{"id":"3d694cbf-501b-4898-94e9-09f9bb362f84","arxiv_id":"1908.09195","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A two-stage variational autoencoder with per-latent-dimension linear regression predicts future glaucoma visual fields more accurately than a classical patient-level spatiotemporal model, especially with few baseline visits.","lead":"This paper applies a variational autoencoder, a deep generative model, to predict future eye damage patterns in thousands of glaucoma patients, and reports better prediction accuracy with fewer clinic visits than classical spatiotemporal statistics. A generalist might read it as a case study in using unsupervised deep learning to make large-scale longitudinal spatial predictions where classical models struggle.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The second-stage assumption that latent dimensions evolve as independent linear functions is asserted without diagnostics; if it fails, the decoded future fields—and the few-visit advantage—could be artifacts.","rationale":"The reader's CONDITIONAL verdict is appropriate. The manuscript has genuine strengths: a patient-level train/test split, a large prospective cohort, code on GitHub, and a concrete held-out prediction evaluation. The central claim, however, is a claim about extrapolation, and the extrapolation is performed by a component the paper does not validate. The first-stage i.i.d.-images assumption is acknowledged and affects the representation, but the second-stage linear/independent latent dynamics are asserted in one sentence (Section 3.5) and are the exact mechanism that produces future latent codes. Since the decoder is fixed after training, any bias in the extrapolated latent code propagates directly into the decoded visual field. A diagnostic on latent-space extrapolation error and a comparison with a richer temporal model would settle whether the reported few-visit advantage is a genuine property of the generative representation or an artifact of the auxiliary linear model. I therefore recommend no change to the reader's verdict: acceptance should remain conditional on such validation.","tokens_in":115929,"tokens_out":8220,"duration_ms":94780,"concrete_test":"On the held-out test patients, freeze the trained encoder and compute the actual latent codes for future visits; compare them to the latent codes obtained by the independent linear regressions from the first 3, 5, and 8 visits, reporting per-dimension RMSE and a joint latent-space distance. Then refit the second stage with a vector autoregressive or Gaussian-process latent model and recompute image-space MAE on the same test predictions. If latent extrapolation error is large but a more flexible temporal model does not reduce MAE, the linear/independent assumption is not the bottleneck; if it does reduce MAE, the current point estimates are systematically biased and the central claim requires qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the two-stage pipeline in Section 3.5: train a VAE under an i.i.d.-images approximation, fit independent linear regressions in the 8-D latent space to each patient's baseline visits, extrapolate, and decode. The paper asserts these regressions 'prove sufficient' but provides no residual diagnostics, no check for cross-dimension coupling, and no evaluation of latent extrapolation error on held-out patients. With T=3 baseline visits each latent regression has one residual degree of freedom, and a 5-step-ahead forecast amplifies slope error by the horizon, so the reported few-visit advantage could be an artifact of the smooth decoder rather than evidence that latent trajectories faithfully represent progression. The Discussion acknowledges the i.i.d.-images limitation but does not address misspecification of the second-stage temporal model. This is load-bearing because the final prediction is precisely the decoded image at the extrapolated latent point; if that latent point is biased, the image-space MAE advantage is not a trustworthy basis for the clinical claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage variational autoencoder (VAE) pipeline for spatiotemporal image data with a large number of subjects. Stage one trains a VAE on all images as independent samples, using an MMD regularizer, a deterministic encoder, and a decoder with constant variance, yielding an 8-dimensional latent representation of 12x12 visual fields. Stage two fits independent linear regressions over time for each latent dimension separately, extrapolates each patient's latent trajectory, and decodes the predicted latent vector to obtain a predicted visual field. The authors compare this approach with a classical spatiotemporal model (Rushworth et al., 2014) and pointwise regression in a simulation study and on 29,161 visual fields from 3,832 eyes, reporting lower mean absolute error (MAE) for the VAE, especially when few baseline visits are used to predict several future visits and in patients with glaucoma.","tokens_in":116174,"tokens_out":4996,"duration_ms":54869,"significance":"If the predictive gains hold up, the paper offers a computationally scalable and clinically relevant alternative for monitoring glaucoma progression, and the two-stage recipe is simple enough to be ported to other large imaging cohorts. The manuscript's strengths are the patient-level train/validation/test split, the large real dataset, the availability of R code, and the explicit focus on forecasting rather than reconstruction. The held-out case study is the most credible piece of evidence. However, the simulation component is weakened by a circular data-generating design, and the second-stage temporal model is asserted rather than validated; these gaps need to be closed before the central claim can be accepted.","major_comments":[{"comment":"The simulation's VAE data generating process uses the encoder and decoder trained on the Section 5 visual field data, and the VAE variants are then trained on data simulated from that same process. This makes the VAE's good simulation performance partly self-fulfilling and does not provide evidence about robustness to misspecification. The authors should either replace this DGP with one that is independent of the evaluated model, or explicitly add misspecified DGPs (e.g., nonlinear latent trajectories, cross-dimension coupling) and report whether the VAE advantage persists.","section":"Section 4, Simulation Study"},{"comment":"The claim that independent linear regressions per latent dimension 'proves sufficient' is not supported by any diagnostics. Because the final prediction is the decoded image at an extrapolated latent point, biased latent extrapolations would directly produce biased image predictions; with T=3 baseline visits each latent regression has one residual degree of freedom, and 5-step-ahead forecasts amplify slope error by the horizon. The authors should provide residual diagnostics for the latent regressions, check cross-dimension residual correlations, compare against an AR or multivariate alternative, and report latent-space prediction error on held-out patients. Without this, the few-visit advantage could be an artifact of the smoothing decoder rather than evidence of faithful latent dynamics.","section":"Section 3.5, A Two-Stage Approach"},{"comment":"The latent dimension is a key user-specified parameter, yet it is fixed at eight based on the authors' own prior work (Berchuck et al., 2019a) without sensitivity analysis. The paper should report whether the MAE comparisons and qualitative conclusions are stable across a range of latent dimensions (e.g., 4, 8, 16), or at least justify why a dimension optimized for a related estimation task transfers to the forecasting task studied here.","section":"Section 5, VAE in Glaucoma: A Clinical Impact"},{"comment":"The case study's central claim is based on visual inspection of MAE boxplots; no formal comparison is provided, such as paired differences with confidence intervals or multiplicity-adjusted tests across horizons. Given the correlated, repeated-measures structure of the data, the authors should quantify the differences and their uncertainty, and report the number of patients contributing to each cell in Figure 6.","section":"Section 5, Prediction results (Figure 6)"}],"minor_comments":[{"comment":"The Gaussian kernel is written as k(x,x') = exp{|x-x'|/(2 tau^2)}; this should be exp(-||x-x'||^2/(2 tau^2)), with the squared norm and the negative sign.","section":"Section 3.4, Equation (3.6)"},{"comment":"The notation for the approximate posterior is inconsistent: Equation (3.5) uses q_phi(z|x_i), while the surrounding text sometimes writes q_phi(x|z); please unify the notation.","section":"Sections 3.2-3.3"},{"comment":"There is a duplicated phrase 'In in this study' that should be corrected.","section":"Section 2, Assumptions for Modeling Visual Fields"},{"comment":"The axis labels are inconsistent between the caption ('Residual Standard Error') and the panels ('Residual Standard Deviation'), and the dense point overlays make the boxplots difficult to read; please align terminology and consider separate panels or clearer legends.","section":"Figure 3"},{"comment":"The paper uses many abbreviations (ST, PW, VAE, TD, MD, PSD, HFA-II) without a single point of definition; a glossary or a table of abbreviations would improve readability.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the applied scope of the journal, but the simulation design and the unvalidated second-stage temporal model are the main concerns. The self-citation for the latent dimension choice and the circular VAE simulation should be handled carefully in revision. The held-out case study is the strongest evidence and should be highlighted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague —\n\nThis paper is worth knowing about if you work on large-cohort imaging prediction. It is a straightforward two-stage recipe: train a VAE on all visual fields as i.i.d. images, regress each of eight latent coordinates linearly on visit time per patient, extrapolate, then decode. What is actually new is not the VAE itself, which is standard, but the specific pipeline and the systematic comparison against a classical spatiotemporal model and pointwise regression on 3,832 patient eyes. The held-out evaluation is real and supports the main claim: VAE predictions are better by MAE, and the margin grows when only three visits are used to predict further ahead, especially in glaucoma patients. That is a clinically relevant result, and the paper is mostly honest about its limitations. The GitHub repository is a plus, though I have not run the code.\n\nThe soft spots, in increasing order of concern. First, the simulation includes a VAE setting that uses the Section 5 VAE to generate the data, which will flatter the method; this is a weak form of circularity, not a decisive one, because the main evaluation is on held-out test patients. Second, the latent dimension of eight is taken from a self-cited bioRxiv paper with no sensitivity analysis. Third, the MAE comparisons have no uncertainty intervals or paired tests, so we cannot tell how stable the reported advantages are.\n\nThe stress-test note about the second-stage model is fair and worth taking seriously. The paper asserts that independent linear regressions per latent dimension \"prove sufficient,\" but gives no residual diagnostics, no check for cross-dimension coupling, and no evaluation of latent extrapolation error on held-out patients. With three baseline visits each regression has one residual degree of freedom, and a five-step-ahead forecast amplifies slope error by the horizon. If the latent trajectory is biased, the decoded future field is biased, and part of the few-visit advantage could be the smoothness of the decoder. I would not say this sinks the empirical claim, because the comparison is still held-out MAE, but it should be addressed with diagnostics rather than asserted.\n\nWho is this for? Applied statisticians and clinical imaging groups. A methods-focused reader gets a useful worked example of two-stage latent forecasting, with known limitations clearly laid out. It deserves serious refereeing and likely a conditional decision with requests for trajectory diagnostics, latent-dimension sensitivity, and uncertainty on the prediction comparisons.","headline":"Useful VAE-based visual-field forecasting with a real held-out evaluation; the unvalidated linear latent trajectory is the main thing to fix before endorsing the clinical claim.","tokens_in":116659,"tokens_out":2732,"would_cite":false,"duration_ms":33625,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"A variational autoencoder trained on thousands of visual fields predicts future glaucoma damage more accurately than classical spatiotemporal models, especially when only a few baseline visits are available.","keywords":["variational autoencoder","spatiotemporal modeling","glaucoma progression","visual fields","deep generative model","scalable spatial statistics","prediction"],"falsifier":"If a reader re-runs the case study and splits test patients according to whether their latent-coordinate trajectories curve (e.g., sudden worsening after stable years), the central claim would fail if mean absolute prediction error is substantially higher for those patients than for linear progressors. A sharper version: replace the independent linear regressions with a latent autoregressive model trained on sequences; if the sequence model beats the two-stage VAE on the same test set, the independence assumption is the bottleneck.","tokens_in":115745,"feed_emoji":"👁","tokens_out":5206,"duration_ms":51449,"temperature":0.7,"pith_summary":"This paper tries to show that a variational autoencoder (VAE) can serve as a scalable method for modeling large collections of spatiotemporal data, and that it predicts future visual fields in glaucoma more accurately than a classical spatiotemporal model or pointwise regression. The proposed two-stage procedure first trains a VAE on all visual fields as independent images to learn an eight-dimensional latent representation of the visual field, then fits a separate linear regression in time to each latent dimension and decodes the predicted latent point back into a visual-field image. The claim matters because glaucoma care often requires forecasting how a patient's vision will worsen from only a few early examinations. The paper reports that the VAE's advantage grows when fewer baseline visits are used and predictions reach further into the future, and is most pronounced in glaucoma patients.","feed_headline":"Model predicts glaucoma damage from just three eye exams","feed_subtitle":"A two-stage variational autoencoder outperforms classical regression at forecasting future visual fields, most in glaucoma.","key_machinery":"The central object is a two-stage variational autoencoder. Stage one uses a convolutional encoder that maps each padded 12-by-12 visual field to an eight-dimensional latent vector, with maximum mean discrepancy as the regularization loss and a decoder that returns a de-noised reconstruction; it treats every visual field in the training cohort as an independent image. Stage two models each latent dimension separately as a linear function of visit number, extrapolates to a future visit, and feeds the predicted latent vector through the decoder to produce a predicted visual field. The mechanism that carries the argument is the generative decoder: because the latent space is trained on the full cohort, the decoder turns a modest extrapolation in latent coordinates into a smoothed spatial image that borrows the spatial structure learned from thousands of fields.","core_discovery":"The central claim is that the two-stage VAE provides more accurate predictions of future visual fields than a patient-level spatiotemporal autoregressive model and pointwise linear regression, particularly when only three to five baseline visits are available. In a cohort of 3,832 eyes with 29,161 visual fields, the VAE was trained jointly on all fields, projected each field into eight latent dimensions, and extrapolated each dimension with an independent linear regression; decoding the extrapolated latent point produced a de-noised predicted field. Across simulation settings generated from three different data-generating processes, the VAE with sufficiently large training data had consistently lower mean absolute prediction error than the spatiotemporal and pointwise baselines, even when the data were generated by the spatiotemporal model itself. In the clinical case study, the VAE predicted five future visits from the third, fifth, or eighth visit with lower MAE than both comparators, and the gap widened for glaucoma-only patients.","pith_inferences":["A direct test of the linear-trajectory assumption: split patients by whether their observed latent paths are approximately linear, and compare VAE prediction error in the two groups; if nonlinear progressors are systematically worse, a latent autoregressive or recurrent stage would be needed.","The two-stage template should transfer to other longitudinal imaging biomarkers (e.g., retinal layer thickness maps or brain MRI morphometry), but the required training cohort size and latent dimension would need recalibration.","Because the VAE is trained on the full cohort, its extrapolation assumes that future disease patterns resemble patterns already present in training; cohorts with rare fast-progressing phenotypes may require targeted augmentation or patient-specific fine-tuning.","If temporal dependence in the latent space matters beyond linear trends, a single-stage temporal VAE trained on sequences could improve both clustering of disease trajectories and forecast accuracy; the paper's current design cannot detect that failure."],"forward_implications":["With at least one thousand training samples, the VAE predicts three future visits accurately even when the data were generated by a spatiotemporal process, a pointwise process, or the VAE itself; the classical spatiotemporal model predicts well only in its own generative setting.","Using three baseline visits, the VAE's prediction advantage over both comparators grows for longer forecasting horizons, so early clinical decisions could be supported by generative forecasts from minimal follow-up.","The decoded predictions are deliberately de-noised: the VAE strips spatial variability from the raw field before forecasting, so predicted images represent the underlying progression pattern rather than visit-specific noise.","The latent space separates healthy, suspect, and glaucoma eyes, and movement through the space corresponds to worsening in particular visual-field regions, which makes the prediction mechanism interpretable rather than a black box."],"supporting_citations":[{"why":"Supplies the variational autoencoder objective and reparameterization trick on which the two-stage method is built.","marker":"Kingma and Welling (2013)"},{"why":"Provides the spatiotemporal autoregressive data-generating process used as one simulation setting and as the classical ST competitor.","marker":"Rushworth et al. (2014)"},{"why":"Defines the spatially correlated precision matrix used in the spatiotemporal simulation model.","marker":"Leroux et al. (2000)"},{"why":"Establishes the eight-dimensional latent space as optimal for visual-field data, fixing the VAE's key hyperparameter.","marker":"Berchuck et al. (2019a)"},{"why":"Introduces the pointwise modeling approach that serves as the second baseline in simulations and the case study.","marker":"Fitzke et al. (1996)"}],"fun_headline_variants":["Scalable VAE outperforms classical models on glaucoma data","Deep generative model improves glaucoma forecasting from few visits","Two-stage VAE predicts visual fields with fewer baseline tests","VAE with linear extrapolation reduces prediction error in glaucoma","Latent-space VAE beats spatiotemporal regression for glaucoma"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that treating every visual field as an independent training image, then extrapolating each of the eight latent coordinates with a straight line in time, yields future latent positions whose decoded images match how glaucoma actually progresses.","fun_headline_variants_meta":{"raw":{"variants":["Scalable VAE outperforms classical models on glaucoma data","Deep generative model improves glaucoma forecasting from few visits","Two-stage VAE predicts visual fields with fewer baseline tests","VAE with linear extrapolation reduces prediction error in glaucoma","Latent-space VAE beats spatiotemporal regression for glaucoma"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00047,"raw_usage":{"total_tokens":2309,"prompt_tokens":887,"completion_tokens":1422,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":1343}},"tokens_in":503,"tokens_out":1422,"duration_ms":11991,"temperature":1.0,"reasoning_tokens":1343,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:19:22.083896+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If a reader re-runs the case study and splits test patients according to whether their latent-coordinate trajectories curve (e.g., sudden worsening after stable years), the central claim would fail if mean absolute prediction error is substantially higher for those patients than for linear progressors. A sharper version: replace the independent linear regressions with a latent autoregressive model trained on sequences; if the sequence model beats the two-stage VAE on the same test set, the independence assumption is the bottleneck.","supporting_citations":[{"cited_title":"A spatio-temporal model for estimating the long-term effects of air pollution on respiratory hospital admissions in greater london","cited_arxiv_id":null,"evidence_quote":"Provides the spatiotemporal autoregressive data-generating process used as one simulation setting and as the classical ST competitor."},{"cited_title":"G., Lei, X., Breslow, N., 2000","cited_arxiv_id":null,"evidence_quote":"Defines the spatially correlated precision matrix used in the spatiotemporal simulation model."},{"cited_title":"W., Hitchings, R","cited_arxiv_id":null,"evidence_quote":"Introduces the pointwise modeling approach that serves as the second baseline in simulations and the case study."}],"review_version":1}