In many scientific fields, we observe systems evolving over time—a patient’s health, a student’s skills, or a researcher’s career. The standard way to analyze this data is to align observations by chronological time: comparing two patients because they are both 70 years old, or two students because they are both in their third week of a course.
However, chronological time is often a poor proxy for the underlying process. As noted in Metabolic Regulation in Progression to Autoimmune Diabetes by Marko Sysi‐Aho and colleagues, biological progression rarely occurs at a uniform rate. One patient might be a "fast progressor" while another is slow, meaning that at the same age, they occupy entirely different biological states. This temporal heterogeneity creates a "blurring" effect in data, where shared patterns are obscured by the fact that they occur at different times and speeds across a population.
The Challenge of Asynchrony
The core problem is that most longitudinal datasets are asynchronous. This manifests in three main ways:
- Mismatched Start Points: In clinical settings, patients are often diagnosed at different stages of a disease's lifecycle. As Dev Goyal and Jenna Wiens argue in Clinically Meaningful Comparisons Over Time, forcing a "start-to-finish" match on these trajectories leads to inaccurate assessments because the initial observations do not represent the same biological baseline.
- Variable Rates of Change: Even if two processes start together, they may diverge in speed. Edward De Brouwer highlights in Deep Ensemble Tensor Factorization for Longitudinal Patient Trajectories Classification that comparing data at fixed chronological intervals is ineffective when the biological state of one individual does not correspond to the same state in another at that same interval.
- Event-Driven Dynamics: In fields like neuroscience, specific stimuli or "event markers" occur at different absolute times. John P. Cunningham and Zoubin Ghahramani demonstrate in Gaussian Processes for time-marked time-series data that traditional alignment to a single reference point fails to capture the dynamics surrounding these multiple, asynchronous events.
Decoupling What from When
To solve this, researchers have moved toward models that explicitly disentangle the state (what is happening) from the timing (when it happens).
A landmark approach in this area is SuStaIn (Subtype and Stage Inference), developed by Alexandra L. Young and Neil P. Oxtoby. In Uncovering the heterogeneity and temporal complexity of neurodegenerative diseases, they show that by treating disease progression as a sequence of stages, one can identify distinct disease subtypes while simultaneously mapping where an individual sits on their specific trajectory. This prevents the model from confusing a "late-stage" patient of one subtype with an "early-stage" patient of another.
Similarly, Alfred O. Hero and Arnau Tibau Puig introduced Order-Preserving Factor Analysis (OPFA) in Order-preserving factor analysis (OPFA). OPFA enforces a consistent sequence of underlying factors across subjects but allows for individual timing delays, uncovering shared patterns that standard methods like PCA would miss due to temporal misalignment.
Mapping the Latent Timeline
If chronological time is unreliable, we must construct a latent timeline—a "disease clock" or "process time" that aligns all individuals.
One way to do this is through probabilistic generative models. In Monotonic Gaussian Process for Spatio-Temporal Trajectory Separation in Brain Imaging Data, Clément Abi Nader and Marco Lorenzi use monotonic Gaussian Processes to align patients onto a common timeline using only single "snapshot" scans. By assuming that certain biological markers (like amyloid buildup) must increase over time, the model can infer the relative stage of each patient.
More recently, researchers have turned to Neural Ordinary Differential Equations (Neural ODEs) to model these transitions as continuous flows. Tingsong Xiao and colleagues developed Temporally Detailed Hypergraph Neural ODEs, which combine the continuous-time modeling of ODEs with hypergraphs to capture the unique clinical contexts and individual rates of change in chronic diseases like Type 2 Diabetes.
Beyond Medicine: A Shared Framework
While much of this work is grounded in healthcare, the problem of asynchronous progression is universal.
- Education: Yu Ito uses synchronization techniques in Extracting Latent Skills from Time Series of Asynchronous and Incomplete Examinations to track how student skills evolve despite irregular testing intervals.
- Social Science: Sandra Servia-Rodríguez explores the "rhythm" of academic success in The Evolution of Your Success Lies at the Centre of Your Co-Authorship Network, noting that career trajectories are difficult to compare because scholars start at different times and vary in longevity.
- Natural Language: The Dynamic Author-Persona (DAP) model by Robert A. Giaquinto and Arindam Banerjee 0001 in Topic Modeling on Health Journals addresses the high temporal variance in patient narratives, where individuals post at irregular intervals and start their journeys at different relative points.
By moving away from the rigid constraints of chronological time and embracing the latent, asynchronous nature of progression, these methods allow us to see the shared structure in seemingly chaotic longitudinal data.