E-scooter related injuries: Using natural language processing to rapidly search 36 million medical notes

Abstract

Background: Shareable e-scooters have become popular, but injuries to riders and bystanders have not been well characterized. The goal of this study was to describe e-scooter injuries and estimate the rate of injury per e-scooter trip.

Methods and findings: Retrospective review of patients presenting to 180 clinics and 2 hospitals in greater Los Angeles between January 1, 2014 and May 14, 2020. Injuries were identified using a natural language processing (NLP) algorithm not previously used to identify injuries, tallied, and described along with required healthcare resources. We combine these tallies with municipal data on scooter use to report a monthly utilization-corrected rate of e-scooter injuries. We searched 36 million clinical notes. Our NLP algorithm correctly classified 92% of notes in the testing set compared with the gold standard of investigator review. In total, we identified 1,354 people injured by e-scooters; 30% were seen in more than one clinical setting (e.g., emergency department and a follow-up outpatient visit), 29% required advanced imaging, 6% required inpatient admission, and 2 died. We estimate 115 injuries per million e-scooter trips were treated in our health system.

Conclusions: Our observed e-scooter injury rate is likely an underestimate, but is similar to that previously reported for motorcycles. However, the comparative severity of injuries is unknown. Our methodology may prove useful to study other clinical conditions not identifiable by existing diagnostic systems.

Fig 1. Overview of neural network process for NLP algorithm*.

Our process incorporated both convolutional and deep neural networks into an ensemble voting process, known as random multimodel deep learning (RMDL) [8, 9]. This combination of deep learning techniques uses global vectors for word representation (GloVe) [11] input into a convolutional neural network (CNN) and term frequency-inverse document frequency (TF-IDF) input into a deep neural network (DNN), converting language into a series of vectors that convey word meaning and interact to represent combinations of words. Each of two deep learning neural network techniques (CNN at top, and DNN at bottom) then identify patterns in vector representations that correspond to notes in the training data that describe e-scooter injuries, and translate the strength of these patterns into scores from each of the two neural networks that are averaged through soft voting into a probability of e-scooter injury.

Fig 2. Flowchart of testing process for the predictions of our NLP algorithm*.

ED notes in panel A, and outpatient notes in panel B. As in S1 Fig in S1 File, “Confirmed Negative” refers to a note that, on review by investigators, did not suffer an e-scooter injury, while “Predicted Negative” refers to a note that our NLP algorithm did not predict as an e-scooter injury. Any cases of possible but not necessarily probable e-scooter injury are not included in our injury tallies, and are treated as non-injuries (negatives) in metrics of diagnostic performance in S2 Table in S1 File (as a secondary analysis, we computed and present diagnostic performance under the alternate assumption that possible cases were e-scooter injuries in S4 Table in S1 File). Counts refer to numbers of notes, rather than injuries or patients.

Fig 3. Temporal trends in e-scooter use and e-scooter injuries*.

*Availability of e-scooter trip count data for nearby jurisdictions is available for April 2019 until February 2020, and monthly injury rate is therefore only calculable over this period. Note the fall in the number of injuries after a temporary interruption in scooter availability around August 2018 and another fall in injuries during the COVID-19 lockdown in 2020. Months prior to September 2017, which had very few injuries, are not shown here.