Foundation Models Are Not Plug-and-Play for CGM Forecasting

Zero-shot time-series foundation models do not reliably outperform classical statistical baselines like Elastic Net or specialized architectures like PatchTST when predicting continuous glucose monitor readings. An empirical evaluation across eight public datasets covering Type 1, Type 2, and non-diabetic populations reveals that these models lack the domain-specific sensitivity required for complex glucose dynamics. Clinical performance depends on how effectively you calibrate these models to physiological reality.

The Failure of Zero-Shot Inference

Foundation models are trained on vast, heterogeneous time-series corpora, learning to predict the next token based on statistical patterns seen in energy demand, traffic, or climate data. When applied to continuous glucose monitoring, the model treats the glucose signal as a generic numeric sequence. Because these models are trained on mean-squared-error objectives that prioritize trend-following and smoothing, they struggle to capture the sharp, discontinuous inflections of a glucose spike. They treat an impending hyperglycemic event as a simple continuation of previous values. The model lacks the internal weights to recognize physiological constraints such as the characteristic lag between carbohydrate intake and blood glucose rise.

Adapting via Low-Rank Matrices

Lightweight fine-tuning acts as a critical bridge. By updating the model's adapter layers via Low-Rank Adaptation, you inject two small, trainable matrices into the attention layers. These matrices decompose the weight updates, creating a lightweight, additive layer that sits on top of the fixed pre-trained weights. By training only these low-rank matrices, you effectively shift the model's attention to prioritize metabolic signals without overwriting its general time-series knowledge. Using the Chronos-Bolt model, this fine-tuning reduces root mean squared error by 6.5% to 18.4% in Type 1 diabetes cohorts and 8.6% to 18.2% in non-diabetes or Type 2 cohorts.

Data Fusion for Clinical Accuracy

For those building medical software, treat these models as feature extractors rather than complete solutions. The CGMacros framework demonstrates that incorporating multi-modal inputs is the path to clinical accuracy. This framework reconciles disparate inputs by embedding visual food data and macronutrient logs into a shared latent space alongside the glucose history. A residual-based fusion layer adds this processed external data to the model's temporal hidden states. This layer uses a learned attention bias to map meal intake to a specific, future timestamp where the glucose response is expected to peak. This fusion improves postprandial RMSE by approximately 15% relative to the CGM-only baseline, while the total system error drops by approximately 3%.

Mapping Metabolic Drivers

Consider an insulin-dependent patient. A raw foundation model sees a steady glucose reading and predicts a flat line, missing the effect of a recently consumed meal that has not yet hit the bloodstream. Because the model lacks knowledge of carbohydrates, it cannot anticipate the coming spike. By integrating the nutrient log into the fusion layer, you inject a physical signal—the carbohydrates—that the model uses as a leading indicator. The temporal model shifts its prediction trajectory before the sensor registers the rise because the fusion mechanism maps the nutrient input to a specific, delayed, and predictable glucose response profile. This transition from ignoring external metabolic drivers to treating them as quantitative inputs is the difference between a model that merely tracks history and one that forecasts metabolic states. Whether these techniques will generalize to even higher-frequency or more complex physiological markers remains to be determined by further clinical implementation.