{"id":"607c1303-dbe4-4b5b-ab82-545a7f88fb0d","arxiv_id":"2607.05585","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Hierarchical PointNet models on MediaPipe 3D face meshes classify 107 HPO facial terms (mean AUROC 0.75) with cascading landmark elimination, performing better at parent nodes than rare leaves and transferring unevenly to unseen disorders.","lead":"FaceMesh2HPO turns 2D patient photos into 3D face meshes and classifies them into Human Phenotype Ontology facial terms with a hierarchy of PointNet models that prune landmarks level by level. The approach aims to give clinicians structured, interpretable phenotype labels instead of black-box syndrome scores.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged MediaPipe/negative-class limits.","rationale":"The reader's weakest_assumption correctly isolates the two softest empirical foundations (mesh fidelity and negative-class construction). Both are already quantified or acknowledged by the authors, and the performance claims are carefully stratified by ontology depth and sample size. No stronger load-bearing flaw (e.g., data leakage across the stratified CV, unaccounted label noise that would reverse the parent/leaf gap, or an ablation that actually fails) is present. Therefore the CONDITIONAL verdict—pending public code and broader multi-ancestry validation—remains the appropriate stance; no adjustment is warranted.","tokens_in":55844,"tokens_out":483,"duration_ms":4727,"concrete_test":"Once the promised GitHub release is public, re-run the single best configuration (3D + outline + metadata, T=0.01, soft-label 0.05) on the independent test set after replacing MediaPipe meshes with a higher-fidelity 3D reconstruction (e.g., DECA or a clinical 3D scanner subset) for the same 75 images; if mean parent-node AUROC drops >0.05 or leaf-node ranking changes materially, the geometric-representation claim weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical and modest: hierarchical PointNet models on MediaPipe 478-point meshes plus cascading feature elimination produce clinically usable HPO-term AUROCs (mean 0.750, parents routinely >0.85, many leaves near chance). The paper's own ablations (Table 3), per-node results (Table 4 / C3), sample-size correlations (Fig. 5), and external multi-expert test (Fig. 6, Tables D4–D22) already document exactly those performance bounds and the heterogeneous generalizability. The two assumptions the reader flags—MediaPipe fidelity on dysmorphic faces (explicitly citing Hellmann et al. [34]) and soft-labeled UTKFace negatives—are real and correctly identified, yet they are already treated as limitations in the Discussion rather than hidden premises. No additional internal inconsistency, missing control, or over-claim appears that would further undermine the reported numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"FaceMesh2HPO frames facial dysmorphism analysis as hierarchical HPO-term classification rather than syndrome-level prediction. From 2D clinical photographs the authors extract MediaPipe 478-point approximate-3D meshes, combine them with clinician-curated positive/negative HPO labels for 10 disorders (107 terms after parent expansion) plus soft-labeled UTKFace controls, and train a cascade of dynamically sized PointNets organized along a simplified HPO tree. After each node is trained, Integrated Gradients point importance is thresholded and the surviving mask is passed to child nodes (cascading feature elimination); optional expert region masks override the cascade for selected anatomical terms. A 72-configuration ablation selects 3D meshes + facial outline + age/sex/ethnicity metadata, threshold 0.01 and soft-label 0.05, yielding mean AUROC 0.750 ± 0.042; parent/compression nodes routinely exceed 0.85 while many rare leaves remain near chance. An independent multi-expert test set of seen and unseen disorders shows heterogeneous F1 transfer, stronger at coarser ontology levels. A local web tool is released for clinical use.","tokens_in":56160,"tokens_out":1304,"duration_ms":10619,"significance":"The work supplies a clinically aligned alternative to black-box syndrome classifiers: ontology-linked phenotype scores that can feed tools such as LIRICAL and remain informative for unseen disorders. Strengths include a large multi-clinician annotation effort (124 annotators), systematic ablation, stratified 5-fold CV that respects patient identity, explicit parent-versus-leaf and seen-versus-unseen reporting, and a privacy-preserving local web application. If the hierarchical mesh approach continues to improve with better landmark detectors and more diverse data, it could become a practical component of next-generation phenotyping pipelines. The contribution is incremental rather than transformative, but the empirical documentation of performance bounds is solid and useful for the community.","major_comments":[{"comment":"§5.2 / Table 4 / Fig. 5: leaf-term sample sizes frequently fall below 200 (e.g., Microphthalmia N=66, Blepharophimosis N=152) and the paper itself shows that AUROC ≥ 0.75 for leaves requires ≥ 988 samples. The central claim of “clinically meaningful classification” therefore holds mainly for parent/compression nodes; the manuscript should either restrict the primary claim to those levels or supply a quantitative decision rule (minimum N / minimum AUROC) for when a leaf model is considered usable.","section":"§5.2 Results / Table 4 / Fig. 5"},{"comment":"§6 Discussion and the authors’ own citation [34]: MediaPipe meshes are known to under-represent traits such as Big Nose and Big Lips on dysmorphic faces. Because the entire cascade depends on these 478 points, the Discussion should quantify how often landmark failure occurs on the GMDB-HPO cohort (or at least report a failure-rate proxy) and discuss whether expert region masks can fully compensate. Without that evidence the geometric premise remains only partially validated.","section":"§6 Discussion / citation [34]"},{"comment":"§4.4 / Table 4 detection-prevalence column: training forces a 50/50 class balance while detection prevalence ranges from ~0.23 to ~0.74. Soft labels of 0.05 mitigate but do not eliminate the resulting miscalibration. For clinical deployment the authors should report calibration metrics (Brier, ECE) per node on the external test set and, if necessary, apply the same beta/temperature calibration already used in the web tool to the offline evaluation tables.","section":"§4.4 Training Strategy / Table 4"}],"minor_comments":[{"comment":"Table 1 and §3.1: two HPO terms (HP:0410030, HP:0004493) were dropped for insufficient data; state the exact sample-size cutoff used.","section":"§3.1 / Table 1"},{"comment":"Fig. 2 caption and §4: clarify that the displayed point-importance colors are Integrated Gradients aggregated across folds, not single-fold attributions.","section":"Fig. 2 / §4"},{"comment":"Appendix Table A1: several IRA values are reported as 1.00 ± 0.00; note whether these terms were annotated by only one rater or truly unanimous.","section":"Appendix Table A1"},{"comment":"§4.3: the dynamic width schedule (p = 6/5/4/3) is sensible but the zero-padding rule for <2 points should be stated more explicitly to aid re-implementation.","section":"§4.3 Model-Architecture"},{"comment":"Minor typos: “norphology” (Table C3), “unseed” (Fig. 6 caption), and inconsistent spelling of “softlabel/soft label”.","section":"Tables / Fig. 6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid empirical systems paper that fits a methods-oriented clinical-genetics or medical-imaging venue. The MediaPipe and negative-class limitations are already acknowledged by the authors; the requested revisions are mainly about tightening the claim language and adding calibration/landmark-failure diagnostics rather than redesigning the method. Code and web-tool release plans are a plus. I see no novelty or citation-pattern concerns that would affect the editorial decision."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean engineering paper that actually ships something usable: a tree of PointNets on 478-point MediaPipe meshes, with cascading feature elimination down the HPO hierarchy, trained on a new clinician-curated GMDB-HPO set (124 annotators, explicit absences, 107 terms). That combination is not in GestaltMatcher, PhenoScore, or the 3D morphometrics literature they cite. The ablation (72 configs), 5-fold stratified CV, and multi-expert external set are thorough; the numbers match the claims (mean AUROC 0.75, parents routinely >0.85, many leaves near chance, heterogeneous transfer to unseen disorders).\n\nWhat they do well is honesty about the hierarchy. Parent and compression nodes benefit from larger effective N and coarser geometry; leaves with N often <200 and sparse mesh support stay weak. Detection prevalence drifts far from the forced 0.5 training balance, which they report rather than hide. The local web tool with per-term importance maps and calibration is a practical plus for the intended users.\n\nSoft spots are real but already flagged by the authors. MediaPipe on dysmorphic faces is imperfect (they cite their own prior work on Big Nose / Big Lips failures). Soft-labeled UTKFace negatives and incomplete presence-only annotations remain noisy. Data are still limited in ancestry and rare-term support. None of these invalidate the central empirical result; they bound it. Code is promised on publication; controlled-access data is the usual GMDB constraint.\n\nThis is for people building next-generation phenotyping pipelines or clinical decision support that wants HPO-level outputs rather than black-box syndrome scores. It is not a foundational methods paper, but it is a serious, reproducible step that a referee should see. I would accept it for peer review and would cite the benchmark and the cascading-elimination idea if I were working in the same space.","headline":"Solid hierarchical HPO phenotyper on MediaPipe meshes; parent nodes work, rare leaves do not, and the authors already document both.","tokens_in":56727,"tokens_out":470,"would_cite":true,"duration_ms":6903,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Cascading 3D face-mesh models can classify clinical facial traits along the Human Phenotype Ontology rather than naming syndromes.","keywords":["Face Mesh","Feature Elimination","Deep Learning","Human Phenotype Ontology","Syndromes","PointNet","Facial Phenotyping","Hierarchical Classification"],"falsifier":"Train and evaluate the same cascade on true 3D facial scans of the same patients and check whether leaf-term AUROCs rise substantially above chance and whether detection prevalence no longer drifts so far from the balanced training prior.","tokens_in":56790,"feed_emoji":"🧬","tokens_out":818,"duration_ms":7879,"temperature":0.7,"pith_summary":"Clinicians diagnose rare genetic conditions by recognizing structured facial traits, not by matching whole syndromes at once. This paper argues that the same reasoning can be automated by turning 2D clinical photographs into 3D face meshes and training a cascade of small geometric classifiers, one for each Human Phenotype Ontology term, that pass pruned landmark masks down the ontology tree. The resulting labels stay in clinical language, remain informative even when the underlying disorder is unseen, and can be inspected by region. Performance is strongest at parent and intermediate terms and weaker for rare, sparsely annotated leaves, so the method is presented as a phenotypic scaffold rather than a universal detector of every fine-grained trait.","feed_headline":"3D face meshes classify clinical traits down the HPO tree","feed_subtitle":"Cascading PointNets prune landmarks term by term, yielding ontology-linked phenotype labels instead of black-box syndrome names","key_machinery":"Cascading feature elimination along the HPO tree: each PointNet classifier is trained on its subtree, Integrated Gradients ranks mesh points, a threshold prunes the mask, and the reduced mask is handed to child terms so that deeper models see only the landmarks that mattered higher up.","core_discovery":"Geometric 3D face meshes combined with a hierarchical PointNet cascade that eliminates unimportant landmarks term by term can produce clinically meaningful, ontology-aligned facial phenotype predictions, with mean AUROC around 0.75 and systematically higher accuracy near the root than at rare leaves.","pith_inferences":["If mesh detectors systematically miss dysmorphic landmarks, the cascade will keep under-performing on the very traits clinicians care about most; refining the front-end mesh extractor may matter more than further architecture tweaks.","Soft labels and 50/50 rebalancing mask the true prevalence of rare traits, so any clinical deployment will need explicit calibration and prevalence-aware decision thresholds before the reported F1 numbers can be trusted at the bedside.","The large performance gap between parent and leaf terms suggests the method may be most valuable as a pre-filter that proposes intermediate HPO terms for human review rather than as an end-to-end leaf-term annotator."],"forward_implications":["Diagnostic tools can output reusable HPO trait lists instead of a single syndrome name, supporting differential diagnosis and tools that already consume HPO terms.","Parent and compression nodes already give usable signal and transfer better to unseen disorders than leaf models, so coarser ontology levels are the practical entry point for clinical use.","Point-importance maps and expert region masks make the decision regions inspectable, giving a path to clinician-auditable geometric explanations.","The same cascade can be extended as more curated HPO annotations become available without redesigning the architecture."],"fun_headline_variants":["Cascading PointNets prune 3D mesh landmarks for HPO phenotypes","Hierarchical elimination maps face meshes to ontology traits","3D meshes plus PointNet cascade yield root-stronger HPO labels","FaceMesh2HPO prunes landmarks term-by-term down the HPO tree","Geometry-driven cascade classifies interpretable HPO face traits"],"cache_read_input_tokens":49280,"weakest_assumption_plain":"That 478-point meshes extracted from ordinary 2D clinical photographs, together with incomplete clinician annotations and soft-labeled control faces, still carry enough reliable geometric signal to train the hierarchy.","fun_headline_variants_meta":{"raw":{"variants":["Cascading PointNets prune 3D mesh landmarks for HPO phenotypes","Hierarchical elimination maps face meshes to ontology traits","3D meshes plus PointNet cascade yield root-stronger HPO labels","FaceMesh2HPO prunes landmarks term-by-term down the HPO tree","Geometry-driven cascade classifies interpretable HPO face traits"]},"model":"grok-4.5","effort":"low","cost_usd":0.004004,"raw_usage":{"total_tokens":1198,"prompt_tokens":700,"num_sources_used":0,"completion_tokens":77,"cost_in_usd_ticks":40040000,"prompt_tokens_details":{"text_tokens":700,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":421,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":700,"tokens_out":77,"duration_ms":4360,"temperature":1.0,"reasoning_tokens":421,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T05:24:19.152440+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train and evaluate the same cascade on true 3D facial scans of the same patients and check whether leaf-term AUROCs rise substantially above chance and whether detection prevalence no longer drifts so far from the balanced training prior.","supporting_citations":[],"review_version":1}