Heart Failure Mortality Cohort Insights

• Heart Failure Mortality Cohorts are clearly defined patient groups selected through diagnostic codes and trial registries to study mortality risk.
• They integrate large-scale EHR data using advanced machine learning and survival analysis to uncover predictors like age, biomarkers, and comorbidities.
• Rigorous feature engineering, imputation, and ensemble modeling strategies ensure precise risk stratification and guide actionable clinical insights.

Heart failure mortality cohorts constitute precisely defined patient populations used to investigate risk factors, heterogeneity, and prognosis for death among individuals with heart failure (HF). Cohort construction, variable selection, modeling strategies, and evaluation frameworks vary based on research goals, healthcare context, and data modality. Contemporary studies leverage large-scale, real-world electronic health records (EHR), administrative claims, and clinical trial registries, employing advanced ML and statistical survival methods to characterize and predict HF mortality risk with unprecedented precision and interpretability.

1. Cohort Definitions and Population Characteristics

Heart failure mortality cohorts are typically drawn from EHR repositories, ICU/hospital registries, national health insurance databases, or multi-site clinical trials. Inclusion and exclusion criteria center on a diagnosis of heart failure, often substantiated by ICD-9/ICD-10 codes, administrative diagnosis positions, or validated phenotyping dictionaries. Many studies impose further restrictions by age, comorbidity profile, ICU admission, biomarker completeness, or absence of malignant disease.

Representative examples include:

• A Swedish EHR-based cohort: N = 42,820 adults (≥18 years) with first in-hospital HF diagnosis between 2015–2022; median age 80 years at initial HF admission; outcome ascertainment extended through 2023 (Dippela et al., 20 Nov 2025).
• A French national sample: 10,051 adults with incident HF hospitalization from 2010–2016, encompassing 85,594 hospitalizations and capturing both in-hospital and out-of-hospital deaths with nearly two years of follow-up (Murris et al., 2024).
• US real-world EHR and claims: UK CPRD validation set with 99,382 HF patients, median follow-up 9 months, 44% 4-year mortality (Rao et al., 16 Mar 2025).
• ICU-focused cohorts: MIMIC-III/IV-derived HF cohorts with inclusion based on ICU admission, age >18, and diagnoses consistent with HF (e.g. N=1,177 in MIMIC-III for in-hospital mortality; elderly HF+diabetes N=1,478 in MIMIC-IV for 28-day mortality) (Ashrafi et al., 2024, Fan et al., 18 Jun 2025).
• COVID-19 settings: 3,193 HF patients hospitalized with PCR-confirmed COVID-19 in Lombardy, Italy, with explicit linkage to hospital identifier and 45-day mortality (Caldera et al., 2024).
• Advanced HF/clinical trial benchmarks: Five cohorts (ESCAPE, BEST, GUIDE-IT, UVA Shock, UVA Serial) featuring NYHA III–IV patients, 6-month composite endpoints (death, LVAD, transplant, rehospitalization), and extensive hemodynamic, biomarker, and intervention variables (Lamp et al., 2023).

Common variables at baseline include age, sex, comorbidities (diabetes, renal disease, atrial fibrillation, COPD, hypertension), vital signs, laboratory measures (creatinine, NT-proBNP), and prior history of cardiovascular events (MI, stroke). Specialized cohorts may further include invasive hemodynamics, functional class/biometrics, and device/procedure exposures.

2. Mortality Endpoint Specification and Follow-Up

Mortality endpoints are defined to match study context and follow-up logistics, varying by time horizon and granularity:

• In-hospital mortality: binary outcome during index admission, with no postdischarge follow-up (Ashrafi et al., 2024).
• Short-term (28–90 day) mortality: used for elderly ICU populations and hospital-level outcome profiling (Khettari et al., 2 Apr 2026, Fan et al., 18 Jun 2025, Caldera et al., 2024).
• One-year mortality: assessed following index or latest HF hospitalization, with explicit event prevalence (e.g., 24.8% at initial, 46.7% post-latest hospitalization in the Swedish cohort) (Dippela et al., 20 Nov 2025).
• Three-month mortality: as per the French hospital cohort for HF admissions (Khettari et al., 2 Apr 2026).
• Multi-year all-cause mortality: up to 36 months (TRisk model) or as censored survival over two years (French EGB cohort), using both in- and out-of-hospital death registries (Rao et al., 16 Mar 2025, Murris et al., 2024).

Composite endpoints are common, often combining all-cause mortality and HF rehospitalization to better capture adverse outcomes in chronic HF populations (Farina et al., 18 Dec 2025). Precise temporal alignment of follow-up is achieved through registry linkage, administrative censoring at last contact, or exclusion of immediate in-hospital deaths to ensure eligibility for risk-window analysis.

3. Variable Extraction, Data Processing, and Feature Engineering

Variable sets encompass structured EHR data, laboratory analytes, vital signs, medication/procedure codes, and, in advanced designs, clinical text and functional or imaging parameters. Processing pipelines standardize, impute, and normalize inputs, while accounting for missingness, data imbalance, and outlier distributions.

Key strategies include:

• Uniform discretization (binned) of numeric variables (e.g., vital signs/lab values into 10 bins) and feature selection via recursive elimination, variance inflation factor (VIF), and ablation (Dippela et al., 20 Nov 2025, Ashrafi et al., 2024, Kia et al., 2023).
• Advanced missing-value handling via Missing-Value Aware Encoding (MVAE), imputations (median/mean for continuous), and explicit integration of missingness as a categorical state (Kia et al., 2023).
• Outlier treatment using methods such as Local Outlier Factor (LOF) and capping at clinically implausible values (Ashrafi et al., 2024, Kia et al., 2023).
• Class-imbalance correction through oversampling/undersampling techniques (e.g., SMOTETomek) and ensemble class weighting (Kia et al., 2023).
• Feature reduction for structured data by penalized regression (LASSO), random forest importance, or clustering-informed selection (Khettari et al., 2 Apr 2026).
• Extraction and fusion of clinical biomarkers (NT-proBNP, creatinine, BUN, RDW, leucocyte count); extraction of spatial (ECG, imaging) or functional (ejection fraction, oxygen flow, Braden scores) variables in intensive care settings (Ashrafi et al., 2024, Fan et al., 18 Jun 2025, Farina et al., 18 Dec 2025).
• For unstructured EHR: entity-level embeddings from clinical text, combining NER-annotated domain categories with transformer-based (BioBERT, CamemBERT) architectures, and various fusion approaches with structured variables (Khettari et al., 2 Apr 2026).

Temporal encoding for longitudinal data includes right-aligned visit sequences, discrete time-gap tokens, age/visit-index embeddings, and comprehensive event ordering—crucial for extracting dynamic trends and historical context (Dippela et al., 20 Nov 2025, Rao et al., 16 Mar 2025).

4. Modeling Frameworks and Statistical Approaches

Heart failure mortality prediction leverages a spectrum of ML and statistical survival modeling approaches, often tailored to the data structure and risk window.

Sequence and Survival Models:

• Transformer-based architectures (Llama, ModernBERT, TRisk) with self-attention mechanisms, rotary or relative positional encoding, and context-aware token embeddings excel in modeling high-dimensional longitudinal HF data (Dippela et al., 20 Nov 2025, Rao et al., 16 Mar 2025).
• State-space models (Mamba/Mamba2) offer linear complexity with respect to context length, utilizing selection mechanisms and state-space duality for efficient sequence representation (Dippela et al., 20 Nov 2025).
• Joint (shared-random-effects) models integrate longitudinal biomarker trajectories (NT-proBNP) with time-to-event survival, using Bayesian inference to estimate dynamic risk conditioned on observed time series (Farina et al., 18 Dec 2025).
• Multi-valued decision diagrams (MVDD, CARNA framework) provide interpretable, logic-based risk stratification that natively accommodates missing data—mapping high-dimensional, multi-modal HF features to discrete risk classes and explicit “phenotype” definitions (Lamp et al., 2023).
• Ensemble and gradient-boosted tree models (XGBoost, CatBoost, Random Forest) achieve high discrimination in both structured and multimodal HF datasets, benefiting from robust preprocessing, feature selection, and explainability tools (e.g., SHAP, ALE) (Ashrafi et al., 2024, Fan et al., 18 Jun 2025).
• Multimodal transformers and entity-aware NLP models fuse structured and text-based features for improved performance, using gates, attention, and cross-modal stacking (Khettari et al., 2 Apr 2026).
• Custom multilevel logistic cluster-weighted models (ML-CWMd) explicitly model latent subgroups (phenotypes), hospital effects, and heterogeneity via expectation-maximization, Ising models for binary features, and hierarchical logistic regression (Caldera et al., 2024).
• Survival ensemble approaches (random survival forests, survival gradient boosting) and distance/k-medoids clustering for trajectory-based risk stratification—especially when proportional hazards assumptions are violated (Murris et al., 2024).

5. Evaluation Metrics and Model Validation

Evaluation of heart failure mortality prediction models employs a set of standardized, rigorously defined metrics:

Metric	Purpose	Examples / Results
AUROC	Discrimination (binary/censored outcomes)	XGBoost AUROC 0.9228 (MIMIC-III, ICU) (Ashrafi et al., 2024); TRisk C=0.845 (UK 36-mo) (Rao et al., 16 Mar 2025)
AUPRC	Discrimination (class-imbalanced setting)	Medium-Llama AUPRC ≈ 0.574–0.845 (Dippela et al., 20 Nov 2025)
Brier Score	Probability calibration	Brier ≈ 0.145–0.146 (Llama variants) (Dippela et al., 20 Nov 2025)
C-index	Rank concordance (survival models)	TRisk C=0.845 (UK), C=0.802 (US transfer learned) (Rao et al., 16 Mar 2025)
Integrated Brier	Combined calibration/discrimination	IBS=0.1009 (main joint), IBS=0.0730 (high-freq) (Farina et al., 18 Dec 2025)
ICI	Absolute deviation from perfect calibration	ICI=0.0283 (joint model), ICI=0.053 (TRisk transfer)
F1, MCC	Balanced classification improvement	+3.6% F1, +2.7% MCC post-preprocessing (Kia et al., 2023)
Bootstrapping, CI	Statistical rigor in reporting estimates	95% confidence intervals for all discrimination/calibration metrics

Internal validation protocols incorporate K-fold cross-validation, outcome-stratified splits, and, in large registries, external hold-out test sets. Reference benchmarks (MAGGIC-EHR, ADHERE, EFFECT, SHFM, GWTG) are used for head-to-head comparison, with formal DeLong or paired tests for statistical significance (Lamp et al., 2023, Rao et al., 16 Mar 2025, Dippela et al., 20 Nov 2025). Calibration is evaluated via calibration curves, binned summaries, and integrated indices.

6. Interpretability, Phenotyping, and Clinical Implications

Interpretability and clinically meaningful stratification are central foci:

• Risk attribution is achieved by SHAP values (tree ensembles), integrated gradients (transformer survival models), and feature ablation/ALE plots (distributional inference frameworks), enabling identification of high-impact variables such as age, hemodynamic instability (APS III), oxygen flow, GCS eye-opening, Braden Mobility, NT-proBNP, RDW, and leucocyte count (Ashrafi et al., 2024, Fan et al., 18 Jun 2025, Farina et al., 18 Dec 2025).
• In MVDD/CARNA, each logical path corresponds to an explicit “phenotype,” providing Boolean threshold-based cluster definitions that are directly auditable by clinicians and align with domain knowledge (e.g., low cardiac index plus high PCWP → high-risk) (Lamp et al., 2023).
• Unsupervised trajectory clustering reveals subtypes with distinct risk kinetics (e.g., early death trajectories vs. repeated decompensations; renal-failure–dominated clusters), supporting differential triage and resource allocation (Murris et al., 2024).
• Joint modeling provides dynamic, individualized risk updates, with falling/rising NT-proBNP trajectories translating into adjusted survival probabilities for care planning (Farina et al., 18 Dec 2025).
• The cluster-weighted logistic approach enables risk stratification concordant with patient- and hospital-level heterogeneity, facilitating fairer health system benchmarking and actionable clustering (elderly “healthy,” “young low-risk,” “multimorbid” strata) (Caldera et al., 2024).

7. Limitations, Generalizability, and Best Practices in Cohort-Based HF Mortality Modeling

Principal limitations identified across cohorts and modeling frameworks include:

• Single-center bias or lack of geographic/ethnic diversity (e.g., PROVE registry Iran, single-site French cohort) (Kia et al., 2023, Khettari et al., 2 Apr 2026).
• Incomplete long-term follow-up or absence of post-discharge event data in some ICU studies (Ashrafi et al., 2024).
• Selection bias from stringent biomarker completeness requirements or exclusion of palliative-intent admissions (Farina et al., 18 Dec 2025, Fan et al., 18 Jun 2025).
• Token/embedding truncation issues for lengthy EHRs or clinical notes in transformer-based architectures (Khettari et al., 2 Apr 2026).
• Proportional hazards violations in traditional CPH modeling on HF cohorts (necessitating ensemble survival approaches) (Murris et al., 2024).
• Limited interpretability for deep multimodal architectures absent dedicated explainability modules (Khettari et al., 2 Apr 2026).

Best practices emerging include:

• Disease-specific HF cohorts with rich temporal, biomarker, and intervention data.
• Integrated use of sequence modeling, feature-aware imputation, robust metric reporting, and dynamic prediction approaches.
• Use of interpretable, phenotype-defining algorithms (MVDD, SHAP, ALE) and ablation studies to align models with clinical reasoning.
• Extensive cross-validation/external testing with stratified subpopulation analysis to mitigate bias and underpin generalizability.

Adopting flexible, dynamically updated, and interpretable risk stratification models is supported as the optimal framework for heart failure mortality cohort research and translational deployment (Dippela et al., 20 Nov 2025, Rao et al., 16 Mar 2025, Farina et al., 18 Dec 2025).