Companion to the December 2025 paper on hybrid Deep Neural Network + XGBoost ensembles. The stack fuses BRFSS and HEART_2020 cohorts, semi-supervised pseudo-labeling, and recall-first threshold tuning to deliver reproducible cardiovascular risk prediction.
The published manuscript consolidates the complete scientific narrative: motivation, literature review, methodology, experiments, and deployment guidance for the hybrid DNN + XGBoost system.
Cardiovascular disease drives 17.9 million deaths globally each year, and U.S. prevalence hovers around one in five deaths. The paper frames this project as a recall-focused clinical decision support tool that improves on Framingham-style risk charts (AUC 0.73–0.76) by modelling nonlinear interactions.
BRFSS 2020 provides >400K labeled survey rows; HEART_2020 contributes ~300K auxiliary rows. After harmonization and pseudo-label filtering, 80K confident examples extend the supervised signal without compromising quality (false pseudo-label rate <3%).
Section 3 of the paper lays out the end-to-end process: aligned data sources, hybrid model architectures, pseudo-labeling, and a recall-constrained decision policy. The highlights below mirror that narrative.
Feature alignment maps BRFSS and HEART_2020 columns, followed by a ColumnTransformer that scales numeric signals and one-hot encodes categoricals before splitting into 70/15/15 train/val/test folds.
XGBoost handles structured interactions, a 3-layer hourglass DNN captures nonlinear manifolds, and a logistic meta-learner fuses both validation logits. The stack yields +2.4% AUC over the best base model.
Only confident predictions become labels: pos_thresh=0.97, neg_thresh=0.03.
This expands the training corpus by 26.7% while keeping estimated noise <3%.
A 37-point grid search enforces recall ≥0.90 on validation data and selects the highest-precision option,
settling on t=0.3425 (test recall 0.9012, precision 0.7039).
Optuna orchestrates hyperparameter trials, artifacts are versioned to artifacts/, and the full
run takes ~15 minutes on a Xeon E5 + NVIDIA V100 workstation (Appendix E).
All metrics are computed on a held-out test set and stored in
artifacts/metrics.json. The hybrid model outperforms individual components.
The complete end-to-end pipeline from raw BRFSS data to production-ready model artifacts.
Load CDC BRFSS 2020 survey data containing 400k+ health indicator records with 18 features related to cardiovascular health.
Run brfss_preprocessing.py to handle missing values, encode categoricals, normalize numerics, and align feature schemas.
Execute prune_features.py for feature selection using importance scores and correlation analysis to reduce noise.
Train DNN for representation learning, then feed embeddings to XGBoost. Optuna tunes both model families for optimal performance.
Compute metrics, generate publication plots via sham_plots.py, and export artifacts to artifacts/ directory.
Understanding how data flows through the hybrid DNN-XGBoost architecture from input to prediction.
Raw features from BRFSS dataset including demographics (age, sex), lifestyle factors (smoking, alcohol), and medical history (diabetes, stroke, kidney disease).
Multi-layer neural network with batch normalization and dropout. Learns non-linear representations and extracts high-level features from tabular data.
Gradient boosting ensemble receives both original features and DNN embeddings. Handles feature interactions and provides robust classification.
Combines DNN probability outputs with XGBoost predictions using weighted averaging. Threshold optimization via recall-constrained precision sweep.
Final binary classification (Heart Disease: Yes/No) with calibrated probability scores. Decision threshold set at 0.3425 for optimal recall-precision balance.
High-resolution hero graphics generated in assets/ showcase the hybrid pipeline and workflow narrative.
assets/_- visual selection.png) now powers the hero and architecture sections, illustrating the hybrid DNN + XGBoost story.
assets/📌 Exact Layout Instructions_ - visual selection.png) highlights the training workflow narrative and supports the pipeline section.
Key excerpts from hybrid_dnn_xgb.py highlighting how stacked learners and recall-constrained thresholding work together.
Trains XGBoost and DNN separately, then feeds their validation probabilities to a logistic meta-learner that delivers the hybrid AUC lift.
def train_base_models_and_meta(X_train, y_train, X_val, y_val):
pos_ratio = y_train.mean()
scale_pos_weight = (1 - pos_ratio) / max(pos_ratio, 1e-6)
xgb_model = build_xgb_model(scale_pos_weight)
xgb_model.fit(X_train, y_train)
xgb_val = xgb_model.predict_proba(X_val)[:, 1]
dnn_model = build_dnn_model(X_train.shape[1])
dnn_model.fit(X_train, y_train, validation_data=(X_val, y_val),
epochs=DNN_EPOCHS, batch_size=DNN_BATCH_SIZE,
callbacks=[early_stop])
dnn_val = dnn_model.predict(X_val, batch_size=DNN_BATCH_SIZE).ravel()
meta_X = np.vstack([xgb_val, dnn_val]).T
meta_model = LogisticRegression(...).fit(meta_X, y_val)
hybrid_val_auc = roc_auc_score(y_val, meta_model.predict_proba(meta_X)[:, 1])Sweeps decision thresholds to guarantee ≥90% recall on validation, then locks in the option with the strongest precision, falling back to F1 if needed.
def tune_threshold_for_recall_precision(y_true, y_proba, min_recall=0.90):
best_thresh = 0.5
best_prec = best_rec = best_f1 = -1
chosen_by = "f1_fallback"
fallback_t = 0.5
fallback_prec = fallback_rec = 0.0
fallback_f1 = -1
for t in THRESH_GRID:
y_pred = (y_proba >= t).astype(int)
prec = precision_score(y_true, y_pred, zero_division=0)
rec = recall_score(y_true, y_pred, zero_division=0)
f1 = f1_score(y_true, y_pred, zero_division=0)
if rec >= min_recall and prec > best_prec:
best_thresh = float(t)
best_prec, best_rec, best_f1 = prec, rec, f1
chosen_by = "recall_constrained_precision"
if f1 > fallback_f1:
fallback_t, fallback_prec = float(t), prec
fallback_rec, fallback_f1 = rec, f1
if chosen_by == "f1_fallback":
best_thresh = fallback_t
best_prec, best_rec, best_f1 = fallback_prec, fallback_rec, fallback_f1
return best_thresh, {
"val_precision": float(best_prec),
"val_recall": float(best_rec),
"threshold_source": chosen_by,
}Statistical comparisons, pseudo-label ablations, and clinical threshold analyses from the paper feed directly into the dashboard below.
Stacked model improves on XGBoost (0.9215) and DNN (0.9093) courtesy of the logistic meta-learner.
Operating at t=0.3425 maintains cardiology-grade sensitivity while precision holds at 70.39%.
Adding 80K confident pseudo-labels increases validation AUC without measurable drift (noise <3%).
Data ingest 30s, XGBoost 120s, DNN 480s, pseudo-labeling 60s, threshold sweep 30s on V100-class hardware.
Publication-ready plots generated by artifacts/sham_plots.py for research papers and presentations.
Receiver Operating Characteristic curves for all three models. The hybrid ensemble (green) achieves the highest AUC by leveraging complementary strengths of DNN feature learning and XGBoost's gradient boosting. The diagonal represents random guessing (AUC = 0.5).
Shows how classification accuracy varies across decision thresholds. Each model has an optimal operating point. The hybrid model maintains stable high accuracy across a wider threshold range, indicating better probability calibration.
Critical for medical applications where missing positive cases (heart disease) is costly. The threshold was specifically chosen to maintain 90%+ recall while maximizing precision. Lower thresholds catch more positives but increase false alarms.
Side-by-side comparison of precision, recall, and F1 scores across all models. The hybrid approach consistently outperforms or matches the best individual model on each metric, demonstrating effective ensemble learning.
Direct comparison between the hybrid ensemble and the strongest single model (XGBoost). Bars show improvements across accuracy, precision, recall, F1, and AUC. The hybrid model achieves gains on all five headline metrics.
Bar chart ranking each model by Area Under ROC Curve. AUC measures the model's ability to distinguish between positive and negative classes across all thresholds. Higher is better, with 1.0 representing perfect discrimination.
Lollipop chart emphasizing the incremental AUC gains from the hybrid approach. The visualization highlights how combining DNN and XGBoost yields a 2-3% AUC improvement over either individual model—significant in medical ML applications.
All-in-one visualization combining ROC curves, accuracy sweeps, and key metrics in a single figure. Designed for academic papers and presentations where space is limited but comprehensive model comparison is needed.
The paper treats clinical deployment as a first-class concern. These cards summarize the governance, infrastructure, and monitoring requirements before the model can enter a hospital workflow.
Production-grade tools and frameworks powering the pipeline.
Clone, install, and run the complete pipeline in minutes.