guest ::
| Home > Publications database > Wearable-Based Machine Learning Score for Fatigue: Initial Clinical Validation |
| Home > Publications database > Wearable-Based Machine Learning Score for Fatigue: Initial Clinical Validation |
| Journal Article/Dataset | FZJ-2026-05920 |
; ; ; ; ; ; ; ;
2026
Thieme
New York, NY
This record in other databases:
Please use a persistent id in citations: doi:10.1055/s-0046-1825717
Abstract: Background: Fatigue is a prevalent and debilitating symptom among cancer patients (pts), yet routine assessment still relies on self-report instruments such as the FACIT-Fatigue (FACIT-F). To support continuous fatigue monitoring, we developed a wearable-derived Machine Learning Fatigue Score (ML-F; range 0–52, aligned with FACIT-F) in an observational study (72 pts, 22 healthy controls; 8 weeks), combining 35 physiological features from a fitness tracker (e.g., heart rate- and respiration-related measures) and anchoring the score to a single baseline FACIT-F assessment.Research Question: This study reports the initial clinical validation of the ML-F.Methods: Data were pooled from two prospective cohorts: (1) observational cohort and (2) an app-based feasibility study (38 pts; 4 weeks). Participants wore a Garmin Vivosmart 5 continuously. FACIT-F assessments and other patient-reported outcomes (PROs) were collected weekly and prospectively; laboratory parameters were retrospectively extracted from electronic health records. To maximize data use, pairwise available-case analyses were performed. Validation included discriminative performance (ROC metrics, Youden-Index optimization) across 187 pooled repeated time-matched FACIT-F/ML-F pairs from Weeks 2-4, associations with related PROs (Pearson correlations with EQ-5D-5L and PHQ-4), and exploratory longitudinal tracking (Delta Week 4-Week 2).Results: The combined patient cohort was highly heterogeneous, encompassing diverse tumor entities and metastatic stages. In the full study sample, including healthy controls, the ML-F showed good discrimination of FACIT-defined fatigue (AUC=0.84). Optimizing the cut-off to≤32.2 via Youden-Index (J=0.48 vs. 0.29) improved sensitivity (73.8% vs. 55.0%), while maintaining specificity (73.8%). In exploratory longitudinal analyses within the patient sample, ΔML-F correlated positively with intra-individual changes in calcium (ΔCa; n=17, r=0.58, p=0.01) and hemoglobin (ΔHb; n=17, r=0.55, p=0.02). Conversely, FACIT-F failed to capture these biological dynamics (p>0.80). Furthermore, the ML-F showed significant associations with the EQ-5D-5L index value (n=53, r=0.48, p=0.0003) and the PHQ-4 total score (n=56, r=-0.40, p=0.0024), indicating higher ML-F values were associated with better quality of life and lower psychological distress.Conclusions: This study suggests the initial clinical validity of the ML-F. By reflecting longitudinal physiological changes that static self-reports may miss, the approach supports continuous monitoring of fatigue. While limited by pooled repeated observations and small sample sizes in specific longitudinal sub-analyses, these findings provide a proof-of-concept for digital fatigue monitoring in oncology.
|
The record appears in these collections: |