Can we trust AI not to further embed racial bias and prejudice?
BMJ 2020; 368 doi: https://doi.org/10.1136/bmj.m363 (Published 12 February 2020) Cite this as: BMJ 2020;368:m363Read all of the articles in our special issue on Racism in Medicine
All rapid responses
Rapid responses are electronic comments to the editor. They enable our users to debate issues raised in articles published on bmj.com. A rapid response is first posted online. If you need the URL (web address) of an individual response, simply click on the response headline and copy the URL from the browser window. A proportion of responses will, after editing, be published online and in the print journal as letters, which are indexed in PubMed. Rapid responses are not indexed in PubMed and they are not journal articles. The BMJ reserves the right to remove responses which are being wilfully misrepresented as published articles or when it is brought to our attention that a response spreads misinformation.
From March 2022, the word limit for rapid responses will be 600 words not including references and author details. We will no longer post responses that exceed this limit.
The word limit for letters selected from posted responses remains 300 words.
Over the last few years’ considerable discussion around the topic of Artificial Intelligence (AI) and Machine Learning (ML) have continued to capture the imagination of physicians, journalists and the lay public. Unfortunately the erroneous use of terms – as is the case with the terms AI and ML - can and has lead to more bias and confusion than clarification [1-4].
Most individuals and companies using the term AI or ML are referring to algorithms coupling a series of tests together in an effort to improve the predictability of these algorithms. Each of these tests has their independent sensitivity and specificity errors corrupting the algorithm [5].
The inclusion of these errors, based upon qualitative interpretation of testing outcomes and associated limitations is nothing more than a collection of human bias – including the bias introduced by race, sex and religion. Such testing is not TRUE AI and does introduce bias and prejudice [5,6] – in addition to more tests, time, money and potential loss of life .
TRUE AI [4] does not and cannot introduce such bias because it is free of human interpretation, error and prejudice [7]. Resistance to transitioning from biased medical tests or merely adding the term AI or ML to these tests by those seeking to profit from the use of AI/ML has appropriately lead to considerable confusion and concern [8] over the potential for AI bias and prejudice. To remove this bias, concern and prejudice, and to achieve TRUE AI we must remove the human element [10].
References:
1. Fleming RM, Fleming MR, McKusick A, Chaudhuri TK. FMTVDM©℗ Nuclear Imaging Artificial (AI) Intelligence but first we need to clarify the use of (1) Stress, (2) Rest, (3) Redistribution and (4) Quantification. Biomed J Sci & Tech Res 2018;7(2):1-4, DOI:10.26717/BJSTR. 2018.07.001489.
2. Fleming RM, Fleming MR, Dooley WC, Chaudhuri TK. From Coronary Arteriography to Stenosis Flow Reserve to FMTVDM. The Sequential Evolution of Artificial Intelligence in Cardiology and Oncology – Removing the Human Error Element. Acta Scientific Medical Sciences 2020;4(1):114-118.
3. Fleming RM, Fleming MR, Chaudhuri TK, Dooley WC. Further Qualitative Anatomic Testing Interpretation – Even by a Machine – Is not True AI and is not the answer for Women with Dense Breasts – or for Women with any Type of Breast Tissue. Sci J Womens Health Care. 2020;4(1):1-2.
4. Fleming RM, Fleming MR, Chaudhuri TK, McKusick A. Machine Learning through FMTVDM Proprietary QCA Equations. J Angiol Vasc Surg 2019;4:026. DOI:10.24966/A VS-7397/100026.
5. Fleming RM, Fleming MR, Dooley WC, Chaudhuri TK. Invited Editorial. The Importance of Differentiating Between Qualitative, Semi-Quantitative and Quantitative Imaging – Close Only Counts in Horseshoes. Eur J Nucl Med Mol Imaging. DOI:10.1007/s00259-019-04668-y. Published online 17 January 2020 https://link.springer.com/article/10.1007/s00259-019-04668-y
6. Fleming RM, Fleming MR, Chaudhuri TK. Coronary Artery Calcium (CAC) Scoring and Treatment Decision Making. J Cardiovasc Med Cardiol 2019;6(4):92-93. DOI:10.17352/2455-2976.000200.
7. Fleming RM, Fleming MR, Chaudhuri TK. Replacing Cardiovascular Risk Factors with True AI and Absolute Quantifiable Measurement (FMTVDM) of Coronary Artery Disease. Inter J Res Studies Med & Health Sci. 2019;4(11):11- 13. ISSN:2456-6373.
8. Fleming RM, Fleming MR, Chaudhuri TK, McKusick A, Dooley WC. Nuclear Imaging: Physician Confusion Over True Quantification and Isotope Redistribution. J Clin Cases Rep 2019;3(2):32-42.
9. Fleming RM, Fleming MR, Chaudhuri TK. The Need to Actually Measure What We’re Talking about before we Put it All Together. Int J Nuclear Med Radioactive Subs 2019;2(1):000114.
10. Fleming RM, Fleming MR, Dooley WC, Chaudhuri TK. From Coronary Arteriography to Stenosis Flow Reserve to FMTVDM. The Sequential Evolution of Artificial Intelligence in Cardiology and Oncology – Removing the Human Error Element. Acta Scientific Medical Sciences 2020;4(1):114-118.
Competing interests: FMTVDM is issued to first author.
Can we trust AI? Depends on the data.
Dear Editor
As a lay patient advocate with a past profession working in IT and now retired, I have taken up the opportunity to input to studies of AI in imaging, studies searching for connections of multi-morbidity, and others. Preparing for this work included familiarisation with the AI environment, reading papers and looking at the work of the British Computer Society who are looking at ways to reduce/remove bias in AI.
If AI uses data, then that data must be truly representative of the population it is intended to serve. A second consideration is that development and programming are tasks mostly undertaken by young (often white) males, who may not appreciate subtle considerations necessary to be fair and appropriate with regard to age, sex, ethnicity and other basic attributes of human beings. These attributes affect symptoms of disease, effectiveness of drugs and treatments, outcomes, and side effects. Past failures in AI have been due to little consideration given to these basic matters. Development teams need to be mixed in terms of ethnicity and sex to try and reduce sub-conscious bias, as was evident in the Deepmind kidney injury app, where only 6.32% of the data was about female patients. Thus the app did not work as effectively for them.
To try and reduce data bias, the British Computer Society is recommending greater involvement of women and minorities in degree and further degree courses in computing and data.
Data companies and bodies such as DATA-CAN, the cancer data hub, can contribute. DATA-CAN is hosting a Black Interns Initiative providing young Black people with the opportunity to experience a range of careers in health data science.
https://www.data-can.org.uk/news-item/meet-noni-anigbo-data-cans-first-i...
The rigorous use of equality impact assessments should be implemented in all AI work, and in the creation and programming of Apps, just as they should be used in clinical studies.
Competing interests: No competing interests