Anomaly Pattern Comparison — Oura Digital Twin

Executive Summary

Henrik ANOMALY DAYS

Normal

8/ 87

9.2% of days flagged

Mitchell ANOMALY DAYS

Normal

46/ 594

7.7% of days flagged

FINGERPRINT SIMILARITY

Info

0.52cosine

1.0 = identical signatures

PCA AXIS SIMILARITY

Info

0.52cosine

Dominant failure mode alignment

Henrik and Mitchell show moderately similar anomaly signatures (cosine=0.52). Henrik has 8 anomaly days vs Mitchell's 46. PCA axis similarity is 0.52, suggesting shared dominant failure modes.

Henrik (post-HSCT): Anomaly Detection Results

TOTAL DAYS

ENSEMBLE ANOMALIES

Normal

Z-SCORE METHOD

2.3%

ISOLATION FOREST

10.3%

PERCENTILE METHOD

11.5%

Top 10 Anomaly Days

Date	Score	Agreement	Top Deviating Metrics	Cluster
2026-02-09	1.013	3/3	HR Average (z=+4.6), Readiness (z=-4.5), HR Lowest (z=+3.6)	-1
2026-03-23	0.607	2/3	Steps (z=+2.8), Active Cal (z=+2.0), HRV Average (z=+1.8)	-1
2026-03-09	0.545	2/3	HRV Average (z=+4.0), HRV Balance (z=+2.0), Recovery Index (z=-1.8)	-1
2026-03-03	0.534	3/3	HRV Average (z=+3.3), HR Average (z=-2.2), HR Lowest (z=-2.1)	-1
2026-02-03	0.523	2/3	Activity Score (z=+2.1), Readiness (z=-1.9), Sleep Balance (z=-1.8)	-1
2026-01-31	0.479	2/3	HR Lowest (z=+2.4), Readiness (z=-2.2), HR Average (z=+1.7)	-1
2026-01-19	0.467	2/3	Sleep Score (z=+2.8), Sleep Balance (z=+2.0), Readiness (z=+1.5)	-1
2026-01-09	0.463	2/3	Readiness (z=-2.3), HR Lowest (z=+2.0), HRV Average (z=-1.6)	-1

Mitchell (post-Stroke): Anomaly Detection Results

TOTAL DAYS

594

ENSEMBLE ANOMALIES

Normal

Z-SCORE METHOD

5.1%

ISOLATION FOREST

10.1%

PERCENTILE METHOD

13.6%

Top 10 Anomaly Days

Date	Score	Agreement	Top Deviating Metrics	Cluster
2022-05-04	1.670	3/3	Temp Delta (z=+21.1), Readiness (z=-6.0), Efficiency (z=-4.7)	-1
2021-12-21	1.409	3/3	Efficiency (z=-6.2), Sleep Score (z=-5.5), Total Sleep (z=-5.1)	-1
2021-09-18	1.255	3/3	Readiness (z=-4.4), HRV Balance (z=-4.1), HR Lowest (z=+4.1)	-1
2021-12-20	1.165	3/3	Sleep Score (z=-5.1), Efficiency (z=-5.0), Total Sleep (z=-4.8)	-1
2026-03-28	1.165	3/3	Sleep Score (z=-4.9), Total Sleep (z=-4.6), Efficiency (z=-4.4)	-1
2021-11-17	0.975	3/3	HR Lowest (z=+4.6), HR Average (z=+3.5), Readiness (z=-3.5)	-1
2021-10-24	0.952	3/3	Readiness (z=-3.2), HR Lowest (z=+2.6), Sleep Balance (z=-2.4)	-1
2021-09-17	0.950	3/3	Breath Rate (z=-4.7), Readiness (z=-3.5), HR Lowest (z=+2.8)	-1
2021-11-21	0.947	3/3	HRV Balance (z=-4.4), Readiness (z=-3.1), HR Average (z=+2.6)	-1
2021-11-20	0.833	3/3	HRV Balance (z=-3.2), Steps (z=+2.5), Readiness (z=-2.2)	-1

Anomaly Timeline

Anomaly Fingerprint Comparison

Cluster Analysis

Metric Severity Comparison

Clinical Implications

Anomaly Signature Differences

Henrik (post-HSCT) typical bad day: Readiness (z=-0.92), Steps (z=+0.66), Activity Score (z=+0.56), HR Average (z=+0.51), HR Lowest (z=+0.50)

Mitchell (post-Stroke) typical bad day: Readiness (z=-1.40), Total Sleep (z=-1.01), HR Lowest (z=+0.94), Sleep Score (z=-0.94), HR Average (z=+0.86)

Cluster Patterns

Severity Comparison

Anomaly severity distributions are significantly different (Mann-Whitney U, p=p=0.005).

Key Takeaways

Both patients show similar anomaly signatures, suggesting shared physiological stress pathways despite different conditions.
Z-score normalization removes raw-value bias, enabling meaningful comparison of 'how bad is bad' relative to each patient's baseline.
Ensemble methods reduce false positives: only days flagged by multiple methods are considered truly anomalous.

Methodology & Limitations

Anomaly Detection Pipeline

Data Collection: 17 metrics from Oura Ring across sleep, readiness, and activity domains.
Z-Score Normalization: Each metric normalized per patient (z = (x - mean) / std). This removes raw-value differences (e.g., HRV 9ms vs 43ms) and focuses on relative deviation.
Three Detection Methods:
- Z-Score Threshold: Day anomalous if 3+ metrics exceed |z| > 2.0 or composite magnitude > 2.0
- Isolation Forest: Unsupervised tree-based method (contamination=0.1, n_estimators=200)
- Percentile-Based: Flag values outside 5th/95th percentile, anomalous if 3+ metrics extreme
Ensemble: Day is anomaly if 2+ of 3 methods agree. Weighted score: z-score (0.35) + Isolation Forest (0.40) + percentile (0.25).
Fingerprinting: Mean z-vector of anomaly days reveals the 'typical bad day' signature.
PCA: Reduces dimensionality; PC1 loadings reveal the dominant failure mode.
DBSCAN Clustering: Groups anomaly days by similarity (eps=2.0, min_samples=2). Clusters labeled by dominant metric group.

Limitations

Henrik has ~81 days of data vs Mitchell's ~518 days. Small sample size limits clustering reliability.
Oura Ring consumer-grade sensors; not clinical-grade measurement.
Temperature delta may show seasonal drift in Mitchell's longer dataset.
Isolation Forest contamination rate (10%) is assumed, not empirically derived.
Cosine similarity treats all metrics equally; clinical weighting may differ.