Probabilistic Modeling Confirms Delayed, But Non-Deficient Locomotor Development in Low Birth Weight Piglets

In the study of vertebrate locomotion, kinematic measures of gait, dynamic posture, and coordination often change when comparing subjects of different body mass, size, and age. Is it, conversely, possible to infer subject characteristics from the kinematic measures? For this study, piglets ( Sus domesticus ) were filmed from lateral perspective during their first ten hours of life, an age at which body mass and size have major consequences for survival. We then apply deep learning methods for point digitization (DeepLabCut), calculate joint angle profiles, and apply information-preserving transfromations (Fourier Series and PCA) to retrieve a multivariate kinematic data set. We train a probabilistic model to predict subject characteristics from kinematics. The model infers subject characteristics accurately for strides from piglets of normal birth weight (i.e. the category it was trained on), but surprisingly predicts the body mass and size of low birth weight piglets (which were left out of training, out-of-sample prediction) to be “normal”. The discrepancies between prediction and observation confirm that dynamic posture and coordination are unaffected by lower birth weight and lower size, which is evidence that low birth weight does not imply locomotor deficits. However, the age of some (but not all) low birth weight individuals is underestimated, supporting the hypothesis that these piglets can experience a delay in locomotor maturation. (218 words)


Introduction
Vertebrate locomotion is a complex phenomenon. The kinematic and dynamic measurements obtainable in experiments represent the collective output of interacting variables of an ensemble of subsystems: the musculoskeletal apparatus, energy supply, metabolism, and multiple levels of neuro-motor control (Nishikawa et al., 2007). All these subsystems are potentially affected by characteristics of the animal (e.g. age) and by external circumstances (e.g. friction) in different, non-trivial ways. Conversely, studying alterations in locomotor patterns under controlled experimental conditions holds diagnostic potential (e.g. Figueiredo et al., 2018). In well-studied species, it is even possible to identify individuals from kinematics (e.g. Patua et al., 2021).
Yet the complexity of the biological system is a major challenge to locomotor research. Experiments might record two animals of different weight, walking at different speeds, and producing two distinct kinematic patterns (e.g. one tending to bend the knee and ankle joints, the other emphasizing bending of knee and hip). Additionally, even for identical external conditions and in a single individual, there is intrinsic variation in the process, owing to the fact that "successive movements [. . . ] never exactly repeat themselves" (Bernstein, 1935). Which of the many "input factors" is responsible for the difference lateral to the camera. After the recordings the piglets were returned to their litter and remained with the sow.

Digitization
We used the software DeepLabCut (DLC, Mathis et al., 2018) for digitization of all video material, in combination with custom made point tracking software (Mielke et al., 2020) to generate a training set.
In total, our dataset contained 180 videos (more than 11 hours, 169 animals) of video. Our goal was to prepare a general DLC network which is capable of automatically tracking piglets at multiple ages, and which can be shared and re-used for subsequent research questions. This is why the full data set was used for digitization and for the calculation of some derived measures (size PCA). However, the analysis focus of this study (see below) was only a subset of the data (58 animals of the youngest age class).
The video processing workflow, applied to the full data set, was as follows. To get a balanced training set, one representative stride of each of the animals was selected, and the video was cut, cropped to runway height, and optionally mirrored horizontally so that movement would always be rightwards. All videos were concatenated and submitted to the DLC training set generation. DLC was set to select 2552 frames from these videos, which were tracked in an external software and re-imported for training (80% training fraction). Seventeen landmarks (i.e. points of interest or "key-points"; usually joint centers) were digitized, representing all body parts visible on the lateral perspective (head: snout, eye, ear; back line: withers, croup, tail base; forelimb: scapula, shoulder, elbow, wrist, metacarpal, fore hoof; hindlimb: hip, knee, ankle, metatarsal, hind hoof). We selected a "resnet 152" network architecture and trained for 540, 672 iterations (16 days of computer workload). The network was then applied to digitize the full data set twice: once in default direction and once horizontally mirrored, because training set was always rightward movement.
Naturally, the trained network would only extract potentially useful landmark traces for episodes which resembled the training set, i.e. in episodes with a piglet moving perpendicular to the image axis, in lateral aspect and rightward direction. We automatically extracted 2597 of these by filtering for high digitization "likelihood" provided by DLC, low noise (i.e. steady landmark movement) and consistent, plausible landmark distances. We further applied an automatic algorithm to find footfalls and label stride cycles in the candidate episodes (4730 cycles). This procedure involved a start-end-matching 8 . CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprint this version posted March 24, 2022. ;https://doi.org/10.1101https://doi.org/10. /2022 optimization (using Procrustes superimposition) to ensure that strides were indeed cyclical. To further assess digitization quality, gait variables were automatically extracted. Definition of these variables was chosen to simplify the automatic procedure, as follows. Stride distance, frequency, and speed are trivial measures of the animal movement. Duty factor is available for fore-and hindlimb, and measures the fraction of time in which the respective hoof is in ground contact. Clearance is approximated by quantifying the ratio of flexion of each limb (one minus the quotient of minimum and maximum absolute hip-toe-distance during the stride). Head and torso angle are the stride-average angles of the snout-ear or withers-croup lines with respect to the static reference frame. Hindlimb phase measures the time between hind-and forehoof touchdown, divided by the stride cycle duration. Where applicable, gait variables were prepared for analysis (see below) by converting them to dimensionless values (Alexander and Jayes, 1983;Hof, 1996) using the cumulated distance of landmarks along the snout-to-tailbase line of the animal as reference, extracted as stride average from the digitized landmarks. Only strides with plausible values (i.e. those which lie within the theoretical distribution of each parameter; 1862 cycles) where processed. Manual inspection further boiled down the data set to 897 stride cycles (the others excluded for digitization errors, multi-animal confusion, non-walking gait, intermittent or sidewards locomotion, or incompleteness). Finally, 368 of the remaining strides from 58 animals were in the youngest age category (< 10 h) and thus selected for the present analysis, the data table is available online (see below).

Data Processing
The landmark data provided by DLC was further processed for analysis. Python code for the whole procedure is publicly available (https://git.sr.ht/~falk/piglet_fcas, Python version 3.10.1 at time of model calculation, https://www.python.org). First, joint angle profiles (i.e. joint angle values over time) were extracted for all relevant joints and for the total forelimb angle (croup-withers-hoof).
Shoulder, elbow, wrist, hip, knee, and ankle were the six joints sufficiently well digitized and therefore considered relevant for analysis. We then applied Fourier Series decomposition in the framework we previously termed Fourier Coefficient Affine Superimposition (FCAS, Mielke et al., 2019), a flexible procedure which subsumes the following steps. Joint angle profiles are cyclic, i.e. periodical, and can therefore be transformed to the frequency domain with a Fourier Series decomposition (8 harmonics 9 . CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprint this version posted March 24, 2022. ;https://doi.org/10.1101https://doi.org/10. /2022 were deemed sufficient by visual comparison of raw and transformed/retransformed profiles). In the frequency domain, the affine components (mean, amplitude, phase) of a joint angle profile are easily accessible (cf. Mielke et al., 2019). The forelimb angle served as reference to temporally align all cycles in the data set (removal of phase differences between different cycles; forelimb angle was not used further). Then, mean and amplitude of the joint oscillations were isolated for all joint angles and are categorized as "dynamic posture" parameters. Mean joint angle is the temporal average, whereas amplitude is related to effective range of motion (eROM). The residual, i.e. relative phase differences captured by non-affine Fourier coefficients, can be categorized as "coordination" sensu stricto (it measures the precise temporal succession of joint configurations). In our case, there were 96 variables of coordination (6 angles, 8 harmonics, real and imaginary) which were submitted to a PCA. Only the first 12 coordination components (CC) were used for statistical analysis, capturing 76.4% of the variability in coordination.
To summarize, FSD and FCAS served three purposes: (i) temporal alignment of the cyclic traces, (ii) separation of meaningful parameter categories (dynamic posture and coordination), and (iii) preparation for multivariate analysis via PCA.
Information retention is generally a strength of this method. FCAS and PCA are mathematical transformations, which means that the information content after transformation is theoretically identical to that prior to transformation (theoretically, because only a finite number of harmonics can be used, yet this is of little concern for continuous, smooth joint angle profiles). The neglected PCs and the residual not captured by 8 harmonics were the only information from kinematics of the given joints to be lost in this procedure, and by definition these contain the least information. Apart from that, all information present in the raw joint angle profiles enters the analysis.
Furthermore, all transformations are reversible, hence any analysis outcome can be translated back to kinematics with high accuracy. This can be of use when presenting raw joint angle profiles and their averages. One crucial aspect of the FCAS procedure is temporal alignment of the joint angle profiles in the frequency domain. As opposed to conventional temporal alignment, where a single characteristic point in the stride cycle is chosen as a reference, transformation to the frequency domain allows a mathematically optimal alignment which considers the whole stride. For example, the ankle profiles ( Fig.   2, lower right plot, green traces) have a distinct, short period of quick flexion-extension (visible as peaks 10 . CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprint this version posted March 24, 2022. ; https://doi.org/10.1101/2022.02.04.479126 doi: bioRxiv preprint on the trace). These peaks occur at different phases of the stride because the raw traces were aligned by the forelimb, and fore-hind phase difference varies across strides. This has consequences for averaging: conventional averaging "per time point" would flatten and widen the flexion-extension peak. Using FCAS, it is possible to get a more representative average of the raw traces which has its amplitude conserved, but phase and mean angle averaged. To re-iterate, this is enabled by temporal alignment and averaging in the frequency domain, followed by inverse transformation back to the time domain. Reversibility further bares a lot of herein unused potential, for example for interpolating unobserved subject states or for inferring kinematics of fossile species by phylogenetic and morphometric bracketing. Though we used a 2D dataset herein, the procedure could be applied equally well to angles measured from 3D coordinate data.

Statistical Modeling
To summarize, four categories of variables were used for analysis: • subject characteristics: age, sex, mass, birth weight category, size • spatiotemporal gait variables: distance, frequency, speed, clearance (fore-/hindlimb), duty factor (fore-/hindlimb), head angle, hindlimb phase • dynamic posture: mean joint angles and eROM for six joints • coordination: the residual after extraction of dynamic posture (see above) Our research question is whether a probabilistic, linear model is able to predict subject characteristics (specifically: age, mass, and size, optionally logarithmized to comply with normality assumption) from raw kinematics (expressed as dynamic posture and coordination) and gait variables (collective variables).
Given the common conception that kinematics are a complex output of an individual motor system, this might be considered an "inverse" modeling approach. The present analysis focused on three outcome variables ( Fig. 1): mass (kg), size (arb. units, from a PCA of marker distances), and age (h, optionally log-transformed). Though these outcome variables were specific per individual and recording session, we analyzed them "per stride" (i.e. there were multiple strides with identical subject measures on the outcome side).

11
. CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprint this version posted March 24, 2022. ; https://doi.org/10.1101/2022.02.04.479126 doi: bioRxiv preprint The model formula is: Herein, θ is either of the outcome subject characteristics, β are slopes associated with the model parameters (s sex, G gait variables, P dynamic posture, C coordination), v are data vectors (e.g. v 1 is a vector of ones for the intercept α, and v s is a boolean vector coding for subjects of 'sex == male'), and ϵ is the model residual. Priors (i.e. a priori assigned distributions) for all slopes were Normal distributions with mean and standard deviation corresponding to the mean and two times standard deviation of the observed values of each parameter; logarithmic transform was applied where necessary.
The observable ("likelihood") prior for θ was a Student's T distribution with a Gamma distributed ν (degrees of freedom; allows for wider-than-normal tails and robust regression), ϵ was modeled to be a  of predictions had to match the original sample size, and though repeated predictive sampling is possible, we found that number sufficient and would thus retrieve 294 mass, size, and age predictions for each stride in the data set.

13
. CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is

Results
The present analysis is centered around a linear model which is designed to infer mass, size, and age (subject characteristics) from an extensive set of kinematic parameters from 2D videos. The numbers provided by the model sampling are equally extensive, and will only be reported in brief. The key purpose of the model is posterior predictive sampling of the LBW strides which were left out of the model, and which are analyzed in detail below.
To assess whether there are qualitative differences between the birth weight categories, one can compare the joint angle profiles (i.e. raw, angular kinematics) on which the present analysis was performed (Fig.   2). The intra-group variablility clearly exceeds the differences between groups, although it must be 14 . CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprint this version posted March 24, 2022. ; https://doi.org/10.1101/2022.02.04.479126 doi: bioRxiv preprint emphasized that groups are inhomogeneous (with regard to age, speed, etc.), which might lead to a bias if composition of LBW and NBW data differs. LBW walk with a more extended hindlimb posture, as indicated by the parallelly offset average hip, knee, and ankle profiles. Additionally, NBW individuals on average seem to move the shoulder at a more flexed angle. No differences in coordination are apparent (which would manifest in altered temporal structure of the profiles). These findings indicate that LBW kinematics are hardly distinguishable from NBW kinematics by qualitative, visual assessment.
A quantitative comparison of variable kinematic measurements can be achieved with probabilistic linear models. For the purpose of predictive sampling (see below), we train models to describe the interrelations of kinematic parameters and subject characteristics in NBW piglets. The outcome of MCMC sampling of a linear model are value distributions for slopes, which in our case indicated how certain kinematic parameters are associated with a change in mass, size, and age (supplementary material 7.1). None of the gait-or coordination parameters were correlated with differences in animal mass. However, mass was associated with changes in the dynamic posture of the hindlimb (hip, knee and ankle). For size, the model inferred associations with head angle, hindlimb duty factor and phase, and some coordination components (CC3, CC7), as well as changes in the fore-and hindlimb posture and an effect of sex.
Finally, age was associated with a change in forelimb clearance and duty factor (both increase with age), potential changes at the hip, ankle, and wrist, and several coordination components (CC6, CC9). Some eROM slope distributions for age were high in average magnitude, but variable (the "credible interval" contained zero). These model results provide detailed insight into parameter interrelations in the present data set and indicate which of the parameters are the relevent ones to infer a given subject attribute in predictive sampling.
Performing in-sample and out-of-sample prediction with the models trained on NBW strides elucidated if and how left-out strides differed from NBW model expectation (Fig. 3). Note that, to capture variance (i.e. uncertainty in the prediction), each stride was predicted repeatedly.
Out-of-sample predictions for the NBW validation set matched those of in-sample NBW prediction in terms of average values and standard deviation for all modeled outcome variables, which confirms that prediction of subject characteristics from kinematics is possible. In contrast, predictions for LBW strides did not match those of the NBW training set. Low birth weight animals were predicted on average 0.39 kg heavier than actual, and their size was overestimated (+ 1.89 units). Both prediction faults matched the 15 . CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is  16 . CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprint this version posted March 24, 2022. ; https://doi.org/10.1101/2022.02.04.479126 doi: bioRxiv preprint actual differences in magnitude (cf. methods, Fig. 1). In contrast, the age predictions for the low birth weight subjects were bimodal: most ages were correctly inferred from stride-wise kinematics, but ages for some strides were underestimated. The underestimation of those strides quantified to less than five hours.
In summary, the NBW-trained model "guesses" the size and mass of the animals producing LBW strides to be "normal" (although they are not), which indicates that these defining features of LBW do not reflect in altered kinematics. However, age prediction is bimodal, i.e. some strides are classified as typical for animals of younger than actual age.
To find out whether the bimodality was related to certain individuals, or certain strides from different individuals, we grouped the predictions per stride or subject and calculated the chance of over-or underestimating age. Of the 8 low birth weight subjects who contributed 39 strides, 4 individuals were consistently underestimated (Tab. 1). Consistently means that more than 75% of all predictive samples were below actual age, and that the ages for a majority of strides were on average underestimated. The magnitude of underestimation was between two and five hours. Curiously, those were the individuals recorded at a slightly higher age (> 5 hours). Overestimation in the other four LBW individuals was also consistent, but less so (less extreme underestimation rate, mean ∆ < 2 h). Standard deviation of the estimates did not vary across individuals or birth weight categories.
We conclude that underestimation of age is consistent over multiple strides of the same individual, and thus individual-specific.

17
. CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprint this version posted March 24, 2022. ; 18 . CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprint this version posted March 24, 2022. ;https://doi.org/10.1101https://doi.org/10. /2022 Birth weight variability in piglets is considerable. The average birth weight of a new born piglet in our data set is just above a kilo, yet the span within a litter is typically above 800 g. Size ranges are accordingly high. Intuitively, one would expect a pronounced difference in how these newborn animals move their more or less heavy bodies. The overarching question is whether low or critically low birth weight can be associated with a deficit in locomotion.
Previous studies reported behavioral, physiological, and other differences for piglets of low body mass which could affect locomotion (Alvarenga et al., 2013;Muns et al., 2013;Quiniou et al., 2002;Roelofs et al., 2019;Vanden Hole et al., 2018a,b). One might therefore anticipate differences in coordination, which is the immediate outcome of motor control. However, we observe little qualitative difference in LBW and NBW kinematics (Fig. 2). There is a small difference when averaging hindlimb posture, however that might be trivially explained by inhomogeneous group composition (but see below). This is further supported by the fact that our models were unable to correctly retro-infer either mass or size from blindly provided kinematic parameters including coordination. Strides from LBW animals were estimated to come from an animal with a "normal" mass and size (Fig. 3). This confirms that low body mass or small size can not be causal for altered 2D kinematics, and it raises doubts whether there are any deficits in coordination and control. Within the range of kinematic variability, we observe LBW animals perfectly capable of normal posture, coordination, and overall stride results (collective variables). Note that this does not rule out problems of balance, stability, or endurance.
We do see differences for LBW compared to NBW recordings, nonetheless. First, a difference in sample size. There are much fewer valid LBW strides in our data set: only 39 of 368 observations are LBW.
This could be interpreted as evidence for a lower capacity (despite equal potential) of LBW to produce normal locomotion. Yet there are proximal, trivial explanations: based on conventions, the 10% lower quantile of birth weights in a litter is considered LBW, and there is a hard cap of 800 g. The resulting share is equal in our training set for video digitization, and in the final data set, because of pseudorandom, opportunistic sampling on-site (i.e. recording work was permanent, yet determined by farrowing and feeding of the subjects). The minority of LBW training videos might lead to an under-learning of those animals, reduced digitization quality and therefore an exclusion bias for "non-normal" individuals.
Though it seems unlikely, we cannot rule out reduced locomotor capacity in LBWs, because the present 19 . CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprint this version posted March 24, 2022. ; https://doi.org/10.1101/2022.02.04.479126 doi: bioRxiv preprint data set is unsuited to count the occurrence of locomotor behavior. On the other hand, the strict stride filtering criteria for "good" kinematics may have involuntarily filtered out deficient individuals. Our conclusion that low birth weight individuals are non-deficient is strictly tied to the definition of the low birth weight category, which is herein based on weight criteria and did not regard phenotypical indicators of intra-uterine growth restriction (which we did not record, cf. Amdi et al., 2013).
A second difference of LBW locomotion is that the age is underestimated for strides of some, but not all individuals. Prediction is consistent per individual, although no subject characteristics except sex entered the model ("blind" inference). This supports the hypothesis that locomotor development is sometimes delayed in LBWs. Delayed development does not necessarily corroborate the hypothesis of locomotor deficiency in LBW: we would expect truly deficient strides to be substantially different from the data trained to the model, making it "unpredictable" (i.e. higher variance of posterior samples). Instead, the predictions are consistent for repeated measures of an individual, without notable increase in variance.
For the affected subjects, we can even quantify a delay of less than five hours, which could nevertheless be critical given the rapid maturation of locomotor behavior in this species (Vanden Hole et al., 2017) and the importance of postnatal competition (Litten et al., 2003).
Note however that the causality might be inverse. We measured age underestimation only in the case of the individuals which were recorded late within our sampling time frame (age > 5 h, Tab. 1). This is consistent with prior evidence that energy reserves are depleted after birth (normal locomotion) but not replenished after four hours in the case of LBW (Le Dividich et al., 2017;Vanden Hole et al., 2019).
Assuming that this is the case, i.e. energy reserves are depleted, we would expect two effects: (i) the animal might succeed in the locomotor task less frequently (not quantified, see above), and (ii) the kinematics might be altered, which we observed. The present model was trained from NBW data and thereby tuned to kinematic development from animals with normal energy levels (therefore it can infer age from kinematics in NBW). That same model quantifies the potentially energy-deficient animals as younger. It might be that energy deficiency coincidentally causes effects which are exactly opposite to the changes that piglet kinematics undergo in normal development. However, the more likely explanation is an actual delay or a temporary halt in development. Failure of the LBW to compete in the first hours are sufficient to explain reduced intake (Amdi et al., 2013), the absolute size and mass difference alone might be crucial, and an immediate question which we cannot address with the present data set is whether and 20 . CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprint this version posted March 24, 2022. ;https://doi.org/10.1101https://doi.org/10. /2022 how (fast, likely) animals would recover from the delay. Alternatively, there might be a technical artifact in probabilistic sampling ("shrinkage", cf. Gelman et al., 2020) which caused the underestimation of "above average age" individuals. Yet this is an unlikely explanation, given that shrinkage would apply equally to NBWs, and inversely to the opposite, younger subjects. Finally, with only eight LBW individuals, we cannot exclude coincidence in which indivuduals are affected.
Neither of these technical explanations puts doubt on the clarity of the initial finding: a subset of the low birth weight individuals produced locomotor behavior which is quantitatively similar to that of younger individuals.
A corollary question is which patterns in the kinematic variables cause the different age prediction. We report high magnitude (but also highly variable, i.e. "non-significant") slopes inferred from the age model (supplementary material 7.1). Note that these slopes solely reflect effects within the NBW data subset.
We also observed slight differences in the average hindlimb dynamic posture (Fig. 2). In fact, a more extended hindlimb is typical for the youngest animals of both birth weight categories. We emphasized potential differences in group composition to explain that, and different age per group might be a proximal explanation. However, the average age of LBW animals ( 5.3 h ) in our data set is nominally above that of NBW ( 3.8 h ), which is a discrepancy with the age underestimation. Yet if we assume that the hypothesis of delayed locomotor development is correct, the nominal age would be misleading, and LBW effectively behave similar to younger animals. This can explain the apparent discrepancy in age group composition and age predictions from kinematics. It also suggests that dynamic posture might be the major proxy for perinatal maturation, though many other parameters also entered the probabilistic model and influenced the model outcome.
The findings discussed above are enabled by inverse-modeling subject characteristics as a function of kinematic parameters, using probabilistic models. The models do reveal various parameter associations, 3). Our probabilistic models implicitly regard "non-significant" parameter slopes, which are usually neglected in standard hypothesis testing (such as the high magnitude, highly variable joint ROM quantities, which might nevertheless have relevance for age prediction in case they are correlated and add up). The data transformations and predictive modeling strategy we applied herein hold further potential for inferring kinematics, for example by morphometric bracketing of extinct taxa. For that purpose, one would train a model to infer dynamic posture and coordination from a given range of morphometrics, predict samples for specific morphometrics in that range, and convert the samples back to (3D) joint angle profiles which could be animated.
There is a conceptual hierarchy, but no clear causality, when modeling parameter dependencies in quantitative studies of locomotion. For example, animals might increase the hip angle (posture) and the temporal pattern in the joint angle profiles (coordination) to reach higher speeds (spatiotemporal gait variables), but the speed they reach might depend on age (subject parameter). But age might also influence the speed without changing dynamic posture or coordination, simply because the animal grows and increases strength. Age might also affect dynamic posture and coordination directly, for example if dimensionless speed stays constant but clearance changes with age. Reducing clearance can increase speed (less unnecessary hoof lifting) or reduce it (higher duty factor), which can be distinguished by the other parameters. This complex interrelation of spatiotemporal quantities complicates the intuitive modeling strategy, which involves using subject characteristics as a factor to predict a multivariate block of gait variables and kinematics. In our experience, model residuals in such models are high, multiple testing can yield putatively coincidental significances, cross-dependencies within the multivariate data set might be underestimated, and sample size requirements are high. And even under ideal circumstances: chances are that such models would yield some age effect, even with random data.
In contrast, the strategy applied herein is related to the question: "given the complete kinematic output of a behavior, can we infer subject characteristics of the animal producing it?" We used probabilistic models, which are able to capture intrinsic variability of the process, and addressed specific categorical questions (NBW/LBW differences) by out-of-sample predictions. We demonstrated that, in the first ten hours of piglet life, (1) kinematics seem to be indifferent to low body mass and size, and (2) locomotion of some LBW individuals could be explained by a delay in locomotor maturation which is initiated post partum.

22
. CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprint this version posted March 24, 2022. ;https://doi.org/10.1101https://doi.org/10. /2022 as well as Laura Buyssens, Georgios Petrellis, Gunther Vrolix, Charlotte Vanden Hole, Denise Vogel and all students who joined for help during recordings. Maja Mielke provided valuable comments on the manuscript text.

23
. CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprint this version posted March 24, 2022. ; https://doi.org/10.1101/2022.02.04.479126 doi: bioRxiv preprint

Supplements: Detailed Modeling Results
Asterisk (*) indicates slopes for which the credible interval did not include zero. FL: forelimb, HL: hindlimb, diml.: dimensionless, d.s.: dimensionless stride, eROM: effective range of motion 24 . CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is 25 . CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprint this version posted March 24, 2022. ; https://doi.org/10.1101/2022.02.04.479126 doi: bioRxiv preprint . CC-BY 4.0 International license made available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprint this version posted March 24, 2022. ;https://doi.org/10.1101https://doi.org/10. /2022