A model-evolution story: from a retinopathy classifier to an oculomics platform

Challenge

A shortage of ophthalmologists meant referable diabetic retinopathy was caught late in high-volume, underserved clinics — and a naïve model scored on plain accuracy looked deceptively strong while missing the patients who mattered.

Approach

Rather than chase a bigger model, the system evolved in deliberate stages: re-engineer the features and metric, then the architecture, then trust, then the platform — each step measured against what actually matters in screening (sensitivity, not accuracy).

Model evolution

1. v0

   Naïve baseline

   An off-the-shelf CNN on raw fundus images, scored on plain accuracy — which looked high only because most images are healthy.

2. v1

   Feature engineering

   Retina-crop + Ben-Graham colour normalisation, and a switch to Quadratic Weighted Kappa and sensitivity — metrics that actually reward catching at-risk patients.

3. v2

   Architecture evolution

   EfficientNet-B3 transfer learning with an ordinal head that respects the 0–4 grade order, LR warmup and test-time augmentation — lifting QWK and steadying predictions.

4. v3

   Trust & explainability

   Grad-CAM attention overlays and a referable-vs-not decision, keeping a clinician in the loop on every read.

5. v4

   Platform evolution → oculomics

   The same imaging-and-deep-learning foundation extended toward oculomics — reading the retina as a window into systemic, cardiometabolic and neurological risk.

Results

• Switched to Quadratic Weighted Kappa + sensitivity — metrics that reward catching sick patients
• EfficientNet-B3 with an ordinal head aligned to the 0–4 clinical scale
• Grad-CAM explainability + referable decision on every read
• Same foundation extended toward systemic (oculomics) biomarkers