When you visit an eye doctor, retinal scans can help diagnose various eye ailments, most notably diabetic retinopathy, a disease in which high blood sugar levels damage retinal blood vessels resulting in blood leakage that can impair vision and eventually lead to blindness. But these scans can provide additional more global information beyond the eye about a person's health. About 5 years ago I wrote about a research group at Google that used retinal images to assess cardiovascular (CV) health. They found that a deep neural network trained on retinal images could predict a major CV event (e.g. heart attack or stroke) within a 5 year window for 70% of the subjects, which matched the sensitivity of other CV event predictors such as those based on cholesterol levels.
More recently the Google team shifted their attention from retinal images to external eye images (Figure 1) which can be taken with a smartphone rather than special cameras needed for the retina. They asked whether these images of the eye exterior could also be used to potentially diagnose disease both in the eye as well as the rest of the body. Training a deep learning model on eye images, they found that the model could predict the presence of diabetic retinopathy (like a retinal scan), elevated HbA1c (a biomarker of diabetes), and elevated blood lipids (a biomarker of cardiovascular risk).
Encouraged by these positive results, they have extended their efforts in a new piece published in Lancet Digital Health with collaborators from UCLA. In this work, the Google AI researchers trained a deep learning system (DLS) that could detect abnormalities in a range of disease biomarkers from external eye photographs. The DLS was trained on a dataset of over 123,000 eye images from patients with diabetes with the objective of predicting common biomarkers typically used to diagnose disease in the liver, kidney, and other organ systems.
The external eye photographs were taken with a fundus camera which is normally used to take retinal pictures, but by increasing the distance to the subject, the focal plane was moved so that the external eye (front) was imaged instead of the retina (back). The hope is that eventually even a selfie could be analyzed by the DLS, but for this study, the eye pictures were taken by a high quality camera in a controlled fashion.
The biomarkers were chosen based on preliminary data exploration experiments that pinpointed the ones with the most promising prediction potential. The 9 selected biomarkers represented five different organ systems spanning the kidney (estimated glomerular filtration rate [eGFR]), urine albumin-to-creatinine ratio [ACR]), liver (albumin, aspartate aminotransferase [AST]), blood (haemoglobin, white blood cells [WBC], platelets), bone (calcium), and thyroid (thyroid stimulating hormone). Rather than predicting a quantitative value for each biomarker, the predictor classified the patients as above or below a threshold value that is indicative of possible disease.
The deep learning system (DLS) was trained on 123,130 eye images from 38,398 diabetic patients in the Los Angeles area across 11 clinic sites. The test (validation) dataset was divided into three groups denoted A, B, and C based on location. Validation set A was also from clinics in Los Angeles, whereas validation sets B and C were from patient populations in Atlanta, Georgia. The patient characteristics of B and C were considered to be significantly different from A.
The DLS was a standard convolutional neural network (CNN) which is typically used to analyze images. The prediction results were compared to a baseline logistic regression model fitted to clinicodemographic variables (e.g. age, sex, race and ethnicity, years with diabetes). Overall, the DLS was able to outperform the logistic regression model on each of the nine different biomarkers. As expected the performance was particularly good for validation set A which was the most similar to the training data. In terms of statistical significance, the DLS was better than the baseline for 7/9 biomarkers in validation dataset A with a 5-20% improvement (Figure 2).
For validation sets B and C. there were fewer points of comparison because a smaller subset of the biomarkers were measured in these populations. Regardless the DLS outperformed the baseline model for most of the biomarkers although the differences were smaller than for validation set A.
A key question is the robustness of these predictions to degradation in eye photo image quality. The researchers found that "even with low resolution images (75 pixels across, corresponding to less than 1% of the pixel count of modern smartphone cameras), the DLS still outperformed the baseline for several systemic parameters." This finding bodes well for taking the eye pictures with the smartphone in which resolution and overall quality may vary substantially.
In summary, this study is a promising first step toward using external eye photos to predict levels of systemic disease markers spanning the kidney, liver, and blood cells. The main result was the deep learning system trained on the eye photographs was able to outperform a baseline logistic regression model on nine different biomarkers. Ultimately, researchers believe that the DLS could be used to develop new non-invasive screening tools for a variety of systemic diseases. A key future objective is to improve the prediction accuracy, and perhaps a more diverse training data would help.
Figure 1. A fundus photo is a picture of the retina at the back of the eye. An external eye photo, which was used in the study described above, is a picture of the front of the eye, which can be taken with any consumer camera (link).
Figure 2. Biomarker level prediction accuracy of DLS compared to logistic regression baseline model. Percent accuracy was scored as the area under the curve (AUC) using different threshold values for the binary prediction. "Results are presented for validation sets A, B, and C. n refers to the total number of datapoints, and N refers to the number of positives. Error bars show 95% CIs [...] ACR=albumin-to-creatinine ratio. AST=aspartate aminotransferase. AUC=area under the receiver operating characteristic curve. DLS=deep learning system. eGFR=estimated glomerular filtration rate. TSH=thyroid stimulating hormone. WBC=white blood cells." The DLS prediction accuracy of various biomarkers (brown bars) was higher than the baseline model (gray bars) across the three validation datasets (Babenko et al. The Lancet Digital Health, 2023).
No comments:
Post a Comment