Home-Cage fTIR Floors for Measuring How Much Follow-Up a Gait Study Actually Needs

Rodent gait studies of spinal cord injury and neurodegeneration usually stop recording after a few weeks and report the gait scores at that point as the result. Their standard tool is frustrated total internal reflection (fTIR) imaging: light is trapped inside a glass plate and leaks out only where a paw touches it, so a camera below sees bright footprints on a black background. That tool has two blind spots. Animals too impaired to step across a glass runway cannot be measured at all, and each session costs handling time, so studies sample sparsely and end early. We propose moving the fTIR plate out of the test room and into the home cage, as a short glass corridor that the animal crosses on every trip between nest and food. Every contact the animal makes is then recorded, including knees, the tops of paws, and a dragged belly. The central question is one that no existing fTIR study can answer: how long does a study have to watch before its early gait measurements reliably predict where the animal ends up months later?

The opportunity comes from pairing a measurement device with a methodological worry. Clinical and translational research keeps finding that short windows capture a "honeymoon" response that later fades. The ADL+ trial saw cognitive gains at 6 months that partly receded within 3 months of follow-up (ADL+). A cardiac imaging marker predicted 6-month remodeling but could say nothing about survival (CMR dyssynchrony). Preclinical gait work has the same structure but has never quantified it, because its measurements are too sparse to rebuild a trajectory. Dense, continuous fTIR recording makes a truncation curve measurable: predictive accuracy for the long-term outcome as a function of how many days of data were seen.

Core Claim

A home-cage fTIR corridor that records every voluntary crossing, together with a classifier that labels the kind of body contact rather than only plantar footprints, can do three things weekly runway sessions cannot. It can measure locomotion during the acute phase after spinal cord injury, when runway assays have no data. It can separate forepaw compensation from hindlimb recovery day by day. And it can yield an empirical estimate of the minimum follow-up at which early gait data predict late outcomes, including late decline. The claim is falsifiable. If weekly CatWalk sessions extended to the same horizon predict late outcomes as well as the dense home-cage record, the home-cage apparatus adds nothing beyond cost, and we will say so.

What Prior Work Supplies and What It Leaves Open

Three fTIR systems supply the optics and the analysis. MouseWalker builds an 80-cm edge-lit acrylic walkway with open software extracting over 20 gait parameters at above 90% auto-tracking accuracy. It was validated on only four wild-type mice at a single time point, so it offers no repeated-measures variance or drift data. FIM shows that fTIR gives higher and more stable contrast than incident or transmitted light across illumination levels. It also shows that fTIR works at any wavelength, which lets us use near-infrared light for recording during the dark (active) phase. CatWalk dynamic footprints supplies 90 runs of α-synuclein wild-type, knockout, and humanized mice at 100 Hz. The data come from a single age, seven months, and the authors name the lack of longitudinal data as their main limitation. We reuse these designs and this dataset for calibration. Our change is where and how often the plate is used, not the optics.

The most important prior work is the CatWalk spinal cord injury index (Combination of Defined CatWalk Gait Parameters). It combines nine large-effect parameters through linear discriminant analysis into one recovery score. That score separates moderate from moderate-severe contusions where the subjective 21-point Basso Beattie Bresnahan (BBB) open-field scale does not. The paper motivates our proposal in two ways. First, it states that the runway is blind in the acute phase because rats must support their weight to cross. Second, four of its nine parameters are forepaw measures, which raises the possibility that a stable composite score hides compensation rather than neural repair. It also constrains our design. The composite score works, so our baseline must be this score computed more often and for longer, not a straw man.

The open-field pose study (Gait-level analysis of mouse open field behavior) shows the alternative route: deep-network keypoint tracking of top-down video across 1,898 mice. It also documents that route's limits. Forepaws are occluded and were excluded from analysis, 30 Hz is too slow for fast gaits, and cohorts were not followed over time. We treat it as the strongest camera-only competitor for home-cage monitoring.

Research Program

The apparatus. Each home cage is split into a bedded nest chamber and a feeding chamber, joined by a bedding-free corridor whose floor is an edge-lit acrylic waveguide with near-infrared LEDs. A high-speed camera sits below the corridor and is triggered by motion. A small side camera records a subset of crossings to produce ground truth. Debris on the glass is the main known failure mode, and it is exactly what MouseWalker flags as requiring manual correction. We address it with a narrow corridor, daily cleaning, and an automatic background-drift check that flags sessions where static bright pixels accumulate.

Acute-phase access. Veterinary rules require that severely injured rats have food and water in the nest chamber during the acute phase. A preferred food is placed only in the feeding chamber, so crossings are voluntary but still frequent. Crossing frequency is itself a measured outcome.

The contact classifier. fTIR records any contact with a refractive index above air, not only plantar steps. We train a small convolutional classifier on contact patches, using labels from the side camera, to assign each patch to one of five classes: plantar forepaw, plantar hindpaw, dorsal paw, knee or limb drag, and trunk or tail. This turns the "floor effect" into a measurement, for example the fraction of hindlimb contacts that are plantar, which can be tracked from day 1. We also reimplement the nine-parameter composite score on plantar crossings so that it is directly comparable with CatWalk.

The trajectory model and truncation analysis. For each animal we build daily feature summaries with three groups: plantar gait parameters, contact-class fractions, and forepaw-versus-hindpaw loading measured as contact area and brightness. From these we fit a predictor of a pre-registered late outcome $Y$ measured at the final time point $T_{\text{end}}$. The quantity of interest is the truncation curve

$$ A(T) = \mathrm{AUC}\big(\hat{Y}(x_{0:T}),\, Y\big), $$

where $x_{0:T}$ is an animal's home-cage record up to day $T$, $\hat{Y}$ is a cross-validated prediction from that record, and AUC is the area under the ROC curve for the binary late outcome. The minimum sufficient follow-up is the smallest $T$ at which $A(T)$ is within a pre-set margin of $A(T_{\text{end}}-\text{last month})$. To test the compensation hypothesis, we ask whether a rising forepaw loading share while hindlimb plantar fraction stays flat predicts later decline better than the composite score does.

Baselines

Endpoint baseline (naive reference). This is the design most studies use now. The composite discriminant score (CatWalk SCI index) is measured once at 6 weeks and carried forward as the prediction of the late outcome. It is computed on CatWalk runway sessions from the same animals, with the same discriminant weights refit on our training folds. Any method that fails to beat it has shown no value for extended follow-up.

Weekly CatWalk to the full horizon (strongest existing method). The same animals get standard CatWalk runway sessions every week until $T_{\text{end}}$, scored with the nine-parameter composite. The predictor gets the whole weekly trajectory, with the same model class and hyperparameter search budget as our method. It is the strongest competitor because it is validated, it is more sensitive than BBB, and extending it only takes more sessions. The reported reference points are a composite score of about 0.12 in uninjured rats against 0.03–0.06 after injury, and separation of moderate from moderate-severe contusion at p = 0.018 where BBB failed (CatWalk SCI index). We found no reported result on Lacuna for this index's ability to predict long-term outcomes. Beating this baseline means a higher $A(T)$ at every $T$ up to 8 weeks, or reaching the sufficiency margin at least two weeks earlier, with confidence intervals from an animal-level bootstrap.

Weekly BBB scoring. Two observers who do not know group assignments score BBB weekly from day 1. This is the conventional way to cover the acute phase that runways miss, and our acute contact-class features must beat it.

Overhead pose estimation in the same home cage. An overhead camera records the same cages, and an HRNet-style keypoint model is fine-tuned on our frames following Gait-level analysis of mouse open field behavior. It gets the same labeling budget as our contact classifier. We found no reported forepaw tracking accuracy on Lacuna for this approach, because that study excluded forepaws.

Ablations. Three ablations isolate our ingredients. First, subsample the home-cage record to one day per week, which separates the effect of density from that of the home-cage setting. Second, remove the non-plantar contact classes, which tests the acute-phase contribution. Third, remove the forepaw features, which tests the compensation hypothesis.

Why not just…?

Why not just run weekly CatWalk sessions for six months? This is the real threat. It keeps a validated apparatus and simply removes the truncation. It still cannot measure rats that cannot cross the runway, so the first weeks after severe contusion stay empty, and weekly sampling may miss short dips. It also measures a handled, motivated crossing rather than voluntary behavior. Whether any of this matters is an empirical question, and the weekly-subsample ablation answers it. If subsampled home-cage data and weekly CatWalk both match the dense record's truncation curve, the recommendation will be "extend follow-up with existing runways," and the home-cage floor is justified only for the acute window.

Why not just use overhead video and a pose network in the home cage? The largest such study excluded forepaws because they were occluded and recorded at 30 Hz (Gait-level analysis of mouse open field behavior). Four of the nine most discriminative spinal-injury parameters are forepaw measures (CatWalk SCI index), so the compensation question depends on exactly what overhead video misses. The head-to-head baseline measures this directly rather than assuming it.

Why not just use BBB during the acute phase and CatWalk afterward? BBB is coarse and subjective, and it failed to separate injury severities that the fTIR composite separated. Splicing two scales with different sensitivities also produces a trajectory with a break at the handover. Our acute comparison tests whether contact-class fractions carry more information than BBB during the weeks when both are available.

Proposed Experiments

The first experiment checks equivalence. On mice and rats walking plantar, gait parameters from the home-cage corridor should match MouseWalker and CatWalk values: step length rising from about 25 to 30 mm with speed, and contralateral phase near 0.5 (MouseWalker). The corridor should also reproduce the genotype separations in the α-synuclein CatWalk dataset when features are computed identically. The contact classifier is evaluated against held-out side-camera labels, with per-class accuracy reported.

The main experiment follows rats for six months after thoracic contusion at moderate and moderate-severe severities, with sham controls. The late outcomes are pre-registered: the final composite score, a sustained drop from each animal's own plateau, and histologically spared tissue. The decisive checks are three. First, whether the home-cage truncation curve reaches sufficiency earlier than weekly CatWalk. Second, whether acute contact-class features in the first two weeks predict late outcome better than BBB over the same period. Third, whether rising forepaw loading share flags later decline among animals whose composite score looked stable at 6 weeks.

A second cohort of humanized α-synuclein and knockout mice is recorded from 3 to 12 months. It extends the cross-sectional seven-month snapshot into trajectories and tests whether age at onset of gait deviation can be estimated from the dense record.

A negative result would also be informative. A flat truncation curve that saturates by 6 weeks would be evidence that, for these models, short follow-up is enough.