Standard deep learning architectures—from Recurrent Neural Networks (RNNs) to Transformers—share a common "uniformity assumption." They treat data as a sequence of equidistant steps ($t_1, t_2, t_3$), much like words in a sentence or frames in a video.
However, Electronic Health Records (EHRs) are fundamentally "messy." A patient might visit a clinic twice in one week, then not again for six months. Vital signs might be recorded every hour in the ICU but only once a day in a general ward. As Zhengping Che and Yan Liu point out in Exploiting Convolutional Neural Network for Risk Prediction with Medical Feature Embedding, ignoring these irregular intervals discards critical context regarding disease progression and urgency.
The research direction emerging from these 198 papers focuses on moving beyond discrete sequences to models that treat time as a continuous, informative dimension.
Time as a Clinical Signal
In medicine, the timing of an observation is often as predictive as the observation itself. This concept, known as "informative sampling," suggests that the frequency of medical encounters reflects a clinician's underlying suspicion. Ahmed Alaa and Mihaela van der Schaar demonstrate this in Learning from Clinical Judgments: Semi-Markov-Modulated Marked Hawkes Processes for Risk Prognosis, where they treat the timing of tests as a vital signal. A sudden "burst" of lab tests often precedes a clinical crash, even if the lab values themselves haven't yet crossed a danger threshold.
Adapting Discrete Models with Time-Awareness
The first wave of solutions involves modifying standard architectures to "sense" the gaps between events.
One common approach is the use of time-decay functions. In ATTAIN: Attention-based Time-Aware LSTM Networks for Disease Progression Modeling, Xi Yang and colleagues introduce a mechanism that discounts the influence of past events based on how much time has elapsed. If a patient had a high fever six months ago, it should weigh less on a current risk prediction than a fever recorded six hours ago.
Similarly, Phuoc Nguyen and Truyen Tran in Deepr: A Convolutional Net for Medical Records treat time gaps as literal "words" in a patient's clinical sentence. By embedding these gaps, the model learns to recognize motifs—like a specific sequence of treatments followed by a long period of stability—that signal recovery or chronic management.
The Shift to Continuous-Time Dynamics
While time-aware gates help, they still operate on a discrete grid. A more robust approach treats patient health as a continuous latent trajectory.
Neural Ordinary Differential Equations (ODEs) and Controlled Differential Equations (CDEs) have become the state-of-the-art for this problem. Instead of jumping from visit to visit, these models define the derivative of the patient's state. Ningtao Liu introduces GRU-TV in GRU-TV: Time- and velocity-aware GRU for patient representation on multivariate clinical time-series data, which uses Neural ODEs to track the "velocity" of physiological change. This allows the model to estimate a patient's state at any arbitrary point in time, effectively "filling in the blanks" between sparse observations.
For counterfactual reasoning—predicting "what if" a different treatment was given—Nabeel Seedat developed TE-CDE in Continuous-Time Modeling of Counterfactual Outcomes Using Neural Controlled Differential Equations. By modeling treatment effects as continuous interventions, the system can reason about varying doses and timings that don't align with the original observation schedule.
Probabilistic Interpolation and Imputation
Another facet of this research focuses on the "missingness" inherent in irregular data. Rather than simple linear interpolation, researchers use Gaussian Processes (GPs) to provide uncertainty-aware adapters.
Benjamin M. Marlin and Steven Cheng-Xian Li pioneered the use of GP adapters in A scalable end-to-end Gaussian process adapter for irregularly sampled time series classification. This framework maps irregular observations onto a fixed-dimensional latent space that standard models can process. Building on this, Michael Moor and Karsten Borgwardt developed MGP-TCN in Early Recognition of Sepsis with Gaussian Process Temporal Convolutional Networks and Dynamic Time Warping. By combining GPs with Temporal Convolutional Networks, they can detect sepsis hours earlier by explicitly modeling the uncertainty caused by irregular sampling.
Why This Works
These methods succeed because they align the model's architecture with the underlying biological reality. Biological processes do not pause between doctor visits; they evolve continuously. By incorporating time-decay, continuous-time dynamics, or probabilistic interpolation, these models stop treating "time" as a mere index and start treating it as a physical dimension where the rate of change is a primary feature of the disease itself.