MMAP-Net: Missing-aware multimodal fusion for multi-label retinal disease classification with score-calibrated deferral

Background and Objective:Deployment-oriented retinal AI systems should explicitly account for incomplete clinical metadata, prediction uncertainty, and inputs acquired under conditions that differ from model development data. Many retinal classification pipelines, including most image-only baselines on the Brazilian Multilabel Ophthalmological Dataset (BRSET), do not evaluate these issues together. We present MMAP-Net, a multimodal retinal classification prototype evaluated on BRSET and complemented by a single external-device shift check on mBRSET. The contribution is framed as empirical characterization of a deployment-aware workflow, not as evidence of clinical translation readiness.Methods:MMAP-Net combines a Swin-V2 image encoder, an FT-Transformer metadata encoder that represents missing values with mask tokens and per-feature missingness indicators, and a Bidirectional Gated Cross-attention with Dual Aggregation (BiGCDA) fusion module with a learnable per-sample modality gate α. The canonical reported output space contains Head A (four primary binary labels) and Head B (five-class diabetic-retinopathy grading). A score-calibrated deferral rule is fitted on the validation split and applied unchanged to the held-out test split and to mBRSET; it is a selective-deferral heuristic rather than a method with formal coverage guarantees. Epistemic uncertainty is estimated with MC Dropout. Six ablation variants are evaluated on BRSET ((Formula presented) images, 8,524 patients), and mBRSET ((Formula presented) portable-camera images) is used only as a frozen-model external-device check without domain adaptation.Results:Under the canonical single-seed Protocol 1 setting ((Formula presented) MC Dropout, validation-selected seed 42), MMAP-Net achieves macro-AUROC 0.9673 and diabetic-retinopathy AUROC 0.9982 [95% CI: 0.9969–0.9992] on the held-out BRSET test split. The clearest architectural effect is task-specific: removing the learnable modality gate lowers optic-disc AUROC by 0.0831 (95% CI: 0.0679–0.0978) and image-quality AUROC by 0.0346 (95% CI: 0.0178–0.0512). Under the matched single-pass ablation protocol, MMAP-Net yields a higher aggregate macro-AUROC than the image-only model ((Formula presented) ). It also meets the predefined equivalence criterion relative to early fusion at the aggregate macro-AUROC level (TOST,  ± 0.01 bound, (Formula presented) ). Under 30% simulated MCAR metadata removal, macro-AUROC changes by only (Formula presented), comparable to early fusion ((Formula presented) ); this result supports stability under controlled random metadata loss, not evidence of a verified MNAR advantage. The score-calibrated deferral rule defers 64.56% of mBRSET samples versus 10.17% of held-out BRSET test samples at the validation-locked 90% setting, showing a larger deferral rate under this single external-device shift. Raw probabilities are miscalibrated for several labels; validation-fitted label-specific isotonic calibration reduces macro-ECE to 0.019 on the held-out test split with no test-set information used during fitting. At validation-locked DR operating points, sensitivity, specificity, PPV, NPV, and referral burden are reported to characterize workload trade-offs.Conclusions:The evidence supports MMAP-Net as a deployment-aware research prototype with task-specific gate benefits, explicit missingness representation, validation-fitted probability calibration, and score-calibrated deferral behavior under one external-device shift check. It does not establish aggregate superiority over early fusion, performance under verified MNAR mechanisms, formal coverage guarantees, quantitative faithfulness of visual explanations, or broad multi-country external generalization. A matched retraining ablation shows that the exploratory auxiliary hypertension branch changes primary-task macro-AUROC by only (Formula presented) ; it is therefore reported only as a negative auxiliary-task result and is not part of the canonical MMAP-Net architecture or reported output space.