Heart Failure Cardiac Arrest Prediction

> Research explainer: This briefing examines verified primary research published 68 days before the briefing date. It is not a same-day research update and does not provide medical advice.

Evidence

This Research explainer examines a retrospective observational registry-based study of whether a machine-learning model could predict cardiac arrest after hospital discharge among adults with a new diagnosis of heart failure. The population comprised patients discharged alive after their first hospitalisation for heart failure and identified in the Swedish Heart Failure Registry (SwedeHF); those receiving palliative care at discharge or an implantable cardioverter defibrillator or cardiac resynchronisation therapy with defibrillator at discharge were excluded. pmid:42350023

The reported cohort included 45 068 patients. Cardiac arrest was defined through registration in the Swedish Registry for Cardiopulmonary Resuscitation through final follow-up, while deaths without resuscitation were treated as competing events. The investigators developed a Random Survival Forest model for competing risk using 82 predictors and assessed it using Brier score, observed-versus-predicted cumulative incidence, C-index, and time-dependent AUC-ROC. pmid:42350023

The study reported that 2399 (5%) patients received cardiopulmonary resuscitation. It also reported a C-index of 0.52 and time-dependent AUC-ROC of 0.63-0.65 for the model. These are reported performance measures, not estimates of a treatment effect, and the supplied evidence does not provide a comparator intervention, confidence interval, or p-value for them. pmid:42350023

The reported cumulative incidence of cardiac arrest increased gradually over follow-up, remained below 8%, and tended to plateau after approximately 10 years. The authors reported a low Brier score for prediction of cumulative incidence at the group level, while concluding that the model could not accurately predict cardiac arrest in individual patients newly diagnosed with heart failure. pmid:42350023

Analysis — Individual-level prediction performance

This study addresses a clinically consequential outcome, but its principal finding is about model performance rather than prevention, treatment, or prognosis for a particular person. Its registry-based design, large reported cohort, and use of competing-risk methods are relevant because a death without resuscitation can prevent observation of a later CPR-treated cardiac arrest. The model’s assessment therefore reflects the outcome definition and the handling of competing events as well as the predictors used. The reported low Brier score supports the authors’ statement that predicted cumulative incidence had group-level performance; it does not establish accurate patient-by-patient risk classification. pmid:42350023

The C-index of 0.52 and AUC-ROC of 0.63-0.65 were reported alongside the authors’ conclusion of limited individual-level accuracy. These metrics should be read as measures of discrimination in this development setting, not as evidence that the model changes outcomes or identifies a clinically actionable threshold. No confidence intervals, p-values, external-validation result, or comparison with an alternative prediction approach is supplied here. Accordingly, statistical significance cannot be assessed from this evidence packet, and even a statistically significant performance difference would not by itself demonstrate clinical usefulness. pmid:42350023

The study’s framing also matters: the operational outcome was cardiac arrests treated with CPR, rather than the biological event of cardiac arrest as such. That distinction narrows what the reported prediction target represents. The evidence supports describing the work as a model-development study in a defined Swedish registry population, with group-level cumulative-incidence performance but limited individual-level prediction; it does not support patient-specific conclusions, recommendations, or predictions beyond that population and outcome definition. pmid:42350023

Limitations

The supplied limitations state that the model predicts CPR-treated cardiac arrests rather than the biological event of cardiac arrest, does not reflect changes in health over time, and lacked information on transitions to palliative care and Do-Not-Attempt-CPR orders. The authors state that the latter missing information limits the model’s clinical relevance. pmid:42350023

The evidence packet also notes that model performance might improve with more decision trees and more repeated cross-validation. This possibility is not a demonstrated improvement and does not alter the reported performance values. Because the supplied material describes a single registry-based development study and does not provide external validation, confidence intervals, p-values, or a comparator model, the briefing cannot establish how the model would perform in another setting or whether it would improve clinical outcomes. pmid:42350023