Detecting Adverse Drug Events in Dermatology Through Natural Language Processing of Physician Notes

Background: Adverse drug events (ADEs) are major drivers of high cost and poor outcomes worldwide; however, the degree to which they are described in dermatology patient notes is poorly understood.

Objectives: To determine the frequency and nature (including causative drugs and drug-ADE combinations) of ADEs described in physician notes for dermatology patients.

Methods: Physician free-text notes recorded during outpatient visits at 5 specialty dermatology networks in the OMNY Health Dermatology Database were used. All notes were deidentified and only non-templated entries were included. A pretrained ADE classifier was used to screen which of the entries contained information about ADEs. Positive entries were passed to a named entity recognition (NER) pipeline for detection of drug and ADE entities. A relation extraction (RE) model was then applied to determine which of the drug-ADE pairs were related. Finally, an assertion status module was applied to classify the nature of the drug-ADE relationship (e.g., present, absent, conditional). Pretrained biomedical text embeddings and natural language processing (NLP) models available in a licensed software package (Spark NLP for Healthcare; Lewes, Delaware, US) were used.

Results: Approximately 448 million free-text entries from 4.1 million distinct patients were collected and deidentified. After removal of templated text, 8.7 million free-text entries from 2.5 million distinct patients and 6.0 million distinct encounters were ultimately used in the analysis. The ADE classifier identified 25,984 entries (0.3%) as potentially containing ADE related information. Further application of the NER pipeline and RE model yielded 7,640 drug-ADE relations across 5,299 distinct patients. Assertion status results identified 5,307 (69%) of the ADEs as being present. Among these present ADEs, the 5 most common causative agents were isotretinoin, spironolactone, doxycycline, tretinoin, and triamcinolone, which together accounted for 17% of ADEs. The top adverse event entity was rash (16%), followed by dryness (1.9%) headaches (1.7%), joint pain (1.6%), and hair loss (1.5%). When specific drug-ADE combinations were tallied, the 5 most common drug-ADE entities all included isotretinoin as the causative agent (dryness, headaches, joint pain, mood changes, and cheilitis).

Conclusions: These results suggest that acne medications are the most common causative agents for many of the ADEs observed in specialty dermatology networks. Additionally, they suggest that pretrained, end-to-end NLP systems are effective for characterizing ADEs in clinical text. Further research is needed to quantify the agreement with diagnostic codes and difference in yield when compared to structured data.