PAD Diagnosis and Estimation of Treatment Effectiveness Using Machine Learning

Ali Al-Ramini, University of Nebraska-Lincoln

Copyright 2023, Ali Al-Ramini

Abstract

The fusion of engineering, medicine, and data science offers transformative solutions to healthcare challenges. This dissertation epitomizes this synergy by focusing on developing novel diagnosis methods for Peripheral Artery Disease (PAD), a condition often underdiagnosed and demanding specialized training for accurate diagnosis and treatment. We first employed machine learning (ML) to classify individuals with PAD using laboratory-based gait features in this work. An in-depth analysis distinguished essential gait features statistically. The ML approach extracted valuable features for classifying PAD, with ground reaction forces (GRF) emerging as a pivotal metric. Models that utilized GRF scored up to 87% accuracy (0.64 Matthew’s Correlation Coefficient). Our data-driven approach demonstrated that ML algorithms could achieve strong performance values, with GRF measurements providing superior information for PAD classification. Next, we delved into the significance of GRF as a diagnostic tool for PAD severity. Valuable, traditional assessments, such as the walking impairment questionnaire (WIQ), short form 36 health survey questionnaire (SF-36), and Absolute Claudication Distance (ACD), had limitations in distinguishing PAD severity levels. In contrast, GRF showcased its ability to classify patients with PAD (low and high severity) and healthy controls based on severity. Moreover, we reduced the number of features to only one GRF feature (Propulsive Peak), which generated an accuracy of 91% with Matthew’s Correlation Coefficient of 0.87. This further proved GRF's (Propulsive Peak) ability to quantify PAD severity, establishing it as a severity measure for PAD. Furthermore, we introduced a 2D health map and the 1D GRF Propulsive Peak scale, presented transformative tools for PAD diagnosis. These tools could enhance PAD severity assessment precision and hint at innovations like wearable technologies to transition PAD management from clinical to home settings potentially. Finally, we explored the potential of ML in predicting PAD treatment outcomes. Our ML approach used pre-intervention data as input to predict post-intervention severity values. The model's capability to accurately predict post-intervention GRF Propulsive Peak emerged as a significant breakthrough, which yielded an R-squared of 0.945 on the test set. This predictive ability promises to revolutionize PAD treatment strategies, allowing clinicians to make more informed decisions and optimize post-treatment care. All in all, this dissertation underscores the potential of ML in PAD diagnosis, severity assessment, and treatment outcome prediction. The findings advocate a paradigm shift towards a data-driven, patient-centric approach to PAD management, integrating ML insights with emerging technologies for improved patient outcomes and tailored treatment strategies.