A closed-loop system for dynamic personality monitoring and judicial risk early warning based on multimodal fusion: --PDSS Model Construction and evidence

Conventional personality assessment methods are constrained by static, self-reporting data collection paradigms, which have fundamental limitations in terms of quantifiability, behavioral prediction validity, and social equity, and are difficult to capture the dynamic evolution process of individual psychological states, which constitutes a significant bottleneck in high-risk and highly dynamic decision-making scenarios such as the judiciary. This study aims to construct and validate a dynamic monitoring system based on multimodal data fusion, with the aim of providing real-time, accurate scientific evidence for judicial decision-making, enabling the advancement and personalization of judicial intervention, and promoting the intelligent transformation of the judiciary. This study systematically integrates multi-source heterogeneous data such as physiological signals, behavioral patterns, social interaction attributes and digital footprints, and employs a hierarchical fusion strategy (early/middle/late fusion) and a deep learning model (LSTM-Attention network), The Personality Dynamic State Space Model (PDSS Model) was constructed to achieve precise and dynamic assessment of individual risk states. The system optimizes predictive performance through dynamic weight allocation, establishes a mapping matrix of risk classification and intervention measures, identifies risks using a dynamic threshold early warning model, and automatically pushes personalized intervention plans. Empirical research shows that compared with traditional static assessment methods, the response time of this multimodal system in capturing emotion fluctuations and predicting aggressive behavior is significantly reduced, and the automated intervention mechanism greatly improves the timeliness of measure initiation. During the six-month judicial practice application, the reoffending rate of high-risk individuals decreased significantly by 45% (p=0.008), the adoption rate of judges' decisions increased to 74% (p<0.001), and the misallocation rate of judicial resources decreased by 66% (p=0.002). By integrating Digital Phenotyping techniques, this study broke through the limitations of traditional assessment in terms of data breadth and real-time performance, not only verifying the application potential of artificial intelligence in psychometrics, but also revealing the necessity and feasibility of the shift from static assessment to dynamic monitoring paradigm. This study provides strong support for building an intelligent, closed-loop judicial management system, but it still faces severe challenges in terms of data generalization, legal evidence fit and ethical governance, which urgently need to be addressed through subsequent research and policy-making.