Metabolic Risk Prediction — NHANES 2019-2020
Can you predict which patients are at risk for metabolic syndrome using real CDC clinical data? This hackathon challenges participants to build the best binary classifier predicting metabolic at-risk status from a cleaned, competition-ready subset of NHANES 2019-2020 data.
The Challenge The Adaptive Health Orchestration Engine (AHOE) previously achieved an AUC of 0.9743 predicting metabolic risk from NHANES 2017-2020 clinical data using a hybrid Bayesian fusion model. Your goal is to match or beat this benchmark using a dataset covering 9,693 adult patients across 29 clinical, demographic, dietary, and activity features.
The Target Variable
Participants must predict whether a patient is metabolically at-risk (target = 1) based on the following definition:
* HbA1c ≥ 5.7% (prediabetes threshold, ADA 2024) OR
* WHO Metabolic Syndrome criteria: ≥ 3 of elevated waist circumference, low HDL, high triglycerides, or elevated blood pressure.
Prevalence in this dataset is approximately 49.6%.
The Dataset Source: CDC NHANES 2019-2020 (National Health and Nutrition Examination Survey). The data includes 29 features across six domains: * Glycemic: HbA1c * Biochemistry: LDL, HDL, total cholesterol, triglycerides, CRP * Body measures: BMI, weight, height, waist circumference * Blood pressure: Systolic, diastolic * Demographics: Age, gender, ethnicity, poverty ratio, education * Dietary & Activity: Calories, carbs, fat, protein, sodium, activity score
The split consists of 7,754 training rows and 1,939 test rows (stratified 80/20).
Tracks and Awards * AHOE Benchmark Beater: Awarded to the top submission that achieves the highest AUC-ROC on the test set and beats the AHOE benchmark of 0.9743. Winners must publish their notebook publicly and share a methodology writeup. This is a non-monetary award.
Evaluation and Submission
Submissions are evaluated on AUC-ROC (Area Under the Receiver Operating Curve). Participants must submit a Kaggle notebook containing a CSV file with two columns: id and target (predicted probability between 0.0 and 1.0). External data is not permitted, but pre-trained models and feature engineering are encouraged. Teams may make a maximum of 5 submissions per day.