Date on Master's Thesis/Doctoral Dissertation


Document Type

Doctoral Dissertation

Degree Name

Ph. D.


Computer Engineering and Computer Science

Degree Program

Computer Science and Engineering, PhD

Committee Chair

Frigui, Hichem

Committee Co-Chair (if applicable)

Calhoun, Aaron W.

Committee Member

Hardin, Tim

Committee Member

Lauf, Adrian

Committee Member

Nasraoui, Olfa

Committee Member

Yampolskiy, Roman Vladimirovich


Medical care--Simulation methods; Medical care--Data processing; Computational linguistics; Speech processing systems


The Simulation for Pediatric Assessment, Resuscitation, and Communication (SPARC) group within the Department of Pediatrics at the University of Louisville, was established to enhance the care of children by using simulation based educational methodologies to improve patient safety and strengthen clinician-patient interactions. After each simulation session, the physician must manually review and annotate the recordings and then debrief the trainees. The physician responsible for the simulation has recorded 100s of videos, and is seeking solutions that can automate the process. This dissertation introduces our developed system for efficient segmentation and semantic indexing of videos of medical simulations using machine learning methods. It provides the physician with automated tools to review important sections of the simulation by identifying who spoke, when and what was his/her emotion. Only audio information is extracted and analyzed because the quality of the image recording is low and the visual environment is static for most parts. Our proposed system includes four main components: preprocessing, speaker segmentation, speaker identification, and emotion recognition. The preprocessing consists of first extracting the audio component from the video recording. Then, extracting various low-level audio features to detect and remove silence segments. We investigate and compare two different approaches for this task. The first one is threshold-based and the second one is classification-based. The second main component of the proposed system consists of detecting speaker changing points for the purpose of segmenting the audio stream. We propose two fusion methods for this task. The speaker identification and emotion recognition components of our system are designed to provide users the capability to browse the video and retrieve shots that identify ”who spoke, when, and the speaker’s emotion” for further analysis. For this component, we propose two feature representation methods that map audio segments of arbitary length to a feature vector with fixed dimensions. The first one is based on soft bag-of-word (BoW) feature representations. In particular, we define three types of BoW that are based on crisp, fuzzy, and possibilistic voting. The second feature representation is a generalization of the BoW and is based on Fisher Vector (FV). FV uses the Fisher Kernel principle and combines the benefits of generative and discriminative approaches. The proposed feature representations are used within two learning frameworks. The first one is supervised learning and assumes that a large collection of labeled training data is available. Within this framework, we use standard classifiers including K-nearest neighbor (K-NN), support vector machine (SVM), and Naive Bayes. The second framework is based on semi-supervised learning where only a limited amount of labeled training samples are available. We use an approach that is based on label propagation. Our proposed algorithms were evaluated using 15 medical simulation sessions. The results were analyzed and compared to those obtained using state-of-the-art algorithms. We show that our proposed speech segmentation fusion algorithms and feature mappings outperform existing methods. We also integrated all proposed algorithms and developed a GUI prototype system for subjective evaluation. This prototype processes medical simulation video and provides the user with a visual summary of the different speech segments. It also allows the user to browse videos and retrieve scenes that provide answers to semantic queries such as: who spoke and when; who interrupted who? and what was the emotion of the speaker? The GUI prototype can also provide summary statistics of each simulation video. Examples include: for how long did each person spoke? What is the longest uninterrupted speech segment? Is there an unusual large number of pauses within the speech segment of a given speaker?