TEDDY: A Pediatric Foundation Model for Risk Forewarning from ICD-Coded Diagnostic Histories
Researchers developed TEDDY, a 1.84-million-parameter transformer model, trained on 73 million pediatric diagnoses to predict disease onset and visit timing with high accuracy, outperforming existing baselines.
Intelligence analysis by Gemini 2.5 Flash

TEDDY, a novel pediatric foundation model, leverages longitudinal diagnostic histories from electronic health records to forecast disease risks in children. It demonstrates superior predictive capabilities for a wide range of conditions, including rare diseases, and can detect potential health issues years before their first recorded diagnosis, using a relatively compact model archite…
Imagine a super-smart computer program called TEDDY that learns from kids' past doctor visits, like a detective reading their health storybook. It can spot tiny clues in their medical records to guess if they might get sick with something like asthma or ADHD a long time before anyone else knows. This helps doctors help kids sooner, keeping them healthier as they grow up!
Analysis
Unlocking Pediatric Health Data
Pediatric electronic health records (EHRs) contain a wealth of developmentally structured clinical trajectories, yet their full potential for generative healthcare foundation models has largely remained untapped. The TEDDY (Temporal Event Decoder for Disease in Youth) model addresses this gap by presenting a 1.84-million-parameter decoder transformer. This model was meticulously trained on an extensive dataset comprising approximately 73 million ICD-10 diagnoses from 1.6 million children, all sourced from a single pediatric institution. By modeling longitudinal diagnosis trajectories and visit timing, TEDDY aims to extract meaningful patterns from complex patient histories, offering a new paradigm for understanding and predicting pediatric health.
Superior Predictive Capabilities
TEDDY's performance in predicting disease onset is notably robust. Across 797 distinct disease-onset prediction tasks spanning 16 ICD-10 chapters, the model achieved a median Area Under the Curve (AUC) of 72.0%. This significantly outperformed several established baselines, including DenseNet (50.0%), CNN (57.2%), RNN (60.1%), and LSTM (62.7%), on 96-99% of tasks. The model's predictive strength was particularly evident among lower-prevalence diagnoses, with 90% of the 225 rarest conditions showing 95% confidence intervals above chance. Furthermore, TEDDY demonstrated a remarkable ability to detect predictive signals more than two years before the first recorded diagnosis, with median AUCs of 59.7% in unrestricted analysis and 64.4% in a fixed-cohort sensitivity analysis. In specific benchmarks like asthma and attention-deficit/hyperactivity disorder (ADHD), TEDDY achieved AUCs of 79.3% and 84.7% respectively, surpassing even a general-purpose language model three orders of magnitude larger.
Implications for Clinical Practice
The development of TEDDY establishes pediatric diagnostic histories as a viable substrate for compact generative models. This approach supports broad, rare-disease, and long-horizon risk forecasting without the necessity for population-scale data or billion-parameter models, which are often resource-intensive and difficult to deploy. The model's ability to accurately predict disease onset and visit timing years in advance could transform preventative pediatric care, allowing clinicians to intervene much earlier. This could lead to improved patient outcomes, more efficient resource allocation, and a proactive rather than reactive approach to managing children's health, particularly for conditions that are challenging to diagnose early.
Key points
- TEDDY is a 1.84-million-parameter decoder transformer designed for pediatric risk forewarning.
- It was trained on approximately 73 million ICD-10 diagnoses from 1.6 million children at a single institution.
- The model achieved a median AUC of 72.0% for disease-onset prediction, outperforming several baseline models.
- TEDDY demonstrated strong performance for lower-prevalence diagnoses and could detect predictive signals over two years before diagnosis.
- It supports broad, rare-disease, and long-horizon risk forecasting using compact models and institution-specific data.
TEDDY's ability to forecast disease risk years in advance, especially for rare conditions, could revolutionize pediatric preventative care, enabling earlier interventions and potentially improving long-term health outcomes for children. Its compact size and reliance on institution-specific data make it a practical tool for widespread adoption in healthcare systems.
While promising, the model was trained on data from a single institution, which might limit its generalizability to diverse patient populations or different healthcare systems. Potential biases in the training data could lead to disparities in risk assessment, and the miscalibration in long-tail visit timing predictions suggests areas for further refinement before clinical deployment.

