Multimodal AI-approach for the automatic screening of cardiovascular diseases based on nocturnal physiological signals

Figures

• 

  Fig. 1. Confusion matrix of the proposed multimodal SleepCVD-Net for the CVD screening. A. Training set, B. Validation set, C. Test set.

• 

  Fig. 2. The receiver operating characteristic curve of the proposed multimodal SleepCVD-Net for the automatic screening of CVD based on nocturnal physiological recordings.

• 

  Fig. 3. Study design of the proposed multimodal SleepCVD-Net model for the automatic screening of CVDs based on PSG recordings. A SHHS database that we analyzed for this study. B Extraction of PSG recordings and building of the physiological signal dataset. C Graphical overview of the proposed multimodal SleepCVD-Net model for automatic screening of CVDs.

• 

  Fig. 4. Graphical overview of the proposed SleepCVD-Net model for the automatic screening of CVDs based on nocturnal physiological recordings. A Multimodal input consisting of nocturnal physiological signals, including ECG, airflow, and SpO2. B Data preprocessing was performed for raw signal normalization using two-layer batch normalization for each modality. C Feature extraction was conducted to generate a comprehensive feature map by combining three sub-feature maps. D The classification network for final discrimination was constructed using a multilayer fully-connected neural network. E The proposed multimodal SleepCVD-Net model produced predictive outcomes based on nocturnal physiological recordings. Mode_ECG is a sub-mode for the single-lead ECG signal processing and feature map generation. Mode_Airflow is a sub-mode for the Airflow signal processing and feature map extraction. Mode_SpO2 is mode of the oxygen saturation signal processing and feature maps.