Accurate prediction of intradialytic hypo- and hypertension in hemodialysis patients: the dual-attention transformer model

Abstract Background and Hypothesis Intradialytic hypotension (IDH) and intradialytic hypertension (IDHTN) are frequent haemodialysis complications; each independently linked to increased cardiovascular risk and all-cause mortality. Existing predictive models often fail to capture the irregular nature of real-world haemodialysis data. The study objective was to develop an advanced machine learning architecture specifically designed for irregular time-series analysis of dialysis. Methods . We conducted a retrospective analysis of 1 300 haemodialysis patients, encompassing 182 111 sessions and 1 201 323 timestamped observations. Patients were randomly assigned to training (80%), validation (10%), and testing (10%) cohorts. IDH and IDHTN were defined by systolic blood pressure changes from pre-dialysis baseline: IDH is defined as a systolic blood pressure lower than 90 mmHg, and IDHTN is defined as an increase of ≥10 mmHg in systolic blood pressure without the occurrence of IDH. A Transformer model with dual-attention mechanism was developed for real-time event prediction. Performance was evaluated using area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), and F1 score. Results In the whole cohort, the incidence of IDH and IDHTN was 0.79% and 25.60%, respectively. The Transformer model demonstrated robust predictive accuracy in the test cohort, achieving AUROCs of 0.96 for IDH and 0.88 for IDHTN (all P < 0.01). The model’s high AUPRCs for both definitions confirmed its sensitivity in detecting early or moderate blood pressure fluctuations. To assess clinical utility, a complementary LightGBM model using pre-dialysis features predicted common symptoms or interventions with an accuracy of 0.926. Feature importance analyses validated established risk factors and identified key time-dependent covariates. Conclusion Our dual-attention Transformer model enables accurate, real-time prediction of intradialytic blood pressure abnormalities by effectively interpreting complex, asynchronous clinical data.