Model performance (regression).

DAS-Net outperforms the four baselines for both prediction tasks. The naive baseline simply reuses the last available DAS. The MLP and XGBoost baselines use the last available values of each feature as input and our model the whole patient history. The LSTM baseline sequentially processes the patients’ histories.

DAS-Net 在两项预测任务上均优于四款基线模型。其中朴素基线仅复用最后可用的DAS数据；MLP（多层感知机，Multilayer Perceptron）与XGBoost（极限梯度提升树，eXtreme Gradient Boosting）基线以各特征的最后可用值作为输入，我们的模型则对完整的患者病史进行建模；LSTM（长短期记忆网络，Long Short-Term Memory）基线则对患者的病史序列进行逐次处理。