ML trained to detect signs of diseases in retinal biomarkers

A research team from Google has trained a deep learning system (DLS) to detect certain diseases by analyzing external eye photos.

A model generating predictions for an external eye photo

Their study shows that machine learning models can be leveraged to identify specific biomarkers – patterns and structures – that may indicate the presence of systemic diseases like diabetic retinopathy, macular degeneration, or elevated glycated hemoglobin, kidney, blood, liver diseases.

This discovery is particularly exciting since it was previously unknown that external eye photos could contain signals for health conditions.

Given that these photos can be taken using consumer devices like smartphones or digital cameras, these algorithms provide an accessible and cost-effective way to detect diseases.

The model

Researchers developed a deep learning system (DLS) that utilizes external eye photographs as input to predict various systemic parameters such as those related to the liver, kidney, bone or mineral, thyroid, and blood.

This DLS was trained with 123,130 images from 38,398 patients with diabetes who underwent diabetic eye screening in 11 different locations throughout Los Angeles County, California.

Evaluation was performed on nine prespecified systemic parameters using three validation sets comprising 25,510 patients with and without diabetes who underwent eye screening in three independent sites located in Los Angeles County, California, and the greater Atlanta area, Georgia.

The baseline logistic regression model used to evaluate the performance of the DLS was trained with the scikit-learn Python library.


The research proves that DLS algorithms can predict systemic biomarkers from external eye photos with greater accuracy than a baseline logistic regression model that solely relies on clinic-demographic variables (such as age and years with diabetes).

The DLS model was compared to the baseline regression model by computing the area under the receiver operator curve (AUC). The DLS model outperformed the baseline one for all but one of the nine prediction tasks.

AUC performance of DLS compared to the baseline model

In many instances, the DLS exhibits a high level of robustness and outperforms the baseline model, even when working with images that have been reduced in size to 150×150 pixels.

Conclusion, further research

The use of DLS to predict biomarkers of systemic diseases from photographs is a non-invasive and accessible way for early disease detection and treatment. This could greatly enhance healthcare accessibility for individuals living in remote or underserved regions.

Additionally, researchers are exploring the use of this imaging modality to identify other biomarkers.

It’s important to note that the research is still in its early stages and more work needs to be done before this technology can be widely adopted.

This study utilized large tabletop cameras to capture images, and hence, additional research is required to evaluate its applicability to images taken with other devices such as smartphone cameras.

Additionally, the datasets analyzed in this research mainly included patients with diabetes and lacked enough representation from a broader population.

Learn more:

Other popular posts