Adoption of machine learning (ML) techniques in healthcare applications is hindered by several challenges, including difficulty in accessing large volumes of retrospective data due to regulatory constraints, or the cost and delay associated with acquiring prospective data through clinical experiments. We observe that prior domain knowledge can enable efficient use of the limited quantity of data, and develop techniques that leverage prior physical and physiological information to enhance ML data efficiency for three illustrative applications arising in healthcare. Specifically, we present a model-based hybrid feature extraction method for improved analysis of blood pressure waveforms; a deep harmonic prior model for separation of mixed signals sensed by optical biosensors; and a deep prior ensemble learning method for boosting signal separation performance for experimental data acquired in large pregnant animal models. The proposed methodologies are validated with relevant clinical or preclinical data, demonstrating significant effectiveness compared to previous work.

The model-based approach combines a simplified cardiovascular model with a rule-based method to augment data, and to extract key features from arterial blood pressure waveforms, significantly enhancing feature detection accuracy in real data. It offers a data-efficient technique for feature extraction in circulatory system studies, offering robustness in face of blood pressure waveform deformations due to pathological conditions, such as aging and vascular stiffness.

We then present a custom deep learning model for separation of quasi-periodic physiological signals in wearable biosensors, integrating harmonic priors within a neural structure to efficiently separate signal sources in time-frequency images using limited data. This approach is validated using both synthetic and in vivo data, showing substantial improvements over existing signal separation techniques. Extending the deep harmonic prior technique, we propose a deep prior ensemble learning method, a boosting approach that optimizes learning bias and variance to enhance the accuracy and stability of signal separation in challenging real-world conditions. The method is developed to address the unique requirement of separating low-SNR biological signals, and its effectiveness is demonstrated using photoplethysmography (PPG) signals that are non-invasively acquired from pregnant ewe models.

The thesis bridges the gap between theoretical advancements in machine learning and practical requirements of a set of healthcare applications, demonstrating that generalizable machine learning performance can be effectively achieved in data-restricted environments, thus showcasing the potential for adoption in a broader set of healthcare applications.