Feeding-Bout Dynamics as a Non-Invasive Clock for Diet-Induced Gut Inflammation in Fish

Replacing fishmeal with sustainable proteins often gives farmed fish an inflamed gut. Single-cell protein made from industrial gas fermentation is one example. Today the only way to see that inflammation is to kill the fish and examine the intestine, usually once, at the end of an 8-week trial. Because of this, the field cannot tell whether a dietary "rescue" additive fixes the problem or only delays it. Nobody can afford to keep sacrificing fish over a 12- to 24-month production cycle.

This note proposes a way to watch the problem develop without killing the animals. Cameras and underwater microphones already classify how hard fish are feeding. We would use them to record the shape of each feeding bout: how fast the fish strike, how quickly the frenzy decays, and how the effort is spread across the tank. We would then calibrate that shape against sparse gut tissue samples, so that the behavioral signal becomes a validated early-warning indicator. With that indicator, a nutrition trial can run far beyond 8 weeks at little extra cost.

Core Claim

The claim is that the time course of a feeding bout carries information about gut inflammation that total feed intake and body weight do not. That information should show up earlier, and it can be read continuously without harming the fish. The claim is specific and falsifiable in two ways. If the curve adds nothing over what a technician gets by weighing uneaten pellets, it is false. It is also false if the curve does not predict inflammation measured two weeks later.

The biological motivation comes from Wang et al.. They replaced 45% of fishmeal protein with Clostridium autoethanogenum protein in diets for large yellow croaker. At 8 weeks this depressed growth and digestive enzymes and raised gut TNF-α, IL-1β, IL-6 and IFN-γ. Tributyrin at 0.05–0.1% largely reversed these effects, but the benefit narrowed at higher doses. That study, like the 8-week shrimp trial of Mu et al. and the 10-week bass trial of Xie et al., is a single snapshot. These papers motivate the proposal: they show a real pathology and a real rescue, but with no time resolution.

The sensing method comes from U-FFIA. It fuses overhead video and hydrophone audio to classify 10-second clips into four feeding-intensity levels. It stays robust when one modality is degraded and reports 0.907 accuracy. U-FFIA never links intensity to internal pathology and never looks at multi-week trends, and those two things are what we add.

HAIL-FFIA extends intensity classification across six species. Its authors saw accuracy collapse on a Sunfish cohort with unexplained health problems, and they could not diagnose the cause. We read that as indirect evidence that feeding behavior shifts when fish are unwell. It also warns us that sick fish break a feeding classifier that was trained only on healthy fish.

The durability question is framed by work on surrogate endpoints. Sohal et al. found that a reproducible imaging marker predicted 6-month cardiac remodeling (odds ratio 5.67), yet they conceded it says nothing about survival. Chew et al. found cognitive gains that regressed within three months after their intervention ended. Together these papers constrain how we should interpret the proposal. A rapid marker is only worth having if it is validated against the later, harder outcome. For that reason our design scores the behavioral signal against inflammation weeks ahead and against 24-week outcomes, not against concurrent measurements.

Research Program

The trial uses juvenile large yellow croaker reared in individual tanks, with four diets taken from Wang et al.:

  • 40% fishmeal as the control;
  • the 45% CAP replacement diet;
  • that diet plus 0.1% tributyrin, the reported best dose;
  • that diet plus 0.8% tributyrin, a dose where the rescue was weaker.

Four tanks receive each diet, and the trial runs for 24 weeks instead of 8. At week 12, two tanks from the unsupplemented CAP arm switch to the fishmeal diet. This crossover creates a recovery event inside those tanks. It gives a causal test: the behavioral signal should move back toward the control pattern after the switch, and it should do so before growth catches up.

Every tank gets a hydrophone and an overhead camera recording each scheduled feeding. A U-FFIA-style distilled student network turns each bout into a sequence of intensity estimates, one per second. From each sequence we extract a small number of pre-registered curve features:

  • latency to the first strong response;
  • peak intensity;
  • the half-life of the decay from peak;
  • the area under the intensity curve.

The largest confound is growth. Fish grow roughly tenfold over the trial, which changes their sounds, their silhouettes and their appetite. Rather than building a continual-learning model to separate aging from disease, we let the fishmeal control tanks act as an age-matched reference. Each tank's features are expressed as deviations from the control-tank trajectory on the same day. We also re-label a few hundred clips at each sampling point and fine-tune the classifier. The fine-tuning keeps the intensity labels calibrated. The deviation-from-control step removes the shared age drift from the diet effect.

Tissue calibration is deliberately sparse. At weeks 2, 4, 8, 12, 16, 20 and 24, three fish per tank are sampled for gut histology, trypsin and lipase activity, malondialdehyde, and RT-qPCR of TNF-α, IL-1β, IL-10 and Nrf2. These measurements are combined into a single pre-registered inflammation index. Three fish per tank per timepoint is few enough that tank density, and therefore social feeding behavior, is not badly disturbed. That disturbance is the concern the research file raises about serial destructive sampling. Pellet intake (offered minus recovered uneaten pellets) is recorded at every feeding. Tank biomass is weighed every two weeks.

Baselines

The naive reference is to do nothing new: run the standard 8-week endpoint protocol and extrapolate. Concretely, the week-8 inflammation index and weight gain of each tank are used to predict that tank's week-24 inflammation index and survival. This is what the field does now. Beating it means only that continuous monitoring flags at least one arm whose week-8 status reverses by week 24. It is a low bar, but if no arm reverses, the whole durability motivation weakens, and we would report that.

The strongest competitor is intake plus growth. This uses daily pellet-weighed feed intake and biweekly biomass, fitted with the same mixed-effects model and the same deviation-from-control normalization as the behavioral features. It is the strongest competitor because feed refusal is the classic, cheap sign of gut distress. A skeptical nutritionist would predict it captures everything. It gets the same tanks, the same tissue labels, the same leave-one-tank-out cross-validation, and the same number of tuned parameters.

We found no reported result on Lacuna for predicting fish gut inflammation from either intake or behavior. There is therefore no published number to beat, and this baseline is defined by the experiment itself. Beating it means a statistically reliable gain in predicting the inflammation index two weeks ahead when the bout-curve features are added to intake and growth. The gain is measured as held-out R² or rank correlation across tanks, with a tank-level permutation test.

The perception baseline is the feeding-intensity classifier itself. We reuse the U-FFIA architecture, whose authors report 0.907 accuracy on AV-FFIA. We also run the audio-only CNN6 from AFFIA3K, which reports 0.65 accuracy and 0.74 mean average precision. These are the reported upper and lower reference points for how well intensity can be read at all. We re-measure both on hand-labeled croaker clips at each growth stage, so that any failure of the pathology prediction can be separated from a failure of perception.

Two ablations isolate the key ingredient:

  • Bout totals only: the per-bout mean intensity, which is essentially what U-FFIA outputs, used in place of the curve features. This tests whether the curve's shape, rather than its average level, carries the signal.
  • No control reference: the features used without subtracting the age-matched control trajectory. This tests whether the simple normalization is enough to handle growth drift.

Why not just…?

Why not just weigh uneaten pellets? This is the real threat, which is why it is the primary baseline rather than a straw man. Inflamed fish may simply eat less, and intake would then track the index as well as any camera. The argument for bout shape is timing. A fish with mild discomfort may still finish its ration while striking more slowly or stopping sooner, and intake is saturated at that stage. If the ablations show that bout totals match the curve features, and that intake matches both, we would conclude that pellet weighing is enough for long trials. That is still a useful, publishable result for extending trials cheaply.

Why not just sacrifice fish at more timepoints? We do sample serially, but thinly. Sampling enough fish for weekly resolution depletes tanks and changes competition for feed, as noted for small cohorts like the 2 g shrimp of Mu et al. It also gives results days later, after RNA extraction. The experiment tests whether dense behavioral data between sparse samples recovers what denser sacrifice would show. The crossover arm checks this: if the signal cannot time the recovery after week 12 more finely than the four-week tissue grid can, the camera adds nothing.

Why not just run a fecal isothermal CRISPR test for cytokine transcripts? RECCT-Cay shows that recombinase-based amplification with a Cas12a readout works on crude stool in under an hour. It reports a detection limit of 7 copies/µL, but only for human parasite DNA. Detecting fragile host mRNA in fish feces that have been diluted in seawater is unproven. The file also notes that per-test enzyme cost is high for daily surveillance. We treat this as a complementary, higher-risk track. Frozen fecal samples collected at each tissue timepoint would allow it to be tested later against the same inflammation index.

Why not just use a continual-learning model like HAIL-FFIA to handle growth? HAIL-FFIA reduces forgetting across species to 9.36%. However, it was built for adding species, not for the gradual drift of one cohort as it grows, and it faltered on sick fish. Our concurrent-control subtraction is simpler. The no-control ablation shows directly whether anything more elaborate is needed.

Proposed Experiments

Three results decide the question.

The first is prediction ahead of time. Under leave-one-tank-out cross-validation, do the bout-curve features plus intake predict the inflammation index two weeks ahead better than intake plus growth alone, and better than bout totals? Success requires a significant gain under the tank-level permutation test, with the tank, not the fish, as the unit of analysis.

The second is the recovery timing test in the crossover tanks. Changepoint detection on the behavioral deviation series should place the recovery earlier than intake and growth do. The tissue index at weeks 16 and 20 provides the check.

The third is durability. For each diet arm, compare the week-8 ranking with the week-24 ranking on the inflammation index and survival. Then ask whether the behavioral trajectory over weeks 8–16 anticipated any reversal, and whether the 0.1% and 0.8% tributyrin arms diverge or converge.

The main statistical limitation is the small number of tanks. With 16 tanks, a modest effect could be missed. We therefore pre-register the feature set, the index, and the analysis, and we report effect sizes with intervals rather than relying on significance alone.