⚡ electrical engineering

GraD-IBD: Graph Representation Learning from Diagnosis Trajectories for Early Detection of Inflammatory Bowel Disease

This paper proposes GraD-IBD, a graph-based model that reformulates irregular ICD diagnosis trajectories into temporally directed graphs with a novel time-decay message passing mechanism to achieve efficient, scalable, and accurate early detection of Inflammatory Bowel Disease.

Original authors: Leo Y. Li-Han, Ellen L. Larson, Elizabeth B. Habermann, Cornelius A. Thiels, Hojjat Salehinejad

Published 2026-05-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Leo Y. Li-Han, Ellen L. Larson, Elizabeth B. Habermann, Cornelius A. Thiels, Hojjat Salehinejad

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a patient's medical history as a long, messy string of beads. Each bead is a diagnosis code (like "stomach pain" or "diarrhea") recorded every time they visit a doctor. For a disease like Inflammatory Bowel Disease (IBD), these beads are scattered irregularly—sometimes you get a whole handful in one week, sometimes none for months. Trying to find the pattern of IBD in this messy string is like trying to read a story where the words are jumbled and the sentences are missing.

The paper introduces a new tool called GraD-IBD to solve this puzzle. Here is how it works, using simple analogies:

1. Organizing the Mess: The "Bucket" System

First, the researchers realized that looking at every single day is too chaotic. So, they invented a "bucket" system.

The Analogy: Imagine you are sorting a year's worth of mail. Instead of looking at every single day, you put all the mail from one week into a single "bucket."
The Result: Instead of a long, messy string of beads, the patient's history becomes a neat stack of weekly buckets. Each bucket contains a list of what was diagnosed that week.

2. Building a Map: The "City" of Diagnoses

Next, they turned these buckets into a map.

The Analogy: Think of each diagnosis code as a city on a map.
- If a patient had "stomach pain" in Week 1 and "diarrhea" in Week 2, the map draws a one-way road from the "Stomach Pain City" to the "Diarrhea City."
- The width of the road depends on two things:
  1. How often the symptom appeared (a busy road vs. a dirt path).
  2. How similar the symptoms are (like connecting two cities that are neighbors).
The Time Factor: The map also has a "freshness" rule. Information from last week is considered "fresh" and carries a heavy weight. Information from six months ago is considered "stale" and fades away, just like a memory that gets fuzzy over time.

3. The Detective: The "Message Passing" Mechanism

Now, they built a digital detective (a Graph Neural Network) to walk this map.

The Analogy: Imagine a detective walking from city to city. When they arrive at a new city (a new week's diagnosis), they don't just look at that city alone. They ask their neighbors: "What did you see last week?"
The Twist: The detective listens more closely to neighbors who are:
1. Similar to the current city (context-aware).
2. Recent (time-decay).
3. Frequent visitors (high occurrence).
By gathering these clues, the detective builds a complete picture of the patient's health journey to decide: "Is this person likely to develop IBD?"

What Did They Find?

The researchers tested this system against other methods (like standard computer models that read the "string of beads" or other map-based models).

The Result: GraD-IBD was better at spotting IBD early than the other methods.
The Efficiency: It was also much faster and lighter. If the other models were like heavy, fuel-guzzling trucks, GraD-IBD was a nimble, electric scooter. It achieved the same (or better) results while using a fraction of the computer power.
The Long Game: The longer the time gap they tried to predict (e.g., predicting 6 months in advance vs. 1 month), the more GraD-IBD pulled ahead of the competition.

The Bottom Line

The paper claims that by turning messy medical records into a structured, time-aware map, they created a smarter, faster way to catch Inflammatory Bowel Disease early. This is especially helpful for people whose medical records might be sparse or scattered, giving doctors a better chance to spot the problem before it gets serious.

Note: The authors specifically mention this was tested on data from Mayo Clinic and validated against other diseases that look similar to IBD, but they caution that the model needs testing on data from other hospitals before it can be used everywhere.

Technical Summary: GraD-IBD

Problem Statement

The International Classification of Diseases (ICD) provides a standardized, globally recognized record of diagnostic events during patient encounters. However, utilizing longitudinal ICD code sequences for disease prediction presents significant challenges due to their irregular, unstructured nature and hierarchical complexity. Traditional sequential modeling methods (e.g., RNNs, Transformers) often struggle with these irregularities, leading to overly complex model designs and high computational costs.

Specifically, this paper addresses the early detection of Inflammatory Bowel Disease (IBD). IBD is difficult to diagnose in non-specialty settings due to nonspecific symptoms (e.g., abdominal pain, bleeding), often resulting in a 3–6 month delay between symptom onset and diagnosis. This delay leads to worse outcomes, particularly for adults under 50 who may have sparse medical records. The goal is to develop a model that can detect IBD risk based solely on diagnosis trajectories, distinguishing it from benign abdominal conditions with similar presentations, while maintaining computational efficiency.

Methodology: GraD-IBD

The proposed GraD-IBD (Graph Representation Learning from Diagnosis Trajectories) model reformulates irregular ICD trajectories into structured, temporally directed graphs to facilitate efficient learning. The methodology proceeds in three main stages:

1. Visit-Bucketized Frequency Encoding

To handle irregular visit intervals, the authors introduce a visit-bucketized frequency encoding paradigm.

Patient diagnostic histories are divided into time buckets of duration $\tau$ (e.g., weekly).
Diagnoses within a short period are treated as a holistic snapshot of the patient's status rather than distinct causal events.
This transforms the irregular time series into a structured matrix $X \in \mathbb{R}^{N \times T}$ , where $N$ is the number of unique ICD codes and $T$ is the number of buckets. The value $x_{c,t}$ represents the frequency of code $c$ in bucket $t$ .

2. ICD-Graph Construction

The matrix is further compressed into a temporally directed graph $G(V, E)$ for each patient:

Nodes ( $V$ ): Represent unique ICD codes observed in specific, chronologically ordered non-empty visit buckets. Recurring codes in different buckets are distinct nodes.
Edges ( $E$ ): Directed connections exist from all nodes in a bucket $b_i$ to all nodes in the subsequent non-empty bucket $b_{i+1}$ , capturing the temporal flow of the diagnosis history.
Edge Weights: Defined by $f(e) = x_{c,b_i} \cdot \text{Sim}(u_c, v_d)$ , where $x_{c,b_i}$ is the occurrence frequency and $\text{Sim}$ is the cosine similarity between the source and target node embeddings. This encodes both the frequency of diagnosis and the semantic similarity between codes.

3. Graph Neural Network with Context-Aware, Time-Decay Message Passing

The core of the model is a novel message-passing mechanism designed to update node features:

Node Embedding: Each node is initialized with a learnable embedding vector representing the ICD code type.
Message Passing: Features from preceding buckets are aggregated via a weighted sum based on edge weights. This ensures that similar and repeated diagnoses in prior visits carry more weight.
Time-Decay Factor: The aggregated message is multiplied by an exponential time-decay factor, $\exp(-\lambda \cdot \Delta t)$ , where $\Delta t$ is the time gap between buckets. This ensures that recent visits weigh more heavily than distant historical events.
Feature Update: The time-decayed message is added to the target node's embedding and passed through a fully connected network (FCN) with ReLU activation.
Classification: A global average pooling aggregates node embeddings into a fixed-length graph vector, which is then fed into a two-layer FCN for binary IBD classification.

Key Contributions

The paper claims three primary contributions:

Visit-Bucketized Frequency Encoding: A novel paradigm for translating irregular ICD time series into structured matrices, simplifying data representation.
Patient-Level ICD-Graph: The design of a temporally directed graph that reformulates diagnosis encodings, capturing both diagnostic co-occurrence and temporal dependencies.
GraD-IBD Model: The construction of a graph neural network featuring a context-aware and time-decay message passing mechanism. This allows the model to learn meaningful features while reducing complexity compared to sequential models.

Experimental Results

The model was evaluated on a real-world dataset from the Mayo Clinic Health System, comprising 1167 IBD patients (cases) and 7871 matched controls with benign abdominal conditions.

Performance: GraD-IBD achieved a mean AUROC of 0.752, AP of 0.597, and F1 score of 0.612 for a 1-month prediction interval. It outperformed state-of-the-art sequential models (Transformer, LSTM) and other graph models (GraphSAGE, GAT, GCN).
Efficiency: The graph-based approach demonstrated significant reductions in computational complexity. GraD-IBD required 23.495 million FLOPs and 0.172 million parameters, compared to the Transformer (419.680M FLOPs, 1.38M params) and LSTM (698.420M FLOPs, 2.31M params).
Ablation Study: Experiments confirmed that the combination of code similarity, occurrence frequency, and time decay yielded the best performance. Removing these context-aware or temporal components resulted in lower metrics, with the uniform-weight baseline (equivalent to standard GraphSAGE) performing the worst.
Sensitivity Analysis: The model maintained superior performance across prediction intervals ranging from 1 to 6 months. The performance gap between GraD-IBD and baseline models widened as the prediction interval increased (e.g., a 44% improvement in Average Precision over GraphSAGE for a 6-month interval).

Significance and Claims

The authors position GraD-IBD as the first work to leverage graph representation learning specifically for IBD management. The significance of the work lies in its ability to:

Enable Efficient Prediction: By reformulating irregular ICD trajectories into graphs, the model achieves high accuracy with significantly lower computational costs than sequential models, making it scalable for large-scale clinical applications.
Address Data Sparsity: The approach is particularly valuable for populations with sparse medical records (e.g., adults under 50), where traditional methods may struggle.
Improve Early Detection: The model's robust performance in predicting risk with lead times of up to 6 months suggests potential for reducing the 3–6 month diagnostic delay currently associated with IBD, thereby improving patient outcomes.

The paper concludes that graph representation learning offers a promising pathway for efficient, scalable, and accurate disease risk prediction from longitudinal ICD codes, though it acknowledges the need for external validation across different healthcare systems to confirm generalizability.