Swedish researchers build AI model to spot heart failure patients at risk of cardiac arrest
Researchers developed a machine learning tool that can identify which heart failure patients are most likely to suffer cardiac arrest, potentially enabling preventive interventions. Since survival rates after cardiac arrest remain dismally low—just 12% for out-of-hospital cases—better prediction could redirect resources and save lives.
Originaltitel: Prediction of cardiac arrest in patients with heart failure in Sweden: a registry study with development of a machine learning model
**Maskinlärningsmodell kan inte identifiera hjärtpatienter i arrest-risk** Ett nyutvecklat prediktionsverktyg för att identifiera hjärtsviktspatienter med högt risk för hjärtstillestånd presterar dåligt på individnivå. Forskare vid Göteborgs universitet och Sahlgrenska sjukhuset byggde en Random Survival Forest-modell med 82 prediktorer från Svenska hjärtsviktregistret, baserad på 45 068 nydiagnostiserade patienter (2005–2021). Endast 5 procent erhöll återupplivning, medan 71 procent dog utan intervention. Modellens C-index nådde endast 0,52 och AUC-ROC 0,63–0,65 — långt under klinisk användbarhet. Forskarna konkluderar att dödsrisken utan återupplivningsförsök överskuggar förmågan att skilja ut arrest-kandidater. För regionala inköpare och kliniker innebär detta att algorithmer baserade på befintliga registerdata inte löser behovet av bättre riskövervakning. Nya datakällor eller biologiska markörer krävs före någon implementering kan motiveras.
<p>Objective: 30-day survival after cardiac arrest is low, 12.4% and 36% for out-of-hospital and in-hospital cardiac arrest, respectively. Heart failure is a known risk condition for cardiac arrest. Improving our ability to identify patients at high risk of cardiac arrest would enable prevention. We aimed to develop a prediction model for cardiac arrest to be used in patients newly diagnosed with heart failure.</p><p>Design: A nationwide registry-based observational study.Setting Data were sourced from the Swedish Heart Failure Registry (1 January 2005 to 31 December 2021).Participants This cohort included 45 068 patients discharged from hospital after first hospitalisation for newly diagnosed heart failure. Patients discharged from hospital with palliative care and/or implantable defibrillators were excluded.Outcome measure and analysis The primary outcome was defined as cardiac arrest registered in the Swedish Registry for Cardiopulmonary Resuscitation until final follow-up (15 November 2022). Patients who died without resuscitation were treated as competing events. A Random Survival Forest model for competing risk was developed using predictors from the heart failure registry. The model was evaluated with Brier score, observed versus predicted cumulative incidence, Concordance-index (C-index) and time-dependent area under the curve of a receiver operating characteristics graph (AUC-ROC).</p><p>Results: In this cohort, 2399 (5%) patients had received cardiopulmonary resuscitation (CPR) (5%), and 31 989 (71%) patients died without resuscitation. Our model with 82 predictors had a low Brier score indicating a capacity to accurately predict cumulative incidence of cardiac arrest on a group level. However, the model also had a low C-index 0.52 and low AUC-ROC 0.63-0.65.</p><p>Conclusion: Our Random Survival Forest model for competing risk could not accurately predict cardiac arrest in individual patients newly diagnosed with heart failure, because the event death without attempted resuscitation was treated as a competing event. The lack of information on transitions to palliative care and Do-Not-Attempt-CPR-orders limits the clinical relevance of any cardiac arrest prediction model.</p>