Page 197 - Computational Retinal Image Analysis

192 CHAPTER 10 Statistics in ophthalmology

7 Biomarkers

A term that has become very common within ophthalmic and vision research is bio-
marker. This is quite a broad term however as is true of disease it means different
things to different people. While there are precise definitions within the literature,
they overlap considerably and it is probable that many who use the term have not
applied its strict definition. Perhaps the most commonly adopted definition is “any
substance or biological structure that can be measured in the human body and may
influence, explain or predict the incidence of outcome of a disease” [42]. Other defi-
nitions require that a biomarker must be able to be measured with certain accuracy
and reproducibility but that measurement may be made using molecular, biochemi-
cal and cytogenic techniques but also by modern imaging methods such as corneal
topography and optical coherence [43]. One of the reasons that biomarkers have
become so prevalent in medical research is the potential they offer to accelerate the
research process from bench to bedside and back to the bench.
Historically clinical trials have relied upon clinical endpoints—endpoints that
capture the patient’s clinical state. Initially these were largely determined by the
clinician—for example, intraocular pressure measurement in glaucoma studies but
over time the importance of relating the outcome to the patient was emphasized so that
outcomes which truly reflected a patient experience (for example loss of sight) were
used. Conditions may take a long time however to lead to such outcomes and if trials
have to rely upon these they may be very costly and time consuming. Technology and
biomarkers offer the potential to reduce this considerably. An example in ophthalmol-
ogy is seen with the UKGTS study—repeat measurement of visual field led to more
precision in estimates and statistical significance was observed with fewer patients
[44]. Biomarkers used in this fashion are termed surrogate endpoints, however, there
are strict criteria which must be considered if a biomarker is being considered for this
purpose. To be used as a surrogate endpoint there must be solid scientific evidence that
a biomarker consistently and accurately predicts a clinical outcome. Obtaining such
evidence is, however, itself time consuming. If such evidence is obtained it is still im-
portant not to take this beyond what has been shown—for example by assuming that
other measures of pathology in a particular condition will show the same relationship
with the surrogate as that between surrogate and clinical endpoint [45].

8 Ophthalmic imaging data challenges on intersection
of statistics and machine learning

In the above text, we focussed on the statistical analysis of data as well as designs
in ophthalmology. Here we try and indicate a link between statistics, retinal image
analysis and machine learning (ML): for instance, is it worth discussing ML tech-
niques for retinal data analysis, from the point of view of statisticians? What is the
difference between ML and statistics especially in relation to retinal imaging and
where does data science fit?