High Precision Cascaded Spatio Temporal Deep Inference for Real Time Histamine Risk Prediction in Tropical Artisanal Fisheries

Abstract

Rapid histamine accumulation in tropical fisheries constitutes a substantial public health hazard, particularly via scombroid poisoning, and underscores the need for rigorous, data-driven cold-chain surveillance. Artisanal vessels (≤ 30 GT), however, predominantly depend on ice-based cooling strategies that are thermally unstable and lack real-time diagnostic functionality, thereby failing to sufficiently suppress microbial growth kinetics under ambient conditions that frequently exceed 30°C. To address this gap, we propose a Cascaded Spatio-Temporal Deep Inference Architecture that couples a Convolutional Neural Network (CNN) for spatial feature denoising with a Long Short-Term Memory (LSTM) network for temporal kinetic modeling. This hybrid architecture assimilates high-frequency thermal measurements from an optimized R404A vapor-compression refrigeration system and predicts histamine risk indices under Arrhenius-based kinetic constraints. Field deployment on a 10 GT vessel demonstrated that the system maintained a highly stable storage temperature of -20.1 ± 0.5°C. The proposed model exhibited high predictive accuracy with an R² of 0.97 and an RMSE of 0.45°C, significantly outperforming a Linear Regression baseline (RMSE = 1.85°C, p < 0.01). Importantly, the system extended the prime-quality shelf life by more than 52 hours while keeping histamine concentrations well below the U.S. FDA limit of 50 mg/kg. Collectively, these findings support a scalable health informatics framework and indicate that AI-driven predictive certification can substantially reduce food safety risks in resource-limited maritime supply chains.