Longitudinal Electronic Health Records

• Longitudinal EHRs are temporally ordered patient data streams that record evolving clinical states, interventions, and outcomes.
• Advanced methods like hierarchical Transformers and multi-task Gaussian process imputation effectively address irregular sampling and missingness.
• Generative and joint modeling techniques facilitate synthetic data creation, precise phenotyping, and fair clinical decision support.

Longitudinal electronic health records (EHRs) comprise temporally ordered, patient-specific data streams that capture the evolving clinical states, interventions, diagnoses, procedures, laboratory results, and other healthcare-related outcomes over time. Unlike cross-sectional datasets, longitudinal EHRs encode disease progression, treatment response, and health trajectories, enabling dynamic modeling but also presenting major analytical challenges, especially regarding irregular sampling, missingness, heterogeneity, and complex dependencies among clinical variables. The increasing breadth and granularity of longitudinal EHRs have catalyzed diverse methodological advances for representation learning, prediction, simulation, clustering, and causality, with critical implications for phenotyping, risk stratification, synthetic data generation, and fair clinical decision support.

1. Representation and Modeling Paradigms

A central challenge in longitudinal EHRs is the faithful representation of irregular, multi-modal sequences. Traditional approaches such as time-bucketed aggregation or fixed-interval matrices typically fail to capture the varied timing and semantic heterogeneity of real-world clinical trajectories. Modern deep models, such as Patient2Vec, introduce personalized, interpretable sequence encoders that combine code-level skip-gram embeddings, hierarchical attention over time, and bidirectional GRU encoders to summarize a patient’s trajectory, yielding state-of-the-art results in hospitalization risk prediction and interpretable risk drivers (Zhang et al., 2018). Hierarchical Transformer architectures—most notably Hi-BEHRT—segment long token sequences into overlapping windows, extracting local features and aggregating over time to capture both fine-grained and decadal dependencies, with explicit positional and age information for multimodal inputs (Li et al., 2021). This enables risk models to exceed classical Transformer baselines both in AUROC and AUPRC, especially for patients with long or irregular medical histories.

Beyond basic sequence encoding, MUSE-Net incorporates multi-task Gaussian process (MGP) imputation to handle missing values, multi-branch classifier heads for label imbalance, and a time-aware self-attention encoder with explicit elapsed time positional encodings, robustly addressing irregularity, incompleteness, and imbalance in high-dimensional EHR time series (Wang et al., 2024).

2. Temporal Alignment, Event Time Annotation, and Subtyping

Temporal misalignment—where disease onset or relevant clinical events precede or do not coincide with EHR observation windows—creates artificial heterogeneity and confounds analyses. Subtype-aware timeline registration jointly estimates individual time shifts (over a pre-specified discrete grid) for each patient by minimizing within-cluster trajectory dispersion after low-dimensional projection via penalized B-splines. A discrete optimization alternates between cluster assignment, trimmed centroid estimation, and per-subject shift re-alignment. This approach improves recovery of true disease trajectories, enhances clustering (measured by silhouette coefficients), and significantly boosts downstream predictive tasks such as severe acute kidney injury risk (Gai et al., 13 Jan 2025).

Precise event time annotation is frequently unavailable. The MATA method employs a two-step semi-supervised approach: it extracts trajectory-level features from unlabeled data via functional principal component analysis (FPCA) on point processes, then fits a penalized proportional odds model with B-spline sieves for event time on a small labeled set. This procedure achieves root-n consistent estimation, is robust to heavy censoring, and outperforms nonparametric and rule-based benchmarks in both simulations and cancer recurrence annotation tasks (Liang et al., 2021). LATTE further improves label-efficiency in event timing by leveraging pre-trained EHR concept embeddings, concept re-weighting via an MLP, a visit attention network, and a Bi-GRU, combining unsupervised pretraining on silver-standard surrogate labels with semi-supervised fine-tuning, resulting in superior phenotyping accuracy across Type 2 diabetes, heart failure, and multiple sclerosis onset (Wen et al., 2023).

3. Generative Modeling and Synthetic Data

Privacy concerns and data scarcity motivate the development of generative models for high-dimensional, longitudinal EHRs. The HALO framework models the full patient visit–code tensor using a two-level, hierarchical autoregressive Transformer: a visit-level masked Transformer captures between-visit dependencies, while a masked linear autoregressive network emits per-code (diagnosis, med, lab) probabilities within each visit (Theodorou et al., 2023). Discrete and continuous variables (labs/time-gaps) are handled via discretization and bucket-based decoding. HALO attains >0.9 $R^2$ on unigram code frequencies, supports extremely high-dimensional codes ($d$ > $10^4$), and yields synthetic datasets enabling downstream disease prediction models to approximate the AUROC of real-data-trained baselines with no evidence of memorization under membership inference attacks.

Conditional generative models such as EVA employ a hierarchical variational autoencoder framework with patient-level and condition-level latent variables, leveraging amortized variational inference and SG-MCMC for global weights. EVA generates disease-specific or multimorbid synthetic cohorts, exhibiting strong statistical fidelity (perplexity, bigram correlation), plausible realism in blinded clinician studies, and successful privacy resistance (Biswal et al., 2020). For joint continuous/discrete synthesis, EHR-M-GAN combines dual VAEs mapping real trajectories into a shared latent manifold with a coupled, bilateral LSTM generator, training adversarially to capture multimodal, correlated dynamics in ICU data. This approach outperforms single-type generative baselines in MMD, correlation structure, downstream utility, and privacy leakage (Li et al., 2021).

4. Clustering, Subgroup Discovery, and Phenotyping

Defining patient subtypes with shared longitudinal patterns is essential for personalized medicine. Spline-based trajectory clustering (clustra) addresses inconsistency in observation timing by modeling cluster centroids as thin-plate regression splines fit via penalized least squares and employing an EM-style K-means on regularized distances (Adhikari et al., 1 Jul 2025). Cluster number selection is guided by silhouette analysis and Adjusted Rand Index stability. This approach efficiently uncovers clinically interpretable phenotype clusters in large-scale blood pressure trajectories.

For causal inference, tree-based subgroup discovery algorithms (SDLD) couple generalized interaction trees with node-specific longitudinal targeted maximum likelihood estimators (L-TMLE), enabling time-varying confounding and dropout adjustment while discovering subgroups with heterogeneous treatment effects. The method recursively splits baseline covariates, estimates local average treatment effects with doubly robust estimators, and validates via held-out evaluation, supporting precise heterogeneity analysis in treatment effect from EHR data (Yang et al., 2022).

5. Statistical and Joint Modeling of Longitudinal Outcomes

Standard mixed effects models fail when EHR observation times are informative (i.e., visit frequency depends on patient health status). EHRJoint introduces a joint modeling framework for the visit process (possibly with frailty), longitudinal outcome process, and time-to-event process, using estimating equations and robust inference procedures for settings with both informative presence and observation (Du et al., 2024). This approach achieves unbiased exposure-biomarker and exposure-outcome effect estimation in complex real-world settings, outperforming non-informative and two-process models as shown in the Michigan Genomics Initiative biobank. Nonparametric Bayesian tree ensemble models, such as BNDS, accommodate sparse, irregular EHR data, informative missingness, and survival outcomes by constructing additive tree models with joint probability over event occurrence, enabling individualized posterior survival curve estimation and credible bands (Bellot et al., 2019).

6. LLM-based Reasoning, Decision Support, and Fairness

Longitudinal EHRs pose unique challenges for LLMs due to document length and semantic heterogeneity. Retrieval-augmented generation (EHR-RAG) addresses context-length limitations by integrating event- and time-aware hybrid retrieval (semantic+temporal scoring of evidence), iterative query refinement, and dual-path (factual/counterfactual) reasoning to optimize LLM inference over structured records. EHR-RAG achieves substantial macro-F1 improvements in long-horizon clinical prediction (Cao et al., 29 Jan 2026). CliCARE extends LLM-based decision support by transforming free-text, longitudinal cancer records into patient-specific temporal knowledge graphs (TKGs), aligning these with clinical guideline KGs, and grounding LLM reasoning in process-oriented recommendations. Output quality is validated against human experts and LLM ensembles, showing marked gains in summary, recommendation, and hallucination reduction compared to RAG baselines (Li et al., 30 Jul 2025).

Ensuring fairness in clinical prediction is increasingly critical. The FLMD model employs a two-stage deconfounder paradigm, first learning per-encounter latent confounders to capture unobserved medical factors, then enforcing counterfactual fairness by training predictions to be invariant to demographic perturbations. Empirical results indicate FLMD consistently improves both accuracy and health disparity metrics, and remains robust under distributional shift and group imbalance (Liu et al., 2023).

7. Multimodal and Continuous-Time Integration

Recent developments bridge the gap between temporally sparse EHR data and dense physiological streams from wearables by learning joint, continuous-time patient representations. Modality-specific encoders map clinical and wearable observations to a shared latent process, with multimodal attention and mixtures-of-exponentials kernels modeling cross-modal dependencies and timescales. Cross-modal self-supervised objectives—reconstructing masked tokens, forecasting signals, and asymmetric modality-to-modality prediction—yield temporally coherent and clinically grounded latent trajectories (Zhang et al., 18 Jan 2026). This multimodal foundation modeling improves long-horizon event prediction, physiological inference, calibration, and robustness under missing data, compared to EHR- or wearable-only approaches.

---

The evolution of analytical and generative frameworks for longitudinal EHRs, spanning interpretable deep sequence models, temporal registration, high-fidelity synthetic data generation, joint modeling under informative missingness, and process-grounded LLM reasoning, continues to expand the feasibility and rigor of clinical research, precision medicine, and deployment of fair, privacy-preserving ML models in healthcare environments.