Predicting Early Functional Decline from Longitudinal Laboratory and Vital Sign Trajectories: A Large-Scale Study Using the All of Us Research Program
Longitudinal trajectories of 12 routine biomarkers over a 3-year window significantly outperform static lab summaries in predicting pre-clinical functional decline (AUROC 0.797 vs. 0.755) The study leverages the All of Us Research Program with nearly 300,000 participants (11.1% cases), demonstrating robust generalizability at scale Trajectory features (slope, variability, delta, mean) provide an independent predictive signal beyond demographics alone (AUROC 0.727 vs. 0.680 in matched analysis) T
Analysis
TL;DR
- Longitudinal trajectories of 12 routine biomarkers over a 3-year window significantly outperform static lab summaries in predicting pre-clinical functional decline (AUROC 0.797 vs. 0.755)
- The study leverages the All of Us Research Program with nearly 300,000 participants (11.1% cases), demonstrating robust generalizability at scale
- Trajectory features (slope, variability, delta, mean) provide an independent predictive signal beyond demographics alone (AUROC 0.727 vs. 0.680 in matched analysis)
- The model enables sustained early detection 3–12 months before functional decline onset, with AUROC ranging from 0.768 to 0.740 across horizons
- Because the model uses only routinely ordered measurements, it supports passive, zero-burden EHR integration for clinical deployment
Why It Matters
This research addresses a critical gap in geriatric and preventive care: functional decline is typically detected too late, after falls or observable gait impairment have already occurred. By demonstrating that existing longitudinal EHR data can identify patients in the pre-clinical phase, this work opens a pathway for early intervention strategies that could preserve mobility and independence in older adults without requiring additional clinical testing or patient burden.
Technical Details
- Dataset: All of Us Research Program with N = 297,861 participants; 11.1% were cases of functional decline
- Feature Engineering: Trajectory features extracted for 12 biomarkers over a 3-year pre-index window, including slope, variability, delta, and mean
- Model Architecture: LightGBM gradient boosting classifier, compared against static laboratory summary baselines
- Evaluation Metrics: AUROC and AUPRC; statistical significance assessed via DeLong test (p < 0.001)
- Validation Approaches: Age- and sex-matched 1:1 analysis to isolate trajectory signal; horizon analysis to assess prediction lead time (3–12 months pre-onset)
- Performance: Trajectory-based model achieved AUROC 0.797 and AUPRC 0.380 vs. static model AUROC 0.755 and AUPRC 0.304
Industry Insight
- EHR-Integrated Screening: The zero-burden design using existing clinical data makes this approach highly deployable in health systems without requiring new data collection infrastructure, accelerating time-to-clinical-impact.
- Longitudinal Over Cross-Sectional: This study reinforces the growing consensus that temporal patterns in routine biomarkers carry predictive information lost in static snapshots—a principle that should extend to other chronic disease prediction tasks.
- Preventive Care Opportunity: Detecting functional decline 3–12 months earlier creates a actionable window for interventions (physical therapy, nutrition, medication review) that could reduce fall-related hospitalizations and healthcare costs at scale.
Disclaimer: The above content is generated by AI and is for reference only.