Continuous Time Modeling of Asynchronous Clinical Observations

This direction is merged into another direction: temporal data sparsity. Its last version stays readable and citable here.

Version 1 of 1120 papers124 conceptsfrozen 2026-10-04

Most machine learning models for time series assume a "metronome" view of the world: data arrives at fixed, predictable intervals, like the frames of a video or the daily closing price of a stock. However, Electronic Health Records (EHR) are fundamentally different. They are "event-driven." A patient’s data is only recorded when they interact with the healthcare system—perhaps a flurry of lab tests during an ICU stay, followed by months of silence, and then a single follow-up visit.

This creates a dual challenge of irregular sampling and extreme sparsity. Because measurements are taken based on clinical need rather than a fixed schedule, the "gaps" between data points are not just missing values—they are varying durations of unobserved physiological change. Researchers are developing new architectures to transform these scattered snapshots into continuous "patient trajectories."

The Problem with the Fixed-Grid Assumption

Standard sequence models, such as Recurrent Neural Networks (RNNs) or standard Transformers, typically treat each new observation as the "next" step in a sequence ($t, t+1, t+2$). In clinical settings, this is misleading. As Jiheum Park and Chao Pang note in CEHR-XGPT, clinical events occur at unpredictable intervals. Treating a gap of two hours the same as a gap of two months causes models to lose the "temporal signal"—the understanding of how much a patient’s condition could have evolved during the silence.

Furthermore, EHR data is "multivariate sparse." At any given doctor visit, a clinician might measure blood pressure but not cholesterol. This results in a 3D tensor (Patients $\times$ Features $\times$ Time) where the vast majority of cells are empty. Edward De Brouwer and Yves Moreau describe this as a "holey" dataset that prevents standard models from maintaining a consistent hidden state (Deep Ensemble Tensor Factorization).

Bridging the Gaps: Interpolation and Imputation

One strategy to handle this irregularity is to "clean" the data before reasoning about it. Satya Narayan Shukla and Benjamin M. Marlin proposed Interpolation-Prediction Networks, which decouple the process into two stages: an interpolation layer that learns to "fill in" a continuous representation of the patient, followed by a prediction network that reasons over this smoothed signal (Modeling Irregularly Sampled Clinical Time Series).

Similarly, David Sontag and Narges Razavian developed differentiable multivariate kernel regressions that learn how to impute missing lab values directly via backpropagation (Temporal Convolutional Neural Networks for Diagnosis from Lab Tests). The intuition here is that the model shouldn't just guess a missing value; it should learn the relationship between different lab tests over time to reconstruct the most likely physiological state.

Continuous-Time Modeling

Rather than forcing irregular data onto a fixed grid, a second school of thought uses mathematical frameworks that are inherently continuous.

  1. Gaussian Processes (GPs): Yanbo Xu and Yanxun Xu use GPs to "borrow strength" across similar patients, allowing the model to estimate individualized treatment-response curves even when an individual's data is sparse (A Bayesian Nonparametic Approach for Estimating Individualized Treatment-Response Curves).
  2. Neural Ordinary Differential Equations (ODEs): For diseases like Alzheimer's, where progression is biologically irreversible, Seungwoo Jeong and Heung-Il Suk use Neural ODEs to model brain atrophy as a continuous-time dynamical system. This allows the model to predict a patient's status at any future time point, regardless of when their last scan was taken (Deep Geometric Learning with Monotonicity Constraints for Alzheimer's Disease Progression).

Time-Aware Architectures

Recent work has focused on adapting the Transformer architecture—the current gold standard for sequences—to the clinical domain. The challenge is that the standard "positional encoding" in Transformers only tells the model the order of events, not the duration between them.

  • Temporal Tokens: CEHR-XGPT treats time itself as a discrete token, allowing the model to learn the "rhythm" of clinical encounters.
  • Axial Attention: Rachael DeVries and Ole Winther introduced the Bi-Axial Transformer (BAT), which applies attention across both the clinical variables and the time points simultaneously. This helps the model capture how a specific lab result (like creatinine) relates to other variables across varying time intervals (Bi-Axial Transformers).
  • Extrapolatable Embeddings: Ziyang Song and Hao Xu developed TimelyGPT, which uses specialized position embeddings to solve the "extrapolation problem," enabling the model to forecast health outcomes much further into the future than the sequences it saw during training (Timelygpt).

The Significance of the "Missing" Signal

A critical insight across this research is that missing data in healthcare is rarely "missing at random." As Alexis Bellot and Mihaela van der Schaar argue, the fact that a doctor chose to order a test is itself a data point. Their BNDS model treats missingness as a meaningful signal of clinical intent, which often improves survival predictions (A Bayesian Approach to Modelling Longitudinal Data in Electronic Health Records). By moving away from simple imputation and toward models that embrace temporal irregularity, researchers are building systems that better reflect the complex, non-linear reality of human health.

Papers in this version

Showing 60 of 120 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_continuous_time_modeling_of_asynchronous_06ba411d,
  title = {Continuous Time Modeling of Asynchronous Clinical Observations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-time-modeling-of-asynchronous-clinical-observations/rdr_06ba411d236a4412bff82b3afef9eaef},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_continuous_time_modeling_of_asynchronous_06ba411d_v1,
  title = {Continuous Time Modeling of Asynchronous Clinical Observations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-time-modeling-of-asynchronous-clinical-observations/rdr_06ba411d236a4412bff82b3afef9eaef/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}