Incorporating Patient Self-Presentation in Digital Phenotyping of Mental Health: A Patient-in-the-Loop Approach to Passive Sensing

Passive sensing of human behavior provides a new lens to understand patients with mental health disorders and substance use disorders, especially outside their clinical visits. The allure of this approach is the nearly negligible effort and continuous scale at which we can capture patient behavior through commodity devices, like smartphones and wearables. However, we are yet to witness long-term adherence of sensing for behavioral health outside of research settings. One of the key barriers to engaging with sensing is a patient’s lack of agency in how their data will be used for clinical inferences. Ultimately, in contemporary paradigms, patients finds themselves removed from the implications of their data, and thus, anticipate challenges of privacy and trust. When patients experience a fissure in trust, connection, and mutual respect, they are more likely to discontinue their treatment. To overcome this limitation, and bring the best of behavioral sensing to patients, we need to explore patient-facing human-in-the-loop approaches to sensing.

Human-in-the-loop approaches are already emerging In non-digital contexts of behavioral health care. For instance clinicians are starting to collaborate with patients to provide transparency on clinical visit notes and documentation. This approach inspires our pilot study to investigate a new paradigm to digital phenotyping where patients will have an opportunity to reflect on their data and provide feedback to adjust corresponding machine learning models. Our formative analysis will involve a longitudinal study, where patients will interact with their data through a prototype dashboard and periodically collaborate with experts to make sense of their data and the model. Through this process, the study will elicit a variety of patient-generated data abstractions of their behavioral data. These artifacts can inform the design of automated interfaces for patients to interact with digital phenotyping models. Moreover, we can also clarify the opportunities for expert intervention and collaboration.

We focus on patients with risk of depression as a representative case of individuals who can benefit from digital phenotyping but also foster anxieties of being misrepresented by their data. By evaluating our patient-in-the-loop approach, we will be able to define (1) an ontology of data abstracts based on patient needs, and (2) a guideline for future collaborative workflows that maintain patient engagement while minimizing stakeholder burden. Taken together, the evidence from this research will help further a new form of digital phenotyping that is privacy preserving, parsimonious, and personalized.