Anish Salvi ¹ Leo Arnal ¹ Kevin Ly ² Gabriel Ferreira ³ Sophia Y Wang ^1,4 Curtis Langlotz ^1,5 Vinit Mahajan ^1,4 Cassie Ann Ludwig ^1,4*

• ¹ School of Medicine, Stanford University, Stanford, United States
• ² Rosalind Franklin University of Medicine and Science, North Chicago, Illinois, United States
• ³ Centro Universitário de Várzea Grande, Várzea Grande, Brazil
• ⁴ Byers Eye Institute, Stanford Healthcare, Palo Alto, California, United States
• ⁵ Department of Radiology, School of Medicine, Stanford University, Stanford, California, United States

Given close relationships between ocular structure and ophthalmic disease, ocular biometry measurements (including axial length, lens thickness, anterior chamber depth, and keratometry values) may be leveraged as features in the prediction of eye diseases. However, ocular biometry measurements are often stored as PDFs rather than as structured data in electronic health records. Thus, time-consuming and laborious manual data entry is required for using biometry data as a disease predictor. Herein, we used two separate models, PaddleOCR and Gemini, to extract eye specific biometric measurements from 2,965 Lenstar, 104 IOL Master 500, and 3,616 IOL Master 700 optical biometry reports. For each patient eye, our text extraction pipeline, referred to as Ocular Biometry OCR, involves 1) cropping the report to the biometric data, 2) extracting the text via the optical character recognition model, 3) post-processing the metrics and values into key value pairs, 4) correcting erroneous angles within the pairs, 5) computing the number of errors or missing values, and 6) selecting the window specific results with fewest errors or missing values. To ensure the models’ predictions could be put into machine learning ready format, artifacts were removed from categorical text data through manual modification where necessary. Performance was evaluated by scoring PaddleOCR and Gemini results. In the absence of ground truth, higher scoring indicated greater inter-model reliability, assuming an equal value between models indicated an accurate result. The detection scores, measuring the number of valid values (i.e., not missing or erroneous), were Lenstar: 0.990, IOLM 500: 1.000, and IOLM 700: 0.998. The similarity scores, measuring the number of equal values, were Lenstar: 0.995, IOLM 500: 0.999, and IOLM 700: 0.999. The agreement scores, combining detection and similarity scores, were Lenstar: 0.985, IOLM 500: 0.999, and IOLM 700: 0.998. IOLM 500 was annotated for ground truths; in this case, higher scoring indicated greater model-to-annotator accuracy. PaddleOCR-to-Annotator achieved scores of detection: 1.000, similarity: 0.999, and agreement: 0.999. Gemini-to-Annotator achieved scores of detection: 1.000, similarity: 1.000, and agreement: 1.000. Scores range from 0 to 1. While PaddleOCR and Gemini demonstrated high agreement, PaddleOCR offered slightly better performance upon reviewing quantitative and qualitative results.