Improving Maternal and Fetal Cardiac Monitoring Using Artificial Intelligence
dc.contributor.committeeMember | Gokaraju , Rama | |
dc.contributor.committeeMember | Saadat Mehr, Aryan | |
dc.contributor.committeeMember | Bui , Francis | |
dc.contributor.committeeMember | Wu , FangXiang | |
dc.creator | Mohebbian, Mohammad Reza | |
dc.creator.orcid | 0000-0003-2645-4425 | |
dc.date.accessioned | 2022-01-06T18:54:06Z | |
dc.date.available | 2022-01-06T18:54:06Z | |
dc.date.created | 2021-11 | |
dc.date.issued | 2022-01-06 | |
dc.date.submitted | November 2021 | |
dc.date.updated | 2022-01-06T18:54:06Z | |
dc.description.abstract | Early diagnosis of possible risks in the physiological status of fetus and mother during pregnancy and delivery is critical and can reduce mortality and morbidity. For example, early detection of life-threatening congenital heart disease may increase survival rate and reduce morbidity while allowing parents to make informed decisions. To study cardiac function, a variety of signals are required to be collected. In practice, several heart monitoring methods, such as electrocardiogram (ECG) and photoplethysmography (PPG), are commonly performed. Although there are several methods for monitoring fetal and maternal health, research is currently underway to enhance the mobility, accuracy, automation, and noise resistance of these methods to be used extensively, even at home. Artificial Intelligence (AI) can help to design a precise and convenient monitoring system. To achieve the goals, the following objectives are defined in this research: The first step for a signal acquisition system is to obtain high-quality signals. As the first objective, a signal processing scheme is explored to improve the signal-to-noise ratio (SNR) of signals and extract the desired signal from a noisy one with negative SNR (i.e., power of noise is greater than signal). It is worth mentioning that ECG and PPG signals are sensitive to noise from a variety of sources, increasing the risk of misunderstanding and interfering with the diagnostic process. The noises typically arise from power line interference, white noise, electrode contact noise, muscle contraction, baseline wandering, instrument noise, motion artifacts, electrosurgical noise. Even a slight variation in the obtained ECG waveform can impair the understanding of the patient's heart condition and affect the treatment procedure. Recent solutions, such as adaptive and blind source separation (BSS) algorithms, still have drawbacks, such as the need for noise or desired signal model, tuning and calibration, and inefficiency when dealing with excessively noisy signals. Therefore, the final goal of this step is to develop a robust algorithm that can estimate noise, even when SNR is negative, using the BSS method and remove it based on an adaptive filter. The second objective is defined for monitoring maternal and fetal ECG. Previous methods that were non-invasive used maternal abdominal ECG (MECG) for extracting fetal ECG (FECG). These methods need to be calibrated to generalize well. In other words, for each new subject, a calibration with a trustable device is required, which makes it difficult and time-consuming. The calibration is also susceptible to errors. We explore deep learning (DL) models for domain mapping, such as Cycle-Consistent Adversarial Networks, to map MECG to fetal ECG (FECG) and vice versa. The advantages of the proposed DL method over state-of-the-art approaches, such as adaptive filters or blind source separation, are that the proposed method is generalized well on unseen subjects. Moreover, it does not need calibration and is not sensitive to the heart rate variability of mother and fetal; it can also handle low signal-to-noise ratio (SNR) conditions. Thirdly, AI-based system that can measure continuous systolic blood pressure (SBP) and diastolic blood pressure (DBP) with minimum electrode requirements is explored. The most common method of measuring blood pressure is using cuff-based equipment, which cannot monitor blood pressure continuously, requires calibration, and is difficult to use. Other solutions use a synchronized ECG and PPG combination, which is still inconvenient and challenging to synchronize. The proposed method overcomes those issues and only uses PPG signal, comparing to other solutions. Using only PPG for blood pressure is more convenient since it is only one electrode on the finger where its acquisition is more resilient against error due to movement. The fourth objective is to detect anomalies on FECG data. The requirement of thousands of manually annotated samples is a concern for state-of-the-art detection systems, especially for fetal ECG (FECG), where there are few publicly available FECG datasets annotated for each FECG beat. Therefore, we will utilize active learning and transfer-learning concept to train a FECG anomaly detection system with the least training samples and high accuracy. In this part, a model is trained for detecting ECG anomalies in adults. Later this model is trained to detect anomalies on FECG. We only select more influential samples from the training set for training, which leads to training with the least effort. Because of physician shortages and rural geography, pregnant women's ability to get prenatal care might be improved through remote monitoring, especially when access to prenatal care is limited. Increased compliance with prenatal treatment and linked care amongst various providers are two possible benefits of remote monitoring. If recorded signals are transmitted correctly, maternal and fetal remote monitoring can be effective. Therefore, the last objective is to design a compression algorithm that can compress signals (like ECG) with a higher ratio than state-of-the-art and perform decompression fast without distortion. The proposed compression is fast thanks to the time domain B-Spline approach, and compressed data can be used for visualization and monitoring without decompression owing to the B-spline properties. Moreover, the stochastic optimization is designed to retain the signal quality and does not distort signal for diagnosis purposes while having a high compression ratio. In summary, components for creating an end-to-end system for day-to-day maternal and fetal cardiac monitoring can be envisioned as a mix of all tasks listed above. PPG and ECG recorded from the mother can be denoised using deconvolution strategy. Then, compression can be employed for transmitting signal. The trained CycleGAN model can be used for extracting FECG from MECG. Then, trained model using active transfer learning can detect anomaly on both MECG and FECG. Simultaneously, maternal BP is retrieved from the PPG signal. This information can be used for monitoring the cardiac status of mother and fetus, and also can be used for filling reports such as partogram. | |
dc.format.mimetype | application/pdf | |
dc.identifier.uri | https://hdl.handle.net/10388/13758 | |
dc.subject | Cardiac, Deep Learning, CycleGAN, Monitoring, Fetal ECG. | |
dc.title | Improving Maternal and Fetal Cardiac Monitoring Using Artificial Intelligence | |
dc.type | Thesis | |
dc.type.material | text | |
thesis.degree.department | Electrical and Computer Engineering | |
thesis.degree.discipline | Electrical Engineering | |
thesis.degree.grantor | University of Saskatchewan | |
thesis.degree.level | Doctoral | |
thesis.degree.name | Doctor of Philosophy (Ph.D.) |