🤖 machine learning

Re-mixing Embeddings for Patient Augmentation in Data Scarce Multiple Instance Learning

This paper introduces RECIPE, a statistically grounded patient augmentation method that generates realistic patients in embedding space by learning disease-specific instance distributions via Gaussian Mixture Models, effectively addressing data scarcity in Multiple Instance Learning across diverse clinical scenarios including missing classes and low-data regimes.

Original authors: Muhammed Furkan Dasdelen, Fatih Ozlugedik, Anastasia Litinetskaya, Nassir Navab, Carsten Marr, Ario Sadafi

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Muhammed Furkan Dasdelen, Fatih Ozlugedik, Anastasia Litinetskaya, Nassir Navab, Carsten Marr, Ario Sadafi

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 a master chef trying to teach a young apprentice how to cook a specific, rare dish. The problem? You only have a few ingredients left in the pantry, and for some versions of the dish, you have zero ingredients at all. If you try to teach the apprentice with so little, they will never learn the true flavor.

This is the exact problem doctors and AI face when diagnosing rare diseases. They have a lot of data for common conditions, but for rare ones, or for "healthy" control groups in specific studies, the data is missing or extremely scarce.

The paper introduces a new method called RECIPE (Re-mixing Embeddings via Clustering for Patient augmEntation). Think of RECIPE as a smart, statistical sous-chef that can invent new, realistic "fake" patients to fill the empty spots in the chef's (the AI's) training kitchen.

Here is how it works, broken down into simple steps:

1. The Problem: The "Empty Pantry"

In medical AI, we often use a technique called Multiple Instance Learning (MIL). Imagine a "bag" of data represents one patient. Inside that bag are hundreds of tiny pieces of information (like individual cells in a blood sample or tiny patches of a tissue slide). The AI looks at the whole bag to guess the diagnosis.

The trouble is, if you only have 5 patients with a rare disease, the AI gets confused. It can't learn the "recipe" for that disease. Worse, sometimes you have no healthy patients to compare them against. Traditional methods (like just flipping images or changing colors) don't work here because they can't create new types of patients; they just make slight variations of the same few.

2. The Solution: The "Statistical Recipe Book"

RECIPE solves this by looking at the data not as individual pictures, but as flavor profiles (called "embeddings").

Step A: The Taste Test (Clustering):
The method takes all the tiny pieces of data from all the patients it has (even from different hospitals or studies) and groups them into clusters based on similarity. It's like sorting a giant pile of spices into jars: "Spicy," "Sweet," "Earthy," etc.
Step B: Writing the Recipe:
For a specific disease (e.g., a specific type of leukemia), the AI counts how many "spices" (data points) usually go into the mix. It creates a statistical "recipe": "To make a healthy patient, you need 60% of this spice, 20% of that one, and a pinch of this other."
Step C: Cooking New Dishes (Augmentation):
Now, the AI can "cook" new patients from scratch. It doesn't need a real patient to start with. It just follows the recipe: it grabs the right amount of "spices" from the existing data pool and mixes them together to create a brand new, realistic-sounding patient.

3. The Magic Trick: Borrowing from a Neighbor

The most impressive part of RECIPE is what happens when you have zero patients for a category (like healthy controls).

Imagine you are in a kitchen with no salt, but your neighbor next door has a huge jar of salt. Usually, you can't just take their salt because it might be a different brand or quality.

RECIPE's approach: It analyzes the neighbor's salt jar to understand the statistical pattern of "saltiness." Then, it uses that pattern to season your ingredients.
In the paper: The researchers took data from a large, healthy dataset (cAItomorph) to learn what a "healthy" blood cell looks like statistically. They then used that "recipe" to generate healthy patients for a different, smaller dataset (AML-Hehr) that had no healthy people in it.
The Result: The AI trained on these "borrowed" healthy patients performed almost as well as if it had been trained on the full, real dataset. It effectively "transferred" the knowledge of health without needing to share the actual private data.

4. Picking the Best "Fake" Patients

You can't just generate infinite fake patients; some might be weird or unrealistic.

The Quality Control: The AI plays a game of "guessing." It looks at the new fake patients and asks, "How confident am I about this?"
The Strategy: If the AI is very confused (uncertain) about a fake patient, that means the patient is interesting and teaches the AI something new. RECIPE specifically picks these "confusing" fake patients to add to the training set. It ignores the ones that are too easy or too obvious.

5. Where Did They Test This?

The authors tested this "sous-chef" in three very different kitchens:

Missing Ingredients: Creating healthy patients for a leukemia study where no healthy people were available.
Tiny Portions: Creating more patients when only 1, 2, or 4 real patients existed for a disease.
Different Cuisines: Using it on non-image data, like single-cell RNA sequencing (genetic data) and flow cytometry (cell counting), proving it works even when there are no pictures involved.

The Bottom Line

RECIPE is a tool that helps AI doctors learn rare diseases by mathematically remixing existing data to create realistic new examples. It allows them to:

Fill in missing groups (like healthy controls) using statistics from other studies.
Learn effectively even when they only have a handful of real patients.
Do all this without needing to share sensitive patient data, just the statistical "recipes."

The paper claims this method consistently beats other ways of making fake data, often reaching performance levels close to having a full, perfect dataset.

Technical Summary: Re-mixing Embeddings for Patient Augmentation in Data Scarce Multiple Instance Learning

Problem Statement
Data scarcity remains a critical bottleneck in medical Multiple Instance Learning (MIL), particularly for rare diseases, specialized expensive modalities, and scenarios where invasive data collection is limited. In MIL, a "bag" (e.g., a patient) consists of multiple instances (e.g., image patches, single cells, or genomic features), and the model predicts the bag-level label without instance-level annotations. Traditional augmentation techniques (rotations, flips) operate at the instance level and fail to generate realistic variations at the bag level. Furthermore, existing bag-level mixing strategies (e.g., PseMix, ReMix) suffer from three primary limitations:

They operate exclusively online, creating ephemeral synthetic bags that cannot be validated or reused.
Their local mixing scope fails to capture global population statistics.
They require real examples from every class, rendering them ineffective when an entire class (e.g., healthy controls) is missing from the dataset.

Methodology: RECIPE
The authors introduce RECIPE (Re-mixing Embeddings via Clustering for Patient augmEntation), an offline, statistically grounded framework for generating new patient bags in embedding space. The pipeline consists of two main stages:

Cluster-based Unsupervised Sampling (Recipe Learning):
- Embedding Pooling: Instance embeddings from all training patients are pooled together.
- Gaussian Mixture Modeling (GMM): A GMM is fitted to the pooled embeddings to define unsupervised clusters.
- Recipe Derivation: For each disease class, the method computes a "recipe"—a statistical distribution of instances across the learned clusters. This is defined by the mean ( $\mu_{c,k}$ ) and standard deviation ( $\sigma_{c,k}$ ) of the number of instances assigned to each cluster $k$ for class $c$ .
- Patient Generation: New patients are generated by sampling the number of instances for each cluster based on the learned recipe ( $n_{c,k} \sim \mathcal{N}(\mu_{c,k}, \sigma^2_{c,k})$ ) and then drawing that many embeddings from the pool of real instances assigned to those clusters. This allows for the generation of patients even when no real examples of a specific class exist, provided the recipe is derived from an external dataset.
Uncertainty-Aware Patient Selection:
- To ensure the generated data improves model performance rather than introducing noise, the method employs uncertainty quantification.
- A pre-trained MIL model evaluates the generated bags using Monte Carlo (MC) dropout at inference.
- Predictive Entropy is calculated to quantify uncertainty ( $U(x) = H[\bar{p}(x)]$ ).
- Only the most uncertain generated patients (those with high entropy) are selected to augment the training set, under the hypothesis that these samples provide the most informative learning signal.

Key Contributions

Offline, Statistically Grounded Generation: Unlike online mixing methods, RECIPE learns global distributional "recipes" to create persistent, analyzable synthetic cohorts.
Missing Class Handling: RECIPE is the first MIL framework capable of generating entirely missing patient classes (e.g., healthy controls) by transferring statistical recipes from external datasets without sharing sensitive patient data.
Cross-Modality Applicability: The method is demonstrated to be effective across diverse data modalities, including histopathology images, single-cell RNA sequencing (scRNA-seq), and flow cytometry.
Architecture Agnosticism: The framework is compatible with various MIL aggregators, including Attention-based MIL (ABMIL), DSMIL, and Transformers.

Experimental Results
The authors evaluated RECIPE across three clinically relevant scarcity scenarios:

Cross-Dataset Transfer (Missing Class):
- Scenario: All healthy controls were removed from the AML-Hehr dataset. A "healthy recipe" was derived from the external cAItomorph dataset and applied to the AML-Hehr embedding pool.
- Outcome: The generated healthy patients closely matched real cell-type compositions (Jensen-Shannon Divergence = 0.12). Training with these generated patients achieved performance comparable to full-dataset training (0.78 vs. 0.83 balanced accuracy) and significantly outperformed direct transfer of external data (0.66).
Low-Data Regimes:
- Scenario: The number of healthy controls in AML-Hehr was reduced to as few as $n=1$ .
- Outcome: With only one real healthy control, the baseline achieved 0.26 balanced accuracy. RECIPE achieved 0.68, substantially outperforming MixUp (0.60) and PseMix (0.64).
Non-Image Modalities (scRNA-seq and Flow Cytometry):
- Scenario: Augmentation on COVID-19 severity datasets with naturally limited sample sizes.
- Outcome: RECIPE improved balanced accuracy by +14% (PCA embeddings) and +26% (scVI embeddings) on scRNA-seq data, and by +22% on flow cytometry data, outperforming all baseline augmentation strategies.

Significance and Claims
The paper claims that RECIPE offers a novel solution to the data scarcity bottleneck in medical MIL by enabling the generation of realistic patient cohorts through the re-mixing of pooled embeddings based on learned statistical distributions. Its primary significance lies in its ability to:

Address the "missing class" problem, allowing for the synthesis of healthy controls or rare disease phenotypes using external statistics without compromising patient privacy.
Improve MIL performance in extreme low-data regimes where traditional augmentation fails.
Provide a generalizable framework applicable beyond imaging to single-cell and cytometry data.

The authors emphasize that their approach moves beyond local, online mixing to model global population statistics, enabling the creation of validated, reusable synthetic data that can bridge gaps in rare disease research and privacy-preserving data augmentation.