Deep learning of longitudinal visual fields predicts glaucoma progression rate and identifies fast progressors
Taiabur Rahman, Siddiqur Rahman, Muhammad Moniruzzaman, Ummay Kawsar, Sayedatunnessa Ratna, Shadman Siddique, Rafsan Siddique, Tausif Ahmad, et al.
cs.CV · cs.AI
Quick verdict
BiLSTM on serial Humphrey VFs predicts glaucoma MD progression rate (MAE 0.139 dB/yr, AUC 0.990); but labels derive from inputs and threshold was test-tuned.
Summary
The paper introduces GLAM, a deep learning framework that predicts glaucoma progression rates from longitudinal Humphrey 24-2 visual field total-deviation sequences alone. A per-visit encoder feeds a two-layer BiLSTM whose output is attention-fused with a five-feature clinical branch (age, sex, visit count, follow-up duration, baseline MD); dual heads predict MD and VFI progression rates with aleatoric uncertainty. On the public UWHVF dataset (4,276 patient-eyes after filtering; 2,992/640/644 splits), GLAM achieves test-set MD-rate MAE of 0.139 dB/yr (R²=0.927) and AUC 0.990 for fast-progressor detection (MD rate < −1 dB/yr; 56/644 test eyes), with sensitivity 98.2% and specificity 90.6% at a Youden threshold. Ablations attribute most accuracy to the BiLSTM (removal raises MAE to 0.218). The authors honestly disclose that labels are OLS slopes of the same TD sequences fed to the model, so high scores partly reflect approximating the labelling process, and call for external validation on GRAPE/Harvard-DR. Key weaknesses: baselines ignore the longitudinal series entirely; the Youden threshold is computed on the test set despite a claim that the test set was never used for calibration; arithmetic in the conservative operating point (PPV 56.0% versus '1.5 false flags per detection') does not reconcile; and the VFI proxy is a linear transform of the MD proxy, making the second task largely redundant. Code and weights are released.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- GLAM predicts MD progression rate from VF sequences plus five clinical features with test MAE 0.139 dB/yr (R²=0.927), a 73.5% MAE reduction over a naive mean baseline (Table 2).
- GLAM detects fast progressors (MD rate < −1 dB/yr) with AUC 0.990 (95% CI 0.982–0.996), sensitivity 98.2% and specificity 90.6% at the Youden-optimal threshold on the held-out test set.
- VF-only deep learning can match multimodal OCT+VF pipelines for progression prognostication (abstract claim, hedged in the Discussion as not directly comparable).
- The model's aleatoric uncertainty estimates are conservative (99.8% empirical coverage at nominal 90%) and correlate with absolute error (Spearman ρ=0.41, P<0.001).
Strengths
Honest disclosure of label circularity
The Limitations section explicitly states that both targets are OLS slopes computed from the same TD sequences fed as input, and names external validation on GRAPE and Harvard-DR as required before generalisability claims — an unusually candid self-assessment.
Complete statistical reporting
Stratified-bootstrap AUC CI (0.990, 0.982–0.996, §2.4), PPV/F1 at two operating points, Anderson–Darling residual check (A²=0.79), residual-vs-truth slope test (0.003, P=0.71), per-quintile MAE (0.092→0.205, Fig. 5) and visit-count tertile analysis (0.171 vs 0.124 dB/yr).
Strong reproducibility posture
Sensible ablation design with verified arithmetic
Robustness checks across thresholds and operating points
Weaknesses
Baselines never look at the trajectory
Test-set threshold selection contradicting the stated protocol
Internal arithmetic inconsistencies
VFI task redundant by construction
Attention interpretability undermined by the placeholder branch
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
GLAM (Glaucoma Longitudinal Analysis Model) is a BiLSTM-based framework that consumes serial Humphrey 24-2 total-deviation (TD) maps plus five clinical features (age, sex, visit count, follow-up duration, baseline MD) and outputs MD and VFI progression rates with aleatoric uncertainty. Evaluated exclusively on the public UWHVF dataset (4,276 patient-eyes; 2,992/640/644 splits), the headline claims are: MD-rate MAE 0.139 dB/yr (R² = 0.927; §2.3, Table 2), fast-progressor AUC 0.990 (95% CI 0.982–0.996; §2.4), sensitivity 98.2%/specificity 90.6% at a Youden threshold, and a 73.5% MAE reduction over baselines. The abstract's claim that 'VF-only deep learning can match multimodal pipelines' rests on comparison with Bowd et al.'s OCT+VF AUC of 0.97 — a cross-dataset, cross-task comparison the authors themselves disclaim in the Discussion. Critically, this is concurrent rate estimation over a completed follow-up, not prospective forecasting: no experiment feeds truncated histories, so the 'early-warning layer' framing in the Discussion is extrapolation.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is GLAM and what data was it trained on?
GLAM is a deep learning model that estimates how fast glaucoma is damaging the visual field, using only sequences of Humphrey 24-2 visual field tests plus five routine clinical variables (age, sex, visit count, follow-up duration, baseline MD). It was trained and evaluated on the open University of Washington Humphrey Visual Field dataset; after quality filtering (at least 3 tests and 1 year of follow-up per eye) 4,276 patient-eyes were analysed.
How accurate is GLAM at predicting glaucoma progression?
On its held-out test set of 644 eyes, GLAM predicted the mean-deviation progression rate with an error of 0.139 dB/year (R² = 0.927) and detected fast progressors (rate worse than −1 dB/year) with AUC 0.990, sensitivity 98.2% and specificity 90.6% at its chosen threshold. Caveat: the labels were computed from the same visual field series the model sees, so independent external validation is still pending.
Does GLAM require OCT or fundus images?
No. The model runs on perimetry (visual field) data and basic clinical variables only; a structural imaging branch exists in the architecture but was fed zeros throughout because the dataset contains no images.
Is the GLAM code and dataset publicly available?
Yes, according to the paper: the UWHVF dataset is openly available under CC BY 4.0 from the original UW authors, and the GLAM implementation (preprocessing, model, training, evaluation, seeds and lockfiles) is released at github.com/taiabrbd/glam.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 are available only as captions in the extracted text; visual content (scatter tightness, ROC shape, calibration plot) cannot be verified.
- Supplementary files (TRIPOD checklist, Nature Portfolio Reporting Summary) are referenced but not included in the extracted text.
- Minor metadata mismatch: the provided metadata lists 8 authors while the full text lists 10 (adding Tahsin Ahmad and Golam Rabbani).
- Reproducibility claims (released repository, seeds, lockfiles) could not be verified against the actual code, as no repository access was performed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.