JMIR Publications

ABSTRACT

Background:

Lifestyle interventions for patients with prostate cancer (PCa) have been shown to improve treatment adherence and quality of life. However, there remains a lack of large language models (LLMs) capable of delivering individualized and professional lifestyle recommendations under clearly defined medical safety boundaries and controlled evidence sources.

Objective:

This study aimed to develop and evaluate a supervised fine-tuned large language model—PCaPLMM_SFT (Prostate Cancer Patient Lifestyle Management Model via Supervised Fine-Tuning)—to support health literacy improvement and lifestyle self-management among PCa patients.

Methods:

We searched English-language literature primarily from PubMed (February 2015–February 2025) to build a structured lifestyle-management knowledge base covering diet, physical activity, weight management, medication adherence, and psychological support. We used a retrieval-augmented generation (RAG) pipeline to generate patient-style question–answer (QA) pairs from retrieved knowledge slices. Bilingual English–Chinese QA data were generated from English-language source evidence through patient-oriented reformulation and RAG-based answer generation, and independent English and Chinese test sets were constructed to assess bilingual QA performance. We trained Baichuan2-7B-Chat using a two-stage strategy, consisting of continued pre-training followed by SFT with low-rank adaptation (LoRA). Model outputs were evaluated in two double-blind rounds by referee LLMs (Qwen3-Max and DeepSeek-R1) and compared with GPT-3.5-Turbo and the base Baichuan2-7B-Chat using 2,500 queries across five lifestyle scenarios. Additionally, three domain experts conducted a blinded review of 50 QA samples (10 per scenario). We used the Mann–Whitney U test with effect size r and Benjamini–Hochberg false discovery rate correction, and examined consistency using intraclass correlation coefficients (ICCs).

Results:

Based on 2,211 included publications, we constructed the PCaPLMM_SFT-Train dataset. The knowledge base yielded >150,000 structured knowledge slices. After two rounds of review, we obtained 42,330 single-turn QA pairs and 3,008 multi-turn dialogues, and the SFT phase utilized 45,338 structured QA samples. In the dual-round referee LLM assessment, PCaPLMM_SFT consistently outperformed Baichuan2-7B-Chat across dimensions and showed comparable or superior performance to GPT-3.5-Turbo across five lifestyle scenarios. Consistency analyses indicated moderate-to-good agreement between referee models across rounds, supporting the robustness of the comparative evaluation.

Conclusions:

PCaPLMM_SFT demonstrates the feasibility of constructing a medical lifestyle-focused LLM by integrating structured medical knowledge, QA-style training data, and a multi-layer evaluation system. This framework provides a reproducible methodological foundation for evidence-based health education and lifestyle management and establishes groundwork for future evaluation in real-world health management settings.