Investigation of bias in an epilepsy machine learning algorithm trained on physician notes

Abstract

Racial disparities in the utilization of epilepsy surgery are well documented, but it is unknown whether a natural language processing (NLP) algorithm trained on physician notes would produce biased recommendations for epilepsy presurgical evaluations. To assess this, an NLP algorithm was trained to identify potential surgical candidates using 1097 notes from 175 epilepsy patients with a history of resective epilepsy surgery and 268 patients who achieved seizure freedom without surgery (total N = 443 patients). The model was tested on 8340 notes from 3776 patients with epilepsy whose surgical candidacy status was unknown (2029 male, 1747 female, median age = 9 years; age range = 0‐60 years). Multiple linear regression using demographic variables as covariates was used to test for correlations between patient race and surgical candidacy scores. After accounting for other demographic and socioeconomic variables, patient race, gender, and primary language did not influence surgical candidacy scores (P > .35 for all). Higher scores were given to patients >18 years old who traveled farther to receive care, and those who had a higher family income and public insurance (P < .001, .001, .001, and .01, respectively). Demographic effects on surgical candidacy scores appeared to reflect patterns in patient referrals.

Prospective validation of a machine learning model that uses provider notes to identify candidates for resective epilepsy surgery

“…Our model achieved good performance using an SVM classifier here and in earlier studies . Recent advances in deep learning and natural language processing, including the use of recurrent neural networks and embedding techniques, used here, may represent an avenue of exploration for further improvements in performance.…”

“…Our model used provider notes and achieved equal classification accuracies to those of epileptologists in a direct comparison . It generated surgical candidacy scores that were not biased by patients’ gender or race . We integrated our model with a pediatric hospital's electronic health record (EHR) and assigned surgical candidacy scores to epilepsy patients who were scheduled to visit the neurology clinic.…”

Early identification of epilepsy surgery candidates: A multicenter, machine learning study

“…Cutoffs were chosen based on prior literature. 15,24 Zip codes were used to determine how far patients traveled to receive care. Median household incomes for zip codes were obtained from the United States Census Bureau 25 and represented as a continuous variable.…”

“…Machine learning methodologies (ML) can be used to identify candidates for epilepsy surgery years before they undergo surgery. [12][13][14][15] ML algorithms are infrequently implemented into care. In pediatrics, one algorithm was fully automated including the provision of decision support to providers 13 ; however, the positive predictive value (PPV) of the alerts from this system was low at only 25%.…”

Can antiepileptic drug efficacy be studied from electronic health records? A review of current approaches

“…[12][13][14][15][16] However, this method can be incompletely sensitive, and missing data tend to be biased, with underreporting of underrepresented socioeconomic demographics. [17][18][19][20] When ethnicity is present in structured fields, Denny et al report over 90% concordance with genetic ethnicity, which can be used as a gold standard for ethnicity extraction when available. [17,21] Sholle et al augmented structured fields with simple NLP to achieve an F1 of 91%.…”