Music based isolated systolic and diastolic hypertension detection with ECG spectrograms using Explainable AI technique Grad-CAM

Abstract Background Hypertension (HT, blood pressure BP140/90) poses cardiovascular risk, and if left untreated can lead to cardiovascular diseases, stroke, and chronic kidney disorders. However, isolated systolic hypertension (SBP:130-140, DBP90) (ISH) and isolated diastolic hypertension (SBP:120-140, DBP:80-90) (IDH) diagnosis are often underestimated. Studies have shown that ECG signals encapsulate information about, and can be used to categorise, BP states 1,2. The presence of music has been found to influence listeners’ cardiovascular function, and improve HT diagnosis 3. Whether music engagement also improves ISH/IDH classification has not been studied. And, the discriminating parts of the music-induced ECG changes that benefit BP sub-stratification is not known. Purpose This study aims to evaluate if the three-way categorisation of ISH, IDH and HT from ECG signals improves with music engagement, and to use XAI (explainable artificial intelligence) to account for the changes brought about by music listening. Methods ECG signals and continuous BP measurements were acquired from a cohort of 66 individuals (38 females, age=43.82±21.56 yrs, weight=63.54±8.97 kg, height=154±10.23 cm) during a 5-minute silence baseline and during ~40 minutes of listening to digitally altered (for tempo/loudness range) western classical music rendered on a reproducing piano 4. The ECG signals were converted to spectrograms after pre-processing. These images were classified using the XAI technique Grad-CAM (Gradient-weighted class activation mapping) 5 based pre-trained SqueezeNet (a Deep Neural Network for image classification) 6. Results The ECG spectrogram dataset was divided into training:validation:test sets in the proportion 75:15:10. The XAI technique Grad-CAM enabled the network, the pre-trained SqueezeNet, to distinguish between the three groups with greater interpretability. It generates heat maps that act as weights to focus the network on specific regions in the image. The classification accuracy was 74.3% (95% CI: 70.1-77.8) with precision 72.9% (68.9-78.6) and recall 75.7% (70.2-80.9) during baseline silence. In contrast, ECG spectrograms obtained during music listening classified hypertensives with accuracy 80.5% (77.2-85.4), precision 81.4% (78.9-84.6) and recall 82.6% (80.2-86.9), a 5-10% improvement on all measures. Ablation studies with the raw spectrograms and XAI Grad-CAM modified spectrograms showed a 20% increase in efficiency. Conclusions Detection of isolated systolic and diastolic hypertension, and hypertension improves with music engagement, and is further improved by the explainable AI Grad-CAM technique to concentrate the network on the most discriminating parts of the ECG spectrogram. The improved categorisation is a step towards music-based BP substratification for diagnosis of isolated systolic and diastolic hypertension for early detection of cardiovascular risk during everyday music listening.Schematic Diagram Graphs