SETU: Optimal Event-Triggered Edge LLMs for Mental Health via Sequential Hypothesis Testing in Low-Resource Settings

Access to mental health support at Indian en- gineering colleges is constrained by structural scarcity: roughly one counselor per 3,000 stu- dents, unreliable network connectivity, and stu- dent reluctance to share sensitive data. Contin- uous passive sensing incurs prohibitive battery overhead and privacy cost; cloud-hosted conver- sational agents fail offline. We propose SETU, an offline, privacy-preserving mental health sup- port architecture for commodity Android devices whose core contribution is the application of the Sequential Probability Ratio Test (SPRT) as a battery-aware gate for an on-device INT4- quantized language model. Unlike Bayesian on- line change-point detection, SPRT provides for- mal Type I and Type II error guarantees via Wald’s optimality theorem, directly linking the false- alarm parameter α to weekly LLM invocation frequency and therefore to battery overhead. A personalized null hypothesis, estimated from each user’s 14-day behavioral burn-in, adapts detec- tion to individual baselines without population- average assumptions. Simulation over 200 syn- thetic behavioral streams demonstrates that SPRT- Personalized achieves the lowest ERDE-5 (0.047) at low noise and a 5×reduction in false alarms versus BOCPD-Fixed at high noise (σ=1.4), while retaining 40% higher sensitivity than the CUSUM baseline. Retrieval-based personaliza- tion and differential privacy aggregation complete a system designed entirely to the hardware and infrastructure constraints of the Global South.