Skip to content

MODEL_CARD.md

Latest commit

 

History

History
168 lines (132 loc) · 6.4 KB

MODEL_CARD.md

File metadata and controls

168 lines (132 loc) · 6.4 KB

Model Card: OpenAge

Model Overview

Field Value
Model type Gradient Boosted Trees (sklearn GradientBoostingRegressor)
Variants Standard (21 features) and Extended (35 features)
Input Lab biomarkers (CBC + CMP) and medical history questionnaire items
Output Estimated biological age (continuous, in years)
Training data NHANES (CDC), ~50,000 records across 2003-2020 cycles
Intended use Research, personal health exploration, benchmarking
Not intended for Clinical diagnosis, medical decision-making without professional oversight
Experimental Autoencoder + 1D CNN (PyTorch) available in models/experimental/

Architecture

Primary: Gradient Boosted Trees

The production model uses sklearn's GradientBoostingRegressor, trained to predict chronological age from blood biomarkers and medical history.

Standard model (21 features):

  • n_estimators=4000, max_depth=8, min_samples_split=30, learning_rate=0.01
  • 16 lab biomarkers + 5 medical history questionnaire items
  • Test MAE: 5.11 years, R²: 0.906

Extended model (35 features):

  • n_estimators=6000, max_depth=10, min_samples_split=30, learning_rate=0.01
  • 30 lab biomarkers + 5 medical history items
  • Test MAE: 6.07 years, R²: 0.873

Experimental: Autoencoder + CNN

An experimental neural network architecture is also provided in openage/models/experimental/. It uses a two-branch design: an autoencoder compresses 42 biomarkers into a 30-dim latent space, and a 1D CNN predicts age from the latent representation. See models/experimental/autoencoder_cnn.py for details.

Input Features

Standard Model — 21 Features (16 lab + 5 questionnaire)

NHANES Code Name Unit Category
LBXMCVSI Mean Cell Volume fL CBC
LBXGH Glycohemoglobin (HbA1c) % Metabolic
LBXSATSI ALT IU/L Liver
LBXRBCSI Red Blood Cell Count million cells/uL CBC
LBXPLTSI Platelet Count 1000 cells/uL CBC
LBXSLDSI LDH IU/L Metabolic
LBXLYPCT Lymphocyte % % CBC
LBDLYMNO Lymphocyte Count 1000 cells/uL CBC
LBXSCK CPK IU/L Metabolic
LBXSCR Creatinine mg/dL Kidney
LBXSAPSI ALP IU/L Liver
LBXSKSI Potassium mmol/L Electrolyte
LBXRDW Red Cell Distribution Width % CBC
LBXMOPCT Monocyte % % CBC
LBXSBU Blood Urea Nitrogen mg/dL Kidney
LBXSGL Glucose mg/dL Metabolic
MCQ220 Cancer history binary Questionnaire
MCQ160D Angina history binary Questionnaire
MCQ160A Arthritis history binary Questionnaire
MCQ500 Liver condition history binary Questionnaire
MCQ550 Gallstones history binary Questionnaire

Extended Model — 35 Features (adds 14 markers)

Includes all 21 standard features plus:

NHANES Code Name Unit
LBXHGB Hemoglobin g/dL
LBXHCT Hematocrit %
LBXWBCSI White Blood Cell Count 1000 cells/uL
LBXSASSI AST IU/L
LBXSTB Total Bilirubin mg/dL
LBXSCA Total Calcium mg/dL
LBXSCLSI Chloride mmol/L
LBDLDLM LDL Cholesterol (Martin-Hopkins) mg/dL
LBDHDD Direct HDL Cholesterol mg/dL
LBXTR Triglycerides mg/dL
LBXSTP Total Protein g/dL
LBXSC3SI Bicarbonate mmol/L
LBXSOSSI Osmolality mmol/Kg
LBXSNASI Sodium mmol/L

Training Data

Field Value
Dataset NHANES (National Health and Nutrition Examination Survey)
Source CDC NHANES
Years included 1999–2020 (multiple survey cycles)
Sample size after preprocessing ~50,000 records
Target variable Chronological age (RIDAGEYR)
Public availability Yes

Data sources within NHANES

Laboratory data was merged from the Biochemistry Profile (BIOPRO) and Complete Blood Count (CBC) datasets, joined on SEQN (respondent sequence number) with demographic data providing the age target.

Preprocessing

Step Method
Missing value handling Mean imputation across features (training set statistics)
Normalization None (raw biomarker values used directly)
Outlier handling None applied at preprocessing; model trained on full distribution
Feature engineering None; raw NHANES biomarker values used as-is
Train/test split 70/30 random split (random_state=3454)

Training Procedure

Standard Model (21 features)

Parameter Value
Model GradientBoostingRegressor (sklearn)
n_estimators 4,000
max_depth 8
min_samples_split 30
learning_rate 0.01
loss least squares ("ls")
Train/test split 70/30, random_state=3454

Extended Model (35 features)

Parameter Value
Model GradientBoostingRegressor (sklearn)
n_estimators 6,000
max_depth 10
min_samples_split 30
learning_rate 0.01
loss least squares ("ls")
Train/test split 70/30, random_state=3454

Evaluation

Standard Model — NHANES Test Set

Metric Train Test
MAE 4.47 years 5.11 years
R² 0.928 0.906
Pearson r 0.963 0.952

Extended Model — NHANES Test Set

Metric Train Test
MAE 2.32 years 6.07 years
R² 0.966 0.873
Pearson r 0.983 0.934

Survival Analysis

Metric Value
Univariate Cox PH HR 1.098 per year (95% CI: 1.095–1.100)
Concordance 0.83

See BENCHMARKS.md for full performance analysis including feature importance rankings and subgroup breakdowns.

Ethical Considerations

  • Demographic bias: NHANES oversamples certain demographic groups (Hispanic, Black, Asian, older adults, low-income) by design. Model performance may vary across subpopulations.
  • Geographic scope: Trained on a US-representative survey. Applicability to non-US populations has not been validated.
  • Over-interpretation risk: Users may interpret biological age as a definitive health indicator. The model estimates a statistical quantity, not a clinical diagnosis.
  • Equity: Biomarker reference ranges vary by ancestry and sex. The model does not explicitly account for these differences beyond what is captured in the training data distribution.