Development and validation of AI-based triage support algorithms for prevention of intradialytic hypotension

Intradialytic hypotension remains one of the most recurrent complications of dialysis sessions. Inadequate management can lead to adverse outcomes, highlighting the need to develop personalized approaches for the prevention of intradialytic hypotension. Here, we sought to develop and validate two AI-based risk models predicting the occurrence of symptomatic intradialytic hypotension at different time points.

The models were built using the XGBoost algorithm and they predict the occurrence of intradialytic hypotension in the next dialysis session and in the next month. The initial dataset, obtained from routinely collected data in the EuCliD^® Database, was split to perform model derivation, training and validation. Model performance was evaluated by concordance statistic and calibration charts; the importance of features was assessed with the Shapley Additive Explanation (SHAP) methodology.

The final dataset included 1,249,813 dialysis sessions, and the incidence rate of intradialytic hypotension was 10.07% (95% CI 10.02-10.13). Our models retained good discrimination (AUC around 0.8) and a suitable calibration yielding to the selection of three classification thresholds identifying four distinct risk groups. Variables providing the most significant impact on risk estimates were blood pressure dynamics and other metrics mirroring hemodynamic instability over time.

Recurrent symptomatic intradialytic hypotension could be reliably and accurately predicted using routinely collected data during dialysis treatment and standard clinical care. Clinical application of these prediction models would allow for personalized risk-based interventions for preventing and managing intradialytic hypotension.