Uplift modeling to determine which fluid-norepinephrine regime results in a postoperative acute kidney injury-free recovery in patients scheduled for cystectomy and urinary diversions

Markus Huber^1*Marc Alain Furrer^1,2,3François Jardot¹Patrick Yves Wuethrich¹

• ¹Department of Anaesthesiology and Pain Medicine, University Hospital Bern, Bern, Bern, Switzerland
• ²Department of Urology, Inselspital, Bern University Hospital, University of Bern, Bern, Switzerland, Bern, Switzerland
• ³Department of Urology, Solothurner Spitäler AG, Olten, Switzerland, Olten, Switzerland

Background:Postoperative acute kidney injury (PO-AKI) remains common after surgery. Although risk prediction models for PO-AKI exist, it is still unknown which intraoperative regime in terms of fluid and norepinephrine administration is beneficial for a specific patient. We thus aim to investigate the potential of uplift modellinga framework combining causal inference and machine learningin identifying patients for which certain fluid and norepinephrine regimes result in a PO-AKI freerecovery.Methods:Data from a prospectively maintained cystectomy database at a single tertiary center (N=1'482, period 2000-2020) were used. Total intraoperative fluid balance (TIFB) and norepinephrine (NE) administration were dichotomized into a high TIFB/low NE and a low TIFB/high NE regime. Primary outcome was PO-AKI. Confounding was addressed with inverse probability of treatment weighting. Uplift was defined as the difference in likelihood of no PO-AKI with a high TIFB/low NE versus low TIFB/high NE treatment regime. We modelled uplift using logistic regression and random forests as outcome models. Model performance was evaluated with the area under the Qini curve (AUQC).Results:The uplift models demonstrated a higher ability (AUQC: 0.30, 95%-CI: 0.26 -0.30) compared to a random sorting strategy (0.06, 95%-CI: 0.02 -0.06) or a traditional prediction model (AUQC: 0.06 , 95%-CI: 0.03 -0.06) for PO-AKI in sorting patients according to the expected treatment benefit from either a high TIFB / low NE or a low TIFB / high NE regime. The performance of the uplift models is robust with respect to the fluid-NE dichotomization.Conclusion:Uplift modelling provides a clinically relevant step towards personalized medicine by considering the incremental benefit of an alternative treatment versus a control treatment on a patient's outcome, thus moving from a predictive towards a prescriptive risk assessment. We demonstrated the overall higher clinical utility of an uplift modelling approach compared to a prediction model of baseline PO-AKI risk in sorting patients according to the expected treatment benefit from either a high total intraoperative fluid balance / low norepinephrine regime or a low total intraoperative fluid balance / high norepinephrine regime with respect to postoperative acute kidney injury.