CIPHER-Twin: An Invariant and Conformal Digital-Twin Graph for Robust Risk and Time-to-Event Prediction on Heterogeneous EHR Cohorts
robust machine learning, dataset shift, digital twin, electronic health records, conformal prediction, anchor regression, federated learning, survival analysis, calibration, social determinants of health
AuthorsAbstractElectronic health record (EHR) models that move between hospitals face strong dataset shift, weak or delayed labels, and tight requirements on calibration, fairness, and privacy. Study presents CIPHER–Twin, a digital-twin inspired but engineering-focused framework that treats these issues as a robustness and reliability problem on heterogeneous tabular EHR data. Each case is encoded as a typed patient–phenotype graph with nodes for patient, social determinants of health (SDoH), organs, biomarkers, and therapies, and then passed through a lightweight encoder that supports self-supervised pretraining, federated simulation, and downstream multi-task prediction. The encoder uses graph-masked self-supervision to learn biomarker embeddings conditioned on SDoH, and then applies an anchor-based invariance objective to reduce environment-specific shortcuts under cohort shift. Per-label conformal thresholds convert probabilistic outputs into set-valued risk predictions with finite-sample coverage guarantees, while a discrete-time survival head supports time-to-event modelling. A simple FedAvg loop with clipped and noise-perturbed client updates simulates privacy-aware federated training without moving raw data, and an extremely compact student model distilled from the full encoder enables low-latency deployment. Experiments on four public cohorts (Framingham cardiovascular, UCI Diabetes, UCI CKD, and Heart Failure Clinical Records) show strong discrimination on the test split (AUROC: CVD 0.999, T2DM 0.9998, CKD 0.973; AUPR: CVD 0.922, T2DM ≈ 1.000, CKD 0.331) while maintaining near-nominal positive-class coverage with compact set-valued outputs (coverage ≈ 0.88– 0.92; average set size ≈ 0.95). Group-wise calibration by sex shows small AUC gaps for CVD and T2DM (≤ 0.003). A tiny student distilled from the full encoder reproduces teacher embeddings with cosine similarities peaked near 1.0 while reducing parameters from ∼ 9,800 to ∼ 1,150 (∼ 8.5×), which supports embedded or edge deployment. Survival modelling on the heartfailure cohort reaches a modest C-index (≈ 0.42), consistent with the cohort size and limited longitudinal depth, and serves as a stress test for robustness under sparse time-to-event information. Overall, CIPHER–Twin acts as an end-to-end engineering recipe for robust EHR modelling under dataset shift: it aligns typed tabular representations, invariant objectives, and conformal prediction with a federated-ready training loop and fully auditable exports for reproducible evaluation.
•••••••••••••••••••••••••••••••• ejprd.org - Published by Riset Publication Services LLC
EJPRD
Copyright ©2026 by Riset Publication Services LLC