Identification of hand-foot syndrome from cancer patients' blog posts: BERT-based deep-learning approach to detect potential adverse drug reaction symptoms

Exploring a method for extracting concerns of multiple breast cancer patients in the domain of patient narratives using BERT and its optimization by domain adaptation using masked language modeling

Narratives posted on the internet by patients contain a vast amount of information about various concerns. This study aimed to extract multiple concerns from interviews with breast cancer patients using the natural language processing (NLP) model bidirectional encoder representations from transformers (BERT). A total of 508 interview transcriptions of breast cancer patients written in Japanese were labeled with five types of concern labels: "treatment," "physical," "psychological," "work/financial," and "family/friends." The labeled texts were used to create a multi-label classifier by fine-tuning a pre-trained BERT model. Prior to fine-tuning, we also created several classifiers with domain adaptation using (1) breast cancer patients' blog articles and (2) breast cancer patients' interview transcriptions. The performance of the classifiers was evaluated in terms of precision through 5-fold cross-validation. The multi-label classifiers with only fine-tuning had precision values of over 0.80 for "physical" and "work/financial" out of the five concerns. On the other hand, precision for "treatment" was low at approximately 0.25. However, for the classifiers using domain adaptation, the precision of this label took a range of 0.40-0.51, with some cases improving by more than 0.2. This study showed combining domain adaptation with a multi-label classifier on target data made it possible to efficiently extract multiple concerns from interviews.

Symptom-BERT: Enhancing Cancer Symptom Detection in EHR Clinical Notes

CONTEXT

Extracting cancer symptom documentation allows clinicians to develop highly individualized symptom prediction algorithms to deliver symptom management care. Leveraging advanced language models to detect symptom data in clinical narratives can significantly enhance this process.

OBJECTIVE

This study uses a pretrained large language model to detect and extract cancer symptoms in clinical notes.

METHODS

We developed a pretrained language model to identify cancer symptoms in clinical notes based on a clinical corpus from the Enterprise Data Warehouse for Research at a healthcare system in the Midwestern United States. This study was conducted in 4 phases:1 pretraining a Bio-Clinical BERT model on one million unlabeled clinical documents,2 fine-tuning Symptom-BERT for detecting 13 cancer symptom groups within 1112 annotated clinical notes,3 generating 180 synthetic clinical notes using ChatGPT-4 for external validation, and4 comparing the internal and external performance of Symptom-BERT against a non-pretrained version and six other BERT implementations.

RESULTS

The Symptom-BERT model effectively detected cancer symptoms in clinical notes. It achieved results with a micro-averaged F1-score of 0.933, an AUC of 0.929 internally, and 0.831 and 0.834 externally. Our analysis shows that physical symptoms, like Pruritus, are typically identified with higher performance than psychological symptoms, such as anxiety.

CONCLUSION

This study underscores the transformative potential of specialized pretraining on domain-specific data in boosting the performance of language models for medical applications. The Symptom-BERT model's exceptional efficacy in detecting cancer symptoms heralds a groundbreaking stride in patient-centered AI technologies, offering a promising path to elevate symptom management and cultivate superior patient self-care outcomes.

Construction of a Multi-Label Classifier for Extracting Multiple Incident Factors From Medication Incident Reports in Residential Care Facilities: Natural Language Processing Approach

BACKGROUND

Medication safety in residential care facilities is a critical concern, particularly when nonmedical staff provide medication assistance. The complex nature of medication-related incidents in these settings, coupled with the psychological impact on health care providers, underscores the need for effective incident analysis and preventive strategies. A thorough understanding of the root causes, typically through incident-report analysis, is essential for mitigating medication-related incidents.

OBJECTIVE

We aimed to develop and evaluate a multilabel classifier using natural language processing to identify factors contributing to medication-related incidents using incident report descriptions from residential care facilities, with a focus on incidents involving nonmedical staff.

METHODS

We analyzed 2143 incident reports, comprising 7121 sentences, from residential care facilities in Japan between April 1, 2015, and March 31, 2016. The incident factors were annotated using sentences based on an established organizational factor model and previous research findings. The following 9 factors were defined: procedure adherence, medicine, resident, resident family, nonmedical staff, medical staff, team, environment, and organizational management. To assess the label criteria, 2 researchers with relevant medical knowledge annotated a subset of 50 reports; the interannotator agreement was measured using Cohen κ. The entire data set was subsequently annotated by 1 researcher. Multiple labels were assigned to each sentence. A multilabel classifier was developed using deep learning models, including 2 Bidirectional Encoder Representations From Transformers (BERT)-type models (Tohoku-BERT and a University of Tokyo Hospital BERT pretrained with Japanese clinical text: UTH-BERT) and an Efficiently Learning Encoder That Classifies Token Replacements Accurately (ELECTRA), pretrained on Japanese text. Both sentence- and report-level training were performed; the performance was evaluated by the F1-score and exact match accuracy through 5-fold cross-validation.

RESULTS

Among all 7121 sentences, 1167, 694, 2455, 23, 1905, 46, 195, 1104, and 195 included "procedure adherence," "medicine," "resident," "resident family," "nonmedical staff," "medical staff," "team," "environment," and "organizational management," respectively. Owing to limited labels, "resident family" and "medical staff" were omitted from the model development process. The interannotator agreement values were higher than 0.6 for each label. A total of 10, 278, and 1855 reports contained no, 1, and multiple labels, respectively. The models trained using the report data outperformed those trained using sentences, with macro F1-scores of 0.744, 0.675, and 0.735 for Tohoku-BERT, UTH-BERT, and ELECTRA, respectively. The report-trained models also demonstrated better exact match accuracy, with 0.411, 0.389, and 0.399 for Tohoku-BERT, UTH-BERT, and ELECTRA, respectively. Notably, the accuracy was consistent even when the analysis was confined to reports containing multiple labels.

CONCLUSIONS

The multilabel classifier developed in our study demonstrated potential for identifying various factors associated with medication-related incidents using incident reports from residential care facilities. Thus, this classifier can facilitate prompt analysis of incident factors, thereby contributing to risk management and the development of preventive strategies.

Using the Natural Language Processing System Medical Named Entity Recognition-Japanese to Analyze Pharmaceutical Care Records: Natural Language Processing Analysis

BACKGROUND

Large language models have propelled recent advances in artificial intelligence technology, facilitating the extraction of medical information from unstructured data such as medical records. Although named entity recognition (NER) is used to extract data from physicians' records, it has yet to be widely applied to pharmaceutical care records.

OBJECTIVE

In this study, we aimed to investigate the feasibility of automatic extraction of the information regarding patients' diseases and symptoms from pharmaceutical care records. The verification was performed using Medical Named Entity Recognition-Japanese (MedNER-J), a Japanese disease-extraction system designed for physicians' records.

METHODS

MedNER-J was applied to subjective, objective, assessment, and plan data from the care records of 49 patients who received cefazolin sodium injection at Keio University Hospital between April 2018 and March 2019. The performance of MedNER-J was evaluated in terms of precision, recall, and F1-score.

RESULTS

The F1-scores of NER for subjective, objective, assessment, and plan data were 0.46, 0.70, 0.76, and 0.35, respectively. In NER and positive-negative classification, the F1-scores were 0.28, 0.39, 0.64, and 0.077, respectively. The F1-scores of NER for objective (0.70) and assessment data (0.76) were higher than those for subjective and plan data, which supported the superiority of NER performance for objective and assessment data. This might be because objective and assessment data contained many technical terms, similar to the training data for MedNER-J. Meanwhile, the F1-score of NER and positive-negative classification was high for assessment data alone (F1-score=0.64), which was attributed to the similarity of its description format and contents to those of the training data.

CONCLUSIONS

MedNER-J successfully read pharmaceutical care records and showed the best performance for assessment data. However, challenges remain in analyzing records other than assessment data. Therefore, it will be necessary to reinforce the training data for subjective data in order to apply the system to pharmaceutical care records.

Adverse Event Signal Detection Using Patients' Concerns in Pharmaceutical Care Records: Evaluation of Deep Learning Models

BACKGROUND

Early detection of adverse events and their management are crucial to improving anticancer treatment outcomes, and listening to patients' subjective opinions (patients' voices) can make a major contribution to improving safety management. Recent progress in deep learning technologies has enabled various new approaches for the evaluation of safety-related events based on patient-generated text data, but few studies have focused on the improvement of real-time safety monitoring for individual patients. In addition, no study has yet been performed to validate deep learning models for screening patients' narratives for clinically important adverse event signals that require medical intervention. In our previous work, novel deep learning models have been developed to detect adverse event signals for hand-foot syndrome or adverse events limiting patients' daily lives from the authored narratives of patients with cancer, aiming ultimately to use them as safety monitoring support tools for individual patients.

OBJECTIVE

This study was designed to evaluate whether our deep learning models can screen clinically important adverse event signals that require intervention by health care professionals. The applicability of our deep learning models to data on patients' concerns at pharmacies was also assessed.

METHODS

Pharmaceutical care records at community pharmacies were used for the evaluation of our deep learning models. The records followed the SOAP format, consisting of subjective (S), objective (O), assessment (A), and plan (P) columns. Because of the unique combination of patients' concerns in the S column and the professional records of the pharmacists, this was considered a suitable data for the present purpose. Our deep learning models were applied to the S records of patients with cancer, and the extracted adverse event signals were assessed in relation to medical actions and prescribed drugs.

RESULTS

From 30,784 S records of 2479 patients with at least 1 prescription of anticancer drugs, our deep learning models extracted true adverse event signals with more than 80% accuracy for both hand-foot syndrome (n=152, 91%) and adverse events limiting patients' daily lives (n=157, 80.1%). The deep learning models were also able to screen adverse event signals that require medical intervention by health care providers. The extracted adverse event signals could reflect the side effects of anticancer drugs used by the patients based on analysis of prescribed anticancer drugs. "Pain or numbness" (n=57, 36.3%), "fever" (n=46, 29.3%), and "nausea" (n=40, 25.5%) were common symptoms out of the true adverse event signals identified by the model for adverse events limiting patients' daily lives.

CONCLUSIONS

Our deep learning models were able to screen clinically important adverse event signals that require intervention for symptoms. It was also confirmed that these deep learning models could be applied to patients' subjective information recorded in pharmaceutical care records accumulated during pharmacists' daily work.

Adverse event signal extraction from cancer patients' narratives focusing on impact on their daily-life activities

Adverse event (AE) management is important to improve anti-cancer treatment outcomes, but it is known that some AE signals can be missed during clinical visits. In particular, AEs that affect patients' activities of daily living (ADL) need careful monitoring as they may require immediate medical intervention. This study aimed to build deep-learning (DL) models for extracting signals of AEs limiting ADL from patients' narratives. The data source was blog posts written in Japanese by breast cancer patients. After pre-processing and annotation for AE signals, three DL models (BERT, ELECTRA, and T5) were trained and tested in three different approaches for AE signal identification. The performances of the trained models were evaluated in terms of precision, recall, and F1 scores. From 2,272 blog posts, 191 and 702 articles were identified as describing AEs limiting ADL or not limiting ADL, respectively. Among tested DL modes and approaches, T5 showed the best F1 scores to identify articles with AE limiting ADL or all AE: 0.557 and 0.811, respectively. The most frequent AE signals were "pain or numbness", "fatigue" and "nausea". Our results suggest that this AE monitoring scheme focusing on patients' ADL has potential to reinforce current AE management provided by medical staff.

Extracting Multiple Worries from Breast Cancer Patient Blogs Using Multi-Label Classification with a Natural Language-Processing Model BERT (Bidirectional Encoder Representations from Transformers): Infodemiology Study of Blogs (Preprint)

Background

Patients with breast cancer have a variety of worries and need multifaceted information support. Their accumulated posts on social media contain rich descriptions of their daily worries concerning issues such as treatment, family, and finances. It is important to identify these issues to help patients with breast cancer to resolve their worries and obtain reliable information.

Objective

This study aimed to extract and classify multiple worries from text generated by patients with breast cancer using Bidirectional Encoder Representations From Transformers (BERT), a context-aware natural language processing model.

Methods

A total of 2272 blog posts by patients with breast cancer in Japan were collected. Five worry labels, “treatment,” “physical,” “psychological,” “work/financial,” and “family/friends,” were defined and assigned to each post. Multiple labels were allowed. To assess the label criteria, 50 blog posts were randomly selected and annotated by two researchers with medical knowledge. After the interannotator agreement had been assessed by means of Cohen kappa, one researcher annotated all the blogs. A multilabel classifier that simultaneously predicts five worries in a text was developed using BERT. This classifier was fine-tuned by using the posts as input and adding a classification layer to the pretrained BERT. The performance was evaluated for precision using the average of 5-fold cross-validation results.

Results

Among the blog posts, 477 included “treatment,” 1138 included “physical,” 673 included “psychological,” 312 included “work/financial,” and 283 included “family/friends.” The interannotator agreement values were 0.67 for “treatment,” 0.76 for “physical,” 0.56 for “psychological,” 0.73 for “work/financial,” and 0.73 for “family/friends,” indicating a high degree of agreement. Among all blog posts, 544 contained no label, 892 contained one label, and 836 contained multiple labels. It was found that the worries varied from user to user, and the worries posted by the same user changed over time. The model performed well, though prediction performance differed for each label. The values of precision were 0.59 for “treatment,” 0.82 for “physical,” 0.64 for “psychological,” 0.67 for “work/financial,” and 0.58 for “family/friends.” The higher the interannotator agreement and the greater the number of posts, the higher the precision tended to be.

Conclusions

This study showed that the BERT model can extract multiple worries from text generated from patients with breast cancer. This is the first application of a multilabel classifier using the BERT model to extract multiple worries from patient-generated text. The results will be helpful to identify breast cancer patients’ worries and give them timely social support.