Re-evaluating the Safety and Efficacy of AKI Electronic Alerts
Source PublicationJournal of the American Medical Informatics Association
Primary AuthorsWissel, Percy, Zachem et al.

Researchers claim that tailoring AKI electronic alerts could reduce all-cause mortality, though this conclusion rests on a secondary analysis of existing datasets rather than a fresh clinical trial. The investigation re-evaluated individual patient data from three randomised controlled trials—ELAIA-1, ELAIA-2, and UPenn—to determine if specific patient phenotypes respond differently to digital prompts.
The central premise of the study is that automated alerts are not benign interventions. By utilising machine learning models trained on the ELAIA-1 trial and validating them against external cohorts (n = 7,453), the team sought to predict which patients would benefit from a notification and which might be harmed. The primary outcome measured was 14-day all-cause mortality.
Heterogeneity in AKI Electronic Alerts
The analysis revealed a stark divergence in outcomes. Patients predicted to benefit from the intervention showed a lower risk of death in validation cohorts. Conversely, the data suggests that applying these alerts indiscriminately may have adverse consequences. The authors estimate that 43 deaths in the external cohorts might have been preventable had the alerts been restricted solely to likely beneficiaries.
Specific subgroups appeared to drive this variance. Machine learning identified reduced mortality among patients presenting with higher blood pressure and lower predicted risk. However, a concerning signal emerged regarding the setting of care. The analysis found increased mortality associated with alerts deployed in non-urban and non-teaching hospitals. This implies that the utility of the tool is heavily dependent on the clinical environment and the resources available to respond to the signal.
Provider behaviour offers a potential explanation. The study noted that actions taken by clinicians following an alert differed across subgroups, suggesting that the alert triggers different distinct workflows depending on the context. While these findings advocate for precision over broad application, one must exercise caution. These are theoretical projections derived from retrospective data. As the authors note, a dedicated prospective trial is required to confirm whether suppressing alerts for specific phenotypes definitively improves survival.