🧬 biology

Contrastive Representation Learning of Longitudinal Disease Trajectories on Temporal Graphs

This paper proposes a contrastive representation learning framework that models multivariate disease trajectories as temporal graphs to generate robust embeddings for clustering patients with similar progression patterns and uncovering latent structures in longitudinal clinical data.

Original authors: Bastian Pfeifer

Published 2026-07-29

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Bastian Pfeifer

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to understand the story of a person's life, but instead of reading a book, you are looking at a scattered pile of snapshots taken at random times. Some photos are taken every day; others are taken once a year. Some show a happy birthday, others a doctor's visit, and some are just blurry pictures of a park bench. In the world of medical science, this is exactly what "longitudinal data" looks like: a collection of measurements taken from the same people over and over again, but often at messy, irregular intervals.

The big challenge for scientists is figuring out how to group these people together based on how their lives (or diseases) are changing. Are two patients following the same path, even if their check-ups happened on different days? Traditional methods often try to force these messy snapshots into neat, smooth lines, assuming everyone's story follows a predictable curve. But real life is rarely that smooth. Sometimes, people have sudden spikes in symptoms, or their conditions change in complex, non-linear ways that don't fit a simple line. This is where the field of "machine learning" steps in, specifically a branch called "representation learning." Think of this as teaching a computer to look at a messy pile of data and figure out the most important "essence" or "summary" of each person's story, so it can easily tell who belongs in which group. The paper you are about to read tackles the problem of how to do this best when the data is a tangled web of time and similarity.

The Problem: The Messy Timeline

Imagine you are a detective trying to solve a mystery by grouping suspects based on their daily habits. You have a list of notes for each person, but the notes are a mess. Some people wrote a diary entry every morning; others only wrote when something exciting happened. Some notes are about what they ate, others about how they felt, and some are just scribbles.

In medicine, doctors collect this kind of data all the time. They track patients over years, recording blood pressure, heart rates, and symptoms. But the data is "longitudinal" (stretched out over time) and "heterogeneous" (mixed up and different for everyone). The goal is to find "clusters"—groups of patients who are traveling down the same disease path. If you can find these groups, you can predict who might get sicker and tailor treatments better.

The problem is that old-school detective work (traditional statistics) often assumes everyone's story is a smooth, straight line. But diseases don't always move in straight lines. They zigzag, they stall, and they sometimes jump. Furthermore, standard methods often look at each patient in isolation, missing the fact that Patient A and Patient B might be moving in sync, even if they are different people.

The Solution: RankWalk and the "Time-Traveling Graph"

Enter the authors of this paper, who propose a new method called RankWalk. Instead of trying to force the data into a straight line, they decide to build a giant, invisible web (a "graph") that connects all the dots.

Here is how they build this web:

The Nodes (The Dots): Every single measurement a patient ever had becomes a dot on the map.
The Temporal Strings (The Time Travel): If Patient A had a check-up on Monday and another on Tuesday, they draw a string connecting those two dots. This preserves the order of time. It tells the computer, "This happened before that."
The Similarity Strings (The Soulmates): This is the clever part. If Patient A and Patient B both had a check-up on a Tuesday, and their health numbers looked very similar at that exact moment, the computer draws a string connecting them, even though they are different people. It's like saying, "Hey, you two are in the same boat right now."

But wait, what if the check-ups happened at weird times? What if Patient A was seen at 10:00 AM and Patient B at 10:05 AM? The authors use a "sliding window" trick. Imagine a window moving across a timeline. If a patient's check-up falls inside the window, they get grouped with everyone else in that window. This smooths out the messy, irregular timing without losing the story.

The Secret Sauce: The Anchor and the Random Walk

Now that the web is built, how does the computer learn which patients belong together? This is where Contrastive Learning comes in. Think of this as a game of "Spot the Difference."

The computer picks a "Anchor" (a specific patient's check-up) and tries to find other dots that look like it. But instead of just looking at the immediate neighbors, the computer sends out a "random walker"—a little explorer that hops from dot to dot along the strings.

Here is the twist: The explorer is anchor-guided. It doesn't just wander aimlessly. It is biased to look for dots that are structurally similar to the Anchor. If the Anchor is a "sick" patient, the explorer is more likely to find other "sick" patients, even if they are far away in the web.

The computer keeps a score of how fast the explorer finds a match. The faster it finds a similar patient, the more "important" that match is. This is called Rank-Weighted Positive Pair Generation. It's like saying, "If you found the twin in the first step of your search, that twin is a perfect match. If you had to search for ten steps to find a twin, maybe they aren't that similar."

Finally, the computer uses a "contrastive" objective. It tries to make the "twin" dots look very similar in its internal language (embeddings) and the "non-twin" dots look very different. Over time, the computer learns a super-efficient way to summarize every patient's journey into a short, powerful code.

What They Found: The Results

The authors tested their new "RankWalk" method against the old detectives (traditional statistics) and other new machine learning tools. They did this in two ways:

1. The Simulation Lab (The Fake Data)
First, they created fake patient data on a computer.

Scenario A (The Smooth Ride): They made up data where patients followed smooth, predictable curves. Here, the old methods (like fPCA) did well, but RankWalk kept up just as well.
Scenario B (The Noisy Ride): They added "noise"—random errors or bad data points—to the mix. The old methods got confused and started grouping people wrong. RankWalk, however, stayed calm. Because it looked at many different "subspaces" (different angles of the data) and used a ranking system, it ignored the noisy parts and found the true groups.
Scenario C (The Chaotic Ride): They created data where patients switched between different "regimes" (like a disease that suddenly changes its behavior) and had non-linear jumps. The old methods completely failed here because they couldn't handle the chaos. RankWalk, however, excelled, finding the groups with near-perfect accuracy. It proved that you don't need to know the shape of the curve in advance; the graph can learn it.

2. The Real World (The Real Data)
Next, they tested RankWalk on four real medical datasets involving heart disease, liver disease, cognitive aging, and AIDS.

They didn't just look at how well the groups were formed; they checked if the groups actually mattered for survival. They used a "Concordance Index" (a score of how well the groups predicted who would live longer) and a "Log-Rank Test" (a statistical check to see if the groups were truly different).
The Result: RankWalk consistently beat the best existing method (fPCA). For example, in the heart disease dataset (HEART), the old method had a survival prediction score of 0.57 (barely better than guessing), while RankWalk jumped it to 0.71. Even more striking, the statistical evidence that the groups were different skyrocketed from a weak 4.81 to a massive 52.25.
This suggests that by treating the data as a connected web of time and similarity, RankWalk found patterns that the old methods missed, leading to groups of patients that were much more clearly distinct in terms of their health outcomes.

The Takeaway

The paper suggests that we don't need to force messy, real-world medical data into neat, straight lines to understand it. By building a dynamic web that respects both the flow of time and the similarities between different people, and by using a smart "search and compare" strategy, we can uncover hidden patterns in how diseases progress.

RankWalk doesn't just group patients; it finds the right groups. It handles messy data, ignores noise, and adapts to complex, changing disease patterns better than the tools we've been using for decades. While the authors note that this is a methodological breakthrough and future work is needed to handle even more complex scenarios (like missing data or multiple types of data at once), the results so far suggest that this graph-based approach is a powerful new lens for viewing the long, winding road of human health.

Technical Summary: Contrastive Representation Learning of Longitudinal Disease Trajectories on Temporal Graphs

Problem Formulation
The paper addresses the challenge of clustering multivariate longitudinal clinical data to identify patient subgroups with similar disease progression patterns. Longitudinal data is characterized by repeated measurements from the same individuals over time, often featuring irregular sampling intervals, high dimensionality, and complex nonlinear dynamics. Traditional methods, such as mixed-effects models, functional principal component analysis (FPCA), and distance-based approaches like Dynamic Time Warping (DTW), often rely on restrictive assumptions (e.g., smooth underlying functions) or fail to explicitly exploit structural relationships between different subjects. While deep learning methods have emerged, they frequently lack explicit mechanisms to integrate temporal dependencies with cross-subject structural similarities in an unsupervised manner.

Methodology: The RankWalk Framework
The authors propose RankWalk, a novel unsupervised graph contrastive learning framework designed to learn robust representations for longitudinal clustering. The method operates through three primary components:

Robust Longitudinal Graph Construction:
- Nodes: Each longitudinal observation is treated as a graph node.
- Temporal Edges: Consecutive observations for the same subject are connected to preserve individual disease trajectories.
- Similarity Edges: To capture cross-subject relationships at specific time points, the method employs a robust rank-aggregation strategy. Instead of calculating distances in high-dimensional space directly (which is noise-sensitive), it samples multiple feature subspaces, computes pairwise distances within each, and aggregates the relative rankings. The $k$ -nearest neighbors based on this rank-aggregated distance form similarity edges.
- Handling Irregularity: For irregularly sampled data, the continuous time axis is partitioned into overlapping sliding windows. Observations within a window are aggregated (averaged) per subject, and similarity edges are constructed within each window, ensuring smooth transitions between temporal regions.
Anchor-Guided Structural Sampling:
- To identify informative positive pairs for contrastive learning, the method computes a structural affinity matrix using the Jaccard coefficient of node neighborhoods.
- An anchor-guided random walk is performed starting from an anchor node. Unlike standard random walks, transition probabilities are conditioned on the structural affinity to the anchor node, biasing the walk toward structurally similar nodes.
- Rank-Weighted Positive Pairs: Nodes visited earlier in the walk are assigned higher importance. The top- $K$ visited nodes form the set of positive pairs, weighted by their discovery rank to emphasize high-confidence structural matches.
Weighted Graph Contrastive Representation Learning:
- A Graph Neural Network (GNN) encoder, consisting of feature encoding, graph convolution layers with residual connections, and a projection head, maps nodes to low-dimensional embeddings.
- The model is trained using a weighted InfoNCE objective. Positive pairs are the anchor-guided samples, while negative pairs are sampled uniformly from the node population.
- The weighting mechanism assigns higher loss contributions to positive pairs discovered early in the random walk, reflecting higher confidence in their structural correspondence.

Key Contributions

Graph-Based Longitudinal Modeling: The paper introduces a framework that represents longitudinal data as a heterogeneous graph, explicitly integrating within-subject temporal evolution and cross-subject structural similarity.
Robustness to Noise and Irregularity: By utilizing rank-aggregation for similarity estimation and sliding windows for irregular time points, the method reduces sensitivity to noisy features and accommodates real-world sampling schedules without requiring predefined trajectory models.
Anchor-Guided Contrastive Learning: The novel sampling strategy leverages structural affinity to generate high-quality positive pairs for contrastive learning, avoiding the need for labeled data or pairwise distance matrices.

Results
The framework was evaluated on two simulation studies and four real-world biomedical datasets (PBC2, HEART, PAQUID, AIDS).

Simulation Studies:
- In a multivariate trajectory setting with smooth mean curves, RankWalk achieved performance comparable to specialized statistical methods (FPCA, DTW) and outperformed deep learning baselines (TS2Vec, VaDER).
- Crucially, under noisy conditions (where a subset of variables was corrupted), RankWalk demonstrated superior robustness compared to FPCA and DTW, maintaining high clustering accuracy due to its rank-aggregation strategy.
- In a nonlinear regime-switching scenario with latent dynamic processes, RankWalk significantly outperformed classical trajectory-based methods (kml3d, FPCA) and matched or exceeded temporal deep learning models, successfully capturing complex dynamic behaviors without explicit model assumptions.
Real-World Benchmarks:
- Evaluated using survival-based metrics (Concordance Index and Log-Rank test statistics) on four datasets.
- RankWalk consistently outperformed FPCA (the primary comparator for irregular data). Notably, on the HEART dataset, RankWalk improved the C-index from 0.57 to 0.71 and increased the log-rank $\chi^2$ statistic from 4.81 (non-significant) to 52.25 (highly significant), indicating a much stronger association between the learned clusters and patient survival outcomes.

Significance and Claims
The paper claims that representing irregular longitudinal data as graphs, which jointly model temporal dependencies and cross-subject similarities, enables the identification of trajectory patterns more strongly associated with clinical outcomes than traditional functional representations. The authors assert that RankWalk provides a flexible, unsupervised alternative that does not require predefined trajectory models, labeled data, or pairwise distance matrices, while remaining robust to measurement noise and irregular sampling schedules common in clinical settings. The work suggests that graph contrastive learning offers a powerful paradigm for uncovering latent structure in complex longitudinal biomedical data.