Attention to early stages: predicting acute kidney injury in a post cardiosurgical ICU setting using an inclusive time-to-event model

Background

Acute kidney injury (AKI) is a critical complication in intensive care units (ICUs) that is known to have multifaceted impacts. However, as AKI is often detected too late, early prediction is crucial for timely intervention.

Methods

We used an attention-based time-to-event model to estimate the risk of a patient's first AKI incidence in a post-cardiosurgical ICU setting, irrespective of commonly employed limitations such as focusing on severe stages (2 & 3). Pre-, intra-, and postoperative data from 8564 adult patients were included, and AKI was defined by adhering to the full Kidney Disease: Improving Global Outcomes (KDIGO) definition. Models were primarily evaluated using the concordance index (CI).

Results

70.4 % of patients developed AKI, with stage 1 being the most frequent initial stage (88.1 %). The attention-based network outperformed our baseline model, achieving CIs of 0.80, 0.72, and 0.69 for ranking event risks up to 6, 12, and 24 h prior to the onset. In terms of converting the task to a classification problem for literature comparison, we obtained areas under the receiver operator characteristic curve (auROCs) of 0.82–0.73. Performance improved for severe AKIs only, yielding CIs of 0.92, 0.85, and 0.75, and auROCs ranging between 0.94 and 0.78.

Conclusion

We demonstrated the importance of early-stage AKI predictions and presented a novel approach to achieve this. Under similar assumptions, our results showed improvement and approached outcomes comparable to the literature. While practical validation is pending, we are confident that our approach proves useful in assisting physicians to prevent AKI development.