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

Figures

• 

  Figure 1:. Flowchart of data retrieval and cohort construction.

• 

  Figure 2:. Development of the Transformer model. (a) Examples of two typical intradialytic BP trajectories. The orange line illustrates a typical hypotensive episode, and the blue line illustrates a typical hypertensive episode. The red parts represent intradialytic hypotension or hypertension period. (b) Architecture of the dual-attention Transformer model. The diagram illustrates the network architecture for predicting Blood pressure (BP) fluctuation at the next time step (t + 1). Input features (Xt), historical timestamps (Tt), the next timestamp (Tt + 1) and the time interval (Δt) are processed through a Rotary Time Embedding layer. The embedded time vectors are used in a multi-head attention block where Tt + 1′ serves as the Query and Tt′ as the key to update the feature values (Xt). The output is then processed through a standard Transformer encoder block (Multi-Head Attention, Add & Norm, Feed Forward) before being concatenated with the time interval embedding and passed to a final Multilayer perceptron (MLP) layer for classification.

• 

  Figure 3:. Model performance on the test set. AUROCs for predicting IDH (Decrease), IDHTN (Increase) and normotensive state (Normal). (a) Performance of our dual-attention Transformer model. (b) Performance of GRU model. (c) Performance of RNN model. (d) Performance of LSTM model.

• 

  Figure 4:. Mean feature importance for the model predicting IDH and IDHTN. (a) Mean feature importance analysed by gradient accumulation method. (b) Mean feature importance analysed by feature perturbation method.

• 

  Figure 5:. Sensitivity analysis of model performance. ROC curves for the multi-class classification task under different data partitioning schemes. Panels (a) and (b) correspond to two settings with data splits of 70:20:10 and 60:20:20, respectively.

• 

  Figure 6:. Mean feature importance of LightGBM Model for Symptom/Intervention Prediction. (a) Top 10 feature importance of symptom prediction by normalized mean gain. (b) Top 10 feature importance of intervention prediction by normalized mean gain.