Adaptive data selection improves wearable prediction under low baseline performance

Researchers demonstrate that adaptive data selection strategies significantly improve machine learning prediction performance in wearable health systems, but primarily benefit individuals with initially poor baseline performance rather than those already performing well. The findings suggest selective deployment of adaptive sensing based on baseline metrics could optimize resource allocation in health monitoring applications.

This research addresses a critical gap in understanding how adaptive sampling strategies perform across different populations in wearable health systems. While adaptive sensing has gained adoption as a method to improve prediction accuracy under constrained measurement budgets, the study reveals performance gains are not universal—a nuanced finding that challenges one-size-fits-all deployment assumptions.

The work builds on decades of machine learning optimization research, where adaptive strategies have shown promise in laboratory settings. However, real-world wearable systems operate within complex constraints involving individual physiological variability, sensor calibration differences, and behavioral patterns. This longitudinal analysis across multiple modalities (heart rate, activity, ecological momentary assessment) reflects the heterogeneous nature of actual health monitoring deployments.

For developers and health system operators, these results carry significant implications. The inverse correlation between baseline performance and adaptive gain (r = -0.67) suggests that deploying adaptive sensing universally wastes computational resources on high-performing participants while maximizing value for struggling cases. This insight enables cost-efficient triage strategies where systems automatically detect underperforming prediction models and activate adaptive sampling only when needed.

The discrepancy between AUROC improvements (substantial for low performers) and F1 score improvements (inconsistent) indicates that metric selection matters in real implementations. Organizations must establish which performance metric aligns with clinical outcomes before deployment. Future research should investigate why adaptive strategies fail for strong baseline performers and whether hybrid approaches combining fixed and adaptive sampling could eliminate the performance ceiling currently observed.

• →Adaptive data selection yields up to 0.7 AUROC improvement for low-performing baselines but provides minimal or negative gains for strong performers.
• →Inverse correlation of -0.67 between baseline performance and adaptive gain enables predictive identification of which participants benefit most from adaptive sensing.
• →60-80% of participants achieve AUROC improvements, but F1 score gains remain inconsistent, requiring careful metric selection for clinical deployment.
• →Selective deployment strategies that activate adaptive sensing based on baseline performance thresholds could optimize computational efficiency in wearable systems.
• →The heterogeneous response across individuals and modalities indicates adaptive sensing requires participant-level customization rather than uniform application.