A major issue in tele-ophthalmology screening is distinguishing sight-threatening Diabetic Retinopathy (DR) from routine cases. Although Deep Learning (DL) enables scalable automation, traditional classifiers face chronic problems: high class imbalance in screening cohorts and low interpretability required in clinical settings. This paper proposes a Hybrid Lesion-Aware Triage System aimed at prioritizing urgent referrals (Grades 2–4) while retaining high sensitivity. The system uses a two-stream pipeline: one stream employs U-Net-generated pseudo-lesion masks to obtain interpretable biomarkers, including lesion count, density, and morphology; the other stream uses an EfficientNet-B0 backbone to extract high-dimensional global retinal descriptors. These complementary signals are summed and weighted with SMOTE, empirically chosen when cost-sensitive weighting was ineffective in the sparse fused feature space. An XGBoost model performs classification, and the decision threshold is adjusted via ROC analysis to ensure sensitivity over 90%. Using a curated subset of the EyePACS dataset, the system achieved Sensitivity of 90.0% and Specificity of 72.5% (F1-Score: 0.76). Ablation experiments indicate that the fused representation outperforms individual feature streams. Results demonstrate that combining verifiable lesion biomarkers with global deep representations, along with safety-driven calibration, provides an effective screening mechanism that substantially reduces clinician workload while preserving detection of urgent cases.