Bayesian framework integrates longitudinal EHR data with genetic discovery

A new Bayesian system called ALADYNOULLI is designed to forecast how multiple diseases unfold over time from longitudinal electronic health record (EHR) histories. The work targets a central challenge in longitudinal modelling: evaluating prediction without “information leakage,” so that future diagnoses never influence training for past timepoints. By explicitly simulating real clinical follow-up, the researchers aim to make risk estimates more trustworthy for decision-making.

In the primary assessment, ALADYNOULLI produced dynamic 1-year forecasts at 10 fixed moments during follow-up, starting at enrollment and extending to years 1 through 9. At each timepoint, models were retrained using only data available up to that moment, then tested on 1-year outcomes. The key metric was the median area under the receiver operating characteristic curve (AUC) across these dynamic evaluations, summarizing how predictive accuracy changes as new diagnoses accumulate.

The study spans 28 diseases and reports robust discrimination for several clinically important endpoints. Performance was strong for atherosclerotic cardiovascular disease (ASCVD), breast cancer, atrial fibrillation, heart failure, and Parkinson’s disease, with AUC values of 0.879, 0.867, 0.801, 0.811, and 0.796, respectively. Importantly, the evaluation excluded people with prevalent disease at prediction time, aligning the comparisons with prospective risk.

ALADYNOULLI was also benchmarked against established clinical risk tools where feasible. At enrollment, static models trained only on baseline information were compared for both 1-year and 10-year horizons. While longer-term predictions naturally degraded, ALADYNOULLI retained meaningful power—for example, for ASCVD at 10 years—highlighting the advantage of learning from longitudinal EHR trajectories rather than relying solely on cross-sectional inputs.

Calibration analyses further supported reliability: predicted and observed event rates matched closely across millions of at-risk patient-time observations. The researchers report a mean squared error on calibration of 4.67 × 10⁻⁷, with mean predicted rates (5.55 × 10⁻⁴) closely tracking mean observed rates (5.45 × 10⁻⁴). This suggests not only ranking accuracy, but also quantitative risk realism.

To test whether results were driven by subtle temporal artifacts, the team ran sensitivity analyses that excluded events occurring shortly before prediction and varied washout windows. They also examined alternative horizons and a dynamic 10-year rolling interpolation strategy, which updates risk estimates as annual information becomes available. Together, these checks help confirm that temporal patterns are being learned in a causally aligned manner.

Overall, the findings position ALADYNOULLI as a clinically relevant framework for multi-disease risk forecasting, combining longitudinal EHR modelling with predictive discrimination and calibration. By rigorously preventing leakage and validating dynamic follow-up scenarios, the approach offers a pathway toward more actionable, time-aware risk prediction in real-world healthcare settings.