Recurrent Architectures for Longitudinal Patient Trajectories

This direction is split into other directions: temporal sequence modeling. Its last version stays readable and citable here.

Version 1 of 1164 papers179 conceptsfrozen 2026-10-04

Traditional clinical predictive models often treat a patient's medical history as a "bag of features"—a static snapshot of diagnoses, lab results, and demographics. However, health is inherently temporal. A single high blood pressure reading means something very different if it is a lifelong trend versus a sudden spike.

To capture this nuance, researchers have shifted toward modeling Electronic Health Records (EHRs) as sequences. By applying recurrent neural network (RNN) architectures to longitudinal data, these models can track the "trajectory" of a patient’s health, maintaining a memory of past events to better forecast future ones.

Medical History as a Language

The foundational intuition behind this direction is that a patient's clinical record functions much like a sentence. Just as the meaning of a word depends on the words that came before it, the significance of a medical event depends on the patient's history.

In Doctor AI: Predicting Clinical Events via Recurrent Neural Networks, Edward Choi and Jimeng Sun pioneered the use of Gated Recurrent Units (GRUs) to process these sequences. By treating clinical codes (diagnoses, medications, procedures) as tokens, the model learns an evolving "hidden state" that represents the patient's current physiological status. This allows the system to predict not just one outcome, but the entire future trajectory of clinical events.

Capturing the Hidden State of Health

The power of recurrent architectures like LSTMs and GRUs lies in their ability to maintain state across time steps. In a clinical context, this hidden state acts as a digital proxy for the patient's underlying health condition.

As Bjoern Eskofier and Jochen Klucken note in Predictive Models for Health Deterioration, these "temporally aware" models identify patterns of deterioration that emerge over time—patterns that are invisible to static diagnostic snapshots. For example, in pediatric critical care, Yuxuan Liu and colleagues used GRUs to synthesize fluctuating vital signs into a coherent assessment of extubation readiness, effectively filtering noise to find the true physiological trend.

Handling Irregularity and Sparse Data

Clinical data is notoriously messy. Patient visits occur at irregular intervals, and many records are sparse. Researchers have developed several strategies to make recurrent models more robust to these challenges:

  • Bidirectional Processing: Rafael T. Sousa demonstrated that using bidirectional LSTMs allows models to look both forward and backward through a patient's history, which is particularly effective for capturing dependencies in sparse financial billing records used for disease prediction.
  • Dynamic Graphs: Rather than treating visits as flat vectors, Chang Lu introduced Chet, a framework that models medical history as a dynamic disease graph. This approach uses modified recurrent units to track how specific diseases transition between "persistent" and "emerging" states for an individual over time.
  • Algebraic Embeddings: In Finding Algebraic Structure of Care in Time, Phuoc Nguyen and Truyen Tran embedded diseases and treatments into a continuous vector space. By feeding these into an LSTM, they could visualize how the cumulative effect of treatments modifies the underlying illness state across sequential visits.

From Risk Assessment to Treatment Recommendation

While early work focused on predicting mortality or readmission, the field has evolved toward Dynamic Treatment Recommendation (DTR). Here, the recurrent model doesn't just predict what will happen, but suggests what a clinician should do.

Lu Wang and Hongyuan Zha proposed SRL-RNN, which combines supervised learning with reinforcement learning. The RNN synthesizes fragmented temporal data into a coherent patient state, which then informs an agent's treatment policy. Similarly, Luchen Li and Aldo A. Faisal argue that modeling patient dynamics as a history-dependent sequence is essential for sepsis treatment, as current observations are often "partially observable" without the context of the preceding hours.

By moving from static data points to continuous trajectories, these models provide a more faithful representation of the human body's evolution, offering a path toward truly personalized, longitudinal medicine.

Papers in this version

Showing 60 of 164 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_recurrent_architectures_for_longitudinal_7f0d1375,
  title = {Recurrent Architectures for Longitudinal Patient Trajectories},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/recurrent-architectures-for-longitudinal-patient-trajectories/rdr_7f0d13754b2541c0917ff00a5f3c7a3e},
  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_recurrent_architectures_for_longitudinal_7f0d1375_v1,
  title = {Recurrent Architectures for Longitudinal Patient Trajectories},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/recurrent-architectures-for-longitudinal-patient-trajectories/rdr_7f0d13754b2541c0917ff00a5f3c7a3e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}