I am a PhD candidate based in the Centre for Clinical Brain Sciences. I’m affiliated with SCONe and VAMPIRE. The big picture of my PhD is to apply artificial intelligence approaches to predicting myopia onset, myopia progression and risk of pathologic myopia.

This site functions as an open laboratory notebook to support the Open Science Initiative. I provide a super high-level overview of my PhD below. If you have at least 20 minutes, however, you may wish to read a more elaborate version here.


My research in less than 250 words…

Machine learning (ML), much like a clinician, relies on input features (X), e.g. IOP, CCT, visual fields, cup-to-disc ratio, disc size, etc., to arrive at a prediction (Y), e.g. glaucomatous or not glaucomatous. But unlike human clinicians, ML has an enormous capacity to model complex relationships between X and Y.

Nonetheless, selecting useful X may not always be possible/ straightforward. Deep learning (DL), a subset of ML, addresses this challenge by obviating the need to specify X altogether. We can just provide raw retinal images, and it will learn to extract relevant features and model the relationship between these learned features and Y.

The availability of a plethora of myopia interventions with different safety profiles makes it important to stratify patients based on future progression rate. Intervention with more/ stronger side effects is justified if a patient is believed to be a fast progressor. Current prediction methods rely on population-based estimates, e.g. centile growth curve of axial length (AL), epidemiologic models (e.g. age, baseline refractive, error, outdoor time), etc., and are therefore subject to inter-individual variability in progression rate. Prediction based on past progression rate is also not reliable.

Can we tap into the high representation power of DL to predict myopia onset and progression rate from retinal images at the individual level? Such granular approach to myopia prediction will enable personalisation of interventions. Given the prospect of prophylaxis for certain cases of pathologic myopia in future, e.g. treatment targeting CCDC102B for myopic maculopathy, I’m also interested in developing image-based models to predict future risk of pathologic myopia.