LLM-Driven User Preference Topic Modeling for Design Decision Support

Online reviews of home medical devices are noisy, heterogeneous, and uneven in information density, which makes it hard to turn the voice of the customer into engineering design decisions. This study proposes and validates a review-mining framework that couples rule-based data governance with a locally deployed large language model (LLM) and complex-network analysis. Taking home oxygen concentrators as the case, we collected 25,893 raw review entries from JD.com, Taobao, and Tmall. Because overlapping collection windows made 80.4% of these rows redundant (5084 unique review texts), a four-stage pipeline combining regular-expression cleaning, domain-relevance filtering, low-value review removal, and exact-plus-prefix deduplication refined them into a 2238-review corpus (44.0% of the unique review texts). A locally hosted Qwen3-8B model then performed aspect-level extraction within nine predefined design aspects, yielding 6068 structured aspect–topic–confidence records from 2214 reviews (98.9% parsing success), and scored sentiment for 6052 unique review–aspect pairs on a [−1, 1] scale. From these records, we built a topic co-occurrence network (202 nodes and 1895 edges after low-frequency filtering) and ranked design requirements with a Composite Centrality Index (CCI; weights 0.4/0.4/0.2 over min–max-normalized degree, betweenness, and closeness). Simple Operation (CCI = 1.000) and Low Noise (0.981) dominate user preference, followed by Solid Build Quality (0.611) and Oxygen Concentration (0.574); sentiment is strongly polarized on acoustic experience (Low Noise +0.755 vs. Excessive Noise −0.560). Five checks support these findings: an LDA baseline produced substantially coarser, partly brand-dominated topics; the CCI ranking is stable under five alternative weightings (Spearman’s ρ ≥ 0.978) and across 16 combinations of network-construction thresholds; bootstrap resampling (B = 1000) places the top two topics within ranks 1–2 with 95% confidence; the leading themes replicate in a JD.com-only subnetwork; and manual validation of a stratified 450-pair sample (two blinded annotators, adjudicated gold standard) gave 87.6% aspect accuracy and 94.4% sentiment-polarity accuracy. The framework converts unstructured reviews into a quantified, traceable design-priority list for oxygen concentrator development; we interpret the resulting associations as descriptive evidence for prioritization, not as causal effects.