OliverDOU776/Few-step-probabilistic-glucose-forecasting-from-continuous-glucose-monitoring-and-meal-images

What it solves

GlucoFlow addresses the challenge of probabilistic time-series forecasting when target datasets are small but larger source datasets are available. Specifically, it targets glucose forecasting where users need reliable uncertainty ranges (predictive distributions) rather than a single point estimate, and where inference must be fast enough for practical deployment.

How it works

The project uses a conditional rectified flow model that generates full predictive distributions in just 1 to 4 Euler steps, avoiding the slow sampling chains typical of diffusion models. It employs a transfer learning approach: temporal dynamics are pretrained on large source datasets and then adapted (fine-tuned) to smaller target datasets. The model can optionally condition its predictions on side information, such as meal images and nutrient data, using a flexible embedding system.

Who it’s for

It is designed for researchers and developers working on physiological time-series forecasting, specifically those dealing with limited target-domain data and requiring calibrated uncertainty estimates for medical or health-related research.

Highlights

• Fast Probabilistic Inference: Generates future trajectories in 1†4 steps using rectified flow.
• Transfer Learning Path: Provides a pretrain-then-adapt workflow to leverage large source datasets for small target cohorts.
• Multimodal Conditioning: Supports optional inputs like images or metadata via a customizable embedding interface.
• Uncertainty Quantification: Produces full predictive distributions, allowing for the calculation of medians, confidence intervals, and event probabilities.