IEEE Technology and Society Magazine - March 2021 - 62

AI for a Fair, Just, and Equitable World

* Low number of ophthalmologists: Some estimations report 42.5 ophthalmologists per million of
people (OPM) [60] in Mexico, in contrast with
other countries such as Spain with 105.5 OPM or
Argentina 103.6 OPM, Brazil 67.4 OPM.
* Lack of national public programs for DR screening: Although existing national guidelines about
DR and isolated efforts that demonstrate the
positive impact of the early DR detection [61], it
does not exist a national screening program in
Mexico.
In this context, the Jalisco government, IADB,
and the ITESM will implement a three-year project
to harness a DR screening program aided with AI.
We expect that the AI models, coupled with specialized equipment, public policies, help to the ophthalmologists to detect DR patients in earlier stages and
reduce the adverse effects of DR disease in Jalisco.
To obtain more detail about the clinical application
of this project, see [62].
In this section, we present the particular ethical
considerations for this use case, and some preliminary DL models results to refer DR patients to a
specific health care level hospital, according to the
Mexican DR guideline.

Ethical considerations
In this use case, we take into account some
recommendations reported at [63]-[65]. The most
important considerations for this use case are presented as follows:
* General project ethical considerations: We have
taken into account as much as possible the seven
main principles proposed in [66] for the whole
project: 1) social and clinical values; 2) scientific validity; 3) fair subject selection; 4) favorable risk-benefit ratio; 5) independent review; 6)
informed consent; and 7) respect for potential
enrolled subjects.
* Privacy by design: We set the control, limits, and
conditions of the use of data developing a web
application for this project with a clear data
pipeline, restricted data access, and disclosures
based on HIPAA privacy rules. As a result, we
ensure the appropriate safeguards to protect
the privacy of personal health information and
work only with the information with patient
authorization.

* Verifiability and replicability: In this use case, we
propose to create an anonymous open data set
for research and other government implementations. Besides, all the code and the model
weights will be shared via Gitlab.
* Professional responsibility: To tackle this issue,
we collaborate with an ophthalmologist committee and with a local retina association to
show and discuss the advances of the project.
Moreover, we are working to achieve at least a
sensitivity of 80% and specificity greater than
95% for the classification model.
* Human oversight: For all the use cases, we
aim to take into account the preservation of
the jobs. The proposed system only gives an
opinion, and this can be reviewed, verified,
and changed by a professional (remedy for
automatic classification). This action helps us
to clarify that, in this project, workstations are
maintained, and technology is used as a tool
for decision support.
* Prevention of bias using representative and
high-quality data. In this case, we try to follow
the recommendations proposed at [67]. In this
case, data concerning ethnicity, gender, or
income was explicitly not included. First, we
draft the Research Ethics Board (REB) review.
Only afterward, the data were collected in a
second stage. The third step is data exploration
and data cleaning only to maintain RD images.
In the fourth step, the patients' identity is preserved by assigning IDs to each of them, guaranteeing their privacy. The quality evaluation of
the images was made using contrast, noise, and
blurring metrics. Besides, we selected the lowand high-quality images visually and trained a
CNN to classify the quality of the images, and
finally, an expert labeled the images.

Background and data
According to [57], there are five stages of DR
(from R0 to R4): no apparent DR (R0), mild nonproliferative DR (R1), moderate nonproliferative
DR (R2), severe nonproliferative DR (R3), and
proliferative DR (R4). Figure 1 presents an example of each DR image stage. The Mexican health
institutions adapted the ICO guidelines to refer DR
patients: the first level for DR control, DR assessment and essential treatments in the second level,

IEEE Technology and Society Magazine

https://www.hhs.gov/hipaa/for-professionals/privacy/index.html

IEEE Technology and Society Magazine - March 2021

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - March 2021