Hyperbolic Radius as a Triage Signal for Infant Movement Screening

Only trained clinicians can screen infants for cerebral palsy with the General Movement Assessment, which means scoring the quality of a baby's spontaneous movements from video. There are too few of these clinicians, and every infant's video must wait for one. Automated classifiers help only partly. On a preregistered held-out set, the best open pipeline reaches a ROC-AUC of 0.77 but a precision-recall AUC of just 0.41, so no clinic would let it diagnose on its own. A more realistic goal is triage: let the model confidently clear the infants whose videos clearly show typical movement, and send everyone else to a clinician. The clinician's workload then shrinks by the fraction the model clears, provided it almost never clears an infant who is actually at risk. This proposal asks whether hyperbolic embeddings can supply the score that decides who gets cleared. A hyperbolic embedding places data in a curved space where distance from the center has a natural meaning.

The hook comes from self-supervised skeleton learning. In hyperbolic space, the model's training updates grow with an embedding's distance from the origin. As a result, the learned radius behaves like a certainty score that comes out of a single forward pass, with no ensembles and no repeated dropout sampling. The complication is that this radius tracks the wrong thing for infants. Adult work found that large, gross motions get a high radius and small, subtle motions get pushed toward the center. Yet the clinical signal in infant screening is the subtle, low-amplitude movement of the wrists and ankles. We propose to show this inversion empirically, correct it, and then test whether the corrected radius beats standard uncertainty methods at clearing infants safely. A matched Euclidean control will tell us whether the curvature itself matters.

Core Claim

A hyperbolic skeleton encoder has two pieces. The first is an input stage that removes gross amplitude and keeps a learnable frequency band. The second is a light supervised term that ties radius to how clinically unambiguous a clip is. With both in place, the encoder yields a radius that ranks infant clips by how safely they can be cleared. Turning that radius into a clear-or-refer decision with a statistically controlled miss rate should clear more normal infants, at the same sensitivity, than Monte Carlo dropout uncertainty, the approach of UDF-GMA. The new element is not "hyperbolic GMA classification", which would be a modest reuse of HyLiFormer. The new element is treating radius as a deferral score, diagnosing and fixing its amplitude bias on clinical data, and pitting it against a Euclidean feature-norm control. If that control matches the hyperbolic radius, we will say so and drop the geometric claim.

Prior Work and Positioning

The method comes from HYSP. It projects the output of a spatio-temporal graph network over the skeleton into the Poincaré ball, which is a model of hyperbolic space as the open unit ball. It shows that training updates scale with radius, so radius acts as a certainty proxy that yields a self-paced curriculum, and it reports nearly 10% over the Euclidean SkeletonCLR baseline with 10% labels. HYSP also motivates the proposal through its failure mode. Its "uncertain" samples are subtle hand motions, which is exactly the class of motion that carries the diagnostic signal in infants. The authors also note that radius is not a calibrated probability and that the model overfits late in training on small data. Both points constrain our design: we add explicit calibration, and we use early stopping chosen on patient-disjoint validation. HyLiFormer supplies an alternative hyperbolic backbone. It maps motion magnitudes through a tanh so that global features sit near the origin and limb features near the boundary. That amplitude-to-radius mapping is the opposite of what infant screening needs, and HyLiFormer does not model uncertainty at all. The Lorentz model of Nickel and Kiela supplies a numerically stable parameterization for use if boundary instability appears. Their WordNet results, and those of Poincaré Embeddings, are evidence about symbolic trees, not infant motion, so we do not lean on them.

On the clinical side, UDF-GMA is the closest competitor. It estimates two kinds of uncertainty on top of a Euclidean graph network: model uncertainty via dropout, and data noise via loss attenuation. On the patient-disjoint Pmi-GMA protocol it reports 69.65% accuracy and 81.73% sensitivity, rising to nearly 99% sensitivity only after excluding the 60% most uncertain samples. That is exactly the coverage-versus-sensitivity trade-off we target. FAIGCN supplies the input-stage idea. It separates low-frequency physiological movement from high-frequency pose-tracking jitter in the Fourier domain. We reuse that separation so that amplitude and jitter cannot drive the radius. The preregistered open pipeline and its longitudinal follow-up set the tabular reference. They use ViTPose-H keypoints, 38 kinematic features and AutoML. They report ROC-AUC of 0.77 and PR-AUC of 0.41 on the lock-box set, and ROC-AUC of 0.74 for predicting a two-year CP diagnosis. Modeling 3D Infant Kinetics provides an open 3D cohort, and its "kinetic age gap" gives a second, continuous target to check that the radius is not just re-encoding age. The research file also proposes hyperbolic entailment cones linking GMA scores to CP outcomes, but we set that aside. With 143 confirmed CP cases in the largest cohort, a partial-order embedding over three diagnostic levels is not identifiable. Hyperbolic Entailment Cones were validated on hundreds of thousands of edges.

Research Program

The work proceeds in three stages, each with a result that stands on its own.

The first stage is diagnostic. We apply HYSP unchanged to unlabeled infant skeleton clips, pooling the 3D Infant Kinetics segments and the Pmi-GMA segments, with patient-disjoint splits held out. We then measure how learned radius correlates with two things: per-clip motion energy at the distal joints, and the clinical label. The prediction taken from HYSP's adult findings is that radius correlates positively with gross motion energy and carries little label information. Normal fidgety-rich clips would be pushed inward. If instead radius already tracks the label, the correction in stage two is unnecessary, and the proposal simplifies to a deferral study.

The second stage builds the corrected encoder. The input to a standard graph backbone is replaced by per-joint trajectories filtered through a learnable frequency mask initialized from FAIGCN's split. These trajectories are normalized by per-clip amplitude, so that "how much the baby moved" is removed and "how the baby moved" remains. After self-supervised pretraining, a small supervised term is added. With $r(x)$ the hyperbolic distance of clip embedding $x$ from the origin, and $a(x) \in [0,1]$ a target for how unambiguous the clip is, the term is

$$\mathcal{L}_{rad} = \big(\tanh(r(x)/\tau) - a(x)\big)^2 .$$

Here $\tau$ is a scale hyperparameter. The target $a(x)$ is set to 1 for clips from infants whose label is consistent across all of that infant's segments, and it decays with segment-level disagreement within an infant. That disagreement serves as a stand-in for rater ambiguity, because we have no multi-rater labels. Classification is done by hyperbolic distance to learned class prototypes.

The third stage converts radius into a decision. For each infant we pool segment radii and the predicted class. Using conformal risk control on a calibration split, we choose the threshold above which an infant predicted "normal" is auto-cleared, so that the expected miss rate among auto-cleared infants stays below a target $\alpha$. We report the fraction of all infants cleared at that $\alpha$. This calibration step answers HYSP's own caveat that radius is not a probability. We do not need radius to be a probability, only to rank infants well.

Baselines

The naive reference is to refer every infant, which gives zero workload reduction and perfect sensitivity. Alongside it we report a prevalence-only rule, which shows that at roughly 10–15% adverse prevalence, clearing anything blindly is unsafe. Any method must clear a positive fraction of infants at the target miss rate to be worth deploying.

The strongest existing competitor is UDF-GMA's approach: a Euclidean graph backbone with Monte Carlo dropout uncertainty and loss attenuation, where infants are deferred when their uncertainty is high. It is the strongest because it was built for this exact question on this exact benchmark. Its reported operating points on Pmi-GMA are 69.65% accuracy and 81.73% sensitivity overall, and nearly 99% sensitivity after excluding the 60% most uncertain samples. We will re-run it on our patient-disjoint splits with the same backbone depth, the same pose inputs, the same pretraining data, and the same hyperparameter search budget. Its uncertainty will pass through the identical conformal step. Beating it means clearing a larger fraction of infants at the same guaranteed miss rate, with the gap exceeding the spread across patient-level bootstrap resamples. As a single-forward-pass reference we also report softmax confidence and a five-member deep ensemble.

The tabular reference is the preregistered AutoML pipeline on 38 kinematic features. It reports a lock-box ROC-AUC of 0.77 and PR-AUC of 0.41, and we use its calibrated probability as the deferral score. It represents what a clinic could deploy today with interpretable features.

The key ablation isolates curvature. It is an identical encoder with the same frequency input, the same radial loss and the same prototype classifier, but embedded in Euclidean space with the feature norm playing the role of radius. Euclidean feature norms are known to correlate with confidence, so this control decides whether curvature contributes anything. Two further ablations remove the frequency and amplitude input stage, which tests the inversion fix, and remove $\mathcal{L}_{rad}$, which tests whether radius becomes informative unsupervised.

Why not just…?

Why not just use Monte Carlo dropout, as UDF-GMA already does? This is a real threat, and our results may simply match it. The evidence that it leaves room is UDF-GMA's own trade-off: near-perfect sensitivity required setting aside 60% of samples, which barely relieves the clinician. The dropout approach also needs many stochastic forward passes per clip. The head-to-head coverage-at-fixed-miss-rate comparison settles the question. If dropout matches the hyperbolic radius, the remaining argument is only compute cost, and we would report that as a modest result.

Why not just use the Euclidean embedding norm with the same losses? This is the most dangerous alternative, because every mechanism except curvature is shared. No paper in our file tests it on clinical data. HYSP's roughly 10% gain over SkeletonCLR compares whole training methods, not norm against radius. The matched Euclidean ablation is therefore decisive. If the two are within bootstrap noise on both datasets, we will conclude that the gain comes from the frequency input and the radial supervision, and we will recommend the simpler Euclidean version.

Why not just threshold the calibrated probability of the tabular AutoML model? It is interpretable and already preregistered. But its PR-AUC of 0.41 means that among top-ranked infants, adverse and normal cases are heavily mixed, which is precisely what limits safe clearing. We run it through the same conformal step, so the comparison is direct. We have no access to the CHOP lock-box data from the file, so on CHOP we can only cite its reported numbers rather than re-run it. On Pmi-GMA and 3D Infant Kinetics we recompute the same 38-feature family.

Proposed Experiments

The primary benchmark is Pmi-GMA: 1,120 segments from 87 preterm infants, labeled normal versus poor repertoire, with the patient-disjoint inter-partition protocol. It is the only benchmark with a reported uncertainty-deferral operating point. The second benchmark is the open 3D Infant Kinetics dataset: 1,843 segments from 173 infants, with a Major Neurological Impairment subgroup. For this cohort there is no reported triage or deferral result, and the published figures are age-prediction RMSE of 16.18 days and Pearson r of 0.61. On it we also check that the radius is not a proxy for chronological age by partialling out age and the kinetic-age gap. The primary metric is the fraction of infants auto-cleared at target miss rates of 5% and 2%, computed at the infant level with patient-level bootstrap intervals. Secondary metrics are ROC-AUC, PR-AUC, and the area under the coverage-versus-sensitivity curve.

Three results decide the proposal. The first is the stage-one correlation: unmodified HYSP radius should track distal motion energy more than label. If it does not, the "inversion" story is wrong and we say so. The second is the hyperbolic-versus-Euclidean-norm comparison at matched architecture and tuning budget. The third is the comparison with UDF-GMA-style dropout under identical conformal calibration. Because both cohorts are small and have few adverse infants, we preregister the splits, the target miss rates and the primary metric before touching the test partitions. We report every ablation whether or not it favors the hyperbolic model.