PaperlayerView exact revision on arXiv

Source-linked AI summary

Longitudinal retinal imaging and auxiliary tasks for discrete-time survival prediction of referable diabetic retinopathy and maculopathy: a retrospective cohort study

Paul Nderitu, Laura Webster, Samantha S. Mann, Christos Bergeles, Timothy L. Jackson

arXiv:2610.37544v1eess.IV

TL;DR

Existing prognostic image models largely omit routine longitudinal fundus images. This study evaluated multi-visit prediction and found improved maculopathy and composite-outcome prediction, while auxiliary supervision added no benefit.

  • Problem

    Routine longitudinal fundus images remain largely unexplored for predicting incident referable diabetic eye disease, while existing risk engines rely on systemic data not routinely available at screening.

    Source: Introduction · p. 3

  • Method

    The study evaluated multi-visit prediction alongside auxiliary objectives for future-visit age and latent prediction.

    Source: Auxiliary objectives · p. 6, Discussion · p. 14

  • Results

    Longitudinal images improved prediction of incident maculopathy and composite rDR/rM, whereas auxiliary future-age or latent supervision added no benefit.

    Source: Discussion · pp. 14–15

  • Takeaways & Limitations

    Attribution analysis found that the multi-visit model used features from earlier visits rather than relying only on the most recent image.

    Source: Discussion · p. 14

    The model lacks external validation, and only 58 test-set rDR events limited the study’s ability to establish improvement for retinopathy.

    Source: Discussion · pp. 14–15

Abstract

from arXiv · show

Deep learning can predict progression to referable diabetic retinopathy (rDR) and referable maculopathy (rM) from screening photographs, but existing models ignore the prior images that screening programmes hold. We assessed whether multi-visit (MV) images and auxiliary objectives that model the future visit improve prognostic prediction. In this retrospective cohort study of macula-centred photographs from the South-East London Diabetic Eye Screening Programme (UK), eyes without referable disease at baseline and with a follow-up at 20-28 months were split into development (72,559 eyes, 36,648 patients) and test (17,917 eyes, 9,071 patients) sets. A time-aware discrete-time survival model predicted incident rDR (R2 or R3), rM (M1), or either (rDR/rM). We compared baseline-only input with up to three visits (MV), alone or with auxiliary supervision of future-visit age (MV+age) or future-visit latent prediction (MV+latent), using the AUROC of 28-month cumulative risk and paired DeLong tests with Benjamini-Hochberg correction. In the test set, 308 eyes (1.7%) developed rDR/rM, 58 (0.3%) rDR, and 275 (1.5%) rM. MV images increased the AUROC for rDR/rM from 0.749 to 0.786 (+0.037, 95% CI 0.012-0.061; q=0.011) and for rM from 0.749 to 0.788 (+0.039, 0.013-0.065; q=0.010). Auxiliary objectives did not improve MV model performance. In eyes with two or three visits, prior visits accounted for 30-38% of positive Grad-CAM attributions. Prior screening images improved prediction of incident maculopathy and the composite rDR/rM outcome, with gains driven by longitudinal images, not auxiliary supervision. External validation and prospective studies are required before the deployment of models to individualise screening.

Introduction

Routine longitudinal fundus photographs may capture prognostic information missed by single-visit models; this study examines multi-visit prediction and auxiliary supervision for incident referable diabetic eye disease.

Introduction

  • Risk-based screening intervals could reduce appointments and costs, but existing risk engines depend on systemic data not routinely available at screening.Screening low-risk attendees also consumes capacity that could be redirected to higher-risk people.Source: Introduction · p. 3
  • Using routine longitudinal fundus photographs to predict incident referable diabetic eye disease remains largely unexplored; this study compares baseline-only and multi-visit prediction.It also assesses future-visit age and latent-prediction supervision and which visits contribute to predictions.Source: Introduction · pp. 3–4

Study design and participants

  • Eligible eyes lacked baseline referable disease and had 20–28-month follow-up; models used up to two earlier screens, with patient-level development and test sets of 72,559 and 17,917 eyes.The sets included 36,648 and 9,071 patients, respectively.Source: Study design and participants · p. 4

Outcomes

  • Outcomes were incident R2/R3 retinopathy, M1 maculopathy, or either, graded at the first follow-up within the 20–28-month window.The window accommodates variation in routine recall around the nominal 24-month interval.Source: Outcomes · pp. 4–5

Discrete-time survival formulation

  • The survival model predicted hazards over two discrete intervals, converted them to cumulative risk, and used 28-month cumulative risk for evaluation.Because outcomes were observed only at screening visits, onset was interval-censored.Source: Discrete-time survival formulation · p. 5, Outcomes · p. 5

Model architecture

  • A shared image encoder and continuous-time visit encoding feed a transformer whose baseline representation summarizes visit history and drives interval-hazard heads for rDR, rM, and their composite.The baseline representation attends to earlier visits; a baseline-only model uses the same architecture with one input visit.Source: Model architecture · p. 5

Auxiliary objectives

This section examines how prior screening images and auxiliary objectives relate to prediction of incident referable maculopathy and the composite outcome, including the contribution of earlier visits. It also notes that external validation and prospective evaluation are needed before multi-visit data can support individualized screening intervals.

Auxiliary objectives

  • The auxiliary branches predicted a future-visit embedding from history and time, aligning it to an EMA-encoder follow-up embedding or using it to predict follow-up age; these targets were training-only.Latent alignment combined cosine and smooth-L1 losses, while inference survival predictions used only visit history.Source: Auxiliary objectives · pp. 5–6, Model configurations and training · p. 7

Model configurations and training

  • BL used baseline images, while MV used up to three visits; training weighted sampling so eyes with incident rDR made up 12% of sampled eyes.MV+age and MV+latent added future-age supervision and future-visit latent prediction, respectively; the survival loss averaged censoring-masked focal losses across three outcomes.Source: Model configurations and training · p. 6

Statistical analysis

  • The primary measure was test-set AUROC of 28-month cumulative risk, compared against BL on the same eyes using paired DeLong tests and Benjamini–Hochberg correction.Within each outcome, q<0·05 denoted significance; direct comparisons between multi-visit configurations were not prespecified.Source: Statistical analysis · p. 8

Results

  • In the test set, 308 eyes (1·72%) developed rDR/rM, 58 (0·32%) rDR, and 275 (1·53%) rM over the 20–28-month window.The development set had 1 351 (1·86%) rDR/rM, 232 (0·32%) rDR, and 1 235 (1·70%) rM cases.Source: Results · pp. 8–9

Auxiliary objectives

  • Neither future-age supervision nor future-visit latent prediction improved discrimination beyond MV for rDR/rM or rM; auxiliary-configuration point estimates were lower than MV for both outcomes.Future-age supervision yielded AUROCs of 0·775 for rDR/rM and 0·781 for rM; latent prediction yielded 0·756 and 0·767, respectively.Source: Results · p. 9

Results

  • 0·037 (95% CI 0·012 to 0·061; q=0·011) and 0·039 (0·013 to 0·065; q=0·010) were the AUROC gains from BL to MV for rDR/rM and rM, respectively.MV AUROCs were 0·786 for rDR/rM and 0·788 for rM; both gains remained significant after multiple-testing correction.Source: Results · p. 9, Table 2 · p. 10, Figure 2 · p. 11

baseline-only model

  • With two or three input visits, prior visits contributed 30–38% of positive attribution; the baseline accounted for 62–70% with two visits and 64–67% with three.For three-visit histories, the visit before baseline contributed 18–20% and the earliest visit 15–16%; attribution declined with distance, but no eye relied on baseline alone (minimum prior-visit share 14%).Source: baseline-only model · p. 11, Figure 3 · p. 12

Discussion

  • The study found no evidence of auxiliary-objective benefit beyond longitudinal images; its ability to establish a retinopathy benefit was limited by 58 test-set rDR events, and external validation remains outstanding.Grad-CAM attributed about a third of positive evidence to earlier visits, with contribution declining by distance from baseline.Source: Discussion · pp. 14–15
  • Longitudinal data, rather than auxiliary supervision, drove improved prediction of incident referable maculopathy and the composite rDR/rM outcome.Source: Discussion · p. 15
  • Prognostic models in diabetic eye screening, and potentially other retinal and systemic diseases monitored by repeated imaging, should use image histories already held by health systems.Source: Discussion · pp. 15–16
  • External validation and prospective model evaluation are next steps toward individualizing screening intervals using multi-visit data.Source: Discussion · p. 16

Role of the funding source

The paper reports funder non-involvement, author interests, code availability, and restrictions on sharing screening data.

Role of the funding source

  • Funders had no access to study data and did not influence design, model development, analysis, or manuscript preparation; the corresponding author decided whether to submit.The study was funded by Diabetes UK and supported by Wellcome.Source: Role of the funding source · p. 16

Declaration of interests

  • Several authors reported industry speaker, advisory, consulting, equity, or research relationships, while Webster and Bergeles declared no conflicts.The disclosures include named companies and commercial trial-related institutional funding.Source: Declaration of interests · p. 16

Code availability

  • Open-source Python, PyTorch, timm, and Captum resources were used to develop and validate the systems.Source: Code availability · p. 16

Data sharing

  • Screening data are unavailable publicly because of information-governance restrictions.Source: Data sharing · p. 16

TRIPOD+AI checklist

The checklist points readers to the paper’s introduction and methods for the prediction-model rationale, population, data, outcomes, and development procedures.

TRIPOD+AI checklist

  • The checklist identifies the study as prediction-model development or evaluation and directs readers to the title and abstract sections.Source: TRIPOD+AI checklist · p. 19

INTRODUCTION

  • The checklist covers healthcare context and rationale, target population, intended purpose and users, known sociodemographic health inequalities, and objectives specifying model development or validation.The cited locations map p4 to Introduction and p54 to Methods: Study design and participants.Source: INTRODUCTION · p. 19

METHODS

  • Methods checklist entries direct readers to study design, participants, predictors, outcomes, data preparation, model architecture, auxiliary objectives, and training details.The checklist notes that sample size calculation was not described and that the number of screening sites was not stated.Source: METHODS · p. 19
  • Methods checklist entries also cover outcome assessment and blinding, predictor definitions, predictor-assessor qualifications, and missing-data handling.Photographs required no subjective predictor measurement; all eligible eyes were used, without a reported sample-size calculation.Source: METHODS · p. 19

Analytical methods

The checklist maps analysis and evaluation reporting to methods pages, while noting that model updating was not applicable.

Analytical methods

  • Analysis checklist items address data partitioning, predictor handling, model type and development, and treatment of heterogeneity across clusters.The checklist marks cluster heterogeneity as not examined in the single-programme study.Source: Analytical methods · p. 20
  • Performance-measure reporting is directed to the statistical-analysis section, while model updating is marked not applicable.Source: Analytical methods · p. 20
  • The checklist directs readers to the discrete-time survival formulation for prediction calculation and model-output details, and notes that predictions were not recalibrated.Class-imbalance handling is linked to model configurations and training.Source: Analytical methods · p. 20

Training versus evaluation

The checklist points to study-design, ethics, and funding sections for reporting evaluation-data differences, approval, and financial support.

Training versus evaluation

Conflicts of interest

The supplied material locates a declaration of interests and lists reporting items for conflicts, protocol, registration, and data and code sharing. It states that study code availability was not reported.

Conflicts of interest

  • The reporting items request conflict disclosures and protocol and registration information.They also allow a statement that no protocol was prepared or the study was not registered.Source: Conflicts of interest · p. 21
  • The declaration of interests is located on page 12.The supplied passage gives no details of the declaration itself.Source: Conflicts of interest · p. 21

Data sharing

  • The data-sharing item requests details of study-data availability.Source: Data sharing · p. 21
  • Study code availability is described as not stated.The code-sharing item asks for details of analytical-code availability.Source: Data sharing · p. 21

Patient & Public Involvement

The supplied passages list reporting items for patient and public involvement, participant flow, study characteristics, and evaluation-data comparisons; they do not provide those study details.

Patient & Public Involvement

RESULTS

  • The reporting item requests participant flow, outcome counts, and, where applicable, follow-up time.It notes that a diagram may help present this information.Source: RESULTS · p. 21
  • The reporting item requests overall and setting-specific characteristics, including dates, predictors, sample and outcome counts, follow-up, missingness, and demographic differences.It suggests a table may help.Source: RESULTS · p. 21
  • For model evaluation, the item requests comparison with development-data distributions of important predictors and outcomes.Source: RESULTS · p. 21

Model development

The supplied passage is a reporting item requesting analysis-specific counts.

Model development

  • The reporting item requests participant and outcome-event counts for each analysis, such as development, tuning, and evaluation.Source: Model development · p. 21

Model specification

The supplied passages identify the model-architecture section and specify what model information should be reported for reuse and evaluation.

Model specification

  • The reporting item requests enough detail about the full prediction model to support new predictions and third-party evaluation or implementation.Examples include a formula, code, model object, or API, along with access or reuse restrictions.Source: Model specification · p. 21
  • The model architecture section is located on page 6.Source: Model specification · p. 21

Model performance

The performance material points to reported estimates and confidence intervals, while the discussion checklist covers limitations and future applicability. Heterogeneity analyses and model updating are marked as not examined or not applicable.

Model performance

DISCUSSION

  • The discussion checklist names limitations, current-care usability, and future research on applicability and generalizability as topics to address.Source: DISCUSSION · p. 22
  • The checklist also asks how poor or unavailable inputs should be handled and what user interaction or expertise is required.Source: DISCUSSION · p. 22
Loading 2610.37544v1…