PhenoDEF: a corpus for annotating sentences with information of phenotype definitions in biomedical literature

Generalizability of machine learning methods in detecting adverse drug events from clinical narratives in electronic medical records

We assessed the generalizability of machine learning methods using natural language processing (NLP) techniques to detect adverse drug events (ADEs) from clinical narratives in electronic medical records (EMRs). We constructed a new corpus correlating drugs with adverse drug events using 1,394 clinical notes of 47 randomly selected patients who received immune checkpoint inhibitors (ICIs) from 2011 to 2018 at The Ohio State University James Cancer Hospital, annotating 189 drug-ADE relations in single sentences within the medical records. We also used data from Harvard's publicly available 2018 National Clinical Challenge (n2c2), which includes 505 discharge summaries with annotations of 1,355 single-sentence drug-ADE relations. We applied classical machine learning (support vector machine (SVM)), deep learning (convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM)), and state-of-the-art transformer-based (bidirectional encoder representations from transformers (BERT) and ClinicalBERT) methods trained and tested in the two different corpora and compared performance among them to detect drug-ADE relationships. ClinicalBERT detected drug-ADE relationships better than the other methods when trained using our dataset and tested in n2c2 (ClinicalBERT F-score, 0.78; other methods, F-scores, 0.61-0.73) and when trained using the n2c2 dataset and tested in ours (ClinicalBERT F-score, 0.74; other methods, F-scores, 0.55-0.72). Comparison among several machine learning methods demonstrated the superior performance and, therefore, the greatest generalizability of findings of ClinicalBERT for the detection of drug-ADE relations from clinical narratives in electronic medical records.

What Patients Say: Large-Scale Analyses of Replies to the Parkinson's Disease Patient Report of Problems (PD-PROP)

BACKGROUND

Free-text, verbatim replies in the words of people with Parkinson's disease (PD) have the potential to provide unvarnished information about their feelings and experiences. Challenges of processing such data on a large scale are a barrier to analyzing verbatim data collection in large cohorts.

OBJECTIVE

To develop a method for curating responses from the Parkinson's Disease Patient Report of Problems (PD-PROP), open-ended questions that asks people with PD to report their most bothersome problems and associated functional consequences.

METHODS

Human curation, natural language processing, and machine learning were used to develop an algorithm to convert verbatim responses to classified symptoms. Nine curators including clinicians, people with PD, and a non-clinician PD expert classified a sample of responses as reporting each symptom or not. Responses to the PD-PROP were collected within the Fox Insight cohort study.

RESULTS

Approximately 3,500 PD-PROP responses were curated by a human team. Subsequently, approximately 1,500 responses were used in the validation phase; median age of respondents was 67 years, 55% were men and median years since PD diagnosis was 3 years. 168,260 verbatim responses were classified by machine. Accuracy of machine classification was 95% on a held-out test set. 65 symptoms were grouped into 14 domains. The most frequently reported symptoms at first report were tremor (by 46% of respondents), gait and balance problems (>39%), and pain/discomfort (33%).

CONCLUSION

A human-in-the-loop method of curation provides both accuracy and efficiency, permitting a clinically useful analysis of large datasets of verbatim reports about the problems that bother PD patients.