Improving Alzheimer's Disease Prediction with Personalized Autoencoders
If we are to understand the role of machine learning (ML) in critical health care applications, then we need to understand when and how ML-based approaches can be applied to solve healthcare-related problems. Specifically, we need to develop and test rationale justifying the use ML approaches, rather than applying them blindly as black boxes. To this end, I will pursue this goal of creating interpretable ML solutions by specifically focusing on ML-based approaches in the domain of disease prediction for healthcare and precision medicine. As a concrete example, I will focus on the problem of missing data and its impact on the personalized prediction of Alzheimer's disease progression given multimodal data.
If we are to solve this problem of missing data, then we need to formulate an approach tailored to the specific scenario. In order for this approach to be useful beyond the confines of my own experiments, we need to understand how the approach impacted our own results and provide an explanation guiding how future experiments and other researchers might apply this approach to their own studies.
Project type
Research proposal.
Steps
I will introduce the core problem motivating my MEng thesis that is, the task of personalized prediction of the progression of Alzheimer's disease, and will provide an overview of the research I have conducted so far.
I will then narrow my focus and describe the approach and results of a recent study I conducted and presented at a conference. I will explain the successful aspects as well as the limitations (e.g. imbalanced, missing, heterogenous data) of the ML approach used in the study.
Next, I will provide a story explaining why we likely encountered limitations when applying the method in the study, and will hypothesize a potential alternative solution using autoencoders.
Then, I will design and propose the specifications required for implementation of the alternative, autoencoder-based solution.
I will explain why I would use an autoencoder-based method to overcome limitations found in the previous study, (i.e. justify its use to help deal with missing data) and speculate the anticipated effects of its application.
Finally, I would outline the next steps required to implement the proposed mechanism to improve the results of the previous study.
Contributions
Articulated overarching problem in ML (a need for interpretable ML algorithms) and described key research problem (personalized prediction of Alzheimer's disease progression using multi-modal data (e.g. cognitive, demographic, neuroimaging, genetic)) motivating my research.
Described my ML-based approach to solving the aforementioned problem, and explained the key results obtained.
Developed story interpreting results.
Identified advantages and limitations of initial approach.
Formulated and proposed a personalized approach to solving the limitations encountered (i.e. the problem of missing data, per-person differences).
Hypothesized why the proposed solution will solve the previously encountered problems.
Demonstrated how solution could be applied to research problem.
Established plan for implementation as a next step.
Mesert Kebed
Empathetic action recognition and Learning
The vision for this proposal is to further McIntyre's research by testing classroom applications for this theory to determine if students are better able to better retain information if it is handwritten in front of them or if it is presented in a slide format.
Project type
Pilot experiment.
Steps
Subjects will be presented with characters from the Amharic alphabet as well as the corresponding pronunciation for each character. All subjects will be informed that there will be a short quiz at the end of the session that will test their recall of the material presented.
Subjects will be split up into four groups and each presented with the same material through different media. I will write and read out the characters to the first group. For the second group, I will be reading off a slideshow that has the characters on it. Similarly the third group will also be using a slideshow but I will present each character at a time. For the last group, I will hand write the characters before the session, and present it to the subjects.
After the end of the session, the subjects will be given a short 5 minute test that will require simple recall of the characters they had seen earlier.
A similar test will be administered 4 days after the session to test if there are any differences in recall capabilities among the 3 groups.
Contributions
Demonstrated that hand-writing material in front of subjects helped (or had no effect) on recall capabilities in the short and long term.
Yasaman Tahouni
Identifying Design Knowledge in making
If we want to create Computer-aided design tools and interfaces that enhance creative design process, we need to understand how physical interaction with material affects design knowledge.
Project type
Pilot experiment
Steps
Subjects will perform a simple design task, consisting of assembly and manipulation of free-form curves and surfaces, using two different interfaces. First group will use CAD software with conventional mouse/keyboard/monitor interface. Second group will use NURBSForms interface, which is a tangible interface that enables free-form curves and surfaces manipulation in physical form.
After finishing the task, users are asked to write an answer to this question: "Tell me your design story":
You are to tell another designer exactly what you did, so that they can replicate your design process from the beginning to the end. Give them a step by step recipe of your thoughts and your actions, including your mistake, misunderstandings and bugs. Try to be as detailed as possible.
I will compare the design stories from the two groups, and I will evaluate design knowledge, expressed through written self-reflection, based on cogency and coherence of their stories. Some of the measurements are: Amount of detail provided, number of -if any- missing steps, and mismatch between what they say they did and what they actually did.
Contributions
1-Identifying X-Design Knowledge that is rooted solely in the multi-sensory interaction with physical material.
2-Making a case for enabling physical interaction with Computer-aided design tools. In case of proving the hypothesis, this can be considered as a guideline for creating CAD tools that enhance creative design process.
Demonstrating a case for applying Winston's story-telling theory in evaluating design knowledge.