Reseach Paper Example on Ethical Machine Learning

“Ethical Machine Learning in Healthcare”, by Irene Y. Chen et al. identifies the pipeline towards ethical machine learning. In their introduction, they highlight current social injustices faced in healthcare. Minority groups are the primary focus of their discussion, where “unjust differences in quality and outcomes of healthcare between groups often reflect existing societal disparities for disadvantaged groups” (1,124). Chen et al. outline the pipeline through their five steps, problem selection, data collection, outcome definition, algorithm development, and post-deployment considerations.

The first step, problem selection, identifies global health injustice, racial injustices, gender injustices, and diversity of the scientific workforce. In the global health injustice portion, the 10/90 gap, states that “the vast majority of health research dollars are spent on problems that affect a small fraction of global populations” (1, 126). Diseases prevalent in impoverished countries see little to no research in developed countries. This same issue is seen in racial injustice, where there is more funding for genetic diseases affecting white populations, and marginalizes racial minorities. This extreme shows in healthcare where “Black patients with [sickle cell disease] who seek treatment are often maligned as drug abusers” (1, 126). Shockingly, in women’s menstrual cycle, basic facts that can differentiate “normal and predictive of pathology -- remain unknown” (1, 126). Finally, there is a huge issue in interdisciplinary research in the scientific community. Underrepresented groups are often the ones who prioritize under researched topics, thus “[producing] more novel research but their innovations are taken up at lower rates” (2). As Chen et al. state, there must be a push to “diversifying the scientific workforce,” (1, 126) in order to adequately address the issues of society as a whole.

In the second step, data collection, they address the disparity in the regulations of data collection. Of which, research data must reach specific queries otherwise be lost to data noise, the “meaningless information added to data that obscures the underlying information of the data” (1, 127). The interpretation of data amongst various research groups, sees “notoriously aggressive exclusion (or inclusion)” (1, 127) for the four main data types Chen et al. address. This bias of information gathering has compounded to non-inclusive policies for the aforementioned marginalized groups, thus furthering the barriers for these underrepresented groups. To combat this, transparent data collection must be implemented in order to avoid censoring due to “errors in data collection and systemic discrimination” (1, 129). 

Chen et al. in the third step, outcome definition, identifies issues in predictive measures which “may cause clinical practitioners to allocate resources poorly” (1, 130). Within a clinical diagnosis, practitioners apply specific phrases to their clinical notes. The exclusion of which, such as when diseases manifest in patients differently, can lead to delayed care as a result.