Physiology-Guided Machine Learning for Improved Reliability of Non-Invasive Assessment of Pulmonary Hypertension

Sammanfattning: Diagnosing pulmonary hypertension (PH) with right heart catheterization (RHC) is associated with a risk for complications and high expenses, leading to late diagnoses [1]. Transthoracic echocardiography can be used to assess non-invasive indicators for PH such as right ventricular systolic pressure (RVSP), which can be estimated by combining the peak tricuspid regurgitation velocity (TRV) with the estimated right arterial pressure (RAP). However, clinical reports have demonstrated a lower correlation between measurement by RHC and non-invasive RVSP [2], making it desirable to train a neural network to reproduce the work of level 3 clinical experts. Assessments by level 3 readers have shown a bias of predicted RVSP in relation to SPAP measured by RHC of 3.3 mmHg and a limit of agreement (LOA) of [-12,19.3] mmHg. As neural networks aim to learn complex relationships between in- and output, this master thesis aims to investigate if the non-invasive assessment of TRV max can be improved by the physiological information used by the expert level 3 readers. Waveform segmentation by an Unet followed by extraction of the waveform was compared to the physiologically informed pre-trained model (TR-CNN) which estimates physiologically important points. It has been shown that cubic spline interpolation between physiologically important points on the TR waveform may improve RVSP estimations [3]. These points reflect physiological events such as the opening and closing of valves and were identified based on the first and second derivatives [3]. The predicted waveforms from the different networks were used for assessing the Vmax and evaluated against clinical ground truth values. Results show that the physiologically informed TR-CNN model outperformed the Unet. The TR-CNN showed a bias of 0.05 m/s and an LOA of [-0.62, 0.72] m/s when comparing predictions with invasive estimates of TRV max. The TR-CNN predictions of the RVSP compared to invasive pressure measurements had a bias of -0.86 mmHg and a limit of agreement of [-22.32, 20.60] mmHg. Hence, reducing the LOA would be advantageous for clinical approval. Furthermore, expanding the dataset could improve the model’s performance, increase robustness, and enhance statistical relevance. The study suggests that the use of piecewise cubic spline representation could ensure the clinical validity of predicted waveforms. This compared to the Unet which struggled to estimate the waveform in the center, where the signal quality was weaker. One way to improve this could be to modify the Unet by making sure that it only predicts convex curve shapes, or to use a loss function that considers the predicted waveform for the TR-CNN. In conclusion, utilizing physiological information has the potential to enhance non-invasive Vmax assessment, but further research is needed.