Deep Learning Model Uses Electronic Health Record Data to Identify Low Vision Patients

Low section of woman walking with cane on platform
Clinician referral rates for individuals needing low vision services often fall short of national guidelines, a study suggests.

Deep learning models that integrate electronic health record (EHR) data may identify patients with low vision prognosis for targeted referral for low vision rehabilitation, according to research published in the International Journal of Medical Informatics.

Researchers built and evaluated deep learning models that integrate EHR data to predict visual prognosis. They retrospectively identified patients with low vision (N=5547, 57,5% women), defined as having a best documented visual acuity (VA) worse than 20/40 on 1 or more encounters from EHR between 2009 and 2018. The team extracted ophthalmology notes and structured data available from the EHR including demographics, billing and procedure codes, medications, and exam findings. 

Among the models researchers developed, a single-modality deep learning model based on structured inputs was able to predict low vision prognosis with an area under the receiver operating curve (AUROC) of 0.80 (95% CI, 0.75-0.85) and F1 score of 0.70 (0.61-0.74). A combined deep learning model using structured inputs and named entity recognition achieved an AUROC of 0.79 (95% CI, 0.73-0.83) and F1 score of 0.63 (0.56-0.70). However, a combined deep learning model using structured inputs and free-text inputs using domain-specific word embeddings outperformed all others, achieving an AUROC of 0.82 (95% CI, 0.76-0.87) and F1 score of 0.69 (0.63-0.75).

In 40.7% (n=2258) of patients, BCVA never improved to better than 20/40 in 1 year of follow-up. Among these patients, the mean BCVA of the OS and OD were 0.63±1.23 and 0.65±1.25, respectively, at best. 

“We ultimately showed that models incorporating text represented by domain-specific word embeddings outperformed the single-modality model using structured inputs and outperformed all single- and multiple-modality models representing text with biomedical concepts extracted through NER pipelines,” according to the researchers.

Limitations of the study included the single center design and the inability of the deep learning model to decipher the meaning of certain abbreviations found across ophthalmology notes.


Gui H, Tseng B, Hu W, Wang SY. Looking for low vision:predicting visual prognosis by fusing structured and free-text data from electronic health records. Int J Med Inform. 2022;159(3):1-7 doi:10.1016/j.ijmedinf.2021.104678