A Control Theoretic Approach to Evaluate and Inform Ecological Momentary Interventions: A Proof-of-Concept Study (Preprint)

Abstract

Ecological momentary interventions (EMI) are digital mobile health (mHealth) interventions administered in an individual's daily life with the intent to improve mental health outcomes by tailoring intervention components to person, moment, and context. The effects of EMI are often assessed in terms of their impact on simultaneously recorded multivariate psychological Likert scales, also termed ecological momentary assessments (EMA), over time. However, assessing effects of EMI on EMA is challenging. This is because EMA variables may be dependent and dynamic, forming an interconnected network. Through their impact on such networks, EMI effects are difficult to disentangle and are likely to propagate into the future. Here, we propose to use network control theory (NCT) to overcome these challenges and study the effects of EMI on EMA from a dynamical systems (DS) perspective.

NCT is a branch of DS theory that deals with the formal analysis of interventions on networks. Our objectives are to apply concepts from NCT to formally quantify and evaluate proximal intervention effects, analyze putative mechanisms of change, as well as to identify optimal intervention approaches given a set of reasonable (temporal or energetic) constraints in the context of EMI and EMA. Using simulation analyses that simulate intervention effects beyond the presented EMI, we further aim to gain deeper understandings into the inferred networks and their response to external inputs.

We infer linear dynamical systems (DS) models in the form of vector autoregressive models of order 1 from the EMA and EMI data of 10 individuals collected over several weeks. EMI are modeled as inputs to these DS models. We compute controllability measures, impulse responses and optimal intervention schemes according to NCT. Controllability measures and impulse responses are used to evaluate intervention effects. Optimal intervention schemes are analyzed to gain insights into possible EMI designs and important future design considerations.

Using this approach, we were able to identify both real and simulated interventions expected to exert a high impact (high average controllability [AC]) on the inferred EMA networks. Importantly, the identified high impact interventions strongly differed across individuals. However, applying a (simulated) intervention strategy that targets EMA variables with high AC across participants, resulted in a favorable mental health outcome across the sample. This suggests that NCT measures may be useful for tailoring personalized intervention strategies. These results were also consistent with the identification of optimal intervention schemes that were found to mainly target high AC variables during realistic energetic constraints.

Looking at EMI as inputs that exert control over a temporally evolving network of EMA variables may prove valuable for the evaluation and design of personalized EMI delivery schemes.