Can GPT-4 provide human-level emotion support? Insights from machine learning-based evaluation framework

The global shortage of mental health services has sparked considerable interest in leveraging generative artificial intelligence (AI) to address psychological health challenges. This study systematically evaluates the emotional support capabilities of GPT-4 and explores ways to enhance its performance through targeted prompt engineering. Initially, natural language processing and explainable machine learning were employed to develop a predictive model for evaluating the empathy of GPT-4's responses. Feature sensitivity analysis identified key linguistic features that significantly influence empathy performance. These insights were then integrated with empathy component theory and helping skills theory to design prompt engineering to enhance GPT-4's emotional support capabilities. Evaluation results show that while unprompted GPT-4 demonstrates substantial empathy in addressing help-seekers' needs, it still lags behind human counselors. However, when guided by targeted prompts, GPT-4's emotional support capabilities improve markedly compared to its zero-prompt version. Notably, in handling emotional issues such as anger, fear, and disgust, prompted GPT-4 performs at a level comparable to human counselors. In summary, this study provides initial evidence of GPT-4's potential in emotional support and introduces an evaluation framework (initial-Evaluation, Enhancement, and re-Evaluation; EEE) that can be used to assess and optimize LLMs' abilities in mental health applications, offering insights into their role in supporting human mental health services.