Shuntaro Yada

Motivation: Accurate TNM staging from lung cancer radiology reports is crucial for treatment planning and prognosis assessment. Manual staging processes are time-consuming and subject to inter-observer variability. Large language models (LLMs) offer opportunities to automate TNM staging with enhanced interpretability and clinical reasoning. Results: We developed two complementary systems for automated TNM staging from English radiology reports. System I employs GPT-4o with reasoning-based few-shot learning and multi-step voting. System II integrates multiple LLMs (GPT-4o and Gemini-2) using DSPy framework with MIPROv2 optimization. In NTCIR-18 RadNLP 2024 English main task, our approaches achieved first (joint accuracy: 0.6543) and second place (joint accuracy: 0.6296), demonstrating superior performance in T, N, and M classification with accuracies of 0.7037/0.9136/0.8889 and 0.7284/0.9383/0.8395, respectively. Availability and Implementation: Source code freely available at https://github.com/nlptmu/multi-expert-tnm-staging under MIT license. An archival snapshot of the version used in this study is deposited on Zenodo at https://doi.org/10.5281/zenodo.20338561. Implemented in Python 3.12+ with PyTorch 2.6 and DSPY 3.0, supporting Linux.

Background: Cancer survivors often experience complex and coexisting emotions throughout diagnosis, treatment, and posttreatment life. Emotion classification of patient narratives may help in understanding survivorship experiences; however, evidence remains limited for multidimensional classification using cancer survivor interview narratives. Objective: This study aimed to develop and evaluate natural language processing–based emotion classification models using Japanese cancer survivor interview narratives and to examine whether polarity and multidimensional emotion labels provide complementary perspectives. Methods: We analyzed verbatim transcripts from 15 cancer survivor interviews published by the Cancer Note, Nonprofit Organization. Survivor utterances were extracted, noninformative conversational elements were removed, texts were segmented at Japanese punctuation marks, and 5 consecutive sentences were grouped into 1 chunk. Two annotators labeled 1998 text chunks with 3-class sentiment polarity labels (positive, neutral, or negative) and multilabel Plutchik 8-emotion labels (joy, trust, fear, surprise, sadness, disgust, anger, and anticipation). Japanese BERT (Bidirectional Encoder Representations from Transformers) and LUKE (Language Understanding with Knowledge-based Embeddings) were fine-tuned to build a multiclass polarity classifier and a multilabel 8-emotion classifier. Performance was evaluated using precision, recall, F ₁ -score, macroaveraged metrics, Micro- F ₁ for polarity, and Hamming loss for multilabel classification. For comparison, the same architectures were fine-tuned on WRIME (writers’ and readers’ intensities of emotion for their estimation), a Japanese social media emotion dataset, and evaluated on Cancer Note texts as a domain-transfer analysis. The 95% CIs were estimated using bootstrap resampling with 1000 iterations. Results: Neutral was the most frequent polarity label, trust was the most frequent 8-emotion label, and anger was the least frequent emotion label. Label distributions were imbalanced, with most-to-least frequency ratios of 3.47 for polarity and 8.10 for 8-emotion labels. In the 3-class sentiment polarity task, interview-trained models outperformed WRIME-trained transfer models. Interview Text-BERT achieved the highest micro- F ₁ of 0.696 (95% CI 0.676‐0.716), whereas Interview Text-LUKE achieved the highest macro- F ₁ of 0.660 (95% CI 0.639‐0.682). In the 8-emotion multilabel task, Interview Text-LUKE achieved the highest macro- F ₁ of 0.427 (95% CI 0.398‐0.453) and the lowest Hamming loss of 0.078 (95% CI 0.073‐0.082). WRIME-trained transfer models showed lower performance, particularly in the 8-emotion task. Sadness and trust co-occurred most frequently, suggesting that positive and negative emotional elements may coexist in the same narratives. Conclusions: This exploratory study suggests the feasibility of domain-specific emotion classification for Japanese cancer survivor interview narratives. Models fine-tuned on target-domain narratives generally outperformed WRIME-trained transfer models, although the best architecture differed by task and metric. Polarity labels and Plutchik 8-emotion labels provided complementary perspectives on complex and coexisting emotions in survivorship narratives. However, performance for rare emotions remained limited, and the models should be regarded as preliminary research tools rather than clinically actionable systems. Larger, more diverse, prospectively or externally validated datasets, imbalance-aware methods, and user-centered evaluation are needed before clinical translation.

BACKGROUND Cancer survivors often experience complex and sometimes coexisting emotions across diagnosis, treatment, and posttreatment life. Capturing these emotions from patient narratives may support clinicians in assessing psychological states and delivering patient-centered psychosocial care, but evidence remains limited on building multi-emotion classifiers from interview narratives. These narratives reflect both personal and social dimensions of survivorship, aligning with the social model of health. OBJECTIVE This study aimed to develop and evaluate emotion classification models using Japanese cancer survivor interview narratives as a formative step toward narrative-informed support. METHODS We analyzed verbatim transcripts from 15 cancer survivor interviews (typically ~90 minutes each) published by the NPO GanNote. From narratives covering diagnosis/disclosure through posttreatment daily life, we preprocessed the transcripts by removing non-informative conversational elements and segmenting text at punctuation into sentence-level chunks, which were then grouped into blocks of five consecutive sentences to fit the 512-token input limit. Two annotators labeled 1,998 text chunks with (1) 3-class sentiment polarity (positive/neutral/negative) and (2) overlapping labels for Plutchik’s eight basic emotions (trust, anticipation, joy, surprise, sadness, fear, disgust, anger). We fine-tuned Japanese BERT and LUKE to build a multi-class polarity classifier and a multilabel eight-emotion classifier, and evaluated performance with F1-scores (including Micro-F1 for polarity and exact match ratio for multilabel). For comparison, we trained the same architectures on WRIME (social media posts with reader-assigned emotions) and tested transfer performance on the interview data. RESULTS The interviews had a mean of 136.9 words per narrative. Label distributions were imbalanced, with the most-to-least frequency ratio of 3.47 (polarity) and 8.10 (eight emotions); neutral and trust were the most frequent labels, whereas anger was the least frequent. Co-occurrence analysis showed sadness and trust most frequently co-occurred, while anger and anticipation co-occurred least. The LUKE model fine-tuned on interview narratives achieved the best performance for polarity classification (Micro-F1=0.73; Macro-F1=0.67; F1: neutral=0.76, positive=0.62, negative=0.63). For eight emotions, the same model achieved Macro-F1=0.42 and exact match ratio=0.51, with higher F1 for trust (0.62), anticipation (0.54), disgust (0.54), sadness (0.52), and fear (0.49) but lower F1 for anger (0.09), joy (0.33), and surprise (0.21). Models trained on WRIME showed lower performance when evaluated on interview narratives (polarity Micro-F1=0.58–0.60; eight-emotion exact match ratio=0.18–0.21). CONCLUSIONS NLP-based emotion classification from cancer survivor interview narratives is feasible and domain-specific models outperformed general-purpose classifiers, suggesting the importance of training data that reflects the target narrative context. These findings illustrate that survivorship narratives may contain coexisting and contrasting emotions not fully captured by polarity alone. Addressing class imbalance and improving detection of low-frequency emotions (eg, anger) will be important for future work. Before clinical translation, user-centered evaluation to assess interpretability, safety, and workflow integration will also be needed.

We propose a large-scale synthetic dataset that correlates structured background information aligned with the actual distribution of patients with cerebral infarction, nurse characteristics, and nurse-patient dialogues across diverse scenarios. Medical dialogue corpora are scarce due to privacy and access restrictions. Even when available, they primarily focus on physician-patient interactions and offer limited metadata (clinical covariates, staff characteristics, etc.). To address this gap, this resource conditions large language models on patient covariate tables and nurse characteristics, generating multilingual, daily-structured dialogues (7 scenarios per day) using a standardized JSONL schema. Potential applications include nursing education under conditions of limited clinical practice time (supporting scenario-based therapeutic communication and formative assessment), research involving controlled experiments on nursing record formats and quality indicators, and training or benchmarking AI models when actual texts cannot be shared (subject to the selected model’s terms of use). All content is synthetic data and does not contain protected health information. Reproducible scripts pair patients with nurses, assign care pathways, and generate conversations on a scale.

This study assessed the effectiveness of natural language processing (NLP) in detecting adverse events (AEs) from anticancer agents by analyzing data from over 39,000 cancer patients. A specialized machine learning model identified known AEs from anticancer agents like capecitabine, oxaliplatin, and anthracyclines, revealing a significantly higher incidence in the treatment groups compared to non-users. While the NLP approach effectively detected most symptomatic AEs requiring manual review, it struggled with rarely documented conditions and commonly used clinical terms. Overall, the method shows promise for automated AE detection in medical records, particularly for symptoms without laboratory markers or diagnosis codes.