JMIR Publications

ABSTRACT

Background:

The rapidly growing elderly population in Japan has increased demand for home care services. As a result, visiting nurses spend approximately 40% of their working time on documentation. Automated documentation using large language models (LLMs) shows potential but faces hallucination risks and lack of patient-specific context. Although retrieval-augmented generation (RAG) has emerged to address these limitations through knowledge embedding, existing healthcare RAG systems focus on single-patient contexts and remain unexplored for Japanese clinical documentation.

Objective:

This study aims to develop and evaluate an Adaptive Cascaded RAG (AC-RAG) system that safely integrates cross-patient knowledge through four-stage hierarchical filtering and adaptive strategy selection for automating Japanese nursing documentation.

Methods:

We developed a four-stage cascaded retrieval pipeline with disease-gated filtering, demographic similarity scoring, adaptive semantic thresholds, and context volume control. The system selects optimal knowledge integration strategies (Hybrid, History-Only, Cross-Patient-Only, No-RAG) based on data availability. We evaluated 89 home nursing consultations across two Automatic Speech Recognition (ASR) systems, comparing AC-RAG against Few-Shot Generated Knowledge Prompting (FS-GKP).

Results:

The conservative extraction achieved 70.8% higher precision than FS-GKP. For RAG-based summary generation, semantic similarity improved 28% (P<.001, Cohen's d=1.69–1.84), TF-IDF cosine similarity increased 24% (P<.001), and character-level BLEU improved 47% (P<.001). Processing speed increased 89–91% with a 59–61% cost reduction. Ablation analysis demonstrated the hybrid strategy achieved the highest performance (cosine similarity: 0.266±0.038). Cross-patient-only showed lower performance than the no-RAG baseline (cosine similarity: 0.175 vs. 0.192, P=.40, d=0.27), suggesting cross-patient knowledge provides benefit when combined with patient history.

Conclusions:

AC-RAG demonstrates superior accuracy, semantic quality, and computational efficiency. The incremental benefit of cross-patient retrieval requires validation in larger samples. At $0.043–0.054 per consultation, the system demonstrates economic feasibility for deployment in Japanese home care settings. However, moderate entity recall (0.493–0.519) indicates the system is best suited for generating draft documentation requiring nurse review rather than fully autonomous operation.