🧬 biology

A Multi-Dimensional Clustering Approach for Identifying Inborn Errors of Immunity

This paper proposes a machine learning pipeline that curates and transforms raw immunologic data from electronic health records into vectors for unsupervised clustering, aiming to identify novel patterns and features of inborn errors of immunity to facilitate early diagnosis and improve rare disease analysis.

Original authors: Nishad Kulkarni, Alexandra K. Martinson, Nicholas L. Rider, Michael Keller, Syed Muhammad Anwar

Published 2026-05-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nishad Kulkarni, Alexandra K. Martinson, Nicholas L. Rider, Michael Keller, Syed Muhammad Anwar

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

The Big Picture: Finding Patterns in a Chaotic Library

Imagine the medical records of patients with rare immune diseases (called Inborn Errors of Immunity or IEI) as a massive, chaotic library. The books are written in different languages, some pages are missing, and the stories are so complex that doctors often struggle to find the right diagnosis quickly.

The authors of this paper wanted to build a smart librarian (a computer program) that could sort through these messy records to find hidden patterns. Their goal wasn't to cure the patients directly, but to organize the data in a way that reveals new "sub-groups" of diseases that doctors might have missed.

The Ingredients: Cleaning the Data

Before the computer could start sorting, the team had to prepare the ingredients. Think of this like prepping a giant salad where every ingredient comes from a different farm with different measuring cups.

The Data Source: They pulled information from a huge national database of hospital records (Electronic Health Records or EHR).
The "Test" Ingredients: They focused on five specific blood tests that measure different types of immune cells (like T-cells and B-cells). These are the "ingredients" they used to describe each patient.
The Age Groups: Since a child's immune system changes rapidly as they grow (like a caterpillar turning into a butterfly), the team split the patients into six different age groups (from babies to adults). They analyzed each group separately because a 1-year-old's "normal" blood test looks very different from a 15-year-old's.
Fixing Missing Pages: Sometimes, a patient's file was missing a specific test result. Instead of throwing the file away, the computer used a smart guessing method (called MICE) to fill in the blanks based on what other similar patients had.
Standardizing the Cups: Different hospitals use different "normal" ranges for blood tests. The team converted all these numbers into a standard scale (0 to 1), so a result from Hospital A could be compared fairly to Hospital B.

The Sorting Machine: The Clustering Algorithms

Once the data was clean and standardized, they used Machine Learning to sort the patients. Imagine you have a giant box of mixed-up LEGO bricks. You don't know the final picture, but you want to group them by color and shape automatically.

The team tried five different sorting algorithms (mathematical rules for grouping):

K-means: Like sorting bricks into buckets based on how close they are to a central "average" brick.
DBSCAN: Like finding clusters of bricks that are huddled tightly together, ignoring the ones that are scattered alone.
Hierarchical: Like building a family tree of the bricks, starting with small groups and merging them into bigger ones.

They tested thousands of different settings (like changing the size of the buckets or how close the bricks need to be) to see which method created the most logical groups.

The Results: What Did They Find?

The computer successfully grouped the patients. Here is what happened:

Grouping by Disease: The algorithm naturally grouped patients who had the same known disease together. For example, patients with DiGeorge Syndrome (DGS) tended to end up in the same "neighborhood" as other DGS patients. This proved the computer was working correctly.
Finding Sub-Groups: Even within the DGS group, the computer found smaller sub-groups.
- Analogy: Imagine a group of people who all have "Headaches." The computer noticed that some of these people also had "Stomach Aches," while others had "Sore Throats." It split the "Headache" group into two smaller, more specific groups.
- In the study, they found that DGS patients in one cluster mostly had stomach/feeding issues, while DGS patients in another cluster mostly had heart or airway issues.
The Winners: The two best "sorting machines" were K-means and Agglomerative clustering. They created the most useful groups based on a special score the team invented to measure how "pure" and distinct the groups were.

The Limitations: What the Computer Didn't Do

The authors were careful to say what this study didn't do:

It didn't diagnose new patients: This was a "proof of concept" study. They didn't use the tool to diagnose a new patient in a hospital yet.
It didn't look at genetics: They only used blood test numbers, not DNA data.
The "Big Fish" Problem: Because they had way more patients with two specific diseases (DGS and Agammaglobulinemia) than others, the computer mostly grouped those two diseases. The smaller disease groups were harder to sort out. The authors admit this might have hidden some finer details.

The Takeaway

This paper is like building a prototype for a new filing system. The authors showed that if you take messy, real-world hospital data, clean it up, and use smart computer sorting, you can organize rare disease patients into meaningful groups.

They demonstrated that this method can:

Confirm that patients with the same disease look similar in their blood work.
Spot hidden "sub-groups" within a disease that might have different symptoms (like heart issues vs. stomach issues).

The ultimate goal is to use this system to help doctors spot these patterns earlier, but for now, the study is a successful test run showing that the "smart librarian" works.

Technical Summary: A Multi-Dimensional Clustering Approach for Identifying Inborn Errors of Immunity

Problem Statement
Inborn Errors of Immunity (IEI) are rare diseases characterized by complex symptomatology that often leads to delayed diagnosis and prolonged healthcare utilization. Current challenges include difficulties in accessing and curating large-scale Electronic Health Record (EHR) data for routine analysis, and a lack of established machine learning (ML) methodologies for systematically processing and integrating complex immunologic data. While clustering algorithms have been applied to medical imaging and specific modalities like flow cytometry, their application to multi-center laboratory data for IEI pattern recognition and subphenotyping remains underexplored. Existing approaches often rely on categorical data or imputation across entire datasets without accounting for disease-specific or age-specific variations.

Methodology
The authors propose a data curation and unsupervised ML pipeline designed to transform raw immunologic lab data into clinically interpretable vectors for pattern recognition.

Data Source and Cohort: The study utilizes data from the Cerner Real-World Database (CRWD), comprising 5,986 laboratory samples from 2,174 unique patients. The cohort includes seven specific IEI diagnoses: Chronic Granulomatous Disease (CGD), Combined Immunodeficiency (COMB), Common Variable Immunodeficiency (CVID), DiGeorge Syndrome (DGS), Leukocyte Adhesion Deficiency (LAD), Agammaglobulinemia (AGAMMA), and Wiskott-Aldrich Syndrome (WAS).
Data Structuring and Preprocessing:
- Encounter Definition: Data is structured by patient encounters. To represent disease progression rather than a single snapshot, the pipeline selects the first occurrence of each available laboratory test during a hospital stay.
- Feature Selection: The analysis is limited to five clinically prominent lymphocyte subset markers: CD3+, CD3+CD4+, CD3+CD8+, CD19+, and CD3-CD16+CD56+ (Natural Killer cells).
- Normalization: Lab values are normalized against institution-specific reference ranges (low and high) to account for variability across clinical sites. The normalization function scales absolute lab values relative to these ranges.
- Missing Data Imputation: To address missing values common in EHRs, the study employs Multiple Imputation by Chained Equations (MICE), applied specifically within disease-specific and age-subgrouped datasets.
- Age Stratification: Patients are divided into six age groups (0-1, 1-2, 2-5, 5-10, 10-16, and >16 years) based on immunology studies regarding lymphocyte subpopulation changes.
Clustering Algorithms: The pipeline tests five unsupervised clustering algorithms: DBSCAN, Hierarchical DBSCAN (HDBSCAN), K-means, Agglomerative, and K-modes.
Evaluation Metrics:
- Internal Validation: Silhouette score and Davies-Bouldin (DB) score.
- External Validation: Homogeneity, Completeness, V-Measure, Adjusted Rand Index, and Adjusted Mutual Information (using known disease labels blinded during the clustering process).
- Composite Score (CS): A custom metric is defined to rank hyperparameter combinations: $CS = (1 / (1 + DB)) + \sum s_i$ , where $s_i$ represents the other validation scores. This formulation penalizes high DB scores (which indicate poor clustering) while aggregating other performance metrics.
Dimensionality Reduction: For visualization, data vectors undergo Principal Component Analysis (PCA) followed by t-SNE to project 5-dimensional data into 2D space.

Key Contributions

Data Processing Pipeline: The study demonstrates a detailed methodology for converting complex, multi-center clinical lab data into a format interpretable by clustering algorithms, addressing issues of data missingness, reference range normalization, and dimensionality.
Algorithmic Application: It leverages a suite of centroid-based, density-based, and hierarchical ML clustering algorithms on laboratory data for patients with presumed IEI.
Subphenotype Identification: The use of unsupervised clustering successfully identifies disease subgroups (sub-phenotypes) that were not explicitly labeled in the source data, suggesting potential for discovering new disease patterns.

Experimental Results
The study conducted 3,144 experiments across six age groups using 524 hyperparameter combinations.

Performance: The best-performing algorithms based on the Composite Score were K-means (CS 2.94) and Agglomerative clustering (CS 2.83). Density-based algorithms (DBSCAN, HDBSCAN) produced fewer clusters or very small cluster sizes relative to the dataset.
Cluster Analysis: In the 0-1 year age group, Agglomerative clustering (with 7 clusters) revealed that most clusters were dominated by DiGeorge Syndrome (DGS) patients.
Clinical Validation: Within DGS-dominated clusters, the authors observed distinct differences in the frequency of non-IEI diagnostic codes (ICD). For instance, Cluster 0 patients had high frequencies of gastrointestinal/feeding codes, Cluster 3 showed upper airway pathology, and Clusters 4 and 6 exhibited cardiac-related codes, with Cluster 6 uniquely showing aortic anomalies.
Population Bias: The majority of formed clusters corresponded to the two largest population groups in the dataset (AGAMMA and DGS), reflecting real-world prevalence imbalances. However, patients with the same disease type tended to cluster together, suggesting the model captures biologically meaningful structure rather than just sample size effects.

Significance and Claims
The paper claims that systematic clustering of large-scale rare disease data has the potential to identify novel, clinically relevant patterns and sub-phenotypes. The study highlights that:

Age-based analysis is critical: Subdividing data by age is essential for pediatric immunologic evaluation due to rapidly changing immune systems.
Imputation strategy matters: Using disease and age-specific imputation (MICE) allows for the preservation of large datasets and disease-specific differences.
Unsupervised learning utility: The approach can group patients by disease and further subdivide them based on diagnostic and symptom code frequencies, potentially aiding in earlier diagnosis and targeted treatment.
Limitations and Future Work: The authors acknowledge that the current pipeline is a first-order stratification driven by major disease classes, potentially masking finer subphenotypic variations due to population imbalances. Future work aims to apply disease-specific clustering to reduce prevalence-driven effects and incorporate expanded feature sets, including genetic data, to improve subphenotype resolution. The ultimate goal is to transition from clustering to classification models to establish risk and prediction models for IEI patients.