Answers are grounded in this paper and cite their sources.
BackgroundRepeated observations and missing participant and time metadata can complicate interpretation of wellbeing classification results from public sensor datasets.ObjectiveTo assess reproducibility and split sensitivity of wellbeing classification on the released DigitalExposome CSV.MethodsWe reconstructed an initially submitted 30-sample, 15-sample-stride windowed Random Forest (RF) analysis and compared matched 70/30 file-order and unstratified random row splits. Ten random seeds quantified partition variability. We recalculated class-level metrics and ran a deduplicated, 30-row positional-embargo sensitivity analysis. The CSV contains no participant identifiers or timestamps.ResultsThe file-order RF holdout gave Macro F1 = 0.8515; random splits yielded mean Macro F1 = 0.9772 (SD = 0.0014). Reconstructed windows yielded accuracy = 0.7887, Macro F1 = 0.7632, and weighted F1 = 0.7549. Recalculated recall was 0.205 for wellbeing level 2. The deduplicated sensitivity analysis produced Macro F1 values of 0.128-0.162 across examined split fractions and leaf settings.ConclusionScores were sensitive to split procedure and repeated feature vectors. Because row chronology and participant membership cannot be recovered, these results do not estimate future-time or participant-independent performance and do not establish clinical utility.
No takes yet. Share an insight, caveat, or question.
Özel Sebetçi (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: