A Lightweight Explainable Expert System for Resource-Constrained Mobile Devices: An On-Device Random Forest Framework for Multi-Condition Risk Assessment

Deploying intelligent expert systems on resource-constrained mobile devices demands simultaneous satisfaction of competing constraints: low memory foot-print, sub-100 ms inference latency, minimal battery draw, and user-interpretable outputs — all without reliance on cloud connectivity. Existing mobile health applications largely delegate inference to remote servers, introducing privacy risks, connectivity dependencies, and latency penalties that exclude populations with limited internet infrastructure. This paper presents a four-layer on-device expert-system architecture — a data management layer, a compact ensemble inference engine, a deployment-cheap explainability module, and a rule-based risk-context engine — validated on a multi-condition health-screening task. The inference engine uses a 50-tree Random Forest (Breiman, 2001) serialized as a 4.2 MB JSON structure, achieving 94.4% classification accuracy on a 41-class benchmark dataset (131 binary symptom features, 4 920 training samples) with mean inference latency of 28 ms on a mid-range Android device. The explainability module generates matched-symptom lists, missing-symptom analysis, and symptom-coverage ratios in under 1 ms, offering a deployment-feasible alternative to post-hoc methods such as LIME (Ribeiro et al., 2016) and SHAP (Lundberg & Lee, 2017), which impose 200–800 ms overhead incompatible with mobile deployment. A rule-based demographic risk engine encodes establishedclinical scoring logic for five conditions at negligible additional cost. A 30-participant user study found that explainability features increased perceived trust by 83% and comprehension by 87%. The architecture is domain-agnostic: the same four-layer pattern applies to on-device activity recognition, smarthome anomaly detection, and other ambient-intelligence tasks.