Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-0509
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-07-22T06:09:28.544643Z
status: published
lens: g_space
sector: health
headline: A Supervised Fine-Tuned Large Language Model for Lifestyle Management in Patients With Prostate Cancer: Development and Evaluation Study
dek: Lifestyle interventions for patients with prostate cancer 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. This study aimed to develop and evaluate a supervised fine-tuned LLM-PCaPLMM_SFT (Prostate Cancer Patient Lifestyle Management Model via Supervised Fine-Tuning)-to supp…
gain_title: A supervised fine-tuned Baichuan2-7B-Chat model trained on 45,338 prostate cancer lifestyle QA samples outperformed its base model and performed comparably or better than GPT-3.5-Turbo across diet, activity, weight, adherence, and psychological support scenarios in dual-round blinded LLM referee evaluation.
problem_title: (none)
trace_subject: (none)
gain_reading: A supervised fine-tuned Baichuan2-7B-Chat model trained on 45,338 prostate cancer lifestyle QA samples outperformed its base model and performed comparably or better than GPT-3.5-Turbo across diet, activity, weight, adherence, and psychological support scenarios in dual-round blinded LLM referee evaluation.
gain_evidence: PCaPLMM_SFT consistently outperformed Baichuan2-7B-Chat across dimensions and showed comparable or superior performance to GPT-3.5-Turbo across 5 lifestyle scenarios.
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers developed PCaPLMM_SFT, a Baichuan2-7B-Chat model fine-tuned for prostate cancer lifestyle management, using a knowledge base built from 2211 PubMed articles and over 150,000 knowledge slices covering diet, physical activity, weight, medication adherence, and psychological support. They generated 42,330 single-turn and 3008 multiturn QA pairs and evaluated the model against GPT-3.5-Turbo and the base model.

The blinded referee evaluation suggests a reproducible method for evidence-bounded lifestyle education, but clinical utility remains unproven because testing was limited to curated QA queries and expert review of 50 samples rather than patient outcomes, adherence, or safety events in routine care.
limitation: Model has not yet been evaluated in real-world health management settings and remains a methodological feasibility demonstration rather than a deployed clinical intervention.
tag: Evidence-backed gain
key_points: Built knowledge base from 2211 PubMed publications from February 2015 to February 2025 yielding >150,000 structured knowledge slices. | Generated 42,330 single-turn QA pairs and 3008 multiturn dialogues with bilingual English-Chinese reformulation for training and testing. | Trained Baichuan2-7B-Chat with 2-stage continued pretraining and supervised fine-tuning with low-rank adaptation. | Evaluated with 2500 queries across 5 lifestyle scenarios by referee LLMs Qwen3-Max and DeepSeek-R1 plus blinded review of 50 samples by 3 domain experts.
rundown: The team constructed PCaPLMM_SFT-Train from 2211 publications and used a retrieval-augmented generation pipeline to create patient-style QA pairs, including bilingual English-Chinese data and independent test sets. Training used Baichuan2-7B-Chat with continued pretraining followed by low-rank adaptation fine-tuning on 45,338 samples.

Evaluation involved two double-blind rounds with Qwen3-Max and DeepSeek-R1 as referees across 2500 queries, plus Mann-Whitney U tests with Benjamini-Hochberg correction and intraclass correlation for consistency, and a separate blinded review of 50 QA samples by three domain experts.
sources:
- peer_reviewed | Journal of Medical Internet Research | https://doi.org/10.2196/92663 | 2026-07-21
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

c112fd97d47bd889c0692f11c541e33b68de9f32fb44e6eb12a71a584bf3c93b

previous

0000000000000000000000000000000000000000000000000000000000000000