Quick Patient Presets:

Clinical Patient Profile

Live parameter inputs for real-time model scoring
Patient #3 Loaded
36.8 kg/m²
<18.5 (Underweight) 18.5-24.9 (Normal) 25-29.9 (Overweight) 30-39.9 (Obese I/II) 40-80+ (Severe Obese III)
Clinical History & Lifestyle Factors
High Blood Pressure Hypertension history
High Cholesterol Diagnosed dyslipidemia
Physical Activity Exercise in past 30 days
Smoker 100+ cigarettes lifetime
Coronary Heart Disease Prior MI or angina
Difficulty Walking Mobility impairment
Interactive Clinical "What-If" Counterfactuals Live Delta

Multi-Class Risk Assessment

Real-time XGBoost soft-probability inference
Class 2: Diabetes
Estimated HbA1c Equivalent: ≥ 7.6% (Diabetic Range)
Clinical Action Recommendation: Diagnostic HbA1c + OGTT Panel
No Diabetes (Class 0) 8.2%
Prediabetes (Class 1) 4.6%
Diabetes (Class 2) 87.2%
Key Local Risk Contributors (TreeSHAP) View Full Waterfall →
GenHlth: Fair (+1.82) High Blood Pressure (+1.15) BMI 36.8 (+1.08) High Cholesterol (+0.92)

Model Consensus Agreement

Model Architecture P(No-DM) P(Pre-DM) P(DM) Decision Stratum
Multi-Class XGBoost 8.2% 4.6% 87.2% Diabetes (Class 2)
Random Forest (200 Trees) 11.4% 3.8% 84.8% Diabetes (Class 2)
Logistic Regression (Balanced) 6.5% 14.2% 79.3% Diabetes (Class 2)

Patient TreeSHAP Waterfall Attribution

Exact Shapley value impact on Diabetes log-odds for active profile
Interactive Local Attribution
Red bars increase predicted diabetes risk  |  Blue bars decrease predicted diabetes risk

Global Clinical Risk Drivers

Mean Absolute TreeSHAP importance across 253,680 CDC BRFSS records
Cohort Level
General Health status, Hypertension, and BMI drive over 67% of total tree decision splits.

Unsupervised Patient Phenotype Discovery (K-Means, K=3)

Cluster profiling across 21 clinical, metabolic, and lifestyle dimensions
Silhouette Evaluated (K=2..6)

Phenotype 0: Low-Risk Normoglycemic Baseline

Cohort Size:52.4%
Mean BMI:25.8
Hypertension (%):18.2%
High Cholesterol (%):22.4%
Diabetes Rate:4.8%
Normative metabolic health, normal/mild BMI, low comorbidity burden, high physical activity adherence.

Phenotype 1: Moderate Metabolic Syndrome & Aging Cohort

Cohort Size:31.8%
Mean BMI:29.4
Hypertension (%):58.7%
High Cholesterol (%):54.1%
Diabetes Rate:16.3%
Elevated cardiovascular risk factors, overweight BMI tier, emerging hypertension and dyslipidemia.

Phenotype 2: High-Risk Multimorbid Comorbid Phenotype

Cohort Size:15.8%
Mean BMI:33.6
Hypertension (%):82.4%
High Cholesterol (%):76.8%
Diabetes Rate:38.9%
High-grade obesity, severe chronic disease clustering (hypertension, CAD, stroke), restricted physical mobility.
Patient Phenotype Alignment:
Based on current clinical attributes (BMI 36.8, Hypertension, CAD, GenHlth 4), this patient aligns with Phenotype 2: High-Risk Multimorbid Comorbid Phenotype (Diabetes Prevalence: 38.9%).

Held-Out Test Partition Benchmark (50,736 Patients)

Stratified multi-class evaluation comparing Macro-F1 and per-class detection sensitivity
Zero Data Leakage
Model Architecture Accuracy Macro-F1 Weighted-F1 No-DM Recall (0) Pre-DM Recall (1) DM Recall (2) Clinical Trade-Off
Multi-Class XGBoost 84.7% 0.458 81.4% 96.2% 1.4% 42.1% Highest Macro-F1 & precision; excels at healthy vs confirmed diabetic separation.
Random Forest (200 Trees) 84.1% 0.442 80.8% 95.4% 1.1% 38.6% Strong non-linear boundaries; robust ensemble averaging across 200 estimators.
Logistic Regression (Balanced) 73.2% 0.419 75.8% 74.1% 28.5% 68.2% Highest sensitivity for minority Prediabetes detection (28.5% recall) via balanced loss weighting.

The Class 1 (Prediabetes) Detection Challenge

Class 1 represents only 1.83% of the epidemiological cohort (4,631 records). Unweighted models optimize global accuracy by predicting majority classes (0 and 2). Balanced weighting dramatically lifts Class 1 sensitivity from 1.4% to 28.5%.

Clinical Utility of Gradient Boosting

Multi-Class XGBoost with softprob loss partitions the non-linear interaction between self-rated general health, BMI, and hypertension with higher fidelity than linear baselines.