Page 200 - Computational Retinal Image Analysis

9 Discussion 195

9 Discussion

Here we highlight several main points and give some further thoughts or references
to statistics in ophthalmology and vision research.
The design of study and data collection are crucial to any clinical study. This de-
termines the data quality. Sophisticated statistical methods will not make up for a bad
quality data. This is true also of machine learning and computational methods. There
is a term that is familiar to many statisticians but needs to be more widely used in big
data—the term being garbage in garbage out. Statisticians and computer scientists
are not alchemists. They cannot turn base metal into gold. So to avoid this problem the
clinical study should be carefully designed with a thoughtful consideration of the unit
of analysis, sample size, randomization, methods of measurement, consideration of
confounders and missing data possibilities. It is highly recommended, for the success
of the study, to have a statistician on board as a partner to contribute to the decision
about the study design, database construction, and coding of categorical variables.
The appropriate analytical strategies to data analysis depend crucially on the pur-
pose of the study as well as on the way the data are collected. We have seen that a
wrong choice of a data analysis method may lead to biased estimates of the studied
associations. Another crucial point is that the decision of the data analysis method
should ideally be done at the point of designing the study. Indeed, in randomized
clinical trials, which are near the top of the pyramid of the evidence, the standard is
to write a statistical analysis plan before the data is locked for the analysis [39].
The reporting of the results of the ophthalmic study is critical. Some funders and
journals impose that authors follow relevant reporting guidelines; in some journals, the
use of guidelines is not mandatory. Each report should lead to a reproducible work. We
listed some of the guidelines in the Section 5. A full up-to-date list of all the guidelines
is being created by the EQUATOR team (https://www.equator-network.org/).
There is a steady rise of advanced complex analytical methods for imaging data:
statistical, machine learning and data science. While statistical modeling methods
may be simpler in the number of parameters (than machine leaning methods) and
they are inherently interpretable; the drawback is that they commonly require a
thoughtful process to build the model and may be less accurate due the lower num-
ber of parameters (though depending on research question they may be sufficiently
accurate). On the other hand computationally complex machine learning methods
require less time and less clinical expertise to be built and may be more accurate
in e.g. diagnosis or prognosis, but they involve a very large number of parameters
(hundreds of thousand or millions) and larger number of patients data such as large
number of annotated images for training. These are the pros and cons to be weighted
for each study separately.
Alongside major investments in strengthening the measurements via imaging,
genotyping, and biomarker assessments, there is clearly a need for methodological
advances in statistics, machine learning and data science; and especially a need for
these disciplines to collaborate. Only then the great promise to accelerate progress
toward successful prevention and treatment of ophthalmic diseases will be delivered.