Physiology-inspired deep learning for improved heart failure management

Autor:	Schlesinger, Daphne E.
Rok vydání:	2024
Druh dokumentu:	Diplomová práce
Popis:	Heart failure is an increasingly prevalent condition, which is associated with significant morbidity and mortality. While there has been profound progress in the development of pharmacotherapy and specialized devices for heart failure in recent decades, challenges remain in disease diagnosis and management. One of the key issues is that central hemodynamics and cardiac mechanics, the quantities that characterize the state of a heart failure patient, are difficult to measure. Deep learning methods have shown promise for addressing problems in clinical medicine but are fundamentally limited by their opacity to interpretation, which inhibits model trust and adoption. In this thesis, we propose physiology-inspired deep learning approaches to improve heart failure management. Central hemodynamic parameters are typically measured via pulmonary artery catheterization — an invasive procedure that involves some risk to the patient and is not routinely available in all settings. In Chapter 2, we sought to develop a noninvasive method to identify elevated mean pulmonary capillary wedge pressure (mPCWP). We leveraged data from 248,955 clinical records at the Massachusetts General Hospital (MGH) to develop a deep learning model that can infer when the mPCWP > 15 mmHg using the 12-lead electrocardiogram (ECG). Of these data, 242,216 records were used to pre-train a model that generates useful ECG representations. The remaining 6739 records contain encounters with direct measurements of the mPCWP from right heart catheterizations (RHCs), which provide gold-standard hemodynamic measurements. Eighty percent of these data were used for model development and testing (4304 in train, 546 validation, and 540 in the test set), and the remaining records comprise a holdout set (1349) that was used to evaluate the model. We developed an associated unreliability score that identifies when model predictions are likely to be untrustworthy. The model achieves area under the receiver operating characteristic curve (AUROC) scores of 0.80 ± 0.02 (test set) and 0.79 ± 0.01 (holdout set). Model performance varies as a function of the unreliability, where patients with high unreliability scores correspond to a subgroup where model performance is poor: for example, patients in the holdout set with unreliability scores in the highest decile have a reduced AUC of 0.70 ± 0.06. These results demonstrate that the mPCWP can be inferred from the ECG, and the reliability of this inference can be measured. When invasive monitoring cannot be expeditiously performed, deep learning models may provide information that can inform clinical care. We extended this work in Chapter 3, and developed a Cardiac Hemodynamic Artificial Intelligence monitoring System (CHAIS) that uses single-lead ECG data to infer when cardiac hemodynamics are abnormal. CHAIS is a deep neural network that was trained to detect abnormal cardiac hemodynamics using just lead I of the 5930 paired ECG recordings and RHCs from MGH used in Chapter 2. CHAIS was tested on the internal holdout set of 1439 paired single-lead ECGs and RHCs (858 patients) from MGH and on an external validation set of 4629 paired ECGs and RHCs (2577 patients) from another institution. We also prospectively collected single-lead ECG data using a commercially available wearable ECG monitor, from 83 patients who were scheduled for a RHC at MGH, and used CHAIS to infer if their left atrial pressures would be elevated at the time of their RHC. CHAIS achieves an AUROC of 0.80 for detecting elevated left atrial pressures on the internal test dataset and 0.76 on the external validation set. On patients who wore a wearable ECG monitor before RHC, CHAIS had an AUROC of 0.70; however, when ECG data are available within 1.25 hours before catheterization, the AUROC is 0.875. These results demonstrate the utility of ambulatory cardiac hemodynamic monitoring with a wearable ECG monitor. Finally, in Chapter 4, we described an approach to directly incorporating knowledge of cardiovascular physiology into a deep learning framework. The framework consists of a neural network encoder, to map from arterial blood pressure (ABP) waveforms to latent cardiovascular parameters, and a mechanistic model which maps from those parameters to hemodynamic waveforms, called Cardiovascular Simulator (CVSim). We trained the model on a synthetic data set, and found that the model achieved a multiclass accuracy of 47.8 percent for placing samples into classes in terms of their left ventricular end diastolic pressure (LVEDP), cardiac output (CO), and left ventricular ejection fraction (LVEF), using clinically relevant thresholds. Here, we also proposed a physiology-inspired trust score, and find that the multiclass accuracy is higher in the subset of samples in the lowest decile with respect to the score, compared to results on the rest of the data. On applying the model to a small clinical data set from patients in intensive care settings at MGH, classification performance in terms of LVEDP, CO, and LVEF was limited, given the various challenges of transfer from synthetic to real clinical data. However, we observed that cardiac mechanical parameters inferred from the clinical data set trended positively with the administration of pharmaceutical agents expected to modulate those parameters. This suggests that the model can glean meaningful information from clinical ABP waveforms and shows promise for future development. With further clinical testing, the suite of methods described in this thesis have the potential to advance heart failure care by enabling non-invasive central hemodynamic monitoring and minimally-invasive inference of cardiac mechanics. Ph.D.
Databáze:	Networked Digital Library of Theses & Dissertations
Externí odkaz:	https://hdl.handle.net/1721.1/155650 Zobrazit plný text záznamu