Article Text

Download PDFPDF

Operationalising fairness in medical algorithms
  1. Sonali Parbhoo1,
  2. Judy Wawira Gichoya2,
  3. Leo Anthony Celi3,4 and
  4. Miguel Ángel Armengol de la Hoz5
  5. for MIT Critical Data
  1. 1Harvard Paulson School of Engineering and Applied Sciences, Harvard University, Cambridge, Massachusetts, USA
  2. 2Department of Radiology and Imaging Sciences, Emory University School of Medicine, Atlanta, Georgia, USA
  3. 3Laboratory for Computational Physiology, Harvard-MIT Division of Health Sciences and Technology, Cambridge, Massachusetts, USA
  4. 4Division of Pulmonary Critical Care and Sleep Medicine, Beth Israel Deaconess Medical Center, Boston, Massachusetts, USA
  5. 5Big Data Department, Regional Ministry of Health of Southern Spain, Sevilla, Spain
  1. Correspondence to Dr Leo Anthony Celi; LCeli{at}

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

The world is abuzz with applications of machine learning and data science in almost every field: commerce, transportation, banking, and more recently, healthcare. Breakthroughs in these areas are a result of newly created algorithms, improved computing power and, most importantly, the availability of bigger and increasingly reliable data with which to train these algorithms. For healthcare specifically, machine learning is at the juncture of moving from the pages of conference proceedings to clinical implementation at the bedside. Yet, succeeding in this endeavour requires synthesising insights from both the algorithmic perspective as well as the healthcare domain to ensure that the unique characteristics of machine learning methods can be leveraged to maximise benefits and minimise risks.

While progress has recently been made in establishing certain guidelines or best practices for the development of machine learning models for healthcare as well as protocols for the regulation of such models, these guidelines and protocols tend to overlook important considerations such as fairness, bias and unintended disparate impact.1 2 Nevertheless, it is widely recognised in other domains that many of the machine learning models and tools may have discriminatory effect by inadvertently encoding and perpetuating societal biases.3

In this special issue, we highlight that machine learning algorithms should not be focused solely on accuracy but should be evaluated with respect to how they might impact disparities in patient outcomes. Our special issue aims to bring together the growing community of healthcare practitioners, social scientists, policymakers, engineers and computer scientists to design and discuss practical solutions to address algorithmic fairness and accountability. We invited papers that explore ways to reduce machine learning bias in healthcare or explain how to create algorithms that specifically alleviate inequalities.

To prevent artificial intelligence (AI) from encoding the disparities that exist, algorithms should predict an outcome as if the world were fair. If designed well, AI may even provide a way to audit and improve the way care is being delivered across populations. There is growing community momentum towards not just detecting bias but operationalising fairness, but this is a monumental task. Some of the encouraging developments that we have seen have been incorporating patients’ voices in AI. Patient engagement is crucial if algorithms are to truly benefit everyone.

The papers in this special issue cover a variety of topics that addressed the objectives laid out in the call, these were:

  • Identifying Undercompensated Groups Defined by Multiple Attributes in Risk Adjustment4

  • A Proposal for Developing a Platform That Evaluates Algorithmic Equity and Accuracy5

  • Can medical algorithms be fair? Three ethical quandaries and one dilemma6

  • Resampling to Address Inequities in Predictive Modeling of Suicide Deaths7

  • Evaluating algorithmic fairness in the presence of clinical guidelines: the case of atherosclerotic cardiovascular disease risk estimation8

  • Operationalizing fairness in medical AI adoption: Detection of early Alzheimer’s Disease with 2D CNN9

  • Global disparity bias in ophthalmology artificial intelligence applications10

  • Investigating for bias in healthcare algorithms: A sex stratified analysis of supervised machine learning models in liver disease prediction11

It has been more than 5 years since the ProPublica investigative report on machine bias was published. The report detailed how a software used in judicial courts across the USA to inform decisions around parole was prejudiced against black people. Everything we have achieved since then has always been geared towards understanding how difficult it is to prevent AI from perpetuating societal biases in algorithms.

There is a long road ahead before we can leverage the zettabytes of data that are routinely collected in the process of care. We should not only invest in storage and compute technologies, federated learning platforms, GPTs, GRUs and NFTs. Machine learning in healthcare is not just about predicting something for the sake of prediction. The most important task is to augment our capacity to make decisions, and that requires understanding how those decisions are made.

Ethics statements

Patient consent for publication

Ethics approval

Not applicable.



  • Twitter @judywawira, @MITCriticalData, @miguearmengol

  • Contributors Initial conceptions and design—SP, JWG, LAC and MAAdIH. Drafting of the paper—SP, JWG, LAC and MAAdIH. Critical revision of the paper for important intellectual content—SP, JWG, LAC and MAAdIH.

  • Funding LAC is funded by the National Institute of Health through the NIBIB R01 EB017205.

  • Competing interests None declared.

  • Provenance and peer review Not commissioned; internally peer reviewed.