The novel BSGM system was designed to accommodate a clinical testing framework comprising a 30-minute fasting baseline, consumption of a standardized meal, and up to 4 hours of postprandial testing (Fig. 1). This timeframe was chosen to reflect the typical period of the gastric meal response.22 The BSGM system (overviewed in Fig. 2) is comprised of a flexible and conformable pre-gelled ‘peel-and-stick’ high-resolution (HR) sensor array, an ambulatory data logger with custom electronics specifically tuned for gastric bioelectrical data, and a native iOS App with HIPAA-compliant cloud connectivity and Bluetooth 5.0 synchronization to the data logger. The App guides the user through test setup, including user-specific algorithm guided array positioning, requests symptom data from the subject during testing according to a validated design, and manages data transfers. Automated algorithms were also developed for filtering and extracting and visualizing clinical biomarkers. Each of these components are discussed in further detail below.
Gastric bioelectrical signals are of weak amplitude and signal strength diminishes exponentially as distance from the source increases,23 meaning that electrodes should be positioned directly over the stomach for capturing reliable data.24 An electrode sensor array was therefore designed to a size that was capable of achieving a position consistently overlying the stomach with high reliability when placed on the epigastrium (196 cm2; data based on a separate proprietary anatomical study). The sensor array (Fig. 2A, 3A) was screen printed in layers on a 21 x 16 cm thermoplastic polyurethane (TPU) substrate. TPU was chosen for its ease of manufacturing, biocompatibility, and high conformability, allowing comfortable adherence to the user’s epigastrium (Fig. 3B), including through a full range of movement without delaminating from the skin.
An 8 x 8 grid of electrodes (11 mm diameter, 20 mm center-to-center electrode spacing) with adjacent reference and ground electrodes was screen-printed onto the TPU substrate using Ag/AgCl ink. Each electrode pad has an associated conductive track coalescing to a ‘mating panel’ (Fig. 2B). An insulating dielectric layer was then applied to coat the entire surface except for the electrode pads and mating panel (indicated by blue areas in Fig 2B). A conversion process was then undertaken (Fig. 3A). First, an adhesive laminate was applied over the dielectric layer. Conductive hydrogel discs were then positioned on each Ag/AgCl electrode to ensure low-impedance charge transfer at the skin interface.25 Finally, the adhesive and hydrogel layers were covered by a backing layer, which is peeled off immediately prior to application. The entire sensor array was mass-fabricated with an automated process involving roll-to-roll screen-printing, die-cutting, and converting machinery, before individual packaging in moisture barrier foil pouches (Alimetry Ltd, New Zealand).
A total of 22 completed arrays underwent comprehensive testing to ensure adequate electrode quality and performance according to the ANSI/AAMI EC12:2000 Standard and a further 30 arrays underwent shelf-life testing. Further details regarding the array testing procedure can be found in the Supplementary Methods.
A compact board-to-board connector design was introduced that overcomes a key design challenge in the field of high-resolution wearable electronics, being the attachment of a flexible array to a rigid circuit without the need for bulky connector attachments or cables. The sensor array interfaces with the data logger device by means of a custom flexible printed circuit cable and high-density compression connector with a 0.2 mm thick PI stiffener on the back of the connector end, placed between a machined stainless-steel clamp (Alimetry Ltd, New Zealand) (Fig. 2C). The other end of the cable attaches to a zero-insertion-force connector on a custom mating PCB that allows repeatable connections to the data logger without additional tools.
An ambulatory data logger (Fig. 2A) was designed and fabricated, employing all custom electronics and firmware, and housed within an off-the-shelf casing (Hammond Manufacturing, USA; 147 x 89 x 25 mm). Bioelectrical signals are recorded at 250 Hz, amplified, and digitized by low-noise programmable gain amplifiers with each input compared against a common reference electrode as shown on Fig. 2A, to provide unipolar recordings for 64 channels. Data is stored on removable internal storage until uploaded to a HIPAA-compliant cloud server via the App. An onboard accelerometer is used to record motion during the recording. Bluetooth connection with the App is maintained throughout the recording session, and to facilitate data upload. After a successful upload, data is securely deleted from the data logger in preparation for the subsequent test. After assembly, the data logger electronics underwent comprehensive electrical performance testing to ensure design criteria were met, as well as electromagnetic compatibility (EMC) testing. Further details regarding these testing methods are provided in the Supplementary Methods.
App and Array Placement Algorithm
The companion App was programmed in Swift v.5.1, being designed for use on an iPad mini (Apple, CA, USA). A password-protected administration section allows the user to register the test and participant details, customize recording variables, and to guide setup. As noted above, reliable placement of the electrode array directly over the stomach location is an essential design requirement for reliable data capture in body surface gastric mapping. The App therefore further incorporated an array positioning algorithm taking into account guided measurements between xiphoid and umbilicus, xiphoid and anterior superior iliac spine (ASIS), and abdominal circumference (Fig. 2Di).26 These measurements were used to calculate a patient-specific array location with reference to the umbilicus, which is displayed to the user. Guided placement of the array by this algorithm, in conjunction with the chosen array size, aimed to reliably capture the gastric field within the recording electrode in the high majority of participants, by accounting for known anatomical variations (verified in a previous proprietary anatomical study).26 The App also undertakes an impedance check of the array prior to test initiation to ensure optimal data quality (Fig. 2Dii).
Clinical and Analytical Methods
Ethical approval for the clinical studies was obtained from the Auckland Health Research Ethics Committee (AHREC, reference AH1130). This study focused on clinical evaluation of the novel system in 24 healthy subjects to demonstrate reliability for outputting each of the stated gastric biomarkers. Healthy subjects were 18 years or older with no known active GI symptoms or pathology, not meeting Rome IV criteria for a functional GI disorder, and not taking any medication known to affect gastrointestinal motility including anxiolytics and antidepressants. Additional exclusion criteria were any of the following: metabolic, neurogenic, or endocrine disorder known to cause gastric dysmotility (including scleroderma, multiple sclerosis and hyperthyroidism), active GI infection, inflammatory bowel disease, previous gastric or esophageal surgery, history of GI malignancy, open abdominal wounds or abdominal skin not intact, fragile skin, allergy to adhesives and pregnancy. Additionally, those with body mass index (BMI) >35 were excluded, as data reliability is currently uncertain above this threshold.24
Subjects were fasted for at least eight hours and were asked to not use nicotine or consume caffeine in the morning of their study. Measures of chronic gut symptoms were evaluated by the validated Patient Assessment of Gastrointestinal Disorders Symptom Severity Index (PAGI-SYM)27 and Patient Assessment of Upper Gastrointestinal Disorders-Quality of Life (PAGI-QOL)28 questionnaires. The anterior abdomen was shaved if required and the skin prepped using a conductive gel (NuPrep; Weaver and Company, CO, USA). The novel flexible sensor array was then placed on the epigastrium, guided by the positioning algorithm. As per Fig. 1, gastric activity was measured over a 30-minute fasted period, followed by consumption of a standardized meal over 10 minutes, including a 232 kcal nutrient drink (230 mL Ensure; Abbott Nutrition, IL, USA) and an oatmeal energy bar (250 kcal with 5 g fat, 45 g carbohydrate, 10 g protein, 7 g fiber; Clif Bar & Company, CA, USA), and a further 4 hour postprandial recording. Subjects remained reclined at 45 degrees for the entirety of the recording duration with their legs elevated in a comfortable position, and were instructed to limit movement and talking, avoid sleeping, and refrain from touching the array.29 At the end of the test, any adverse events were recorded and subjects were given a short questionnaire regarding system usability. Comfort during the test and any discomfort on removal of the array were assessed using electronic 100-point visual analogue scales (0 - “very uncomfortable” to 100 “very comfortable” for the test; and 0 “not painful” to 100 “most painful imaginable” for the array removal).
Signal Processing and Analysis
Data collected using the BSGM device were processed using an automated proprietary algorithm that enabled filtering, biomarker outputs, and visualizations. In brief, each of the 64 channels were analyzed to first remove segments of significant artifact based on the methods of Gharibans et al.21 Further steps in the algorithm then generated the biomarkers of gastric function. Spatial heat maps were generated to show the predicted gastric location within the mapped field according to a power spectrum. Spectral analyses were performed using a composite of channels located centrally over the gastric position in the heat map, by a short-time Fourier transform (4 minute windows with 75% overlap), visualized as frequency-amplitude and amplitude-time plots.21 Dominant frequency (cycles per minute; cpm), mean amplitude (μV), and variance in the dominant frequency were calculated for each participant and as summary statistics for the whole cohort. Meal response was characterized by the increase in the power of the spectral analysis after a meal (power ratio) and was calculated separately for the first 2 hours postprandially (PR2h) and the entire postprandial period (PR4h)30. Average dominant frequency was calculated for the PR2h phase when signal power is high. The duration taken to return to a stable baseline was also calculated in each period with reference to the fasting period. The frequency-amplitude spectrograms were also averaged, after normalizing amplitude for each participant, to define overall trends in the meal response power curve across the cohort. Mean amplitude was correlated against BMI.
Spatiotemporal metrics were derived for each subject using methods similar to those described by Gharibans et al.16,20 Wave patterns were visualized as propagation animations,23 and their directionality was defined by manual classification. This was performed by 5 independent reviewers with conflicts resolved by consensus panel. Each reviewer visually assessed the animated data in 15 m epochs and classified these as antegrade, retrograde, indeterminate, or low-amplitude noise, with the latter being excluded from subsequent percentage calculations. Summary data on wave directions were also computed via the algorithm and displayed as polar histograms.16,20
Normality was assessed by visual inspection of Q-Q plots. Continuous independent normal variables were compared using Student’s t-test, and continuous independent non-normal variables using the Mann-Whitney U test. More than two sets of continuous dependent normal variables were compared using the repeated measures ANOVA, with a Bonferroni post-test correction applied. More than two sets of continuous dependent non-normal variables were compared using Friedman’s test, with a Dunn correction for multiple comparisons applied. Strength of association between variables was determined using Pearson’s rank correlation coefficient (r). Sample size calculations for Array testing can be found in the Supplementary Methods. The statistical significance threshold was p<0.05. All statistical analysis was performed using GraphPad Prism v.8 (GraphPad Software, CA, USA) and R version 4.0.3 (R Foundation for Statistical Computing, Vienna, Austria). Values are reported as the mean ± standard deviation (SD) unless stated otherwise.