Source-linked AI summary
Predicting Early Functional Decline from Longitudinal Laboratory and Vital Sign Trajectories: A Large-Scale Study Using the All of Us Research Program
Rashmita Kudamala, Aravind V. Kuruvikkattil, Lalitha Pranathi Pulavarthy, Saptarshi Purkayastha
TL;DR
Functional decline is often recognized only after falls or observable impairment, motivating earlier detection from routinely collected data. The study models longitudinal trajectories of twelve biomarkers in All of Us participants over a three-year pre-index window. Trajectory models outperformed static summaries and retained prediction up to 12 months before decline diagnosis, supporting passive EHR-based monitoring within the study’s documented outcome and validation boundaries.
Problem
Functional decline is typically recognized after observable impairment or falls, and routine biomarker data are rarely analyzed longitudinally for pre-clinical prediction.
Method
The study derived trajectory features from twelve routine biomarkers over a three-year pre-index window and evaluated them using EHR-based prediction analyses.
Results
AUROC 0.797 vs. 0.755 and AUPRC 0.380 vs. 0.304 showed that trajectory features outperformed static laboratory summaries; prediction remained detectable 3–12 months before onset.
Takeaways & Limitations
Because the model uses measurements already ordered in routine primary care, it supports passive, zero-burden EHR integration for early detection of pre-clinical functional decline.
Takeaways & Limitations
ICD-10-coded outcomes capture documented recognition rather than objective functional performance, and evaluations used cross-sectional random splits rather than calendar-time validation.
Abstract
from arXiv · showhide
Functional decline in older adults is typically recognized only after falls or observable gait impairment, closing the window for prevention. We investigated whether temporal trajectories of routine biomarkers, already recorded but rarely analyzed longitudinally, can identify patients in the pre-clinical phase of mobility decline. Using the All of Us Research Program (N = 297,861; 11.1% cases), we derived trajectory features (slope, variability, delta, mean) for twelve biomarkers over a three-year pre-index window. LightGBM models incorporating trajectories significantly outperformed static laboratory summaries (AUROC 0.797 vs. 0.755; DeLong p < 0.001; AUPRC 0.380 vs. 0.304). A 1:1 age- and sex-matched analysis confirmed an independent trajectory signal (AUROC 0.727 vs. demographics-only 0.680). A horizon analysis demonstrated sustained prediction 3-12 months before decline onset (AUROC 0.768-0.740). Because the model uses only measurements already ordered in routine care, it supports passive, zero-burden EHR integration for early detection of pre-clinical functional decline.
Introduction
Functional decline is common but usually recognized after impairment or falls, while routine EHR measurements may contain earlier longitudinal signals. This study evaluates whether biomarker trajectories alone can predict pre-clinical mobility decline at scale.
- More than one-third of adults over 65 experience functional decline, contributing to approximately three million emergency visits and $50 billion in annual U.S. healthcare expenditures.
- Routine biomarkers can reflect pathways linked to mobility decline, but clinical interpretation typically evaluates isolated values against reference thresholds rather than longitudinal change.
- A biomarker’s trajectory may distinguish risk when individual measurements appear unremarkable, such as hemoglobin declining from 13.5 to 11.5 g/dL versus remaining stable.
- Existing temporal EHR models predict several clinical outcomes, while fall models report AUROCs of 0.70–0.87 using broad clinical feature sets.
- The study addresses whether features derived exclusively from routine laboratory measurements can predict pre-clinical functional decline without diagnosis codes, medications, or functional assessments.
- The study uses trajectory descriptors for twelve biomarkers, matched analyses, utilization-feature removal, held-out SHAP interpretation, and horizon analysis to assess independent and sustained prediction.
Methods
The study uses harmonized longitudinal EHR data from the All of Us Research Program to define mobility decline cases and construct controls with bias-aware index dates.
- The dataset is a de-identified longitudinal EHR repository harmonized to the OMOP Common Data Model, including eligible participants aged 50–125 years with laboratory data.
- Mobility decline cases were identified using ICD-10-CM codes for walking difficulty, gait abnormalities, unsteadiness, reduced mobility, muscle weakness, and falls.
- Controls had laboratory data without a qualifying mobility decline diagnosis and received pseudo-index dates sampled within their laboratory observation windows.
- The pseudo-index design preserves natural measurement history while preventing immortal time bias in the control cohort.
Biomarker Selection and Observation Windows
Twelve biomarkers with established links to functional decline were harmonized across concept-name variants and measured only before the index date within a three-year window.
- Twelve biomarkers were selected a priori based on links to anemia-related fatigue, frailty, gait instability, neuromuscular function, neuropathy, inflammation, and nutritional status.
- Concept-name variants for hemoglobin A1c, vitamin D, RDW, and CRP were harmonized to canonical labels before feature aggregation.
- Measurements on or after the index date were excluded to prevent data leakage.
- The primary analysis used a three-year pre-index window containing 297,861 participants, including 33,088 cases.
Feature Engineering
The analysis compares static biomarker means with trajectory representations that encode level, variability, directional change, measurement frequency, and rate of change while controlling preprocessing leakage.
- Baseline features represented each biomarker by its mean value within the observation window.
- Trajectory features included mean, standard deviation, delta, count, and linear slope for each biomarker.
- Linear slope measured each biomarker’s rate of change per day using an ordinary least-squares estimator.
- The features yielded 80 trajectory variables per participant, with age and one-hot encoded gender added to combined models.
- Training-fold medians were used for imputation, and measurement counts explicitly modeled informative observation driven by patient health status.
Model Development and Evaluation
Four algorithms were evaluated using stratified five-fold cross-validation, with model-specific handling of class imbalance and feature scaling.
- Four algorithms were evaluated: penalized logistic regression, random forest, XGBoost, and LightGBM.Gradient boosting models used cost-sensitive learning, while logistic-regression features were standardized before fitting.
Sensitivity Analyses
Sensitivity analyses examined demographic matching, utilization-related count features, missingness, interpretability, and data governance.
- 1:1 greedy matching on gender and age isolated trajectory signal from demographic confounding.Matching used exact gender and age within ±2 years, without replacement.
- Removing measurement counts directly tested whether performance exploited informative observation rather than physiologic dynamics.The comparison used a DeLong test between full and count-removed models.
- Missingness rates were stratified by case/control status to quantify differential measurement availability.
- SHAP TreeExplainer values were computed on a held-out 20% test set of 5,000 observations to avoid training-data leakage.Mean absolute SHAP values ranked features globally and were summarized by trajectory descriptor type.
- The study used the All of Us Controlled Tier Dataset v8 within the authorized Researcher Workbench under a central IRB protocol.No additional ethics review was required for analyses conducted within the Researcher Workbench.
Study Population and Missingness
The primary cohort included 297,861 participants, and cases had consistently lower biomarker missingness than controls, indicating informative observation in the EHR.
- 297,861 participants comprised the primary modeling cohort, including 33,088 cases representing 11.1%.Mean age at index was 50.6 years; 61.7% were female and 36.5% male.
- Cases had lower missingness across all biomarkers, with the largest differentials for albumin and hemoglobin.Albumin missingness was 32.1% versus 51.4%, and hemoglobin missingness was 21.3% versus 37.0%.
- Figure 1 stratifies laboratory-feature missingness by outcome across the three-year window.The consistently lower missingness among cases is consistent with more frequent monitoring in patients developing functional impairment.
Trajectory Features Improve Discrimination
Longitudinal trajectory features improved discrimination, positive-class detection, calibration, and prediction up to one year before documented decline, while revealing physiologic and utilization-related predictors.
- AUROC increased from 0.755 with baseline laboratory means to 0.797 with trajectory features (∆AUC = 0.042; DeLong p < 0.001).Demographics alone achieved AUROC 0.680.
- AUPRC increased from 0.304 for baseline to 0.380 for trajectories, a 25% relative improvement over the 11.1% prevalence baseline.The trajectory AUPRC was 3.4× the no-skill rate of 0.111.
- AUROC remained 0.727 after age- and sex-matched analysis, confirming trajectory signal not explained by age and sex.
- AUROC declined from 0.768 at the 3-month horizon to 0.740 at the 12-month horizon, while AUPRC remained 0.361–0.353.Measurements within each specified horizon were excluded to simulate prospective screening before clinical recognition.
- BMI mean, systolic blood pressure mean, diastolic blood pressure slope, and RDW mean were among the top physiologic predictors.Count features ranked highest overall, while BMI mean was the most important non-count feature and diastolic blood pressure slope led trajectory-specific features.
- Removing count features yielded AUROC 0.789, a reduction of 0.009, while the model primarily exploited physiologic dynamics.
- Isotonic calibration reduced Brier score from 0.174 to 0.083 while preserving AUROC at 0.797 versus 0.796.
Discussion
Longitudinal trajectories from routine biomarkers improve prediction of functional mobility decline while requiring no new patient data collection. Their predictive signal persists before diagnosis, although validation and outcome measurement constrain deployment.
- Discussion: AUROC 0.797 vs. 0.755 and AUPRC 0.380 vs. 0.304 show trajectory models outperform static laboratory summaries.The comparison was statistically significant (p < 0.001).
- Discussion: AUROC 0.727 in age- and sex-matched analysis supports an independent trajectory signal beyond demographic confounding.
- Discussion: AUROC 0.768 at 3 months and 0.740 at 12 months show predictive signal up to one year before diagnosis.Corresponding AUPRC values remained stable at 0.361–0.353.
- Discussion: Using only twelve routine biomarkers enables passive EHR monitoring without new tests, patient interaction, or workflow modification.The approach is positioned as a passive digital biomarker compatible with standard tabular classifiers.
- Discussion: BMI mean, diastolic blood pressure slope, and RDW mean provide mechanistically interpretable signals linked respectively to body composition, autonomic dysregulation, and erythropoietic instability.BMI mean was the highest-ranked non-count feature (mean |SHAP| = 0.141), while diastolic blood pressure slope was the leading trajectory-specific feature (0.104).
- Discussion: AUPRC improved by 0.076, or 25%, compared with a 0.042, or 6%, AUROC gain in this 11% case-prevalence setting.The trajectory model’s AUPRC was approximately 3.4× the no-skill baseline of 0.111.
- Discussion: Measurement-count features ranked highly, but removing them reduced AUROC only to 0.789, preserving 99% of discrimination.This indicates the model can exploit informative observation without depending on utilization-proxy features.
- Discussion: Prior models report AUROCs of 0.87–0.92 using broader clinical features, whereas this model uses structured biomarkers without free-text processing or specialist assessments.The All of Us cohort deliberately oversampled Black or African American, Hispanic, and rural participants.
Conclusion
Longitudinal features from twelve routine biomarkers improve prediction of functional mobility decline over static summaries and retain an independent signal after matching and utilization-feature removal. Their exclusive use of routinely ordered measurements supports passive EHR integration for early detection.
- Conclusion: AUROC 0.797 vs. 0.755 and AUPRC 0.380 vs. 0.304 show trajectory features significantly outperform static laboratory summaries.
- Conclusion: AUROC 0.727 vs. demographics-only 0.680 confirms an independent trajectory signal in matched sensitivity analysis.
- Conclusion: AUROC 0.789 after removing utilization-proxy features preserves 99% of discrimination.
- Conclusion: Exclusive reliance on routinely ordered measurements supports passive, zero-burden EHR integration for pre-clinical functional decline detection.