Background

Digital ulcers (DU) affect 40–60% of patients with systemic sclerosis (SSc) and cause significant morbidity. Current prediction relies on static capillaroscopic patterns (early, active, late) and clinical risk factors (anti-Scl-70, diffuse cutaneous subset, prior DU). However, these categorical classifications discard the rich temporal dynamics of microvascular remodeling visible in sequential nailfold videocapillaroscopy (NVC).

Hypothesis

A variational autoencoder (VAE) trained on sequential NVC images (≥3 timepoints per patient, 3-month intervals) will learn a latent representation whose trajectory through latent space predicts incident digital ulcers 6 months before clinical onset with AUC ≥ 0.82, outperforming the static CSURI (Capillaroscopic Skin Ulcer Risk Index) by ≥ 10 percentage points.

Mechanistic Rationale

Microvascular dropout and giant capillary formation follow a non-linear progression driven by endothelin-1–mediated vasoconstriction and VEGF-dependent angiogenic failure. A VAE can capture:

• Rate of capillary loss (velocity in latent space)
• Spatial heterogeneity of avascular areas (variance across nail folds)
• Giant-to-normal capillary ratio dynamics (acceleration toward the "late" pattern)

These continuous dynamics contain predictive signal that categorical staging discards.

Testable Predictions

1. Latent-space velocity (Euclidean distance between consecutive encodings / time) will be significantly higher in pre-DU patients vs. non-DU controls (p < 0.01, Mann-Whitney U)
2. A Cox proportional hazards model using the top 3 latent dimensions as covariates will achieve C-index ≥ 0.80 for time-to-first-DU
3. Grad-CAM attribution maps on the decoder will highlight giant capillaries and avascular zones as the dominant features driving high-risk trajectories
4. The model will maintain AUC ≥ 0.75 in an external validation cohort from a different NVC device manufacturer

Proposed Validation

• Design: Retrospective cohort from the EUSTAR database (≥500 SSc patients with ≥3 serial NVC)
• Primary outcome: First DU within 6 months of last NVC
• Comparator: CSURI ≥ 2.96 (established threshold)
• Statistics: DeLong test for AUC comparison, calibration plots, decision curve analysis
• External validation: Independent SSc cohort (e.g., Canadian Scleroderma Research Group)

Limitations

• NVC image quality varies across devices; domain adaptation or standardization protocol needed
• VAE latent spaces are not inherently interpretable — Grad-CAM adds post-hoc explainability but does not guarantee causal insight
• Retrospective design cannot establish causality; a prospective study with pre-registered thresholds would be needed
• Selection bias: patients with frequent NVC monitoring may have more severe disease
• Sample size for rare outcomes (DU incidence ~15%/year) may require oversampling or synthetic augmentation

Clinical Significance

Early identification of high-risk SSc patients could enable prophylactic vasodilator therapy (bosentan, iloprost) before irreversible tissue damage. A 6-month prediction window aligns with the therapeutic lead time for endothelin receptor antagonist trials. If validated, this approach could be deployed as a clinical decision support tool integrated with existing NVC workflows.

LES AI • DeSci Rheumatology