than the posterior brain delay, and shorter than the anterior brain delay. The fact that posterior brain had the shortest delay while the anterior had the longest delay corresponds with the fact that the posterior brain IMU was closest to the inner wall of the skull back and the rotation axis of the VOR chair, therefore the motion coupled better to the skull, which was fastened on and moved along the chair (figure 3(d)). Although the IMUs on the eyeballs were further away from the rotation axis than the anterior brain, the confinement of the sockets, which have much smaller cavities than the skull does, acted in a way to accelerate the eyeball rotations to match the pace of the brain. One should also notice that the left eye had shorter delay than the right eye due to the greater stiffness of the 4% gelatin than the 2% gelatin. These findings in the processed data indicate that the rotational PER captured by the MEMS IMUs have potentials to reflect the rotational accelerations of the brain in real time when rotational accelerations are applied to the head.

A human volunteer was tested under the non-concussive conditions of hit-motion (figure 6(a)), so as to further validate the idea of applying MEMS IMUs to sensing PER during head rotation. (The study was approved by the Institutional Review Board of UAB and conducted in accordance with the Declaration of Helsinki.) The volunteer sat in the VOR chair with one flexible IMU sensor inserted in the left eye, held in position by the lower eye lid, and the other IMU attached on the forehead as reference.

Zoom In Zoom Out Reset image size

Similarly as in section 2.3, angular accelerations recorded by the gyroscopes were integrated to get time-varying rotation angles. A typical set of results is plotted in figure 6(b). Notice that, at the rotation start in the zoomed-in part (highlighted in green), the left eye rotation (red solid line) experienced a delay relative to the head rotation (black dashed line) as a PER. It appears that the ocular muscles then took control to correct the orientation of the eyeball by accelerating the rotation. The active muscle control allowed the eye to catch up with the head motion at approximately 1.19 s. Later on, the eye motion oscillated around the rotation speed of the head, until both stabilized at approximately 4.25 s when the hit-motion task was completed. All these phenomena correspond with the observations provided in Bush and Miles (1996), Collewijn and Smeets (2000). Also notice that, in the zoomed-in part, the three manually selected thresholds were reached by the eye before an apparent reflex interference was observed, which further validated that the previous selection was legitimate to study the rotational accelerations on the head phantom.

Although such subtle behavior of human eye could not be observed in the phantom test, both experiments showed that the 3D-printed human head phantom behaves similar to the real human head, and at the same time, proved that the rotational PER captured by the MEMS IMUs is different from the head rotation, and it has potentials to reflect the rotational accelerations of the brain in real time. The results from the human testing also showed that the IMUs were sensitive enough to register PER even under non-injurious head acceleration.

Although the gyroscopes were capable to sense the delays of rotation starting times among the roughly aligned sensors, a more controlled experimental apparatus is needed to explore the correlation between the PER and brain acceleration under linear head accelerations. In the experimental apparatus of this section, the same LSM6DS3 IMUs, customized DAQ and 3D skull model were used. The skull-brain-eye system first introduced in Prather et al (2019) was further improved in accuracy and employed in this test. As illustrated in figure 7, the reference IMU was attached on the nose of the phantom skull, two flexible IMUs were attached on the right and the left eye replicas, and three flexible IMUs were inserted in the phantom brain at three typical positions on a diagonal: anterior brain, middle brain, and posterior brain. All the IMUs were aligned with their corresponding sensing axes parallel to one another, and with their centers on the same horizontal sagittal plane, resulting in their z-axes in a direction pointing to the front of and vertical to the face of the skull, x-axes pointing to the left side of the skull, and y-axes upward. Red circles in the cross-section view of figure 7(a) highlight relative positions of the six IMUs on the x–z plane. Notice that the IMUs at AB, MB and PB were evenly separated on a diagonal, also the z-axes of IMUs at RE and PB, LE and AB, reference and MB, overlapped in x-axes respectively. Such alignment of IMUs in the phantom head was achieved by the approach introduced in appendix A.

Zoom In Zoom Out Reset image size

The finished experimental apparatus is shown in figure 8. Many of the settings were identical to the ones introduced in section 2.2. The reference IMU was inserted in the holder which was fixed on the phantom skull. The two eye IMUs were placed on the PDMS eye replicas, held by 1-mm-thick PDMS eyelids. Gelatin mixtures of 4% and 2% mass ratios of gelatin to water were used for the emulation of the right and the left socket tissue respectively. For this experiment, a football helmet was used to better emulate the real-world situation.

Zoom In Zoom Out Reset image size

For testing, the phantom skull inserted in the football helmet was attached to a PVC emulated neck anchored on a portable and adjustable platform with leveling feet as shown in figure 8. This setup allowed the phantom skull to fall in an inverted pendulum orientation, resulting in impact with the platform surface. Repeatable drop-and-impact tests were performed by lifting the phantom skull in helmet using a brass rod and a tiered 3D-printed platform which provides three dropping heights: low (135 mm), medium (155 mm), and high (175 mm). Drops of the skull were initiated by removing the brass rod supporting the neck, which allowed the phantom skull in helmet to fall and collide with the platform surface. The brass rod also served as a triggering mechanism to signal the starting time stamp of the fall, as two electrodes connected to the PVC spine would be shorted while in contact with the brass rod and opened while losing the contact. When impact between the back of the football helmet and the platform surface occurred, another two electrodes on the platform surface were shorted in a similar manner by the conductive copper tape on the back of the helmet. This trigger signal was recorded and synchronized with the recorded IMU data by the DAQ introduced in section 2.1 and logged in micro SD cards for further processing in MATLAB and R language.

A number of drop-and-impact experiments on this skull-brain-eye system were performed with the testing apparatus. Relative accelerations of the eye and brain IMUs to the phantom skull along the impact direction were calculated applying the data fusion technique developed in Meng et al (2017). Instead of relative displacements, only relative accelerations were computed to further diminish errors introduced by integration. Figure 9 plots the relative accelerations from a typical result, where a vertical green solid line marks the impact time, a dashed blue line marks the 20 ms after the impact, and a red dashed line marks the impact end time. Since PER happens in approximately 20 ms after impact, the data sets within such periods were studied.

Zoom In Zoom Out Reset image size

The statistical method of one-way, fixed effects, normal analysis of variance (one-way ANOVA) was employed to study the associations/correlations between the accelerations measured by the IMUs at RE/LE/SK and at AB/MB/PB (Anscombe 1973, Tamhane 1977, Stoline 1981). The usage of one-way ANOVA can be considered reliable since the variable residuals are all approximately normally distributed, which is an assumption of this method. A typical histogram given by the assumption check is plotted in figure 10. For each data set mentioned above, 33 samples were processed with the degree of freedom of 1 and residuals of 30. Probability values (p-values) were calculated to evaluate the association. The levels listed in table 1 indicate the statistical significance or association or correlation for the accelerations between two sensor positions. Specifically, the null hypothesis in the ANOVA tests is not rejected if a p-value > 0.05, labeled as 'null' in table 1. The null hypothesis is rejected if a p-value < 0.05, and the closer the p-value is to 0, the higher level of statistical significance the association has, which is shown by the number of asterisks in table 1 with more asterisks representing greater association.

Table 1. P-values to reject null hypothesis for associations between eye/skull sensors and brain sensors using one-way ANOVA.

Data Set	P-values
	Right eye			Left eye			Skull
	AB	MB	PB	AB	MB	PB	AB	MB	PB
1	null	***	***	null	***	***	***	*	*
2	***	***	***	***	***	***	***	null	null
3	***	null	***	***	null	***	***	***	null
4	***	***	***	***	***	***	**	null	null
5	***	*	***	***	***	***	***	null	**
6	*	***	***	***	***	***	**	null	null
7	***	***	***	***	***	***	null	null	null
8	**	***	***	***	***	***	***	**	*
9	***	***	***	***	***	***	***	**	*
10	***	***	***	***	***	***	***	null	null
11	***	***	***	***	***	***	***	null	null
12	***	***	***	***	***	***	**	null	null
13	null	***	***	null	***	***	***	null	null
14	***	***	***	***	***	***	***	null	null
15	***	***	***	***	***	***	***	null	*
16	***	***	***	***	***	***	***	null	null
17	***	***	***	***	***	***	***	null	null
18	***	***	***	***	***	***	***	null	*
Group	***	***	***	***	***	***	***	null	*
Significance code		p-value is 0 $\lt$ '***' $\lt$ 0.001 $\lt$ '**' $\lt$ 0.01 $\lt$ '*' $\lt$ 0.05 $\lt$ 'null' $\lt$ 1

Zoom In Zoom Out Reset image size

Notice that, in table 1, all the p-values for the correlations are the lowest between the right eye and the posterior brain (RP), and between the left eye and the posterior brain (LP), indicating that the associations in these pairs of data have the strongest statistical significance. This may result from the fact that all the three IMUs at RE, LE and PB were located closer to the inner wall of the skull/socket where the impact power was transmitted faster and earlier than to AB and MB. Statistical significance levels for the correlations between RE/LE and AB/MB (RA, RM, LA, and LM, respectively) are relatively lower, yet generally higher than the ones between the skull and the brain parts (SF, SM, and SP, respectively). Further observations show that, the eyes are better correlated to PB while the skull is better correlated to AB. The correlations of SM and SP are especially weak and unstable, which may explain the inability of helmet/mouthguard mounted sensors to accurately assess mTBI in real-time on the field as discussed in section 1. These observations suggest that sensing the linear PER using MEMS IMUs could provide better outcomes than sensing the skull for real-time on-field objective concussion monitoring and diagnosis.

Although the results provided by the one-way ANOVA confirmed the hypothesis of this paper, p-values to reject the null hypothesis have their limits of generalization. Details about the performance of each eye sensor are unknown. More quantified assessments of correlations are explored in the following two sections, utilizing R² values provided by linear regression analysis, and R and p-values provided by specialized correlation model, respectively.

The data within periods of 20 ms when passive eye response (PER) happens were studied by polynomial linear regression analysis to study correlations. Statistical model fitting coefficients, such as coefficients of determination (R²'s) and p-values, were calculated to evaluate the fit (Glantz and Slinker 1990). R² values for 5th order polynomial regressions of the data plotted in figure 9 were calculated and recorded in table 2. Since the R² values are coefficients of determination which provides a measure of how well observed outcomes are replicated by the model, based on the proportion of total variation of outcomes explained by the model (Draper and Smith 1998), the closer they are to 1, the larger portion of data can be expressed by the regression model, and the better two studied accelerations are correlated. Notice that, even though the relative accelerations of the three brain IMUs appear to be quite different from those of the two eye IMUs over a longer time (figure 9), they have a stronger correlation during the 20-ms time window after impact which we consider as the time for PER, than the correlation of the skull and brain has (table 2). In this case, the strongest correlation happened to be between the left eye and the posterior brain, with an R² of 0.983 in the linear regression analysis, while the weakest correlation was between the skull and the posterior brain with an R² of only 0.176.

The averages and STDs of the R² results from 18 repeatable tests including the one studied above are listed in table B2 in appendix B. The averages are visualized in figure 11. Frequency spectra are plotted as histograms in figure 12 with dashed blue lines marking the mean values and solid blue lines showing the beta distribution fit (Johnson et al 1995). All the plots indicate that the linear accelerations of both eyes are better correlated to the brain than those of the skull, among which the left eye and the posterior brain have the best match, followed by the right eye and the posterior brain. Not only the LP correlation is high in R² value, the concentrated distribution also reflects good repeatability of the experimental procedure. In contrary, the correlations between the brain parts and the skull are very week and rather random in distribution.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Notice the interesting fact that the mass ratios of gelatin used in LE, RE and PB were 2%, 4% and 4% respectively. With most of the other parameters identical, LP had stronger correlation than RP did, even though RE and PB had the same concentration of gelatin. It is possible that this is a result of similar natural mechanical frequencies of LE and PB: one in an environment with less stiff gelatin (2%) but smaller space (eye socket), the other with stiffer gelatin (4%) yet larger space (skull cavity). Although the exact causation of this phenomenon is unknown, both eye linear accelerations were well correlated to the brain parts linear accelerations. The results thus again suggest that sensing the linear PER using MEMS IMUs could provide better outcomes than sensing the skull for real-time on-field objective concussion monitoring and diagnosis.

The limit of linear regression study is that it only reflects one possible regression fit for the data pairs, however good the results are. Multiple statistical models to study correlation exist as bivariate analyses that measure the strength of association between two variables and the direction of the relationship. Typical models use Pearson's linear correlation coefficient, Kendall's tau coefficient, and Spearman's rho, etc, to represent the correlations. Based on the characteristics of our data sets discussed in sections 3.2 and 3.3 and the assumptions of these correlation methods, Pearson R correlation was selected (Benesty et al 2009).

The method of Pearson R correlation is the most widely used correlation statistic to measure the degree of the association between linearly related variables. In Pearson R correlation, the value of the correlation coefficient R varies between −1 and +1. A value of −1 indicates a perfect negative correlation, while a value of +1 indicates a perfect positive correlation, and a value of 0 indicates no correlation between the two sets of data.

The PER data were processed with Pearson R correlation, and the calculated R values with the corresponding p-values are plotted in the box plots in figure 13, where the horizontal red lines mark the median, blue boxes represent the first quartile (upper edge) and the third quartile (lower edge), horizontal black lines label the maximums (upper) and minimums (lower), and red crosses mark the outliers for each data pair. In figure 13(a), the R values for RA, RM, RP, LA, LM, LP, and SF are all close to 1, indicating better and positive correlations. Among them, LP has the largest R values with the best convergence, followed by RP. Even the outliers for LP are close to 1. Although LP and RP have the best correlations, the eyes seem to show little advantage in correlating to the anterior brain compared with the skull. However, the skull is weakly correlated to MB and PB with R values closed distributed around 0. P-values in figure 13(b) confirm the statistical significance of the corresponding Pearson R correlation analysis, among which most of the p-values of RA, RM, LA, LM and SF are lower than 0.05, and the p-values of RP and LP are all well below 0.05. In contrast, the p-values of SM and SP suggest that these correlations are not reliable even some of them may have a good R.

Zoom In Zoom Out Reset image size

Again, the left eye was found as the best indicator of the accelerations of the posterior brain which is closest to the impact point, followed by the right eye. The correlations between the accelerations of the skull and of the brain phantom were either too weak or too random in distribution. This makes sensors on the skull less reliable in reflecting phenomena that occurs in the brain. The merits of using the eyes as surrogates to the brain for real-time on-field objective mTBI monitoring and diagnosis are proved in all the three sets of analysis above. The details of data can be affected by the factors such as how and where the sensors are placed and how acceleration is applied. The key finding is that this is the first time that PER is shown to be time-locked and highly correlated with brain responses. Therefore, we can learn brain motion/deformation from PER.