Multimodal Sentiment and Emotion Analysis Framework for Personalized Health Coaching Messages: Proof-of-Concept Study.

Background: Text generation approaches in health care communication have evolved along 2 major paths. The first path involves generative adversarial networks, progressing from basic architectures to specialized variants like Text-to-Text Generative Adversarial Network (TT-GAN) and Time and Frequency Domain-Based Generative Adversarial Network (TF-GAN), which address challenges in discrete text generation through techniques such as Gumbel-Softmax and reinforcement learning. The second path emerges from transformer-based architectures, particularly Generative Pretrained Transformer-2 (GPT-2), which uses extensive pretraining and self-attention mechanisms to generate contextually appropriate text. GPT-2's transformer architecture enhances persuasive health communication by generating personalized messages using various strategies like task support, dialogue support, and social support for effective health interventions. Objective: This study aimed to use GPT-2 as a generative method to construct persuasive text in a dataset and compare the performance of sentiment analysis and emotion detection analysis. Methods: We combined sentiment analysis tools (VADER [Valence Aware Dictionary and Sentiment Reasoner] and TextBlob) with emotion detection methods (Text2Emotion and NRCLex [National Research Council Lexicon]) to analyze health coaching messages across different persuasive types: reminder, reward, suggestion, and praise. Results: TextBlob and VADER achieved accuracies of 57% and 69%, respectively, while RoBERTa (robustly optimized BERT approach)-sentiment outperformed them with an accuracy of 88%. Emotion detection showed a high prevalence of "joy" and "happy" labels (93.69% positive skew). While transformers excel in accuracy, lexicon-based models like VADER offer a better performance-efficiency balance for real-time health communication systems. For emotion detection, all categories showed perfect accuracy (1.0), while trust showed mixed results, with precision, recall, and F1-score values ranging from 0.81 to 0.96. The emotion detection analysis revealed varying success rates across different emotions, with some categories, such as anger and neutral, showing reasonable performance and others, such as trust, showing mixed performance. Conclusions: This research contributes to understanding the emotional dynamics of persuasive health communication and highlights both the capabilities and limitations of current natural language processing tools in analyzing health-related persuasive messaging. This proof-of-concept study using synthetically generated data establishes a methodological framework for multimodal sentiment and emotion analysis. The findings require validation with real-world health coaching messages before clinical deployment.