--- title: 'UDF-GMA: Dual Perspectives in Clinical and Geometric Domains' url: https://www.emergentmind.com/topics/udf-gma type: topic --- # UDF-GMA: Dual Perspectives in Clinical and Geometric Domains UDF-GMA is an acronym used in two distinct ways in the supplied arXiv literature. In "UDF-GMA: Uncertainty Disentanglement and Fusion for General Movement Assessment" it denotes a pose-based automated GMA framework that explicitly models epistemic uncertainty in model parameters and aleatoric uncertainty from data noise, and then fuses these uncertainties with an embedded motion representation to predict poor repertoire from infant pose sequences [2507.04814]. In the detailed presentation associated with "OffsetAxis: UDF Mesh Reconstruction via Offset-Volume Medial Axis Extraction," the same acronym is used for an "Unsigned Distance Field – Generalized Medial Axis" formulation that restates UDF $0$-level set extraction as medial-axis extraction of an $\alpha$-offset volume [2605.15369]. The shared acronym therefore spans Bayesian deep learning for clinical movement analysis and geometry processing for mesh reconstruction. ## 1. Terminological scope and disambiguation The duplication of the acronym is a substantive feature of the record rather than a typographic coincidence. One usage is clinical and probabilistic; the other is geometric and topological. | Expansion | Domain | Core task | |---|---|---| | Uncertainty Disentanglement and Fusion for General Movement Assessment | Pose-based automated GMA | Predicting poor repertoire | | Unsigned Distance Field – Generalized Medial Axis | UDF mesh reconstruction | $0$-level set extraction via medial axis of the $\alpha$-offset volume | This suggests that unambiguous citation requires the full title or arXiv identifier. Without that disambiguation, the acronym alone is insufficiently specific for scholarly reference [2507.04814], [2605.15369]. ## 2. UDF-GMA for General Movement Assessment: clinical setting and uncertainty model General Movement Assessment is an established clinical tool for evaluating the neurological integrity of infants by qualitatively inspecting their spontaneous whole-body movements—arms, legs, trunk, and neck—at preterm, writhing ($0$–$2$ months corrected age), and fidgety ($2$–$5$ months corrected age) stages. Early "Poor Repertoire" movements in preterm and writhing periods are known predictors of later neurodevelopmental impairment, including cerebral palsy. Automating GMA with video- or pose-based deep learning is motivated by the scarcity of expert assessors, but the approach is hampered by two fundamental forms of uncertainty: aleatoric uncertainty arising from noise in 2D/3D pose estimation, and epistemic uncertainty due to limited training data and lack of knowledge about model parameters [2507.04814]. The probabilistic formalization is cast in Bayesian neural network terms. Network weights $\omega$ are treated as random variables with posterior $p(\omega \mid \mathcal{D})$ for training data $\mathcal{D}=\{(x_i,y_i)\}$. For a test pose sequence $x^\*$, the predictive distribution is $$ p(y^\*=1 \mid x^\*,\mathcal{D}) = \int p(y^\*=1 \mid x^\*,\omega)\,p(\omega \mid \mathcal{D})\,d\omega . $$ The total predictive variance is decomposed into an epistemic term, $\mathrm{Var}_{\omega \sim p(\omega \mid \mathcal{D})}[E[y^\* \mid x^\*,\omega]]$, and an aleatoric term, $E_{\omega \sim p(\omega \mid \mathcal{D})}[\mathrm{Var}[y^\* \mid x^\*,\omega]]$. Because exact Bayesian inference is intractable for deep nets, the method employs MC Dropout as an approximate posterior $q_t(\omega)$. With dropout-thinned weights $\omega_1,\ldots,\omega_T$, logits $z_t=f_{\omega_t}(x^\*)$, and probabilities $\hat{Y}_t=S(z_t):=1/(1+e^{-z_t})$, the predictive mean is $\bar{p}=(1/T)\sum_{t=1}^{T}\hat{Y}_t$, and the epistemic uncertainty estimate is $\mathcal{U}_e=\mathrm{Var}(\{\hat{Y}_t\})=(1/T)\sum(\hat{Y}_t-\bar{p})^2$. A central claim of the framework is that only by modeling both uncertainties can an automated system provide clinically reliable predictions, attenuating overconfident errors on noisy or out-of-distribution inputs and flagging high-uncertainty cases for expert review. A common misconception is that uncertainty in automated GMA is a single scalar nuisance; the formulation explicitly rejects that view by separating data uncertainty from model uncertainty. ## 3. Architectural formulation and objective in the GMA variant The input is a 2D pose sequence $x$ with dimensions $(J \text{ joints} \times C \text{ coords} \times T \text{ frames})$. A compact motion embedding $h$ is first extracted via CTR-GCN, identified as a state-of-the-art spatio-temporal graph convolutional backbone. UDF-GMA then introduces an Uncertainty Disentanglement Module and an Uncertainty-Guided Fusion Module [2507.04814]. Within the Uncertainty Disentanglement Module, the branch $f_e(h)$ is a small MLP with Dropout that outputs a scalar logit $z'$. Applying MC Dropout at inference yields $\{z'_t\}$ and, via the sigmoid-based estimates above, the epistemic uncertainty $\mathcal{U}_e$. The branch $f_a(h)$ is an MLP without Dropout, ending in a Softplus that produces the aleatoric estimate $\mathcal{U}_a=\sigma^2$. The two uncertainty quantities are then combined through an MLP $f_\sigma$ to obtain total predictive variance $\sigma^2=f_\sigma(\mathcal{U}_e,\mathcal{U}_a)$. Aleatoric uncertainty is trained through a heteroscedastic formulation that assumes the final logit follows a Gaussian with data-dependent variance: $$ z \sim \mathcal{N}(\mu,\sigma^2), $$ where $$ \mu = E_{\omega \sim q}[f_\omega(x)] \approx (1/T)\sum f_{\omega_t}(x). $$ Using $N$ samples $z_i=\mu+\sigma \cdot \epsilon_i$ with $\epsilon_i \sim \mathcal{N}(0,1)$, the average probability is $$ p=(1/N)\sum_{i=1}^N S(z_i), $$ and the uncertainty loss is $$ \mathcal{L}_{unc} = -[\,y\log p + (1-y)\log(1-p)\,] + \lambda_0 (e^{\sigma^2}-1). $$ The penalty term is designed so that $\sigma^2$ becomes large only when the data are indeed noisy. The Uncertainty-Guided Fusion Module injects the disentangled uncertainties into the motion representation. Epistemic fusion is defined by $$ h_e^{(1)} = f_r^{e1}(h \parallel \mathcal{U}_e), \qquad h_e' = f_r^{e2}(h \parallel h_e^{(1)}), $$ while aleatoric fusion is $$ h_a = h \odot \mathcal{U}_a, \qquad h_a' = f_r^{a}(h \parallel h_a). $$ The final representation is $$ h_f = h_e' \parallel h_a', $$ which is passed to a classification head $f_c$ to produce $$ p_f = S(f_c(h_f)). $$ Training combines the standard video-level binary cross-entropy loss $$ \mathcal{L}_{cls} = -[\,y\log p_f + (1-y)\log(1-p_f)\,] $$ with the uncertainty regularizer: $$ \mathcal{L} = \mathcal{L}_{cls} + \lambda_1 \mathcal{L}_{unc}. $$ During inference, Dropout remains active only in $f_e$ to estimate $\mathcal{U}_e$; all other Dropout layers are disabled. ## 4. Experimental profile of the GMA variant The experimental benchmark is Pmi-GMA, consisting of 1120 thirty-second clips from 87 preterm infants, with 567 PR and 553 normal clips, captured at 10 FPS as 1280×720 overhead smartphone videos. Two evaluation protocols are reported. The intra-infant split uses 999 train, 111 val, and 121 test clips, with clips from the same infants appearing in all sets. The inter-infant split is stratified 65%/15%/20% by infant with no overlap. Five random seeds are used and mean±$\sigma$ is reported. Data augmentation, applied with 80% chance each sample, includes horizontal flip, time reversal, Gaussian noise, scaling, and magnitude/time warping. Evaluation metrics are Accuracy, Sensitivity, Specificity, and AUC-ROC. Baselines include FM-based methods—CA, STAM, WO-GMA, and smartphone-app—and WM-based methods—WMD with LDA, SVM, or RF, plus CTR-GCN representation [2507.04814]. On the inter-infant test, UDF-GMA reports ACC $69.65 \pm 1.39$, SN $81.73 \pm 8.55$, SP $56.61 \pm 9.27$, and AUC $75.72 \pm 1.95$. The next best method, WO-GMA, reports ACC $62.15 \pm 4.75$, SN $76.54 \pm 9.85$, SP $46.57 \pm 13.57$, and AUC $60.92 \pm 8.19$. The CTR-GCN baseline reports ACC $59.65 \pm 4.68$, SN $72.99 \pm 13.18$, SP $45.68 \pm 19.05$, and AUC $59.37 \pm 8.77$. On the intra-infant test, UDF-GMA reports ACC $93.56 \pm 1.10$, SN $94.83 \pm 2.44$, SP $92.38 \pm 1.56$, and AUC $98.35 \pm 0.54$—state-of-the-art. Compared to the CTR-GCN baseline, the method yields $+10.00\%$ ACC, $+8.74\%$ SN, $+10.93\%$ SP, and $+16.35\%$ AUC on inter-infant evaluation. Low run-to-run variance is reported as evidence of robustness of uncertainty-guided learning. t-SNE visualizations of $h$, $h$ refined by [Ji ICCV ’23], and $h_f$ show dramatically improved class separation only with UDF-fusion. Epistemic-threshold analysis further shows that by discarding test clips with $\mathcal{U}_e$ above a threshold $\tau$, one can trade off data-retention versus sensitivity; at $\tau=0.4$, SN reaches $98.9\%$ while retaining approximately $70\%$ of PR clips, enabling a high-sensitivity triage set for fully automated diagnosis while referring high-uncertainty cases to clinicians. The limitations are explicit: reliance on 2D pose quality, no direct use of pose-estimation confidences, a single-center dataset of preterm/writhing GMs, and the computational cost of MC sampling. Future directions include validation on fidgety and longitudinal GMA datasets, incorporating keypoint confidence maps, investigating lighter uncertainty approximations such as Flipout and SWAG, and deploying a clinical decision-support tool that uses a $\mathcal{U}_e$ threshold to automatically accept high-confidence predictions while routing uncertain cases to expert review. ## 5. UDF-GMA as "Unsigned Distance Field – Generalized Medial Axis" In the OffsetAxis presentation, UDF-GMA refers to a reconstruction formulation for unsigned distance fields rather than a clinical assessment model. The foundational object is a UDF $$ \phi:\mathbb{R}^3 \to \mathbb{R}_+ $$ defined by $$ \phi(x)=\min_{p \in S}\|x-p\|, $$ where $S \subset \mathbb{R}^3$ is the target shape. By definition, $\phi(x)\ge 0$ everywhere and $$ S=\{x \mid \phi(x)=0\}. $$ Unlike signed distance fields, UDFs make no inside/outside choice and can represent open or non-manifold geometries [2605.15369]. Given a scale $\alpha>0$, the $\alpha$-offset volume is $$ \Omega_\alpha=\{x \in \mathbb{R}^3 \mid \phi(x)\le \alpha\} = \bigcup_{p \in S} B_\alpha(p), $$ with boundary $$ \partial \Omega_\alpha = \{x \mid \phi(x)=\alpha\}. $$ If $\alpha$ is smaller than the shape’s weak feature size, then $\Omega_\alpha$ is homotopy-equivalent to $S$. The medial axis of a volume, $$ MA(\Omega)=\{c \in \Omega \mid \exists r>0 \text{ such that } B_r(c)\subseteq \Omega \text{ and } \partial B_r(c) \text{ touches } \partial \Omega \text{ in } \ge 2 \text{ distinct points}\}, $$ becomes, for $\Omega=\Omega_\alpha$, a thickened skeleton of the original shape. For small $\alpha$, $MA(\Omega_\alpha)$ recovers $S$ up to homotopy and captures curves, junctions, and surface patches in one unified object. The central reformulation is that the $0$-level set extraction problem can be restated as the extraction of the medial axis of the $\alpha$-offset volume of the UDF. This avoids sign-propagation heuristics and addresses a common misconception that UDF extraction can simply reuse classical grid-based iso-surfacing. The OffsetAxis account states that classical grid-based iso-surfacing techniques are not applicable because they require a way to distinguish the inside from the outside of the shape. ## 6. Variational pipeline, reconstruction, and empirical behavior in the geometry variant The geometry-oriented UDF-GMA approximates $MA(\Omega_\alpha)$ by a finite set of spheres $(c_i,r_i)$ and optimizes them against oriented offset-surface samples. For each sphere $m_i=(c_i,r_i)$ and its assigned cluster $C_i$ of samples $v_j=(x_j,n_j)$, the per-sphere energy is $$ E_i = \sum_{v_j \in C_i}\left[ Q_{v_j}(c_i,r_i) + \mu \cdot L_{v_j}(c_i) \right], $$ where $$ Q_{v_j}(c,r) = [n_j^\top (x_j-c)-r]^2 $$ is the Spherical Quadric Error Metric, and $$ L_{v_j}(c)=\|(I-n_j n_j^\top)(c-x_j)\|^2 $$ is the line-quadric. Because both terms are quadratic in $[c;r]$, minimizing $E_i$ reduces to solving normal equations $A_i s_i=b_i$ in closed form for $s_i=[c_i;r_i]$. To avoid degenerate or negative radii, the method enforces $0