Clustering-Based Outcome Models for Clinical Studies: A Scoping Review

This scoping review systematically examines and categorizes clustering-based outcome models for clinical studies into informed and agnostic approaches, highlighting their utility in handling high-dimensional data and heterogeneous populations for applications such as risk stratification, rare disease research, and subgroup-specific treatment effect estimation.

The Gist

Imagine you are a doctor trying to predict how a patient will recover from an illness. In the past, doctors might have treated everyone with the same disease roughly the same way, assuming they were all "average." But we know that's not true. One patient might recover quickly, while another struggles, even if they have the same diagnosis. Why? Because their bodies, lifestyles, and genetics (their "covariates") are different.

This paper is a scoping review, which is like a giant map of a new territory. The authors went out and found 55 different mathematical methods that try to solve this problem by grouping patients together before trying to predict their outcomes.

Here is the simple breakdown of what they found, using some everyday analogies.

The Big Idea: Grouping Before Guessing

Instead of trying to build one giant, complicated rulebook for every single person, these methods say: "Let's first sort people into smaller, more similar groups (clusters), and then write a specific rulebook for each group."

Think of it like a tailor making suits:

• The Old Way: Trying to make one "one-size-fits-all" suit that somehow stretches to fit a 5-foot person and a 6-foot person. It never fits perfectly.
• The New Way: First, you measure everyone and sort them into "Small," "Medium," and "Large" piles. Then, you make a perfect suit for the "Medium" pile, a different one for "Large," and so on. The suits fit much better because the people in each pile are actually similar.

The Two Main Strategies

The paper splits these methods into two distinct camps based on when they look at the patient's final health outcome (the "result").

1. The "Agnostic" Approach (The Blind Sort)

• How it works: The computer looks only at the patient's starting data (age, blood tests, genetics) to sort them into groups. It doesn't know who got better or worse yet. Once the groups are formed, the doctor looks at the results and builds a prediction model for each group separately.
• The Analogy: Imagine a librarian sorting books by their cover color and thickness without reading the story inside. Once the books are sorted into "Red/Thick" and "Blue/Thin" piles, the librarian then reads the stories to see which pile has more happy endings.
• When it's good: This is great when you want to be fair and unbiased, or when you are using historical data (like old medical records) to help predict outcomes for new patients. It's like using a map of the terrain to plan a hike before you even start walking.

2. The "Informed" Approach (The Smart Sort)

• How it works: The computer is allowed to peek at the final health outcome while it is sorting the patients. It tries to find groups where the patients not only look similar on paper but also behave similarly in the end.
• The Analogy: Imagine a teacher sorting students into study groups. Instead of just looking at their names or grades, the teacher looks at who actually passed the final exam. They might realize that "Students who like math but hate history" form a specific group that needs a special teaching style. The sorting happens with the knowledge of the result.
• When it's good: This is often more accurate because it finds patterns that purely data-driven sorting might miss. However, it's trickier to use in new situations because it relies heavily on the specific data it was trained on.

Why Do We Need This? (The "Too Many Variables" Problem)

The paper highlights a major problem in modern medicine: Too much data.
Today, we can measure thousands of things about a patient (genetics, proteins, lifestyle). If you try to build a prediction model using all 10,000 variables at once, the model gets confused, overfits (memorizes the noise instead of the signal), and fails on new patients.

Clustering is the "Compression" Tool:
Think of the 10,000 variables as a messy room with 10,000 toys.

• Without Clustering: You try to describe the position of every single toy. Impossible.
• With Clustering: You say, "Okay, all the blocks are in the red bin, all the dolls are in the blue bin." You've reduced 10,000 toys down to just "Red Bin" and "Blue Bin."
• The Result: The prediction model becomes much simpler, faster, and more accurate because it's looking at the groups of toys rather than every single toy individually.

Where is this useful?

The authors point out three main places where this "Grouping" magic shines:

1. Rare Diseases: When you only have 50 patients but 5,000 data points per patient, you can't use standard math. Clustering helps you find the few patterns that exist in that small crowd.
2. Personalized Medicine (Precision Medicine): Instead of saying "This drug works for 60% of people," we can say, "This drug works for the 'High-Inflammation' group, but not the 'Low-Inflammation' group."
3. Using Old Data to Help New Patients: If you have a massive database of old patients (but no outcome data for them), you can use clustering to define "types" of patients. Then, when a new patient comes in, you see which "type" they match and borrow the knowledge from that group to predict their future.

The Bottom Line

This paper is a guidebook for researchers. It says: "Stop trying to treat everyone as an individual statistic. Group them first."

Whether you sort them blindly (Agnostic) or with the help of the answer key (Informed), grouping patients based on their similarities allows doctors to build better, simpler, and more accurate predictions. It turns a chaotic crowd of unique individuals into manageable, understandable teams, making the path to better healthcare clearer.

Technical Summary

Here is a detailed technical summary of the scoping review "Clustering-Based Outcome Models for Clinical Studies" by Vilsmeier et al.

1. Problem Statement

Clinical outcomes often exhibit prognostic heterogeneity (systematic variation in outcomes based on baseline characteristics) and predictive heterogeneity (variation in treatment effects). Traditional outcome models (e.g., regression) struggle when:

• The number of candidate covariates is large relative to the sample size (common in rare diseases and "omics" data), leading to overfitting and unstable estimates.
• High-order interactions between covariates are complex and difficult to specify a priori.
• Longitudinal or functional covariates (irregularly spaced measurements) need to be incorporated.

Standard variable selection methods (e.g., LASSO) may fail to capture complex subgroup structures. The paper addresses the need for methods that use clustering of observational units (patients) based on covariates to create low-dimensional representations for outcome modeling, thereby capturing complex patterns without explicitly modeling high-order interactions.

2. Methodology

The authors conducted a scoping review following the PRISMA guidelines.

• Search Strategy: Systematic search of Web of Science and PubMed, supplemented by 5 manually added records.
• Inclusion Criteria: Records proposing or evaluating methods that combine covariate-based clustering of patients with outcome modeling (regression, survival analysis, etc.).
• Exclusion Criteria: Studies focusing solely on clustering without outcome modeling, clustering of covariates (rather than units), image processing, or software-only papers.
• Screening: 738 records identified $\rightarrow$ 98 full-text assessed $\rightarrow$ 55 records included.
• Classification Framework: The authors classified the 55 identified methods into two primary categories based on the role of the outcome variable in the clustering process:
  1. Informed-Cluster Models: The outcome variable contributes to cluster formation. Clustering and outcome modeling are estimated jointly.
  2. Agnostic-Cluster Models: Clustering is performed solely on covariates in a first step (unsupervised). The resulting cluster variables (membership, probabilities, or distances) are used as covariates in a second step to model the outcome.

3. Key Contributions and Technical Framework

A. Taxonomy of Models

The review provides a rigorous mathematical and conceptual breakdown of the identified approaches:

1. Informed-Cluster Models (Joint Estimation)

• Product Partition Models with Covariates (PPMx):
  • Mechanism: Uses a Bayesian framework where the prior probability of a partition depends on a cohesion function and a covariate-similarity function. The number of clusters is random.
  • Application: Handles zero-inflated counts, time-to-event data, and variable selection.
• Finite Mixtures of Regression Models (FMR):
  • Mechanism: Assumes a finite mixture distribution for the outcome where mixing probabilities depend on covariates (e.g., via multinomial logistic regression). The number of clusters is usually fixed.
  • Extensions: Includes functional covariates (e.g., EEG signals) and joint models for longitudinal biomarkers and time-to-event outcomes.
• Cluster-Aware Supervised Learning (CluSL):
  • Mechanism: A deterministic optimization approach minimizing a loss function that balances cluster-specific prediction error and the dissimilarity of covariates to cluster centroids.
  • Feature: Tuning parameter $\lambda$ controls the trade-off between outcome fit and covariate similarity.

2. Agnostic-Cluster Models (Two-Step Procedures)

• Algorithmic Clustering:
  • Methods: K-means, hierarchical clustering, spectral clustering.
  • Process: Cluster patients based on covariates $\rightarrow$ Fit separate models per cluster or use cluster indicators as covariates.
  • Innovation: Ensemble approaches (e.g., averaging predictions across different $k$) and feature augmentation (adding distances to centroids as new covariates).
• Model-Based Clustering:
  • Methods: Finite mixture models for covariates (e.g., Latent Class Analysis).
  • Process: Estimate posterior probabilities of cluster membership $\rightarrow$ Use these probabilities as covariates in a regression model.
  • Utility: Effective for dimensionality reduction in high-dimensional categorical data (e.g., diagnostic codes).

B. Key Findings from the 55 Records

• Distribution: Agnostic-cluster models (32 records) were slightly more prevalent than informed-cluster models (28 records).
• Disciplinary Divide: Informed-cluster models are almost exclusively published in statistical journals (methodological focus), while agnostic-cluster models dominate biomedical/public health and computer science literature (application focus).
• Objectives:
  1. Subgroup Identification: The most common goal (31 records), aiming to find homogeneous patient groups.
  2. Dimensionality Reduction: Compressing high-dimensional covariates into cluster indicators.
  3. Feature Extraction: Creating new variables (e.g., distance to centroids) to improve prediction.
• Data Characteristics: Real-world applications often involve high-dimensional data where the number of covariates ($d$) exceeds the sample size ($n$), validating the utility of clustering for dimensionality reduction. Simulation studies covered a wide range of $n$ and $d$.
• Outcome Types: Models were developed for metric, time-to-event (survival), and categorical outcomes.

4. Results and Applications

The review highlights several successful applications in biomedical sciences:

• Rare Diseases & High-Dimensional Data: Clustering allows for risk stratification in small cohorts with rich biomarker panels (e.g., gene expression) where traditional regression fails.
• Longitudinal Data: Methods like CluSL and FMR successfully handle irregularly spaced measurements (e.g., EEG, lactate dehydrogenase trajectories) by clustering based on trajectory shapes.
• Historical Data Borrowing: In rare disease research, clustering can be derived from large historical registries (covariates only) to define subgroups, which are then used to fit outcome models in smaller prospective trials (e.g., using power priors).
• Clinical Trial Design:
  • Covariate Adjustment: Using cluster membership as a super-covariate to increase statistical power.
  • Subgroup-Specific Treatment Effects: Estimating heterogeneous treatment effects by including treatment-by-cluster interactions.

5. Significance and Implications

• Bridging the Gap: The review bridges statistical methodology and clinical application, clarifying when to use joint (informed) vs. sequential (agnostic) approaches.
• Handling Heterogeneity: It demonstrates that clustering is a viable alternative to variable selection for managing prognostic and predictive heterogeneity, particularly in "small $n$, large $d$" settings.
• Future Directions:
  • Uncertainty Quantification: Most agnostic approaches treat clusters as fixed estimates, ignoring the uncertainty in subgroup definition. Future work should quantify this uncertainty.
  • Validation: There is a lack of direct comparative studies between informed and agnostic models; such comparisons are needed to determine if the added complexity of joint modeling yields better predictive performance.
  • Clinical Trials: The potential for using these models to design adaptive trials and improve the precision of treatment effect estimates is significant, provided the clusters are stable and interpretable.

In conclusion, the paper establishes that clustering-based outcome models are a powerful, flexible tool for precision medicine, particularly for risk stratification and handling high-dimensional, heterogeneous clinical data.