The central question in this research is whether an AI agent built with cultural and linguistic context can help close the gap between knowing what to do and actually doing it in chronic disease management. EVAA is studying how voice, familiarity, memory, and trust interact with patient behaviour, and what role a relational AI could play alongside human clinicians and community health workers. This brief frames the research questions and the literature our thesis builds on. It is a thinking document, not a results report.
Why information alone is not enough
In medicine, patients are often told exactly what to do, how to take medication, what to eat, and when to exercise, yet many still fail to follow through. The gap between knowing and doing has long been recognised in the public health literature as a behavioural and relational problem rather than a knowledge problem.
EVAA's research asks whether an AI agent, built with deep cultural and linguistic context, can play a meaningful role in closing that gap. The aim is not to provide more information, but to study whether the kind of relationship that makes a patient feel seen, heard, and supported can be partly approximated by software.
What the published evidence around adherence suggests
Across chronic conditions such as diabetes, hypertension, tuberculosis, and HIV, non-adherence is one of the costliest problems in global health. The WHO's 2003 report Adherence to Long-Term Therapies estimates that roughly half of patients with chronic illness in developed countries do not adhere to prescribed treatments, and notes that rates are often worse in low- and middle-income settings.
Community health worker programs such as ASHA workers in India and CHWs across Sub-Saharan Africa have been studied extensively. Reviews including Kallander et al. and Priya et al. consistently report that regular human contact, delivered in a patient's own language by someone who understands their context, is associated with improved adherence. The mechanism in those studies appears to be relationship rather than information delivery alone.
If relationship is the mechanism, the open question is how much of it is mechanisable. Bickmore et al.'s work on relational agents in health argues that continuity, memory, persona, and conversational style are levers software can pull. Whether those levers are sufficient to produce clinically meaningful behaviour change at scale is what EVAA wants to study.
What we want to test
Voice versus text. We want to study whether voice-based interactions produce higher adherence than text-based nudges, particularly for lower-literacy users. The hypothesis is grounded in Donner's work on mobile-internet usage patterns and Nutbeam's framing of health literacy as a public health goal.
Depth of cultural context. We want to compare agents trained on regional diet, local illness beliefs, and familiar language patterns against generic AI health assistants, to see whether and how context changes patient trust and reported compliance.
Relationship continuity. We are interested in whether agents that remember prior conversations, missed doses, or self-reported readings build more trust than stateless interactions. This continuity question is the closest direct analogue to Bickmore et al.'s relational-agents framework.
Frequency and timing. We want to map what cadence of AI check-ins is associated with sustained behaviour change, and at what point check-ins start to feel intrusive, across different chronic conditions and demographics.
What this could mean for chronic disease care
If a culturally tuned, relational AI agent can replicate even part of the adherence-supporting effect that the literature attributes to community health workers, the public-health implications are large. CHW programs are constrained by labour supply and geography. Software is not.
We are equally interested in the limits. Trust, escalation to a human clinician, and clinical safety are not problems that vanish because an agent is well-tuned. The point of this research is to map both what works and where the boundaries are, so that any product built on this thesis is grounded rather than aspirational.
Further reading
- WHO, Adherence to Long-Term Therapies: Evidence for Action (2003).
- Don Nutbeam, Health Literacy as a Public Health Goal (2000).
- Ambady Ramachandran et al., Trends in Prevalence of Diabetes in Asian Countries (2012).
- Karin Kallander et al., Mobile Health Approaches and Lessons for Increased Performance and Retention of Community Health Workers (2013).
- Timothy W. Bickmore et al., Maintaining Continuity in Longitudinal, Computer-Based Health Interventions (2010).
- Ritu Priya et al., Community Health Workers in India: A Review (2019).
- Jonathan Donner, After Access: Inclusion, Development, and a More Mobile Internet (2015).
Share