PHTS Graft Loss Risk Calculator

Pediatric Heart Transplant Survival Risk Assessment Tool

Model assumption: This calculator assumes the donor-recipient size ratio is between 70–200%. If the actual ratio falls outside this range, results may be less accurate.

Donor ischemic time: If donor ischemic time is not provided, the model will assume it is less than 240 minutes (4 hours). Please enter the most recent donor ischemic time when available.

Note: This calculator uses clinical features to predict graft loss risk. Enter the most recent patient values below (at transplant is preferred over at listing). Calculate risk score and view top causal factors.

Cohort:

Model variant:

Kidney Function

eGFR at Transplant (mL/min/1.73m²)

BUN at Transplant (mg/dL)

Creatinine at Transplant (mg/dL)

Cardiac Support

Left Ventricular Assist Device (LVAD)

ECMO at Transplant

Mechanical Circulatory Support Device

Diagnosis & Demographics

CHD: Single Ventricle

Donor Ischemic Time (hours)

Lab Values

Serum Albumin at Transplant (g/dL)

AST at Transplant (U/L)

Other model features

Calculating risk score...

Risk Score

Top Causal Factors

Explore importance and causal responsibility per model (cohort × variant).

Cohort:

Model variant:

Causal Factor Importance

Causal Responsibility

Interactive Causal Factors

Adjust causal factor values to see how they affect patient risk in real-time.

Risk Comparison

Reference Risk

Current Risk

Risk Change

Mean feature importance across CatBoost, XGBoost, and XGBoost RF (MC-CV).

Cohort:

Aggregated Feature Importance (Top 25)

Loading…

Review and compare the 15 reverse feature sets (cohort × variant): what drives errors (over/under prediction) per model.

Summary across all 15 models

Compare over-prediction count, under-prediction count, and feature profile size for each model. Use the dropdown below to review one model’s full reverse feature set.

Loading…

Review one model’s reverse feature set

Select a model to see its missed-prediction drivers (support per feature for over/under/correct) and feature profile (support + IR).

Model:

For your last risk calculation (current cohort × variant)

Calculate risk in the Risk Calculator tab first; then this section will show missing or protective data elements for that model.

Missing or default data

Top causal factors that were not provided or were left at default/empty. Providing these can improve prediction.

Values that went against typical risk direction

Features where your value was in the "protective" or risk-lowering direction (e.g. no ECMO, lower donor age). The model used these as entered.

PHTS Risk Calculator – Model per Cohort

The calculator uses one model per cohort (CHD, Myocardio, Combined), each trained on the top 15 causal/importance features. Model directories: CHD_top/, Myocardio_top/, Combined_top/. Each cohort has its own sec_dx (secondary diagnosis) options and causal factors from that cohort’s SHAP/FFA run.

Workflow

Train (per cohort): train_python_models.py --cohort CHD --top_features_only (repeat for Myocardio, Combined) → best model by C-index (then AU-PRC) → outputs/models/{cohort}_top/
SHAP/FFA (per cohort): run_shap_ffa_workflow.py --cohort CHD --model-variant top (repeat for Myocardio, Combined) → cohort-specific dashboard_data.json (incl. sec_dx_dropdown_options) in outputs/shap_ffa/{cohort}_top/
Deploy: prepare_lambda_dir_phts.py copies all *_top/ models and dashboard data for Lambda; dashboard uses the selected cohort’s model and sec_dx options.

Final Model Metrics

Best model and MC-CV metrics per cohort are written to outputs/models/{cohort}_top/best_model.txt. Key metrics:

Metric	Description
C-index	Concordance index (primary selection). Mean ± SD and 95% CI across MC-CV splits. Higher = better ranking of risk.
Recall	Recall at evaluation time horizon (e.g. 1 year). Fraction of events correctly identified.
AUC	Area under the ROC curve at the time horizon.
AU-PRC	Area under the precision-recall curve (tiebreaker for best model selection).

Deployed values (loaded from the API). Best model chosen (algorithm + variant) and all metrics (C-index, Recall, AUC, AU-PRC) per cohort:

Loading metrics…

Performance summary – all models

C-index, Recall, AUC, and AU-PRC for every cohort × variant available in the deployed package (base, enhanced, top, wisotzkey, FULL).

Tabs

Risk Calculator: Enter features → risk score and top causal factors.
Top Causal Model Features: Importance and causal responsibility; adjust factors and see risk update in real time.
Aggregated Feature Importance: Mean importance across CatBoost, XGBoost, and XGBoost RF (MC-CV), per cohort.
Reverse Feature Importance: Review and compare all 15 reverse feature sets (cohort × variant). Summary table of over/under prediction and profile size per model; select a model to see its full feature profile (support, IR). After a risk calculation, see which data were missing or in a protective direction for that model.
Documentation: This reference.