Practical Learning Modeling Techniques with Personalized Actionable Intervention for In-the-field Prediction

Many chronic diseases, if not well-managed, are major healthcare problems globally with two-third of the total US healthcare spending going toward chronic-related conditions in 2015. To avoid solely relying on expensive treatments, sensing systems and learning modeling techniques have been implemented in healthcare fields to focus on the prevention, early detection, and minimally invasive management of diseases in-the-field. As opposed to the old-fashioned method where a patient only sees a single point of data when he/she meets the clinician, the sensing system can continuously collect data to identify signs of disease anywhere and anytime. However, symptom management, especially informing the patient of symptom prevention and mitigation suggestions, using sensing and modeling systems faces many real-world problems such as weak symptom labeling and providing users actionable information. Thus, this work proposes three practical approaches addressing fundamental data challenges found in real-world sensing and learning modeling systems. The proposed approaches are also implemented in two real-world use cases: dementia caregiver empowerment and improving cancer pain management.

First, the implementation of machine learning models to time-series data collected in the real-world is presented. This work evaluates and implements different learning algorithms, as well as suggests techniques that can address real-world data problems such as imprecise user annotations, small training samples, and imbalanced label distributions. The results suggest that a machine learning algorithm can be applied to predict health events, dementia-related agitations, and cancer-pain episodes from ambient environmental stimuli.

To create appropriate prevention and mitigation suggestions (e.g. symptom intervention) based on the learning model’s prediction, some interpretability of the model is needed. Unfortunately, many real-world applications use complex learning models such as deep learning for their prediction making model interpretation difficult. In this dissertation, a black-box model interpretation technique is presented using the already learned model, predictor importance analysis, and cross-correlation to learn and suggest actionable information to users in real-time. Real-world actionable suggestions can be extracted such as notifying the in-home dementia caregiver to intervene before dementia-related agitation escalates by turning the lights on because the ambient light level is decreasing, which has triggered agitation episodes in the past.

User engagement and domain expert contribution is also important for many in-the-field applications, especially those that involve interventions. Here, a personalized intervention suggestion selection technique is presented that involves user engagement from surveys, self-reports, and sensing systems in combination with domain expert assessments to create scalable and automated prevention and mitigation intervention suggestions for real-world applications. This approach has been used to notify dementia caregivers with personalized agitation interventions. The results suggest that the interventions may help caregivers minimize the stress associated with dementia caregiving tasks.