Preemie-Predict

Data-Driven Neuroprognostication for Children Born Preterm

Preemie‑Predict is a publicly accessible clinical outcome prediction tool developed to support clinicians with a standardized, data‑driven approach to neuroprognostication for children born preterm.  

The incidence of preterm birth is 8 per cent in Canada and is a leading cause of childhood disability. Children born preterm are at increased risk for neurodevelopmental challenges, including cognitive, language, and motor impairments, as well as cerebral palsy. These outcomes can vary as they are influenced by multiple clinical and environmental factors, making neuroprognostication difficult. The variability in outcomes and limitations of neuroprognostication pose significant challenges for risk-stratification and clinical decision-making for children born preterm.    

About Preemie-Predict

What is Preemie‑Predict?  

Preemie‑Predict is an innovative clinical outcome prediction tool that uses supervised machine‑learning models trained on a well‑characterized prospective cohort of children born preterm (N = 545). The model was trained on a preterm cohort dataset, which includes detailed clinical, neuroimaging, and developmental data, to capture the complex, nonlinear relationships that influence neurodevelopmental outcomes.   

Access Preemie-Predict

Who is Preemie‑Predict for?  

Preemie‑Predict is intended for healthcare providers involved in the care and follow‑up of children born preterm, including: 

Development of Preemie-Predict 

This clinical outcome prediction tool was developed using data from a prospective preterm cohort (N=545), followed from birth to 4 years of age. The cohort is well characterized with clinical, neuroimaging and developmental data. The Preemie-Predict algorithm was developed using supervised machine learning with the preterm cohort data. Machine learning models included regression, random forests, neural networks, with an emphasis on gradient-boosted decision tree models (XGBoost). XGBoost performs well on clinical data and can model nonlinear relationships and interactions. Model performance was evaluated using leave-one-out cross-validation for internal validation, and external validation is underway.