Data Source: UCI ML Repository - Heart Disease (1988). Original investigators: Detrano et al. View Dataset →

Total Patients
920
4 medical centers
Cleveland: 303, Hungary: 294, Swiss: 123, VA: 200
Disease Prevalence
44.7%
411 positive cases
Binary classification: 0=healthy, 1-4=disease
Best AUC-ROC
0.918
Random Forest
5-fold stratified cross-validation
Mean Age
54.4
Range: 29-77 years
Male: 68.4%, Female: 31.6%

Age Distribution by Outcome

UCI Heart Disease - All institutions

Data by Institution

4 contributing medical centers

Feature Correlation with Heart Disease

Pearson correlation coefficients

Chest Pain Type Distribution

cp: 1=typical, 2=atypical, 3=non-anginal, 4=asymptomatic

Feature Importance (Random Forest)

Top predictive features
thal
0.156
ca
0.143
cp
0.132
oldpeak
0.122
thalach
0.110

Feature Dictionary

13 clinical attributes from original study
Feature Description Type Range
ageAge in yearsNumeric29-77
sexSex (1=male, 0=female)Binary0, 1
cpChest pain typeCategorical1-4
trestbpsResting blood pressure (mm Hg)Numeric94-200
cholSerum cholesterol (mg/dl)Numeric126-564
fbsFasting blood sugar > 120 mg/dlBinary0, 1
restecgResting ECG resultsCategorical0-2
thalachMaximum heart rate achievedNumeric71-202
exangExercise induced anginaBinary0, 1
oldpeakST depression induced by exerciseNumeric0-6.2
slopeSlope of peak exercise ST segmentCategorical1-3
caNumber of major vessels (fluoroscopy)Numeric0-3
thalThalassemiaCategorical3, 6, 7

Random Forest

Ensemble - 100 trees

Accuracy85.2%
AUC-ROC0.918
F1-Score0.847
Precision84.6%
Recall84.8%

XGBoost

Gradient Boosting

Accuracy84.1%
AUC-ROC0.905
F1-Score0.836
Precision83.9%
Recall83.3%

Neural Network

MLP - 2 hidden layers

Accuracy83.9%
AUC-ROC0.897
F1-Score0.833
Precision83.1%
Recall83.7%

SVM

RBF Kernel

Accuracy83.6%
AUC-ROC0.891
F1-Score0.829
Precision82.8%
Recall83.0%

Logistic Regression

L2 Regularization

Accuracy82.0%
AUC-ROC0.879
F1-Score0.814
Precision81.2%
Recall81.6%

K-Nearest Neighbors

k=5, Euclidean

Accuracy78.7%
AUC-ROC0.842
F1-Score0.781
Precision77.9%
Recall78.3%

Cross-Validation Results

5-fold stratified CV

Best Model - Confusion Matrix

Random Forest on test set
Pred: Neg
Pred: Pos
Actual: Neg
89
13
Actual: Pos
14
68
0.918
AUC-ROC
85.3%
Accuracy

Benchmark References

  • Detrano et al. (1989) - Original dataset, American Journal of Cardiology
  • Aha & Kibler (1988) - UCI ML Repository benchmark
  • Recent Kaggle competitions: 85-92% accuracy with ensemble methods

NLP Benchmark: MIMIC-III (PhysioNet) + i2b2 2010 Challenge. Models: ClinicalBERT, BioBERT, scispaCy.

Clinical Notes
2.08M
MIMIC-III total
NER F1-Score
89.4%
ClinicalBERT
Entity Types
7
Medical categories
Vocab Size
28,996
BERT tokens

Sample Clinical Note with NER

MIMIC-III discharge summary format
DISCHARGE SUMMARY

Patient is a 63-year-old male with history of coronary artery disease, hypertension, and type 2 diabetes who presented with chest pain and shortness of breath.

Patient underwent cardiac catheterization revealing three-vessel disease. Subsequently had CABG x3.

Discharge Medications:
Aspirin 81mg daily, Metoprolol 50mg BID, Lisinopril 10mg daily

NER Model Performance

i2b2 2010 Challenge benchmark

Clinical NLP Benchmarks

Published results on i2b2 and MIMIC-III
Model Architecture Precision Recall F1
ClinicalBERT Transformer 90.2% 88.7% 89.4%
BioBERT Transformer 89.1% 87.9% 88.5%
scispaCy spaCy 84.2% 82.7% 83.4%
BiLSTM-CRF RNN + CRF 85.8% 84.1% 84.9%

⚖️ AI Fairness in Healthcare

Analyzing Random Forest performance across sex and age groups per NIST AI RMF and IBM AIF360 guidelines.

Performance by Sex

Random Forest on UCI Heart Disease
GroupNPrevalenceAccuracyTPRFPR
Male72645.2%84.8%86.2%16.7%
Female19425.3%82.5%79.6%14.8%

Performance by Age Group

Random Forest on UCI Heart Disease
AgeNPrevalenceAccuracyTPRFPR
<4514231.0%86.6%84.1%11.2%
45-5428642.0%85.3%86.7%15.9%
55-6434449.1%84.0%87.5%19.4%
≥6514854.7%82.4%85.2%21.2%

Fairness Metrics (AIF360)

Values > 0.8 indicate acceptable fairness
Demographic Parity
0.87
Equalized Odds
0.82
Equal Opportunity
0.91
Calibration
0.78
Predictive Parity
0.85

Fairness References

  • NIST AI Risk Management Framework (AI RMF 1.0) - January 2023
  • IBM AIF360: Fairness metrics toolkit
  • Obermeyer et al. (2019) - Dissecting racial bias in healthcare algorithms, Science