📄 other

Phenotypic Heterogeneity in Ischemic Heart Disease Uncovered by Unsupervised Clustering of Routine Laboratory Parameters

This study demonstrates that unsupervised clustering of routine, low-cost laboratory parameters can identify three reproducible phenotypic subgroups within ischemic heart disease patients, revealing a distinct lipid-metabolic and hepatic-inflammatory axis that enables interpretable risk stratification in resource-limited settings.

Original authors: Anam Ijaz, Sara Aslam, Shabana NA

Published 2026-07-27

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Anam Ijaz, Sara Aslam, Shabana NA

Original paper licensed under CC BY 4.0 (https://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

The Hidden Patterns in Your Blood

Imagine your body as a bustling city. For years, doctors have been like traffic controllers, looking at the main highways to see if there's a jam. In the world of heart health, that "jam" is Ischemic Heart Disease (IHD), a condition where the heart muscle doesn't get enough oxygen-rich blood. Traditionally, to understand a patient's specific traffic problem, doctors have relied on expensive, high-tech tools like X-ray cameras (angiography) or complex genetic maps (omics) to see exactly what's happening inside.

But what if the city's daily routine reports—the simple logs of trash collection, water usage, and electricity bills—could tell you just as much about the traffic jams? In medicine, these "routine reports" are the standard blood tests everyone gets: checks for red blood cells, liver function, kidney health, and fats in the blood. These tests are cheap, quick, and available almost everywhere. The big question scientists have been asking is: Can we use these simple, everyday numbers to find hidden groups of patients who are actually very different from each other, even if they all have the same heart disease diagnosis? This paper dives into that mystery, using a special kind of computer brain called "unsupervised machine learning" to sort patients into teams based solely on their blood work, without needing to look at the heart's plumbing or DNA.

The Great Blood Test Sort-Out

In this study, researchers in Lahore, Pakistan, decided to play a game of "organize the crowd" with 306 people who already knew they had Ischemic Heart Disease. They didn't have a control group of healthy people to compare them to; they just wanted to see if the sick people were all the same or if they were actually a mix of different types. They took a huge list of routine lab results—things like how much fat is in the blood, how well the liver is working, and how many white blood cells are fighting inflammation—and fed it into a computer.

The computer used a technique called clustering. Think of it like a DJ at a party who doesn't know anyone's names but listens to the music everyone is dancing to. Instead of asking, "Who likes pop?" or "Who likes rock?", the DJ just watches the dance floor. If a group of people is doing the same weird dance move, the DJ groups them together. The computer did this with blood data, looking for patterns where certain numbers always went up or down together.

After trying out five different ways to sort the data, the computer found that the best way to group these patients was into three distinct teams. It wasn't a random guess; the groups were so stable that the mathematical "silhouette score" (a measure of how well-defined the groups are) was high and consistent between the initial data and the new test data. To make these invisible groups visible to doctors, the researchers then trained a separate "detective" computer program (a Random Forest classifier) to learn the patterns of these groups. When tested on new patients, this detective program could correctly identify which team a patient belonged to about 78% of the time, proving that the groups were real and distinct.

The Three Teams Discovered:

The "High-Risk Mix" Team (12.7% of patients): This was the smallest group, but the most intense. These patients had a "perfect storm" in their blood: high levels of bad fats (specifically VLDL and triglycerides), signs that their liver was struggling (high AST and ALT enzymes), and a high number of neutrophils (a type of white blood cell that signals inflammation). The researchers call this an "atherogenic-hepatic-inflammatory" signature. It's like a city where the trash trucks are overflowing, the power plant is smoking, and the police are on high alert all at once.
The "Pure Fat" Team (26.5% of patients): This group was larger. They had high levels of fats in their blood (VLDL, triglycerides, and LDL), but their liver enzymes were normal, and they didn't show the same signs of inflammation. They were like a city with a traffic jam caused purely by too many cars, but the rest of the infrastructure was running fine.
The "Lower-Risk" Team (60.8% of patients): This was the biggest group, making up the majority of the patients. They had the lowest levels of the bad fats and the highest levels of lymphocytes (a different type of white blood cell). They seemed to be the "calmest" group, with the least amount of metabolic and inflammatory stress.

The Detective Work: Finding the Clues

Once the computer sorted the patients, the researchers asked: "Which specific numbers were the most important for making these groups?" They used a digital detective tool called a Random Forest (which is like a team of decision-makers voting on the answer). They found that the main clues were VLDL (a type of fat), Total Bilirubin (a waste product from the liver), Triglycerides, and the liver enzymes AST and ALT. Surprisingly, things like kidney function or standard blood cell counts weren't the main drivers. It turned out that the key to telling these patients apart was the relationship between how their bodies handled fat and how their liver was working.

The researchers then tried to turn these complex computer findings into simple rules that a doctor could read on a piece of paper. They used a Decision Tree (which looks like a flowchart) and a Fuzzy Rule system (which uses language like "High," "Medium," or "Low" instead of exact numbers).

They came up with rules like: "IF the VLDL is High AND the AST is Low, THEN the patient belongs to the 'Pure Fat' team."
While the computer could identify the "High-Risk Mix" group, the simple written rules were less precise for this specific group due to its small size and complex biology, meaning doctors can't yet rely on a single short sentence to spot them as easily as the other groups.

What This Means (and What It Doesn't)

The paper suggests that even without expensive heart scans or genetic tests, we can find hidden differences in heart disease patients just by looking at their routine blood work. It proves that the "one-size-fits-all" view of heart disease might be too simple. There are subgroups, and some (like the 12.7% in the High-Risk Mix) might need extra attention because their bodies are dealing with both fat problems and liver inflammation.

However, the authors are very careful not to say they have solved the mystery of heart disease. They explicitly state that this is a hypothesis-generating study. They found the groups, and the computer is very good at sorting people into them, but they do not know yet if these groups actually predict who will get sicker or die sooner. They didn't track the patients over time to see what happened to them. They also didn't compare these patients to healthy people; they only looked at people who already had the disease.

So, while the study shows that these three groups are real and distinct based on blood chemistry, it's still a "proof of concept." It's like finding three different types of clouds in the sky and realizing they look different, but we haven't yet proven that one type of cloud always brings a hurricane while the others just bring a light drizzle. The researchers hope that future studies will track these patients to see if the "High-Risk Mix" group really does have worse outcomes, which would allow doctors to use these simple, cheap blood tests to give better, more personalized care, especially in places where fancy machines aren't available.

Technical Summary: Phenotypic Heterogeneity in Ischemic Heart Disease Uncovered by Unsupervised Clustering of Routine Laboratory Parameters

Problem Statement
Ischemic heart disease (IHD) remains the leading cause of global mortality, yet current risk stratification often relies on reactive diagnostic tools (imaging, angiography, high-sensitivity biomarkers) that are inaccessible in resource-limited settings. While machine learning (ML) has advanced predictive medicine, most existing models depend on specialized data sources or focus on distinguishing IHD from non-IHD populations. A critical gap exists in understanding the phenotypic heterogeneity within an already-diagnosed IHD population using only routine, low-cost laboratory parameters. This study addresses whether unsupervised learning can reveal clinically meaningful subgroups within confirmed IHD patients without imaging, omics data, or a non-IHD control group.

Methodology
The study employed a cross-sectional design involving 306 confirmed IHD patients recruited from hospitals in Lahore, Pakistan. Participants were recently diagnosed and had not yet commenced lipid-lowering or antihypertensive medications to avoid pharmacological confounding. Exclusion criteria included pre-existing liver disease, kidney disease, or malignancy.

Data Collection: The dataset comprised 26 continuous laboratory parameters (hematological, hepatic, renal, and lipid profiles) and four binary clinical variables (gender, diabetes, hypertension, family history).
Preprocessing: Data were split into training (70%) and test (30%) subsets. Missing values were imputed using a skewness-aware strategy (median for skewed, mean for normal distributions) derived solely from the training set. Outliers were winsorized, and continuous variables were min-max normalized.
Unsupervised Clustering: The 26 continuous features were reduced to five dimensions using UMAP. Five clustering algorithms were compared: K-means, Gaussian Mixture Models (GMM), Fuzzy C-means, Hierarchical Clustering, and DBSCAN. Configurations with $k=3$ and $k=4$ were evaluated. The optimal model was selected based on the Silhouette coefficient on both training and held-out test sets to ensure generalizability and minimize the train-test gap.
Interpretability & Rule Extraction: To translate clusters into clinical insights, three interpretable models were trained on the training set to predict cluster membership:
1. Random Forest (RF): Used for feature importance ranking via Recursive Feature Elimination (RFECV).
2. Decision Tree (DT): A shallow tree (max depth 4) to generate transparent splitting rules.
3. Evolutionary Fuzzy Classifier (ExFuzzy): To generate linguistically accessible IF-THEN rules using three linguistic terms (Low/Medium/High).

Key Results

Optimal Clustering: K-means with $k=3$ yielded the most stable partition, achieving a training Silhouette of 0.361 and a test Silhouette of 0.370. This configuration demonstrated a minimal generalization gap, indicating robust cluster geometry.
Identified Phenotypes: The analysis resolved the cohort into three distinct phenotypes:
1. Cluster 0 (12.7%): An atherogenic-hepatic-inflammatory phenotype characterized by elevated VLDL, triglycerides, transaminases (AST, ALT), neutrophils, and total bilirubin, with low lymphocytes.
2. Cluster 2 (26.5%): A pure dyslipidemia phenotype with elevated VLDL, triglycerides, and LDL, but without significant transaminase elevation.
3. Cluster 1 (60.8%): A lower-risk phenotype with the lowest atherogenic lipid burden and the highest proportion of lymphocytes.
Feature Importance: Random Forest identified VLDL, total bilirubin, triglycerides, AST, ALT, and HDL as the principal discriminating features. Hematological and renal markers contributed negligibly, highlighting a dominant lipid-metabolic and hepatic-inflammatory axis.
Model Performance:
- Random Forest: Achieved a test accuracy of 78.3% (AUC 0.926), effectively distinguishing the clusters.
- Decision Tree: Achieved a test accuracy of 59.8%, demonstrating a trade-off where interpretability came at the cost of generalization, particularly for the intermediate dyslipidemia cluster.
- Fuzzy Classifier: Achieved consistent training and test accuracy of ~44%, limited by linguistic discretization rather than overfitting. It successfully generated interpretable rules, such as "IF VLDL is High AND AST is Low" for the atherogenic phenotype.

Key Contributions

Unsupervised Phenotyping in IHD: The study demonstrates that routine, low-cost laboratory parameters can reveal reproducible phenotypic heterogeneity within a confirmed IHD population, independent of imaging or omics data.
Resource-Limited Applicability: By relying exclusively on standard hematological and biochemical panels (CBC, LFTs, lipid profiles), the framework offers a scalable approach for risk stratification in low- and middle-income countries where advanced diagnostics are scarce.
Explainable AI Framework: The study bridges the gap between complex clustering and clinical utility by translating data-driven clusters into clinician-readable rules using Decision Trees and Fuzzy logic, addressing the "black box" limitation of many ML models.
Biological Axis Identification: The research identifies a specific lipid-metabolic and hepatic-inflammatory axis (centered on VLDL, triglycerides, and liver enzymes) as the primary driver of phenotypic variation in IHD, rather than classical hematological or renal indices.

Significance and Claims
The authors position this work as a proof-of-concept study. They claim that unsupervised, explainable machine learning can stratify an already-diagnosed IHD population into biologically distinct, laboratory-defined phenotypes using universally accessible data.

The paper explicitly states that these findings are hypothesis-generating. It does not claim to have established a definitive diagnostic or prognostic tool. The significance lies in demonstrating that:

Clinically meaningful heterogeneity exists even within a single diagnostic category (IHD).
This heterogeneity can be captured without expensive infrastructure.
The identified phenotypes (particularly the small, high-risk atherogenic-hepatic-inflammatory cluster) warrant prioritized clinical attention and future prospective validation against hard outcomes.

The authors conclude that while the framework is promising for scalable risk stratification in resource-limited settings, external validation in independent cohorts and prospective studies linking these phenotypes to clinical outcomes are required before these groups can be considered validated clinical entities.

The Hidden Patterns in Your Blood

The Great Blood Test Sort-Out

The Detective Work: Finding the Clues

What This Means (and What It Doesn't)

Technical Summary: Phenotypic Heterogeneity in Ischemic Heart Disease Uncovered by Unsupervised Clustering of Routine Laboratory Parameters

More like this