🤖 machine learning

xMICD: Explainable Representation of Multiple ICD Codes

The paper introduces xMICD, a novel method that bridges the gap between predictive performance and interpretability in clinical machine learning by constructing low-dimensional patient representations from ICD codes that leverage pre-trained embedding similarities while remaining anchored to clinically meaningful diagnostic groups.

Original authors: Pat Vatiwutipong, Kumkup Keeratisiwakul, Albert Phuoc Kien Van Truong, Nutcha Yodrabum, Wasin Pansiritanachot, Marvin N. Wright, Thanapon Noraset

Published 2026-08-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pat Vatiwutipong, Kumkup Keeratisiwakul, Albert Phuoc Kien Van Truong, Nutcha Yodrabum, Wasin Pansiritanachot, Marvin N. Wright, Thanapon Noraset

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 you are trying to teach a super-smart robot how to be a doctor. To do this, you feed it millions of patient records, hoping it can spot patterns to predict who might get sick or need emergency care. But there's a catch: the language doctors use is a massive, messy codebook called ICD (International Classification of Diseases). It's like trying to teach the robot by handing it a dictionary with 70,000 words, where most of the pages are blank for any single patient. If you just show the robot the raw codes, it gets overwhelmed by the sheer size and emptiness of the data.

On the other hand, if you try to simplify things by just grouping similar diseases together (like putting all "broken bones" in one bucket), the robot gets confused because it loses the tiny, crucial details that make one fracture different from another. This creates a frustrating tug-of-war in the world of medical AI: you can have a model that is super accurate but acts like a "black box" (nobody knows why it made a decision), or you can have a model that is easy to understand but isn't very good at predicting the future. The big question is: Can we build a system that is both a genius at predicting health outcomes and clear enough for a human doctor to trust?

This is exactly the puzzle a team of researchers tackled in their new paper, introducing a clever new method called xMICD. Think of xMICD as a "translator" that sits between the messy codebook and the robot's brain. Instead of just saying "Yes, this patient has a broken bone" or "No, they don't," xMICD creates a nuanced score. It asks, "How much does this patient's collection of symptoms feel like a specific category of illness?"

Here is how it works in plain terms: Imagine you have a set of reference points, or "anchors," representing different types of health issues, like "Respiratory Trouble" or "Heart Woes." When a patient walks in with a list of diagnoses, xMICD doesn't just check boxes. Instead, it looks at how similar the patient's specific codes are to those anchors using a special mathematical map (called an embedding space). If a patient has a diagnosis that isn't exactly "Heart Woes" but is very close to it on the map, xMICD gives them a high score for that category, even if they don't have the exact code. It's like saying, "This patient isn't a perfect match for the 'Heart' group, but they are leaning heavily in that direction."

The researchers tested this idea on huge databases of real hospital records, including over 400,000 emergency room visits and data from intensive care units. They found that xMICD is a real game-changer. It manages to be just as good at predicting serious outcomes (like whether a patient will die in the ICU or need to be readmitted) as the complex, "black box" deep learning models that usually win these competitions. But unlike those black boxes, xMICD's results are easy to explain. When the model flags a patient as high-risk, a doctor can look at the output and say, "Ah, the system is worried because their symptoms are very similar to the 'Respiratory Failure' group."

The paper suggests that this approach successfully bridges the gap between raw data power and human understanding. It proves that you don't have to sacrifice accuracy to get clarity. By using these "relative assignment" scores instead of simple yes/no switches, xMICD keeps the rich, hidden relationships between diseases that smart computers love, while organizing them into buckets that human doctors recognize. The result is a tool that helps machines make better guesses while keeping the reasoning transparent enough for humans to trust and verify.

Technical Summary: xMICD – Explainable Representation of Multiple ICD Codes

Problem Statement
Electronic Health Records (EHRs) rely heavily on International Classification of Diseases (ICD) codes to summarize patient diagnoses. However, integrating these codes into machine learning (ML) models presents a fundamental trade-off between predictive performance and interpretability.

High-Dimensional Sparsity: Direct one-hot encoding of ICD codes results in extremely high-dimensional and sparse feature spaces (e.g., ~70,000 codes in ICD-10-CM), which increases computational costs and hinders model generalization.
The Trade-off: Existing solutions generally fall into two categories, each with significant limitations:
1. Grouping-based representations: These map codes to clinically defined categories (e.g., ICD blocks, comorbidity indices like CCI or ECI) using binary indicators. While highly interpretable, they lose granular information by treating heterogeneous diseases within a group as equivalent, often leading to suboptimal predictive performance.
2. Embedding-based representations: Methods like ICD2Vec learn dense vector representations that capture semantic relationships and co-occurrence patterns. While these achieve strong predictive accuracy, the resulting dimensions are abstract latent variables lacking direct clinical meaning, rendering them "black boxes" that are difficult for clinicians to interpret.

Methodology: xMICD
The authors propose xMICD (Explainable Representation of Multiple ICD Codes), a method designed to construct low-dimensional patient representations that retain the semantic richness of embeddings while anchoring features to clinically meaningful concepts.

The core mechanism involves a relative assignment process rather than binary membership:

Anchors: A set of reference points (anchors) is defined. These can be clinically defined groups (e.g., ICD-10-CM blocks, CCSR categories, ECI groups) or data-driven clusters (e.g., medoids from farthest point sampling).
Similarity Computation: For a patient visit containing a set of ICD codes $\{C_1, \dots, C_n\}$ , the method computes the similarity between each code $C_i$ and every anchor $A_j$ within a pre-trained embedding space (specifically using ICD2Vec embeddings and cosine similarity in this study).
Normalization: To focus on relative importance rather than absolute similarity, the similarities for each code are normalized via min-max scaling across all anchors. This maps the most similar anchor to 1 and the least to 0 for that specific code.
Aggregation: The final representation for a visit is a vector where each dimension $S_j$ corresponds to an anchor. The value is calculated as the maximum normalized similarity across all codes in the visit for that anchor:
$S_j = \max_{i} \left( \text{normalized\_sim}(C_i, A_j) \right)$
This ensures that if a patient's diagnosis is strongly related to a specific clinical group (even if not strictly inside it), that dimension reflects a high value.

Key Contributions

Novel Representation: Introduction of xMICD, which converts multiple ICD codes into a low-dimensional, clinically structured feature vector suitable for standard ML models.
Hybrid Integration: Successful integration of pre-trained ICD embeddings (capturing latent semantic relationships) with clinically defined groupings (providing interpretability).
Empirical Validation: Comprehensive evaluation on large-scale EHR datasets (MIMIC-IV-ED and eICU-CRD) demonstrating that xMICD balances predictive power with feature-level transparency.

Results
Experiments were conducted on two datasets using XGBoost and Multilayer Perceptron (MLP) models for tasks including critical outcome prediction, hospitalization, 72-hour revisit, in-ICU mortality, and prolonged stay.

Predictive Performance: xMICD achieved performance comparable to the state-of-the-art embedding-based method (ICD2Vec) across most tasks.
- In the MIMIC-IV-ED dataset (high code diversity), xMICD significantly outperformed binary grouping methods and matched ICD2Vec.
- In the eICU-CRD dataset (lower code diversity), raw ICD codes performed similarly to embeddings, but xMICD still maintained competitive performance.
- The only exception was xMICD based on ECI anchors (31 dimensions), which underperformed due to excessive dimensionality reduction and information loss.
Similarity Preservation: Analysis of patient-patient similarity matrices showed that xMICD variants preserved the relational structure of ICD2Vec much better than binary grouping methods (Spearman correlation of ~0.79 for xMICD-FPS vs. ICD2Vec, compared to ~0.21–0.33 for binary groups).
Interpretability: Qualitative assessment based on readability, understandability, meaningfulness, trackability, and simulatability revealed:
- Unlike ICD2Vec, xMICD features correspond to recognizable diagnostic groups.
- Unlike binary groups, xMICD captures "diagnostic proximity," allowing clinicians to see how a patient's condition relates to a group even without a direct code match (e.g., identifying latent critical states).
- SHAP value analysis confirmed that xMICD explanations are structured around clinical concepts (e.g., "respiratory failure," "infections") rather than abstract latent dimensions.

Significance and Claims
The paper claims that xMICD offers a practical solution to the "interpretability vs. accuracy" dilemma in clinical ML. By anchoring embedding-based semantic relationships to predefined clinical groups, xMICD allows models to leverage the predictive power of deep learning while producing features that clinicians can reason about.

The authors emphasize that high interpretability does not necessarily require sacrificing predictive power. xMICD serves as a structured projection of embedding spaces onto clinically meaningful domains, facilitating the development of transparent and trustworthy AI-assisted decision support systems. The method is particularly valuable in settings with high ICD code diversity where dimensionality reduction is critical, yet clinical explainability remains a prerequisite for adoption.

Technical Summary: xMICD – Explainable Representation of Multiple ICD Codes

More like this