In the era of data-driven decision-making, healthcare stands as a critical domain where machine learning (ML) techniques can bring transformative changes. However, the application of ML in healthcare faces unique challenges due to clinicians' limited understanding of intricate ML processes, the diverse and unstructured nature of healthcare data, high computational costs, and the “black box" problem associated with ML algorithms. The recent advent of large language models (LLMs) further introduces the challenge of developing appropriate prompts to guide these models to provide meaningful and contextually relevant responses.

This dissertation grapples with these challenges across a series of studies. First, we analyze multiple ML configurations for the prediction of multiple organ failure in trauma patients, highlighting the impact of classifier choice on performance. Next, we propose a multimodal Transformer model for early sepsis prediction, demonstrating its efficacy over competitive baselines. To address the computational costs, we propose an efficient model for multivariate time series classification. Reinforcement learning is then applied to predict the need for blood transfusion in intensive care units, offering a decision support tool for effective treatment recommendations. Lastly, we conduct a comparative study on the readiness of LLMs for healthcare, introducing a novel prompting strategy to maximize their effectiveness.

The primary objective of this dissertation is to facilitate the advancement, comprehensive evaluation, and systematic optimization of machine learning applications specifically in the healthcare domain. Our work aims to connect complex ML methodologies with practical healthcare applications. As our work progresses, we remain committed to the continuous refinement and enhancement of these models. Our approach aims to balance technical sophistication with ease of use, minimizing the trade-off between the two. We believe that our ML advancements, tailored to the unique needs of healthcare applications, can improve patient outcomes and streamline healthcare delivery.

The skin’s ability to sense its environment is vital to activities of daily living. Cutaneous sensory perception diagnostics allow for the early detection and symptom tracking of tactile dysfunction. However, lack of access to healthcare and the limited frequency of current screening tools can leave skin sensation impairments undiscovered or unmonitored. This work presents a smartphone application for Cutaneous Hand Assessment with a Smartphone Interface (CHASI) to establish Vibrational Perception Thresholds (VPT). CHASI’s vibrational output and force measurement abilities are also characterized. An 18-participant cross-sectional study, with both normative subjects and subjects with sensation impairment, compares the monofilament test with smartphone established VPT (SE- VPT). We find a high positive correlation between SE-VPT and monofilament scores (rs=0.83, p = 0.00014). We also investigate the sensitivity of our proposed SE-VPT method to the motions and forces applied to the touchscreen. We find that variations in force do not alter the practical significance of the monofilament correlation. These results further the smartphone as a potential diagnostic and monitoring tool for hand health.