Abstract:

Psychological crisis intervention faces challenges due to dynamic, complex, and highly personalized nature of mental states. Traditional methods often suffer from delayed monitoring and a lack of closed-loop optimization. While Large Language Models (LLMs) and multi-agent technologies offer new tools, existing research remains fragmented. This paper introduces Parallel Intelligence into mental health and proposes a conceptual framework: Parallel Psychological Crisis Intervention. Based on the ACP approach (Artificial Systems, Computational Experiments, and Parallel Execution), this framework aims to transform intervention from an experience-driven paradigm to a computable, experimental, and iterative one.

The framework begins with the construction of Artificial Psychological Systems (A), which are individual digital twins built from multi-modal data such as physiological signals, behavioral trajectories, and LLM-parsed subjective reports. These artificial systems serve as a virtual laboratory for Computational Experiments (C), allowing practitioners to conduct high-frequency simulations of counterfactual scenarios. This process enables prediction of crisis events and pre-validation of intervention strategies' safety and efficacy, effectively mitigating ethical risks. Finally, Parallel Execution (P) synchronizes the virtual and real worlds, using real-world feedback to continuously calibrate the artificial system. This creates a closed-loop process where the virtual system guides real-world intervention while the real system refines the virtual model.

To demonstrate feasibility, the paper develops PsyRescueGPT, an LLM-driven multi-agent system. The architecture integrates five specialized agents: a Perceptive Agent for data fusion, a Predictive Agent for quantifying crisis probabilities via graph neural networks, a Strategic Agent for generating evidence-based plans, an Experimental Agent for virtual validation of emotional responses, and a Decision Agent as a “Human-in-the-loop” hub for clinician supervision. In a proof-of-concept scenario involving acute suicide risk, PsyRescueGPT executes a cohesive collaborative pipeline. The workflow initiates with digital twin updates and risk identification, followed by the formulation of response plans which are virtually tested through counterfactual simulations. The validated outcomes are then presented to clinicians for final approval, proving that the framework can transform fragmented tools into a seamless, verifiable crisis response workflow.

The theoretical contribution of this work is a systematic modeling language for psychological complexity, while its practical value lies in providing safe, explainable, and supervised AI-led interventions. Future research will focus on the evolution of parallel agents, cross-cultural assessment validation, and rigorous clinical testing of the PsyRescueGPT prototype in diverse scenarios.

Key words: psychological crisis intervention, parallel intelligence, ACP, Large Language Models, multi-agent systems, personalized intervention