Hayato Kizaki

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.

Purpose: Urothelial carcinoma (UC) is a common malignant tumor of the urinary tract, with bladder cancer as the most frequent type. Radical cystectomy (RC) is the standard treatment for muscle-invasive bladder cancer (MIBC), and perioperative chemotherapy is recommended in addition to RC in guidelines. However, no established evidence defines the optimal regimen and number of cycles, and treatment varies by institution. Therefore, this study aimed to clarify perioperative treatment patterns for MIBC using a hospital-based claims database. Methods: Patients with claims records for bladder cancer recorded during the case-identification window (2012–2020) were identified using the Medical Data Vision Co., Ltd. database. The database spans April 2008 to July 2021, enabling longitudinal assessment. Patient characteristics and prescribed perioperative chemotherapy regimens were analyzed. Factors associated with neoadjuvant chemotherapy (NAC) were examined using univariable analysis and an exploratory multivariable logistic regression model. Results: Among 6,723 patients, 78.5% were men, with a median diagnosis age of 71 years. The treatment groups included 2,324 patients who received NAC, 598 who received adjuvant chemotherapy (AC), 364 who received NAC and AC, and 3,437 who underwent RC-only. NAC utilization rate increased over time ( P for trend < 0.0001). Gemcitabine plus cisplatin was the most frequently administered regimen. Univariable analysis revealed an association between clinical stage 3 or higher and the NAC administration. In contrast, older age and comorbidities were associated with non-administration of NAC. Furthermore, the hospital size and cancer hospital designation influenced NAC administration. Conclusion: Perioperative chemotherapy utilization for MIBC has increased annually. Patient characteristics and treatment facility factors influence NAC administration.

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.

Analyzing medication-related incident reports is crucial for patient safety; however, systematically extracting the underlying factors contributing to incident occurrence remains challenging. We developed a multi-label classifier that automatically identified incident factors from 1,212 drug-related incident reports using the Bidirectional Encoder Representations from Transformers and its derivatives. Based on the P-mSHELL model, a comprehensive framework for incident factor analysis, we established seven distinct factor categories and evaluated various pre-trained models through five-fold cross-validation. Almost all models achieved macro F1 scores exceeding 0.6, with the lightweight A Lite BERT model showing comparable performance to BERT. This study demonstrates the practical feasibility of natural language processing techniques for systematic incident factor analysis, supporting enhanced patient safety management.

PURPOSE: Capecitabine, an oral anticancer agent, frequently causes hand-foot syndrome (HFS), affecting patients' quality of life and treatment adherence. However, such symptomatic toxicities are often difficult to detect in structured electronic health record (EHR) data. This study primarily aimed to validate a natural language processing (NLP) approach to identifying capecitabine-induced HFS from unstructured clinical text and demonstrate its application in evaluating medication-associated adverse event trends in real-world settings. METHODS: We conducted a retrospective cohort study using EHRs from the University of Tokyo Hospital (2004-2021). HFS cases were identified using the MedNERN-CR-JA NLP model. After propensity score matching, we compared capecitabine users with and without celecoxib and assessed time to HFS onset using Cox proportional hazards models. NLP-based HFS detection was validated through manual annotation of aggregated clinical notes. Negative control and sensitivity analyses ensured robustness. RESULTS: Among 44,502 patients with cancer, 669 capecitabine users were analyzed. HFS incidence was significantly higher among capecitabine users (hazard ratio [HR], 1.93 [95% CI, 1.48 to 2.52]; P < .001) compared with nonusers. Celecoxib use showed a suggestive association with a reduced HFS risk (HR, 0.51 [95% CI, 0.24 to 1.07]; P = .073). The NLP model demonstrated high accuracy in identifying HFS, achieving a precision of 0.875, recall of 1.000, and F ₁ score of 0.933, based on manual annotation of patient-level clinical notes. Outcome trends remained consistent when using manually annotated HFS case labels instead of NLP-detected events, supporting the method's robustness. CONCLUSION: These findings demonstrate the effectiveness of NLP in detecting HFS from real-world clinical records. The application to celecoxib-HFS detection illustrates the potential utility of this approach for retrospective safety analysis. Further work is needed to evaluate generalizability across diverse clinical settings.

Background: Declining medication adherence remains a critical health care issue, often assessed through unreliable self-reporting methods. Wearable devices (WDs) may offer an objective means to improve adherence monitoring by continuously recording physiological and activity data. Objective: This study aimed to develop and internally validate personalized predictive models, utilizing objective physiological and activity data from WDs, for identifying missed medication doses. Methods: A 30-day prospective observational study was conducted with 8 participants who wore Apple Watches and used a dedicated iOS app. The app collected demographics, medication details, psychological factors, mealtimes, and daily missed dose events. WDs recorded time-series data (ie, activity, heart rate, sleep) at 3-minute intervals. Data were aggregated into 1-hour segments, and lag (6 and 12 h) as well as rolling (24 h) features were generated. Light Gradient Boosting Machine models were constructed for each individual’s dosing regimen if the missed dose rate exceeded 20%. Two modeling approaches were compared: a group cross-validation (CV) model that grouped data by day to avoid data leakage from rolling features, and a nonrolling feature model that excluded rolling features and used leave-one-out CV. F ₁ -score, accuracy, recall, and precision were assessed between the 2 models. Results: Of the 15 enrolled participants, 8 completed the study; 4 had a missed dose rate above 20%. In these 4 individuals, the group CV model achieved F ₁ -scores of 0.435 to 0.902, with accuracy ranging from 0.711 to 0.911, recall from 0.278 to 0.822, and a precision of 1.000 for the most robust regimens. The nonrolling feature model yielded F ₁ -scores of 0.667 to 0.910, with accuracy ranging from 0.800 to 0.906, recall from 0.500 to 0.835, and a precision of 1.000. Morning dosing regimens generally showed higher predictive performance than evening or afternoon. Time-series features, particularly those reflecting 6-, 12-, and 24-hour patterns, emerged as key predictors, indicating that physiological and lifestyle variations prior to dosing strongly influenced missed dose events. Conclusions: Personalized predictive models using WD-derived data demonstrated high precision for detecting missed medication doses, especially in morning and evening regimens. These findings underscore the feasibility of employing continuous, objective physiological and activity data from WDs to forecast nonadherence events. Although the sample size was limited, restricting the generalizability of the results, this study demonstrates the potential of WD-based personalized prediction of medication adherence. Future work should involve larger populations for external validation, strategies to improve recall, especially for clinically critical medications, and careful consideration of real-world implementation challenges.

BACKGROUND Medication-related incidents remain a major concern in healthcare safety and often result from complex, multifactorial causes. The systematic analysis of incident reports is crucial for preventing serious medical accidents; however, manual reviews are time-consuming and resource-intensive. Natural language processing (NLP) offers a promising approach for automating such analyses; however, few studies have explored the integration of human factor engineering frameworks, such as the P-mSHELL model, into NLP-based approaches. OBJECTIVE This study aimed to develop and evaluate a multilabel classifier using NLP to automatically extract contributing factors from medication-related incident reports based on the P-mSHELL framework. METHODS We analyzed 5058 sentences extracted from 1192 medication-related incident reports collected between 2020 and 2022. Eight contributing factor categories were defined based on the modified P-mSHELL model: medical professional (MP), patient family (PF), patient (P), management (M), software (S), hardware (H), environment (E), and liveware-liveware (L-L). Two medical experts annotated the sentences, and inter-annotator agreement was assessed using Cohen's kappa coefficient. A multilabel classification model was developed by fine-tuning a pretrained Japanese BERT model and its performance was evaluated using 5-fold cross-validation, F1-scores, and exact match accuracy. RESULTS Substantial inter-annotator agreement (kappa coefficient > 0.67) was observed across all categories. Multiple labels were assigned to 67.0% of sentences, with MPs (62.9%) and S (59.9%) being the most frequently used. The classifier attained F1-scores exceeded 0.57 for all evaluated labels, with particularly high performance for MPs (0.90) and S (0.83) factors. The exact match accuracy was 0.51, indicating the successful multilabel prediction of all labels in more than half of the test cases. CONCLUSIONS This study demonstrated the feasibility of using NLP technologies to automate the extraction of multiple contributing factors from medication-related incident reports within the structured P-mSHELL. The developed classifier exhibited practical performance levels, particularly for the major factor categories. The proposed approach offers a promising tool for enhancing the scalability, efficiency, and consistency of incident report analysis, ultimately contributing to more proactive and data-driven medication safety management.

Background: Although hypertension is a common side effect of tyrosine kinase inhibitor (TKI) treatment for renal cell carcinoma (RCC), there is limited evidence regarding its occurrence and related risk factors. Preliminary studies suggest that proton pump inhibitors (PPIs) may mitigate the risk of TKI‐induced hypertension; however, their clinical effectiveness remains unclear. Aims: In this study, we examined the prevalence of TKI‐induced hypertension and the patterns of antihypertensive prescriptions among patients with RCC in Japan. Additionally, we investigated factors associated with TKI‐induced hypertension to assess the potential impact of PPIs. Methods and Results: Data from patients diagnosed with RCC who were prescribed TKIs between April 2008 and July 2021 were retrospectively gathered from a Japanese administrative database. TKI‐induced hypertension was detected following the diagnosis of hypertension and subsequently the prescription of an antihypertensive agent during TKI therapy. The prescription details for antihypertensive agents were organized in a tabular format. Cox proportional hazards regression analysis was conducted to examine factors contributing to TKI‐induced hypertension. Among the 225 patients analyzed, 36.4% experienced hypertension, and calcium channel blockers were the most prescribed antihypertensive agents. Pre‐existing hypertension was identified as a risk factor for TKI‐induced hypertension, while the concurrent use of PPIs did not show a tendency to reduce the risk of TKI‐induced hypertension. Conclusions: These findings indicate the importance of blood pressure management in patients with elevated baseline blood pressure.