Label scarcity in biomedicine: Data-rich latent factor discovery enhances phenotype prediction

Marc-Andre Schulz, Bertrand Thirion, Alexandre Gramfort, Gaël Varoquaux, Danilo Bzdok

Introduction: The Challenge of Small-N in Big Data

In the era of "Big Data," biomedical research faces a paradox. Large repositories like the UK Biobank (UKBB) contain data for hundreds of thousands of participants, yet the number of subjects with specific disease labels—such as rare neurological disorders or specific depression subtypes—remains small. This scarcity hinders the application of "data-hungry" deep learning models that typically require thousands of labeled examples to generalize effectively.

The central question addressed here is: Can we use abundant data from the general population to inform models for rare or label-scarce medical conditions? This semisupervised approach involves learning the "language" of human biology through unsupervised latent factor discovery before fine-tuning for specific health outcomes.

Methodology

The researchers implemented a two-stage pipeline to test this hypothesis:

  1. Latent Factor Discovery (Unsupervised): High-dimensional brain imaging data was compressed into a 50-dimensional latent space using three methods: PCA (linear), Isomap (manifold-preserving), and Variational Autoencoders (VAE) (non-linear deep learning).
  2. Phenotype Prediction (Supervised): These 50-dimensional embeddings served as inputs for Logistic Regression and Random Forest classifiers to predict phenotypes like age, sex, and lifestyle factors.

The experiments varied the volume of unlabeled data (100 to 7,000 samples) while keeping the labeled sample size small (100 to 1,000).

The Semisupervision Effect

A primary finding is the "semisupervision effect": increasing the amount of unlabeled data used for latent factor discovery improves the performance of subsequent supervised classifiers, even when the number of labeled examples is fixed.

Out-of-sample accuracy improves as the number of unlabeled samples increases across MNIST, Fashion-MNIST, and UK Biobank datasets.

While this effect was most pronounced in simpler datasets like MNIST (closing ~90% of the accuracy gap), it remained significant in the UKBB brain imaging data (~50% gap closed). This discrepancy highlights that biomedical data is intrinsically more complex, requiring more sophisticated latent representations to capture meaningful biological structure.

Methodological Performance and Scaling

The choice of unsupervised method is critical. While PCA showed minimal gains from additional unlabeled data, VAEs were the most effective at capturing complex, non-linear relationships within the data.

Comparison of PCA, Isomap, and VAE methods across various phenotypes. VAEs consistently show better scaling with additional unlabeled data.

As shown in Figure 2, VAEs yielded superior prediction performance across all targets. For instance, with 1,000 labeled samples, the VAE-based approach achieved a semisupervised gain of approximately 23.6% compared to models trained without extra unlabeled data. Furthermore, while simple targets like age and sex were well-handled by linear classifiers, complex social phenotypes (e.g., household income) benefited from the non-linear capabilities of Random Forests.

Modality and Phenotype Insights

The study also revealed that data modality significantly impacts predictive power. T1-MRI (anatomical brain volume) was the most predictive modality, achieving a weighted cross-target accuracy of 64.8%, outperforming rfMRI (53.5%) and dMRI (58.5%). This suggests that anatomical features currently provide a more stable signal for phenotype prediction than functional or connectivity-based features in these latent spaces.

Conclusion and Future Directions

This research demonstrates that unsupervised latent factor discovery is a powerful tool for overcoming label scarcity in biomedicine. VAEs, in particular, excel at "soaking up" information from unlabeled samples to enhance downstream predictions.

Future work should explore applying these techniques directly to raw 3D MRI scans rather than pre-computed features and extending the framework to clinical diagnostic categories. The ultimate goal is to determine if these latent factors can reveal novel biomarkers for complex diseases where high-quality labels are most difficult to obtain.

Summary updated

Abstract

International audience
BibTeX
@misc{schulz2017label,
  author = {Marc-Andre Schulz and Bertrand Thirion and Alexandre Gramfort and Gaël Varoquaux and Danilo Bzdok},
  title = {{Label scarcity in biomedicine: Data-rich latent factor discovery enhances phenotype prediction}},
  year = {2017},
  eprint = {2110.06135},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2110.06135},
}
Figure