Deep learning models accurately predict glaucoma progression requiring surgery
Click Here to Manage Email Alerts
Newly developed deep learning models using text and numerical input features better predicted when glaucoma progression would lead to surgery compared with an ophthalmologist’s prediction.
“We do feel predictive models can be helpful eventually in providing clinical decision support or automatically identifying high-risk patients for clinical trials, but performance still must be improved before any deployment,” Sophia Y. Wang, MD, said at the virtual American Glaucoma Society annual meeting.
Researchers developed several deep learning models to predict glaucoma progression requiring surgery. Two models were developed, one using unstructured inputs such as text and neural word embeddings to represent clinical notes and one using structured inputs such as ICD and CPT codes. Their predictive glaucoma surgery performances and a third model that combined both unstructured and structured inputs were compared with an ophthalmologist’s predictive performance.
Researchers used ophthalmology clinical notes of patients with glaucoma at a single center from 2008 to 2020. Structured data included demographic information, diagnosis codes, prior surgeries and clinical information. In addition, words from patients’ first 120 days of notes were mapped to ophthalmology domain-specific neural word embeddings, Wang said.
Researchers used area under the curve (AUC) and F1 scores of the deep learning models as evaluation metrics.
In total, 748 patients out of 4,512 underwent glaucoma surgery. The deep learning model that incorporated both combined inputs achieved the highest AUC of 0.731 compared with an AUC of 0.697 for the unstructured input model and an AUC of 0.658 for the structured model.
“We also did have an ophthalmologist review 300 charts and their associated notes to provide an ophthalmologist’s human level prediction as to whether the patient would progress to surgery or not. You can see that all the models outperformed the ophthalmologist’s,” Wang said.
Deep learning models that incorporated text performed better than models that included only non-free-text data, she said.
Collapse
Wang SY, et al. Deep learning approaches for predicting glaucoma progression requiring surgery using electronic health records data and natural language progression. Presented at: American Glaucoma Society annual meeting; March 4-7, 2021 (virtual meeting).
Read more about
Click Here to Manage Email Alerts