Inside the DVC analysis pipeline: the metrics, the methods, and the metadata gap

1. From a list of papers to a map of methods

The starting point was the public Tecniplast DVC® scientific-papers catalogue — the single best index of studies that use the Digital Ventilated Cage [16]. A catalogue, though, is a list of findings; for anyone who actually has to analyse DVC data, the load-bearing detail lives one level down, in the methods sections — the binning, the thresholds, the entropy parameters, the software [16]. This note collapses that level into one referenced reference sheet, so that a DVC user can see, metric by metric, what produced each number.

2. How we built this — a small agent team, in plain terms

Here is the honest behind-the-scenes. Rather than read two dozen papers end to end, we pointed a small team of AI research assistants at the open-access full texts behind that catalogue. Each one was given a handful of papers and a single job — pull out exactly how the data were processed — and they worked in parallel, like five people in a library each taking a shelf. Everything they found was checked against the original PDFs and merged into the Neuronautix knowledge base — a plain set of structured notes the assistants read before writing and add to after. That is the "second brain" idea in practice: the knowledge compounds, so the next question starts where the last one finished rather than from scratch. The result is meant to be two things at once: a genuine scientific resource for DVC users, and a worked example of why a curated, cited knowledge base beats a folder of unread papers.

3. One raw signal underneath everything

Before any metric, there is one stream of numbers. A board of twelve capacitive electrodes, arranged as a 4×3 grid, sits under the floor of each cage; every electrode's capacitance is sampled four times a second (every 250 ms), which adds up to roughly 2.5 MB per cage per day [1][2]. An electrode is said to "activate" when the change in its capacitance between successive samples exceeds a noise threshold calibrated on empty cages [1]. Every named metric below — activity, rest, circadian, spatial, even bedding wetness — is some transformation of that twelve-channel time series, which is why getting the signal right is the whole game [1][2]. That signal is not lab-specific: the activation readout reproduces across independent sites — CNR Rome, The Jackson Laboratory, and Karolinska — which is what licenses treating the recipes below as portable rather than one facility's artefact [3].

From that one stream, the rest of this note follows three branches, so the metrics read as a pipeline rather than a glossary:

• The activity branch — how much the animal moves (activation / locomotion index, distance, rest and activity bouts) and the rhythm and place of that movement (rest-disturbance, circadian parameters, spatial preference).
• The environmental branch — what the same electrodes sense about the cage rather than the animal (bedding moisture → Bedding Status Index → Urination Index).
• The model branch — classifiers and indices layered on top of the first two (severity scoring, cage-change prediction).

One rule cuts across every branch and is worth stating once: metrics computed from the amount or rhythm of the activity signal — the Animal Locomotion Index (ALI) / activation, the rest-bout split, the Rest Disturbance Index, circadian parameters, and the electrode-region spatial metrics — stay valid for group-housed cages of any numerosity, because they read the cage-level signal; only the metrics that reconstruct an individual trajectory — distance walked and the per-animal bout decomposition — require single housing [2]. So ALI is the workhorse that travels across every cage type, and tracking is the high-resolution option you buy by housing one animal per cage.

4. The activity branch (1): activation for any cage, tracking for single housing

From the same raw stream, the literature derives activity in two distinct ways. The first, electrode-activation density (EAD), compares two short averaging windows of the capacitance signal and flags an activation when the absolute change clears the lowest threshold that still rejects empty-cage noise — λ = 1.25 — yielding a binary per-electrode signal that is then averaged per second; crucially, this works for group-housed cages [2]. The second, centroid tracking, reconstructs a position: each electrode's baseline is the maximum capacitance over a one-minute window, the centroid is the electrode coordinates weighted by their "signal drop" to about 1 mm resolution, motion is called when the centroid shifts at least 1 mm between samples, movement-on-the-spot is separated from true locomotion at a 65 mm stride length, and a long rest is any bout of at least 40 seconds — but this only works for a singly housed animal [2]. Reassuringly, the two agree closely — EAD tracks the distance measure at r > 0.95 and the two coincide about 97% of the time — so the cheaper, group-housed-friendly activation metric is a sound proxy for the richer tracking one [2].

5. The activity branch (2): rhythm and place

Still on the same activity series — so these all extend to group-housed cages — the famous indices turn out to be precise algorithms, not vibes. The Rest Disturbance Index — the rest-fragmentation biomarker that recurs across disease models — is computed from minute-binned activity by first zeroing everything below λ = 0.005, then applying a fourth-order Butterworth band-pass filter (normalised cut-offs of 1/2000 and 1/300), and finally taking the sample entropy with parameters m = 2 and r = 0.2 [4]. It has a single mathematical origin and is scale-invariant — every later paper that uses "RDI" points back to that one definition rather than redefining it [4][5]. Circadian phenotyping follows a different chain: a 30-minute running average, then a chi-square periodogram in the ImageJ plug-in ActogramJ, with inter-daily stability, intra-daily variability, and relative amplitude hand-coded in R from the Witting equations — all under a per-cage LED system delivering about 100 photopic lux so each cage is its own light-controlled chamber [6]. And the spatial metrics are read straight off the grid geometry: "frontality" is the share of activity over the six frontal electrodes (numbers 7–12), "wall activity" sums the left and right electrode columns as a thigmotaxis proxy, and a Gini index across the twelve electrodes captures how concentrated the activity is in one spot [8].

6. The activity branch (3): from cage to animal — and why that step is contested

The cage-level signal travels everywhere; turning it into a per-animal number is the step that does not come for free. The common approach is disarmingly simple — divide the hourly cage activity index by the number of mice in the cage, requiring only that each cage holds a single genotype or treatment — and it was validated against a conventional TSE LabMaster infrared beam frame by converting both systems to z-scores and comparing effect sizes [7]. It is worth being candid that this normalisation is not universally accepted: a cage aggregate cannot, in general, be cleanly attributed to one animal — group activity is super-additive when mice interact and sub-additive when they huddle — so dividing by occupancy is a defensible average, not a per-individual measurement, and any study leaning on it should justify the assumption for its own design rather than treat it as settled [7]. For phenotypes that bulk activity misses, an optional GYM500 running wheel logs voluntary-running distance on the very same electrode board, reading the wheel's magnets rather than the mouse [13].

7. The environmental branch: from bedding moisture to a metabolic readout

The most distinctive DVC metrics do not watch the animal at all; they watch its bedding. Increasing bedding moisture is sensed as a decline in the electrodes' electromagnetic-field strength and recorded as the Bedding Status Index (BSI) [10]. From that, the Urination Index is built as a fully deterministic transform — invert each moisture increment, drop the deltas in a window around every cage-insertion event (to reject handling and water-bottle artefacts), take a cumulative sum, and divide by housing density — calibrated at about 1.4 index units per millilitre of added water [10]. Its sibling, the cage-change model, instead trains machine learning on human annotations of soiled bedding, reaching over 90% accuracy at high density and supporting 3–6-week change intervals [9]. The Urination Index is also, as far as this corpus goes, the happy exception: its code is public — UrinatoR, an open-source, MIT-licensed R/Shiny app you can actually run [10][15].

8. The model branch: the transparent classifier worth copying

Not every "AI on DVC data" result is a black box, and one in particular is a model of restraint. A colitis-severity classifier used nothing more exotic than a logistic regression on two features — DVC activity and body weight — trained on a single day's data (day 7) with a probability cut-off of 0.5, and it reproduced the graded disease course [11]. At the more bespoke end, a narcolepsy study added a nest-identification algorithm and found that an inability to sustain activity for more than 40 minutes was itself a robust biomarker [12]. In our view the colitis model is the more reusable template precisely because it is simple, legible, and reimplementable from the paper alone [11].

9. The real bottleneck: raw data and its metadata

Pulling the whole map together exposes a gap that matters more than any single threshold — and it is not the mathematics. The raw electrode stream is, in principle, fully exportable — and that openness is genuinely valuable. In practice it is rarely usable: it tends to sit in per-facility silos, and, more decisively, it arrives without the metadata — strain, sex, age, housing density, cage-change schedule, light programme, interventions and their timestamps — that every recipe above quietly depends on. Raw values with no context are close to inert: you cannot choose the right baseline, exclude cage-change days, normalise per occupancy, or even tell whether a cage was single- or group-housed. The code gap compounds it: across essentially the entire corpus the analysis code is unreleased — statements say "in the article" or "on request," and apart from UrinatoR no public repository exists — so the transform is left as prose and the input it would consume is left under-described [10][15]. And the very first transformation — raw capacitance into the activity and bedding indices — runs inside the proprietary DVC Analytics platform, so even a motivated, well-provisioned user already starts one step downstream of the signal [1]. The practical consequence is that the binding constraint on reuse is data plus metadata, not algorithms: the recipes reproduce fine, but a recipe is worthless without well-described ingredients — exactly the FAIR-metadata, standard-format, and API discipline that recent home-cage-monitoring reviews call for [14].

10. The takeaway: one signal, a connected pipeline, a metadata-shaped gap

Read as a whole, the DVC is not a bag of metrics but a single capacitance signal that forks, by deterministic recipe, into activity, rhythm, place, and environment — and which branch you may legitimately use is set first by your cage configuration: ALI and the series-based indices for any cage and numerosity, trajectory metrics for single housing only. The recipes here — the EAD threshold, the RDI entropy parameters, the electrode-region maps, the Urination Index transform — are reproducible enough to re-implement; what actually decides whether a result is trustworthy and reusable sits upstream of all of them, in the raw data and the metadata that says what the numbers describe. Treat every metric as a provenance record — store the recipe, the baseline, the cage configuration, and the software version with each value — and the gap this note keeps returning to is the one worth closing first [14]. The meta-point is the one we set out to show: a curated knowledge base turned a vendor list of papers into a connected, cited methods map — and because it was written back into that base, the next DVC question we are asked already has this as its memory [16].