Visual Analytics for Smartphone Health Phenotyping
WPI – Computer Science
Thursday , May 13, 2021
Time: 12:00 pm – 1:00 pm
Zoom link: https://wpi.zoom.us/j/92335112887
Prof. Emmanuel Agu (primary advisor)
Prof. Elke Rundensteiner (co-advisor)
Prof. Lane Harrison (committee member WPI CS)
Prof. Angela Rodriguez (external committee member, WPI Psychological Science)
The US healthcare system is schedule-driven, with many patients getting infrequent assessments. Patients often receive little care between hospital visits, which often results in late diagnoses and poor outcomes. Consequently, methods to continuously monitor patients' health and detect ailments early, are desirable. Smartphones are sensor-rich, ubiquitously owned (by over 85% of adults in the USA) and users' interactions with them can provide clues on their health. The emerging smartphone health sensing or smartphone phenotyping paradigm uses apps to passively gather sensor data as the patients live their lives in-the-wild, which are then analyzed using machine learning to infer their health. Prior work has analyzed data using machine learning methods. However, several issues arise at various stages of smartphone phenotyping. During data gathering for machine learning model development, user-provided health and context labels may be wrong or missing, leading to weak supervision of machine learning models.
In this dissertation, we present multiple Interactive Visual Analytics (IVA) frameworks to assist analysts and clinicians in smartphone phenotyping using sensed data including addressing issues such as 1) Correction of wrong or missing health and context labels, 2) Visual steering during development of machine learning health inference models, and 3) Visual support for anomalous health detection and scalability issues for analyzing large populations. Prior IVA work analyzed structured data such as electronic health reports and medical images with high quality labels but not noisy, real world, smartphone-sensed data. Smartphones provide additional challenges because they are not dedicated health assessment devices. Users' lives are intricately interwoven with and confound health predictive behaviors. The specific IVA frameworks we have previously proposed include:
- COMEX and DEFLI: to correct wrong or missing labels using multiple interactive panes and visual linking of data with similar feature values.
- PLEADES: for exploratory smartphone data analysis during machine learning model development that uses visual clustering and projection.
- ARGUS and INTOSIS: for augmenting health inferences from smartphone sensed data, utilizing visual metaphors, and displaying semantic information and smartphone-sensed contextual factors such as geo-locations and movement patterns to enhance clinician's health inferences.
We are proposing additional integrated IVA frameworks to assist the smartphone ailment phenotyping pipeline using novel concepts such as visual feature-similarity linking to discover potentially predictive, symptomatic user activities of daily living and behaviors, as well as multiple linked population-level views of health data for large scale smartphone phenotyping.