Are Foundation Models the Future of Diabetes Care? A Critical Review

The Shift to Foundation Models

In recent years, the field of diabetes care has witnessed a major transformation in the deployment of artificial intelligence (AI). Unlike traditional models designed for single functions, innovations now include expansive foundation models pre-trained on a multitude of data sources, including images, glucose readings, and patient records.

A critical review published in Frontiers in Artificial Intelligence outlines this evolution and strikes a cautionary note, highlighting that the advancements in model architecture have surpassed the evidence substantiating their effectiveness.

Historical Context of Diabetes Models

The review categorizes the development of diabetes-related AI into three distinct eras:

• Era I: This initial phase was characterized by fixed risk scores like the UKPDS risk engine, relying on small cohorts and manually engineered factors such as age, HbA1c, and blood pressure.
• Era II: From 2016 to 2020, the focus shifted to task-specific deep learning systems, exemplified by the IDx-DR, which became the first autonomous AI diagnostic tool approved by the FDA.
• Era III: Starting in 2021, the spotlight has turned toward self-supervised pre-training methods using extensive data sets, enabling models to adapt to various tasks with minimal additional labeling.

Evaluating Foundation Models

One prominent example in this new era is RETFound, a Vision Transformer that leverages 1.6 million retinal photographs for pre-training. Its developers claim it outperforms task-specific models in multiple areas, including grading diabetic retinopathy and predicting systemic health events.

However, the review questions the validity of these assertions, noting that the validation data often come from the same source as the training set, raising concerns about overfitting and generalizability. Additionally, the predictive performance metrics reported are below clinical deployment thresholds.

The review also points out that comparative studies often reveal conflicting results. For instance, when ample labeled data is available, traditional models perform comparably to RETFound, highlighting the need for independent validation and broader testing.

Continuous Glucose Monitoring Developments

In the domain of continuous glucose monitoring (CGM), advances using pretrained temporal transformers are emerging for predictive capabilities, such as hypoglycemia forecasting. While some progress has been made, the evidence remains largely based on benchmark datasets without direct clinical applicability.

A concerning trend identified in the review is the predominance of type 1 diabetes focus within CGM foundation model research, with little evaluation of performance across different populations or conditions.

The Promise and Pitfalls of Clinical Large Language Models

Large language models, like Med-PaLM, have shown promise in medical benchmarks, but experts caution that these scores may not accurately reflect real-world operational capabilities. Many of the models are susceptible to benchmark contamination, leading to inflated performance claims.

Additionally, the authors observe a significant gap in safety evaluations, particularly for AI that delivers patient-facing content, such as dosage recommendations for insulin. Various complications arise when data becomes sparse, risking erroneous, yet confident, AI suggestions.

Call for Rigorous Evaluation Standards

The review advocates for a shift in the culture surrounding model evaluation rather than solely developing advanced architectures. It proposes a comprehensive nine-point standard that includes:

• Use of pre-registered, temporally separated test sets
• Disclosure of potential data contamination
• Mandatory reporting of calibration curves
• Default subgroup analyses
• Version controls for closed models
• Implementation of silent trials for real-world validation
• External validations incorporating multimodal data
• Testing for explanatory reliability
• Utilization of tuned baseline comparisons

This framework could guide the integration of foundational models into clinical practices effectively and safely.

Ultimately, establishing accountability in AI-driven diabetes care remains a pivotal challenge, necessitating urgent attention to ensure health equity and patient safety.