Moving the Workload, Watching the Gut: A Crossover Test of Whether Academic Load Perturbs the Human Microbiome

Educational reforms rarely remove work. They move it. Flipped classrooms push lecture content into required pre-class preparation. Homework caps push supervision onto teachers and parents. This "workload displacement" is well documented in surveys and learning-platform logs. Whether the moved work has any bodily cost is unknown. Separately, gut microbiome researchers study communities by reading one bacterial marker gene (16S rRNA sequencing, which counts which bacteria are present from a DNA barcode). They have a matching problem. They suspect everyday psychological stress reshapes the gut community, but in humans stress is tangled up with diet, sleep, and fixed person-to-person differences, so observational studies cannot cleanly show it.

This proposal uses one field's phenomenon to fix the other field's identification problem. A course can be run so that the same students carry roughly the same total work but at different times in the semester. That gives a scheduled, randomizable shift in when load falls on each person. The question is whether each student's gut community moves in step with that shift. If it does, we have within-person, quasi-experimental evidence that naturalistic cognitive load perturbs the microbiome. We would also learn whether the response tracks peak weekly load or cumulative load. If it does not, that null is informative and cheap to obtain honestly.

Core Claim

The obvious bridge between these fields would be a regression of taxon abundances on burnout scores, pitched as a biological measure of the workload that learning-platform data miss. We think that version is weak. The "unexplained 64%" in Borchers & Pardos includes survey noise that no biomarker should be expected to recover. Cross-sectional microbiome–stress correlations would also be swamped by between-person variation.

Our claim is narrower and testable. Workload displacement provides an exogenous, time-shifted load signal. A randomized crossover between flipped and conventional teaching changes the timing of effort while holding course content fixed. Therefore, if academic load affects the gut, each student's deviation from their own microbial baseline should shift in time along with their own measured effort curve. It should not simply follow the shared calendar.

The research file reports no paper connecting institutional workload to 16S data. The closest work tracks gut communities under artificial physical stressors in animals (Frontiers in Medicine 2025; Biology 2024). It also includes longitudinal 16S models that have no measured human exogenous input (Gibson & Gerber; Zhang et al.). Our work differs in three ways. The stressor is naturalistic cognitive load. It is measured weekly per person. And it is shifted in time by design.

Research Program

Setting and design. The setting is a required course with two parallel sections. Examples include an undergraduate medical block like the ophthalmology course in which Wu et al. randomized 93 students between flipped and didactic teaching, or a programming-heavy course where assignment effort is known to exceed nominal credit (Foster & Wagner). Sections are randomized to flipped-then-conventional or conventional-then-flipped, switching at mid-semester.

Why the crossover shifts effort. Flipped instruction front-loads effort into pre-class preparation. Per Liu et al., uncalibrated preparation materials produce preparation fatigue, and Fei-E Fu notes that "effective preview" requirements turn passive attendance into high-effort pre-class work. So the two arms should show effort curves that are phase-shifted against each other. We will not assume this shift occurs; we will measure it.

Measuring load. Weekly per-person load comes from three sources: learning-platform telemetry built with the feature set of Borchers & Pardos, a short time diary, and a brief exhaustion item.

Biological sampling. Students self-collect stool weekly at home. All samples receive V4 short-read 16S sequencing plus universal 16S qPCR (quantitative PCR, which gives total bacterial load). The qPCR is needed because 16S alone gives only relative proportions that sum to one, which distorts naive covariance. A stratified subset, including peak-load and baseline weeks, also receives full-length PacBio sequencing of the whole ~1,500-base gene. Johnson et al. show that V4 leaves 56% of sequences unclassified at species level, whereas full-length sequencing recovers species and strain distinctions.

Measuring confounders and mediators. Diet and sleep are the obvious channels through which workload could reach the gut. We record them with a brief food-frequency log and a wrist wearable. We also collect weekly salivary cortisol.

Dynamical model. We extend the modular Bayesian generalized Lotka–Volterra model of Gibson & Gerber. That model describes how taxa grow and inhibit one another, groups taxa into interaction modules to keep the parameter count manageable, and couples 16S with qPCR counts. We add a measured workload forcing term:

$$ \frac{dx_i}{dt} = x_i\Big(r_i + \sum_j a_{ij}x_j + b_i\,u_s(t) + c_i^\top z_s(t)\Big), $$

The symbols are defined as follows:

  • $x_i$ is the absolute abundance of taxon $i$ in student $s$.
  • $r_i$ is its baseline growth rate.
  • $a_{ij}$ is the effect of taxon $j$ on taxon $i$, shared within modules.
  • $u_s(t)$ is that student's measured weekly workload.
  • $b_i$ is the taxon's sensitivity to load.
  • $z_s(t)$ collects diet and sleep covariates, with coefficients $c_i$.

What the model tests. The scientific question is whether the load sensitivities $b$ are nonzero and module-structured. The comparison between arms asks whether fitting $u_s(t)$ outperforms a shared calendar term. We also fit structural topic models (Holmes et al.), which group bacteria into co-occurring subcommunities, and let the topic proportions depend on weekly load. This gives a second, assumption-lighter read-out.

Baselines

Naive reference: calendar-only model. Microbial deviation from each student's own baseline is predicted from week of semester alone, shared across arms. This captures season, holidays, and exam timing that both sections experience. Any claimed load effect must beat it on held-out weeks. Beating it means lower held-out prediction error for community composition, measured as Aitchison distance on qPCR-scaled profiles. That improvement must specifically come from weeks where the two arms' measured effort curves diverge.

Strongest workload-measurement competitor: learning-platform model. This is the course-load analytics model of Borchers & Pardos. It reports that platform features explain "up to 36%" of variance in self-reported workload, versus about 6% for credit hours, with a 22% improvement in mean absolute error. We reuse it as the load input $u_s(t)$ rather than competing with it on workload prediction. It becomes a competitor in the secondary question of whether microbial state adds anything for predicting next-week self-reported exhaustion. There, the microbial features must improve held-out prediction beyond platform features plus cortisol and sleep. The comparison uses the same students, the same weeks, and nested cross-validation with identical tuning budgets.

Strongest physiological competitor: cortisol plus wearable sleep. For human stress-to-microbiome responses, we found no reported result on Lacuna. So there is no microbiome number to beat. The bar is internal to the study: the load-to-microbiome association must survive conditioning on these channels, or its mediation through them must be quantified.

Microbiome-analysis baselines. Taxon-wise mixed-effects regression on V4 amplicon variants serves as the simple analysis. The load-forced dynamical model and the topic model must beat it at recovering the arm-by-week signal.

Key-ingredient ablation. We refit the dynamical model with $u_s(t)$ replaced by the student's arm-average workload, and separately with the arm label alone. This isolates whether individually measured, time-shifted load carries the signal, or whether a crude "flipped vs not" contrast does.

Resolution ablation. On the full-length subset, we compare V4-only analyses against full-length analyses. This tests whether the strain-level resolution shown by Johnson et al. changes which taxa appear load-sensitive.

Why not just…?

Why not just correlate burnout scores with gut taxa in a large cross-sectional cohort? Person-to-person differences in gut composition are large, and diet varies with personality and circumstance. A cross-sectional association cannot separate "stressed people eat differently" from "stress changes the gut." Tamai et al. also show that self-reported burden inflates non-linearly under other pressures, so the exposure itself is biased. The crossover uses each student as their own control and randomizes timing. That is precisely what a cross-sectional study cannot do.

Why not just measure cortisol and skip sequencing? This is a real threat. If cortisol and sleep explain every load-linked microbial change, the gut adds nothing for predicting burden, and we will report that outcome. But the core question here is whether naturalistic load reshapes the gut community at all, not whether the gut is a better stress meter. Cortisol cannot answer that. The mediation analysis will show how much of the microbial response passes through sleep, diet, and cortisol. It is also plausible that workload displaces eating behaviour and the gut follows. That would be a finding about displacement, and we will not hide it.

Why not just run cheap V4 sequencing on everyone and forget full-length reads? V4 is our default for all samples, precisely because it is cheap. Full-length sequencing is confined to a subset. If the resolution ablation shows the same load-sensitive modules under both, the recommendation for follow-up studies is V4 plus qPCR, which is a useful result in itself. The 56% species-level failure rate of V4 reported by Johnson et al. is the reason not to assume this in advance.

Why not just use exam weeks as the natural stressor? Exam weeks are shared by everyone and coincide with calendar effects such as travel, holidays, and changes in dining. The crossover gives two groups whose load peaks fall at different times within the same calendar. That separation is the whole identification argument.

Proposed Experiments and Decisive Checks

Manipulation check. The first check is whether displacement actually happened: the measured weekly effort curves of the two arms must be phase-shifted. If the flipped arm does not front-load effort, the study reverts to an observational within-person design, and we say so.

Decisive test: phase shift. Estimate the cross-correlation lag between each student's workload curve and their microbial deviation, measured as Aitchison distance from their own pre-semester baseline. The key question is whether the peak of the microbial response moves with the arm switch. A load effect predicts that each arm's microbial peak follows its own effort peak. A calendar artifact predicts that both arms peak together.

Negative controls. We use two. First, permute arm labels within calendar week. Second, a small cohort of non-enrolled students sampled in parallel should show the calendar term but no arm-specific shift.

Mechanistic check. Test whether the load sensitivities $b_i$ cluster within interaction modules rather than scattering across taxa. Then check whether load-sensitive topics from the Holmes et al. topic model overlap those modules.

Model-assumption check. Compare the forced dynamical model against the network model of Zhang et al. under a load split. The research file notes that the Zhang model assumes a common temporal dampening rate across taxa, which contradicts taxa that recover at different speeds. The test is whether network structure differs between high-load and low-load weeks. If it does, that assumption is violated under load.

Pre-registration. All three primary outcomes are registered in advance, with nulls reported as such: the phase-shift test, survival of the load effect after adjusting for sleep and diet, and the incremental prediction of exhaustion over the Borchers & Pardos features. Because no effect size for naturalistic human load on the gut has been reported, a pilot term will set the final sample size and sampling density.