Posture Sweeps as a Probe of Gut Luminal Content, Validated Against 16S Sequencing

Electrical bioimpedance sensing works by passing a tiny, harmless alternating current through the body and measuring the voltage. It is cheap, wearable, and continuous. Its best-known weakness is that it is very sensitive to posture: when a person sits up or lies down, organs and fluids shift and electrode contact changes, and the readings move a lot. The field has treated this as noise to be avoided. We propose the opposite. Inside the abdomen, posture moves gut contents in physically predictable ways. Gas is non-conductive and rises to the uppermost part of the bowel. Conductive fluid settles to the lowest part. So the way an abdominal electrode belt's readings change across a fixed sequence of postures may report on what is inside the bowel, which is the material gut bacteria ferment.

The test is direct. We record this "posture response" daily while a dietary fiber challenge deliberately shifts each person's gut microbial community. We then ask whether day-to-day changes in the posture response track day-to-day changes in that community. The community is measured by 16S rRNA gene sequencing, which reads a bacterial marker gene from stool DNA to identify which bacteria are present and how abundant they are. The aim is modest and checkable. We are not trying to diagnose dysbiosis from a belt. We want to find out whether posture-resolved abdominal impedance carries any information about microbial community change that ordinary body-composition and diet covariates do not already contain. We set explicit kill criteria in case it does not.

Core Claim

Our claim is that the within-person change in abdominal impedance between postures is a better probe of luminal gas and fluid than any single-posture reading. If microbial fermentation shifts luminal content, this posture response should covary with microbial community shifts measured by sequencing, within individuals and beyond diet and adiposity covariates. The new element is turning postural sensitivity from a nuisance into the measurement itself, and anchoring it to a molecular ground truth. The research file contains no work connecting bioimpedance with 16S sequencing. The closest physical precedent, Braun et al.'s 3D thoracic EIT study, found that sitting versus supine had a "dramatic impact" on images. It discarded the non-supine data rather than modelling it, and it worked on the thorax with no molecular readout.

Research Program

The hardware is an abdominal belt with 16 electrodes in one ring around the umbilical level. It injects current at several frequencies. Low frequencies flow mainly around cells and through fluid, while high frequencies also cross cell membranes, so the ratio between them helps separate fluid from tissue. Each morning, after a standardized wake routine, the participant performs a five-minute sweep: supine, left-lateral, right-lateral, sitting, and standing, holding each for about 45 seconds.

We use the boundary voltages directly rather than first reconstructing an image. This choice is deliberate. Electrical impedance tomography (EIT) reconstruction, the process of turning boundary voltages into a cross-sectional conductivity map, is badly ill-posed. Learned reconstructors show the problem with generalizing to human anatomy. CDMVC refines a total-variation prior with a conditional diffusion model and a learned voltage-consistency network. BCR-Net learns forward and inverse boundary maps with multiscale convolutions. Both were validated only on fixed synthetic or rigid domains, and BCR-Net shows that deep interior structure is poorly resolved without well-spread electrodes. We therefore treat reconstruction as a secondary analysis. For that analysis we capture a depth-camera surface scan per posture to build posture-specific meshes, following Braun's subject-specific Kinect meshes.

The primary feature is the posture-response matrix. For each pair of postures, it records the change in each electrode-pair voltage at each frequency, normalized to that day's supine reading. This differencing cancels much of the electrode contact impedance and the fixed anatomy, which change slowly. What remains emphasizes what moved under gravity.

A small model maps this matrix to a low-dimensional summary. We expect anterior–posterior and left–right asymmetries that flip between lateral postures, because gas migrates to the uppermost side. The model is trained with leave-one-subject-out splits. In the spirit of multi-scale feature fusion, which ballistocardiography work found necessary when single-posture models failed to generalize (Lyons et al.; Jiang et al.), the model sees both the electrode-level and the ring-averaged responses.

The microbial side runs on the same timeline. We plan 30 healthy adults in a crossover design: one baseline week, one week of an inulin fiber supplement, and one washout week. Stool is sampled every second day. All samples get short-read V4 sequencing processed into amplicon sequence variants, which are exact error-corrected sequences. A subset gets full-length PacBio sequencing. The full-length subset matters because Johnson et al. report that V4 fails to classify 56% of sequences at species level. Fiber responders may differ at species or strain level, which is exactly where short reads fail. The outcomes are within-subject distance from that subject's baseline community, measured as Bray–Curtis dissimilarity, and the summed relative abundance of fermentative taxa. As a secondary, reference-free outcome, we use sequence-based biomarkers, which group sequences by alignment-derived nucleotide patterns rather than database labels.

Baselines

The naive reference point predicts each day's microbial change from the subject's own running mean and the study week alone. In other words, it uses only the knowledge that the fiber intervention happened. Any impedance signal must add information beyond the design itself. Otherwise the belt is redundant with the diet calendar.

The covariate baseline is the strongest competitor and the main threat. It predicts microbial change from:

  • standard whole-body bioimpedance body composition,
  • daily body weight,
  • self-reported stool form,
  • a diet and fluid log,
  • morning urine specific gravity as a hydration marker.

It is the most dangerous comparison for two reasons. Abdominal impedance tracks adiposity and hydration. And microbial composition is already known to separate lean from obese people: sequence-based biomarkers reach ROC-AUC 0.84 on an obesity twin cohort in the SBB paper. An impedance signal that merely repackages fat and water would look predictive while saying nothing about luminal content. We fit this baseline with the same leave-one-subject-out splits, the same model families (ridge and gradient boosting), and the same hyperparameter search budget as our method. Beating it means two things. The posture-response features must add significant held-out explained variance in within-subject Bray–Curtis change when stacked on the covariates, in a nested model comparison with subject-level permutation. And they must keep that gain after the covariates are regressed out.

The supine-only baseline uses the same belt and model but only the supine segment, which is the protocol Braun adopted. For context, Braun's supine-only thoracic stroke-volume tracking reached a relative-change concordance rate of 80.9%, which the authors judged insufficient for most clinical uses. That shows that controlling posture alone does not guarantee a usable signal. For our actual task, predicting 16S community change from bioimpedance, we found no reported result on Lacuna.

The key ablation keeps all five postures but destroys the posture structure. We either average across postures or shuffle the posture labels within each day. This preserves data volume and signal-to-noise ratio while removing the gravitational contrast. If it performs as well as the full posture response, the claim that posture is the probe fails, even if impedance turns out to be predictive.

Why not just…?

Why not just measure supine, as everyone does? A single posture cannot separate a gas pocket from a fat layer or from altered contact, because all three change the static reading. Only the change under reorientation isolates mobile contents. Braun's supine restriction did not reach clinical adequacy even in the simpler thorax. Still, this is an empirical question, and the supine-only baseline above settles it directly on our data.

Why not just use body composition and diet logs, skipping the belt? This is the real threat, and we may lose to it. That is why the covariate model is the main baseline rather than a straw man. The within-subject crossover is designed to blunt it, because adiposity barely changes over three weeks while fermentation does. If the posture response adds nothing beyond covariates within subjects, we will report that as a negative result.

Why not just measure fermentation directly with breath gas or fecal inflammation assays? These are discrete measurements. Stool-based readouts also lag luminal events by transit time. The belt's value would be continuous, at-home measurement. We found no reported comparison on Lacuna, so we will collect morning breath hydrogen on the same days as an external check. If the posture response tracks breath hydrogen but not 16S change, the device measures fermentation gas, not community composition. That narrower finding is still useful, and we would report it as such.

Why not reconstruct EIT images with a learned solver first? CDMVC reports SSIM 0.9819 and relative error 0.0634 on EIDORS phantoms, against 2.62 for traditional inversion. But those phantoms have fixed boundaries, and posture changes the boundary itself. Reconstruction adds a second ill-posed problem on top of an uncertain biological link. We run posture-meshed reconstruction only as a secondary analysis, and ask whether it beats raw differenced voltages on the primary outcome.

Proposed Experiments

Before any microbiome claim, we run two decisive physical checks.

The first uses the public Zenodo thoracic EIT dataset of 242 one-minute sequences from 10 volunteers, released with Braun et al.. We apply our supine-normalized differencing to the sitting–supine pairs. We then ask whether it reduces between-posture variance relative to raw amplitudes while preserving within-posture cardiac modulation. This tests the normalization on real posture data at zero cost.

The second is a controlled luminal challenge in 10 participants. They drink 500 mL of water on one day and 500 mL of carbonated water on another, in randomized order. We measure the posture response before and at 15 and 45 minutes after. The kill criterion is that the posture response must separate the gas challenge from the water challenge in held-out subjects, at AUC above 0.8 with subject-level bootstrap. If it cannot detect half a liter of known gas or fluid, it cannot detect fermentation, and the study stops there.

If both checks pass, the main crossover study proceeds. The primary endpoint is incremental held-out explained variance in within-subject Bray–Curtis change from baseline, comparing covariates plus posture response against covariates alone, tested by subject-level permutation. Secondary endpoints are:

  • correlation with fermentative-taxa abundance,
  • agreement with breath hydrogen,
  • whether full-length sequencing sharpens the association relative to V4.

We will also report the posture-shuffling ablation and the supine-only comparison on identical splits. A positive result would be the first evidence that a wearable electrical signal follows gut microbial dynamics. A clean negative result would close off a tempting but untested idea with a properly controlled design.