Peak GVHD Composite
72.9
on 2026-03-02
rSLDS Event Detection
PARTIAL
Feb 9 classified as: Recovery
Alert Lead Time
30d
RED alerts before Feb 9 event
Combined GVHD + BOS
36.6
LOW
Clinical Context: Chronic GVHD affecting skin, liver, mouth (NIH 2014: moderate). Known acute decompensation on Feb 9, 2026 (validation target). Temperature deviation from Oura is the primary inflammatory signal. This analysis tests whether wearable data can provide early warning of GVHD flares.
1. Temperature Fluctuation Analysis
Temperature deviation from Oura measures deviations from personal baseline body temperature. GVHD causes micro-inflammatory signatures visible as increased night-to-night variability. CUSUM change-point detection identifies regime shifts in temperature dynamics.
Regime changes detected: 2026-01-31, 2026-02-11, 2026-03-02
2. Multi-Stream GVHD Composite Score
Six biometric streams weighted by clinical relevance: temperature (25%), HRV (20%), SpO2 (15%), sleep fragmentation (15%), resting HR (15%), activity (10%). Each component z-scored against first 14 days baseline, normalized to 0-100 (higher = more GVHD-like).
3. Recurrent Switching Linear Dynamical System
4-state rSLDS (Remission, Pre-flare, Active Flare, Recovery) fitted via Laplace-EM (Linderman et al. 2017). Each discrete state governs linear dynamics in a continuous latent space, with recurrent transitions that depend on the latent state. Observations: composite score + temperature deviation + HRV (3 features). Falls back to Gaussian HMM (hmmlearn) if ssm library is unavailable.
rSLDS State Distribution
| State | Days | Percentage |
|---|
| Remission | 3 | 4.0% |
| Pre-flare | 27 | 36.0% |
| Active Flare | 23 | 30.7% |
| Recovery | 22 | 29.3% |
4. Early Warning Alerts
YELLOW alert: P(pre-flare) > 0.3 for 3+ consecutive days.
RED alert: P(pre-flare) > 0.5 OR P(active flare) > 0.2.
Retrospective validation against Feb 9, 2026 acute decompensation.
Alert History (50 total)
| Date | Level | P(Pre-flare) | P(Flare) | Composite |
|---|
| 2026-01-10 | RED | 0.0 | 1.0 | 59.4 |
| 2026-01-11 | RED | 0.0 | 1.0 | 58.8 |
| 2026-01-12 | RED | 0.0 | 1.0 | 42.1 |
| 2026-01-13 | RED | 0.0 | 1.0 | 66.7 |
| 2026-01-14 | RED | 0.0 | 1.0 | 40.0 |
| 2026-01-15 | RED | 1.0 | 0.0 | 52.1 |
| 2026-01-16 | RED | 1.0 | 0.0 | 44.5 |
| 2026-01-17 | RED | 1.0 | 0.0 | 44.3 |
| 2026-01-18 | RED | 1.0 | 0.0 | 55.2 |
| 2026-01-19 | RED | 1.0 | 0.0 | 22.1 |
| 2026-01-20 | RED | 1.0 | 0.0 | 56.3 |
| 2026-01-21 | RED | 0.0 | 1.0 | 54.3 |
| 2026-01-22 | RED | 0.0 | 1.0 | 52.2 |
| 2026-01-23 | RED | 1.0 | 0.0 | 49.7 |
| 2026-01-24 | RED | 1.0 | 0.0 | 39.6 |
| 2026-01-25 | RED | 1.0 | 0.0 | 40.4 |
| 2026-01-26 | RED | 1.0 | 0.0 | 28.7 |
| 2026-01-27 | RED | 1.0 | 0.0 | 36.6 |
| 2026-01-28 | RED | 1.0 | 0.0 | 34.5 |
| 2026-01-29 | RED | 1.0 | 0.0 | 62.5 |
5. Predictive Feature Importance
Features ranked by combined score: point-biserial correlation (30%), mutual information (30%), composite correlation (20%), Cohen's d effect size (20%). Binary target: composite score > 65th percentile.
Top Predictive Features for GVHD State
| Feature | Importance | Correlation | Mutual Info | Cohen's d |
|---|
| HRV Median (RMSSD) | 0.520 | -0.548 | 0.266 | -1.45 |
| Sleep Heart Rate | 0.503 | +0.594 | 0.124 | +1.46 |
| Lowest HR | 0.502 | +0.571 | 0.170 | +1.41 |
| Readiness Score | 0.353 | -0.410 | 0.079 | -0.89 |
| Temperature Deviation | 0.282 | +0.326 | 0.127 | +0.65 |
| Temperature Gradient | 0.209 | +0.255 | 0.064 | +0.53 |
| HRV Variability | 0.200 | -0.244 | 0.044 | -0.57 |
6. BOS Risk Integration
Bronchiolitis obliterans syndrome (BOS) is the pulmonary manifestation of chronic GVHD. Existing BOS risk score (24.9) integrated with systemic GVHD composite. SpO2 trend analysis provides supplementary pulmonary assessment.
- BOS Risk Score: 24.9 (MODERAT)
- Combined Risk: 36.6 (LOW)
- SpO2 Mean: 96.07%
- SpO2 Trend: -0.0047 %/day
Disclaimer: This analysis is for research purposes only and should not be used as a sole basis for clinical decisions. N=1 retrospective case study - all detection metrics are descriptive, not inferential. Sensitivity/specificity cannot be computed from a single patient. Validation requires an external multi-patient cohort. Temperature deviation from consumer wearables has limited precision compared to clinical thermometry. All clinical decisions should be made in consultation with the treating hematologist.