Mobile Expert System for Aggression Detection and Prediction: Pilot Evaluation of a Fuzzy–LSTM Model

This study presents a mobile expert system for on-device detection and short-horizon forecasting of aggression using affordable edge hardware. The proposed framework combines lightweight on-body and ambient signals, compact sequential predictors, and an interpretable fuzzy decision layer that converts calibrated probabilities into actionable and auditable alerts. In a subject-held-out pilot study with 10 independent participants, the system achieved a macro-averaged F1 score of 98.3% and an area under the receiver operating characteristic curve of 0.998 on the held-out test split. These results should be interpreted as pilot-scale held-out estimates rather than as definitive evidence of broad superiority across settings, because only 10 independent participants were available for subject-level evaluation and residual optimism or overfitting at the between-subject level cannot yet be excluded. Since the dataset belongs to a completed feasibility-oriented pilot phase, no additional participant-level test cases could be incorporated within the scope of the present study. An exploratory external check on a small independent cohort of 15 cases yielded performance of similar magnitude; however, these findings are presented strictly as preliminary and should not be interpreted as robust evidence of generalization across settings or populations. The compact Long Short-Term Memory forecasters also often reached their best validation region after relatively few effective epochs; in this pilot, that behavior is interpreted as a fixed-cohort optimization characteristic rather than as evidence that the available training data are already sufficient for deployment-oriented generalization. Ablation analyses indicate that short-horizon sequential predictors and weapon-related cues contribute most strongly to predictive accuracy, whereas camera-derived person and weapon cues should be understood as local field-of-view evidence rather than complete scene observability. Beyond pointwise latency, the prototype also demonstrated pilot-stage sustained-load feasibility on Raspberry Pi 3B+ hardware during a continuous 6 h profile, with mean central processing unit utilization of 68.4% ( $\pm$4.2%), mean throughput of 0.798 records/s, and a battery-based mean power proxy of 5.18 W. The design prioritizes calibrated probability estimates, robustness to missing data, and transparent alert generation for non-specialist operators. Aggressiveness labels were defined through an a priori, expert-informed operational codebook intended to stratify short-horizon security risk into Low, Medium, and High levels rather than to provide a clinical diagnosis. Data collection was conducted under written informed consent, ethics approval, and de-identified data-handling procedures. Limitations include the pilot scale, single-site acquisition, and controlled distribution shifts; broader assessment of generalization and fairness will require larger, multi-session, and multi-site cohorts.