Background

Pericarditis occurs in 25–50% of SLE patients during their disease course, yet diagnosis is frequently delayed because subclinical presentations lack classic pleuritic chest pain and early effusions fall below echocardiographic detection thresholds (<50 mL). Current monitoring relies on intermittent clinical visits, missing the gradual inflammatory crescendo preceding overt serositis.

Hypothesis

Continuous multi-modal wearable sensor data — specifically, (1) PR-interval variability and ST-segment morphology entropy from single-lead ECG, (2) precordial-to-peripheral skin temperature gradient trajectories from thermistor arrays, and (3) accelerometer-derived respiratory rate variability reflecting early splinting behavior — when fused via a temporal convolutional network (TCN) with attention mechanisms, will detect subclinical pericarditis in SLE patients 4–10 weeks before echocardiographic confirmation of pericardial effusion ≥10 mm, with AUROC >0.85 and sensitivity >80% at specificity ≥75%.

Testable Predictions

1. PR-interval entropy will increase significantly (>1.5 SD above personal baseline) 6–10 weeks before echo-confirmed pericarditis, reflecting early pericardial inflammation affecting atrial conduction.
2. Precordial skin temperature gradient (ΔT sternal − dorsal forearm) will show a rising trend >0.3°C above rolling 30-day baseline 4–8 weeks pre-diagnosis, reflecting localized inflammatory hyperemia.
3. Respiratory rate variability coefficient of variation will increase >25% above baseline 3–6 weeks before diagnosis due to subclinical pleuritic splinting.
4. The TCN fusion model will outperform any single-modality predictor by >10% AUROC, demonstrating synergistic information capture.
5. False positive rate will be <20% when validated against confirmed non-pericarditis SLE flares (renal, articular), ensuring cardiac specificity.

Proposed Methodology

• Design: Prospective observational cohort, 200 SLE patients (ACR/EULAR 2019 criteria, SLEDAI ≥4), 12-month continuous monitoring
• Devices: FDA-cleared single-lead ECG patch (e.g., Zio XT) + custom thermistor chest patch + wrist accelerometer
• Reference standard: Scheduled echocardiography q3 months + event-triggered echo within 72h of clinical suspicion
• Analysis: TCN with multi-head self-attention, 5-fold temporal cross-validation respecting patient-level splits, calibrated via Platt scaling
• Statistical framework: Time-dependent AUROC with bootstrap 95% CI, net reclassification improvement vs. SLEDAI serositis component alone

Limitations

• Wearable compliance over 12 months is challenging; anticipated 15–20% dropout requiring sensitivity analysis
• Skin temperature gradients may be confounded by ambient temperature, vasculitis-related Raynaud phenomenon, and medication effects (hydroxychloroquine, corticosteroids)
• PR-interval changes could reflect myocarditis rather than isolated pericarditis; concurrent troponin monitoring needed to distinguish
• Single-lead ECG may miss posterior/lateral ST changes visible only on 12-lead
• TCN models require substantial training data; 200 patients with ~15% pericarditis incidence yields only ~30 events, potentially necessitating transfer learning from general cardiac datasets

Clinical Significance

Early detection of subclinical pericarditis would enable preemptive escalation of anti-inflammatory therapy (colchicine, corticosteroid dose adjustment) before hemodynamically significant effusion develops, potentially reducing pericardial tamponade risk, hospitalization rates, and cumulative organ damage (SDI serositis domain). Integration into existing lupus wearable monitoring platforms could provide continuous cardiac surveillance without additional clinical visits, particularly valuable in resource-limited settings where echocardiography access is constrained.

LES AI • DeSci Rheumatology