Three calibrated XGBoost models for predicting risk of Diabetes, Chronic Kidney Disease (CKD), and Anemia from standard laboratory blood tests.

Model Details

Property	Value
Algorithm	XGBoost (Gradient Boosting)
Calibration	Platt Scaling (CalibratedClassifierCV)
Features	35 (13 base + 22 trend)
Training Data	NHANES 2017-2018 (~5,000 adults)
Explainability	SHAP TreeExplainer
Framework	scikit-learn, XGBoost, joblib

Performance (Test Set, n=4,909)

Model	AUROC	Brier	F1	Threshold
Diabetes	1.0000	0.0077	0.9918	0.402
CKD	0.9999	0.0026	0.9962	0.465
Anemia	1.0000	0.0010	0.9981	0.892

Files

File	Description	Size
`xgb_diabetes.joblib`	Raw XGBoost diabetes model	~563 KB
`xgb_ckd.joblib`	Raw XGBoost CKD model	~449 KB
`xgb_anemia.joblib`	Raw XGBoost anemia model	~279 KB
`calibrator_diabetes.joblib`	Platt-calibrated diabetes model	~3.3 MB
`calibrator_ckd.joblib`	Platt-calibrated CKD model	~2.6 MB
`calibrator_anemia.joblib`	Platt-calibrated anemia model	~1.7 MB
`evaluation_report.txt`	Full evaluation metrics	—

Usage

python

import joblib
import numpy as np

# Load calibrated model
data = joblib.load("calibrator_diabetes.joblib")
model = data["calibrated_model"]
features = data["features"]

# 35-feature input vector (13 base + 22 trend features)
# Base: age, sex_encoded, bmi, hba1c, creatinine, albumin, egfr,
#        wbc, rbc, hemoglobin, hematocrit, systolic_bp, diastolic_bp
# Trend: {param}_slope, {param}_delta for each lab parameter
X = np.array([[50, 0, 31.4, 8.2, 1.0, 4.0, 91.7, 7.2, 4.9, 14.5, 43.0, 142, 88,
               0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]])

prob = model.predict_proba(X)[:, 1]
print(f"Diabetes risk: {prob[0]:.1%}")  # ~94%

SHAP Explainability

python

import shap

raw = joblib.load("xgb_diabetes.joblib")
explainer = shap.TreeExplainer(raw["model"])
shap_values = explainer.shap_values(X)
# Positive SHAP = pushes toward high risk
# Negative SHAP = pushes toward low risk

Feature Importance (by SHAP)

Diabetes	CKD	Anemia
HbA1c	eGFR	Hemoglobin
Age	RBC	Hematocrit
BMI	BMI	Sex
Creatinine	Creatinine	Albumin
WBC	Age	RBC

Training Pipeline

Fully reproducible 9-step pipeline in the MedIntel repository:

Download NHANES 2017-2018 from CDC
Merge 8 tables, compute eGFR (CKD-EPI 2021), impute missing values
Construct labels (ADA, KDIGO, WHO criteria)
Synthetic longitudinal augmentation (15% of patients)
Train XGBoost with stratified 5-fold CV
Platt scaling calibration
Full evaluation (AUROC, PR-AUC, Brier, subgroup analysis)
Download BC5CDR corpus for NER
Fine-tune BioBERT for biomedical NER

Intended Use

Clinical decision support for healthcare providers
Risk screening in telemedicine settings
Research on multi-disease prediction

Limitations

Trained on US population (NHANES) — may not generalize globally
Trend features are synthetic (NHANES is cross-sectional)
Not a substitute for clinical diagnosis

Citation

bibtex

@software{medintel2026,
  title={MedIntel: AI-Powered Clinical Risk Intelligence Platform},
  author={Jha, Sushantak Parashar},
  year={2026},
  url={https://github.com/Sushantak17/MedIntel}
}

Sushantak17/MedIntel-clinical-risk-models

MedIntel — Clinical Risk Prediction Models