Data Analysis Methodology Report

1. Data Sources

1.1 Primary Dataset: UCI Diabetes 130-US Hospitals

1.2 Secondary Dataset: CMS Hospital Readmissions Reduction Program

State-level and hospital-level readmission rates and penalty data based on CMS HRRP patterns. Geographic data covers all 50 states plus District of Columbia.

2. Data Cleaning and Preprocessing

2.1 Missing Value Analysis

2.2 Patient Deduplication

The raw dataset contains multiple encounters per patient. To ensure independence of observations and prevent data leakage:

• Records sorted by encounter_id
• Duplicates removed based on patient_nbr, keeping first encounter only
• Before: 101,766 encounters
• After: 71,518 unique patients

2.3 Target Variable Transformation

Original target had 3 categories. Transformed to binary for logistic regression:

3. Feature Engineering

3.1 Age Transformation

Original age values are categorical ranges. Converted to numeric midpoints:

age_mapping = {
    '[0-10)': 5,   '[10-20)': 15,  '[20-30)': 25,
    '[30-40)': 35, '[40-50)': 45,  '[50-60)': 55,
    '[60-70)': 65, '[70-80)': 75,  '[80-90)': 85,
    '[90-100)': 95
}

3.2 Derived Features

3.3 Final Feature Set

4. Machine Learning Pipeline

4.1 Train-Test Split

Stratified split used to maintain class proportions in both sets.

4.2 SMOTE Oversampling

Synthetic Minority Over-sampling Technique (SMOTE) applied to training data only:

from imblearn.over_sampling import SMOTE
smote = SMOTE(random_state=42)
X_train_balanced, y_train_balanced = smote.fit_resample(X_train, y_train)

4.3 Model Selection and Training

from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression

scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train_balanced)
X_test_scaled = scaler.transform(X_test)

model = LogisticRegression(max_iter=1000, random_state=42, C=0.1)
model.fit(X_train_scaled, y_train_balanced)

4.4 Model Evaluation

5. Risk Score Calculation

5.1 Probability to Score Conversion

Risk Score = P(readmission | features) x 100

Where P(readmission | features) is the predicted probability from logistic regression's predict_proba() method.

5.2 Cost Exposure Estimation

Estimated Cost = (Risk Score / 100) x $15,000

The $15,000 figure represents the average cost of a hospital readmission based on industry benchmarks.

5.3 Risk Stratification Tiers

6. Feature Importance Analysis

6.1 Top Risk-Increasing Factors

6.2 Protective Factors

7. Geographic Data Generation

7.1 State-Level Data

State-level readmission rates and penalties are generated based on actual CMS HRRP patterns:

for state_code, info in STATE_DATA.items():
    # Add realistic variation
    rate_variation = np.random.uniform(-0.8, 0.8)
    avg_rate = info['base_rate'] + rate_variation

    # Calculate penalty based on rate
    if avg_rate > 15.5:
        penalty_pct = np.random.uniform(0.5, 2.0)
    elif avg_rate > 14.5:
        penalty_pct = np.random.uniform(0.2, 0.8)
    else:
        penalty_pct = np.random.uniform(0, 0.3)

    # Estimate total penalty
    total_penalty = hospitals * 5_000_000 * (penalty_pct / 100)

7.2 Penalty Calculation Logic

CMS penalizes hospitals with Excess Readmission Ratios (ERR) greater than 1.0. Penalties are calculated as a percentage of Medicare payments:

• States with avg rate > 15.5%: 0.5-2.0% penalty range
• States with avg rate 14.5-15.5%: 0.2-0.8% penalty range
• States with avg rate < 14.5%: 0-0.3% penalty range
• Estimated Medicare payments per hospital: $5,000,000

8. Output Data Verification

8.1 Data Integrity Checks

8.2 Output Files

9. Limitations and Considerations

1. Data Age: UCI dataset is from 1999-2008; healthcare patterns may have evolved significantly
2. Model Performance: AUC of 0.564 indicates modest predictive power, typical for readmission models
3. Geographic Data: State-level metrics are simulated based on CMS patterns, not actual current data
4. Cost Estimates: Based on $15,000 average; actual costs vary widely by condition and facility
5. External Factors: Model cannot capture social determinants, patient compliance, or outpatient care quality
6. Single Disease Focus: Dataset is diabetes-specific; generalization to other populations requires validation

10. Reproduction Steps

# 1. Install dependencies
pip install pandas numpy scikit-learn imbalanced-learn

# 2. Download UCI Diabetes dataset to data/raw/

# 3. Run analysis pipeline
python run_analysis_v2.py

# 4. Copy outputs to dashboard
cp data/processed/*.json dashboard/lib/

# 5. Start dashboard
cd dashboard && npm run dev