A “numerical autopsy” of out-of-hospital cardiac arrest: reclassifying unknown etiologies from five years of nationwide health data
P Cezard, T Laurenceau, L Wang, U Meli, E Menant, R Chocron, X JouvenAbstract
Background
A large proportion of out-of-hospital cardiac arrest (OHCA) are labeled as having an unknown etiology, mainly because most patients die at the scene after unsuccessful or non-initiated resuscitation and therefore do not undergo a complete diagnostic work-up. In addition, conventional autopsy remains costly, logistically complex, and not always feasible, further limiting etiologic ascertainment. This uncertainty hinders epidemiological understanding and may complicate family counseling and screening when an inherited or otherwise detectable condition is suspected. We developed a numerical autopsy approach to identify the most probable cause of OHCA using machine-learning models applied to longitudinal healthcare data.
Methods
We included 22,170 adults OHCA cases recorded in the Sudden Death Expertise Center registry between 2011 and 2020, of whom 18,821 (84.9 %) had an unknown etiology at the time of inclusion. Using the French National Health Insurance Database (SNDS), we extracted all hospital discharge diagnoses, outpatient drug dispensations, and reimbursed diagnostic or therapeutic procedures over the five years preceding the OHCA. We trained a multiclass LightGBM model on cases with known etiologies to estimate the probability of belonging to one of the seven medical categories: cardiac ischemic, cardiac other, acute pulmonary edema, pulmonary, neurologic, embolic and others. Model performance was evaluated using a temporally distinct validation set.
Results
The proportion of cases classified as unknown decreased from 84.9% before numerical autopsy to 9.3% after model-based reclassification. In the temporally distinct validation set, discrimination was good overall, with a weighted average AUC of 0.82, a log-loss of 1.12, and an F1-score of 0.33. Using a predicted probability threshold of 0.5, overall accuracy reached 64.0%. After reclassification, the distribution of etiologies was as follows: cardiac ischemic 61.9% (vs. 7.6% initially); pulmonary 19.5% (vs. 2.6%); cardiac other 2.9% (vs. 1.7%); neurologic 1.6% (vs. 1.1%); other 1.6% (vs. 1.0%); embolic 1.6% (vs. 0.7%); and acute pulmonary edema 1.6% (vs. 0.4%).
Conclusions
Using our algorithm, we could reclassify a large proportion of initially unknown OHCA cases using routinely collected longitudinal health data. These “numerical autopsies” could inform family counseling and screening strategies and strengthen population-level surveillance.Etiology distribution before vs after NA