Generative Class Balancing for Rare Disease Detection

56 concepts57 papers1 proposal

When training machine learning models for clinical diagnostics, data is rarely distributed equally. In datasets containing thousands of examples of healthy tissue or common conditions, a critical, life-threatening anomaly might appear only a handful of times. This extreme class imbalance forces classifiers to bias heavily toward the majority classes, leading to dangerously low recall on rare conditions.

To fix this, traditional machine learning relies on oversampling techniques like SMOTE or basic geometric transformations (such as rotating or scaling images). However, these methods fail in complex medical domains because they do not understand the underlying biological or physical constraints of the data. Geometric transforms merely duplicate existing pathologies at different angles, while linear interpolation between rare samples can create physically impossible artifacts that confuse the classifier's decision boundary.

Generative class balancing solves this problem by shifting the strategy from data duplication to data estimation. Instead of manipulating existing points, deep generative models learn the high-dimensional probability distribution of the minority class, allowing them to populate underrepresented regions of the data space with entirely new, biologically plausible synthetic samples.

Mapping the True Data Manifold

To understand why generative balancing works, it helps to view datasets geometrically. Real-world data does not scattered randomly across a high-dimensional space; instead, it concentrates on a lower-dimensional surface called a manifold.

Traditional oversampling draws straight lines between existing points on this manifold. If the manifold is highly curved—as is common with complex medical data like brain structures or electrocardiogram (ECG) signals—these straight lines cut directly through empty space, creating unrealistic synthetic data.

As Yun Yang et al. demonstrate in GAN-Based Semi-Supervised Learning Approach for Clinical Decision Support in Health-IoT Platform, generative models overcome this by learning the underlying probability density function of the data. Rather than interpolating, they map the true shape of the manifold. This allows the model to generate synthetic samples that maintain the structural integrity and complex class boundaries of real clinical data.

This manifold-learning approach can be implemented through various statistical and neural architectures:

Controllable Synthesis and Anatomy Arithmetic

A major advantage of generative balancing over simple duplication is the ability to perform controllable synthesis. Because generative models map data to a continuous latent space, researchers can manipulate specific variables within that space to "design" rare pathological scenarios on demand.

One powerful manifestation of this is disentanglement, where a model separates an image's style (such as scanner noise or patient age) from its underlying anatomy. Spyridon Thermos et al. exploit this in Controllable cardiac synthesis via disentangled anatomy arithmetic by performing "anatomy arithmetic." By swapping the latent representations of different heart structures, the authors synthetically inject rare pathologies into otherwise healthy cardiac MRIs. This isolates the exact feature changes the downstream classifier needs to learn, stripping away confounding variables.

[Healthy Patient Anatomy]  ──┐
                             ├──► [Anatomy Arithmetic] ──► [Synthetic Rare Pathology]
[Rare Disease Structure]   ──┘

Alternatively, researchers can guide generation using semantic masks or image-to-image translation:

Multi-Modal Guidance and Diffusion

While GANs proved the viability of generative class balancing, the field has increasingly shifted toward Latent Denoising Diffusion Probabilistic Models. As highlighted by Gustav Müller‐Franzes et al. in A multimodal comparison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image synthesis, diffusion models offer significantly higher image diversity and fewer structural artifacts than GANs, making them exceptionally reliable for synthesizing rare medical conditions.

Diffusion models excel at integrating multi-modal prompts, allowing clinicians to use text reports alongside visual masks to guide the generation of rare features. For example, Shiyi Du et al. show in Boosting Dermatoscopic Lesion Segmentation via Diffusion Models with Visual and Textual Prompts that conditioning diffusion models on both visual masks and textual descriptions allows for the precise creation of infrequent skin lesion patterns.

Similarly, Jia Yu et al. demonstrate in Robust Polyp Detection and Diagnosis Through Compositional Prompt-Guided Diffusion Models that combining clinical text reports with traditional image masks generates highly realistic, targeted synthetic data for colorectal cancer screenings, drastically improving detection sensitivity for dangerous, underrepresented lesions.

Closed-Loop and Dynamic Balancing

The earliest implementations of generative balancing operated offline: researchers generated a fixed batch of synthetic data, appended it to the training set, and trained the classifier. However, this static approach risks introducing generative artifacts or over-representing certain synthetic features.

The state-of-the-art has evolved toward dynamic, closed-loop synthesis, where the generation process is tightly coupled with the classifier's real-time performance.

┌────────────────────────────────────────────────────────┐
│                                                        │
▼                                                        │
[Generative Model] ──► [Synthetic Data] ──► [Classifier] ┘
                                                 │
                                                 ▼
                                      [Real-Time Error Signal]

This intuition is realized by Shuhan Li et al. in Iterative Online Image Synthesis via Diffusion Model for Imbalanced Classification through the Iterative Online Image Synthesis (IOIS) framework. Instead of creating a static dataset, the diffusion model generates tailored synthetic data during the training process.

By monitoring the classifier's real-time error signal, the system dynamically adjusts both the quantity of synthetic images (generating more data for classes the model is currently struggling to identify) and the quality (using classifier guidance to push the generated images toward the ambiguous decision boundaries). This closed-loop feedback loop ensures that the generative model places synthetic data exactly where the classifier needs it most, maximizing diagnostic robustness on the rarest corner cases.

Key Foundational Work

Opportunities and Challenges of Deep Learning Methods for Electrocardiogram Data: A Systematic Review (2019) — contributed 1 concept elements to this direction.
Generative Image Translation for Data Augmentation of Bone Lesion Pathology (2019) — contributed 1 concept elements to this direction.
Heartbeat Anomaly Detection using Adversarial Oversampling (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_synthetic_data_augmentation_class_imbala_7c38dbb5,
  title = {synthetic data augmentation (class imbalance mitigation)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/synthetic-data-augmentation-class-imbalance-mitigation/rdr_7c38dbb505ef4982b4db832a15e4b39b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_synthetic_data_augmentation_class_imbala_7c38dbb5_v1,
  title = {synthetic data augmentation (class imbalance mitigation)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/synthetic-data-augmentation-class-imbalance-mitigation/rdr_7c38dbb505ef4982b4db832a15e4b39b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake