Main Parameters of Value (MPVs) are core indicators in the Trends of Engineering System Evolution (TESE) framework, representing the performance attributes that drive technological adoption. Yet, MPVs are often abstract and inconsistently articulated in patent texts, making their automated identification a complex task. This study benchmarks seven AI-powered and traditional approaches for MPV extraction from patent titles and abstracts, including TF-IDF, spaCy NER, KeyBERT with SciBERT, BERT-base, PatentSBERTa, GPT-4, and DeepSeek-R1. Using a semi-automated gold standard aligned with TRIZ/TESE principles, we evaluate these models on a curated USPTO dataset of wound care medical device patents. Results show that instruction-tuned LLMs (GPT-4 and DeepSeek-R1) achieve the highest precision (0.93 +) and F1 scores (0.94), accurately capturing value dimensions with strong contextual awareness. Despite their computational cost, these models outperform others in semantic generalization and MPV diversity. This work illustrates how large language models can power scalable, TESE-compliant analysis for strategic foresight in medical innovation and calls for developing cost-effective alternatives that replicate their accuracy.