Summary

The incorporation of AI in healthcare is marked by notable accomplishments but is equally plagued by issues such as privacy invasion, inherent bias, and inaccuracies. Mismanagement and misuse of AI have not only exposed sensitive patient data but have also led to diagnostic and treatment errors, diminishing the public’s trust in this revolutionary technology. Despite efforts to implement shared standards, there exists a conspicuous absence of a universal framework governing AI's ethical, secure, and effective use in healthcare.We aim to present a comprehensive approach to refine the existing guidelines and regulatory frameworks, ensuring that AI is implemented responsibly and effectively.

Introduction

Background

Technology has had quite an impact on the dynamic field of healthcare, especially in the past few years. Developments such as new machinery, sensors, and wearable devices have changed the medical field for the better. However, the most recent development is that of artificial intelligence. Some examples of the use of AI within the medical field are digital pathology and algorithmic prediction tools. Digital pathology involves the computerized interpretation and analysis of pathology information such as slides and data of people. This process saves time as well as mitigates human error and cost in the clinical area [2].

More generally though, AI is concerned with taking data and creating algorithms based on it to make predictions. In healthcare, this can mean gathering images of patient tumors to predict early signs of them in others, gathering information about patients and various treatments given to predict the best treatments for future patients, or gathering data about patient demographics and their existing medical conditions to predict potential health risks for patients with similar demographics [7]. Artificial intelligence is making its mark in the medical field, but not without concerns.

While AI has its benefits such as efficiency and consistency, it also comes with drawbacks. The handling of the data itself is a concern. Due to the large amount of data required to create AI, lots of patient information must be collected and stored. Within the health field specifically, this invades patient privacy and could cause the patients harm especially if the data is leaked. Further, the datasets AI is trained on are unlikely to be representative of the people AI is used upon, so there is often bias in the results which reflect the bias in the dataset. Adding on to that, the inner workings of AI are rarely understood, yet the public treats results as if they are factual because they come from a machine. The combination of bias and factual interpretation can lead to many inaccuracies caused by AI. These drawbacks of confidentiality, bias, and inaccuracy have a great impact in the medical field where AI tools can be life-changing, so we must avoid them as best we can.

Existing Confidentiality Concerns

One important factor to consider is confidentiality. Artificial intelligence tools within healthcare require sensitive patient data to train and be effective. However, patients get little agency over the use of their data and are rarely explained the potential impact the sharing of their data might have [4]. Much of their data is spread among corporations involved in the development of the AI tool, and patients do not know how that data is being used. Patients have little confidence in technology companies and are unlikely to share their data with them directly, but most are open to sharing data with physicians. Unfortunately, physicians and hospitals share their patient data with technology companies without directly informing their patients, thus betraying their trust and violating their privacy [4].

Besides consent to sharing data, another issue regarding privacy is using the data to trace patients. Although companies claim to anonymize data, studies have found that new strategies can be used to identify individuals from health data [4]. The usage of private health data makes patients vulnerable to data breaches, even more so as technology companies decide to make datasets public to increase transparency in their development process. In 2006, a research group gathered profiles of 1,700 college Facebook users and released their anonymous data to the world for research purposes. Researchers quickly learned they could de-anonymize parts of the dataset, thus compromising students’ security, none of whom even knew their data was being collected [5]. It is already harmful to disclose patient data, but even more so if the patients are unaware of it and the data is sensitive. Given the nature of healthcare and the data necessary for work in the field, ensuring data privacy is of utmost importance if AI tools are to be used.

Existing Bias

Another factor of concern is bias. Bias in terms of artificial intelligence means errors within a system or algorithms that create unfair outcomes for various user groups, usually protected groups such as minors and people of color [2]. A large cause of this issue is unrepresentative datasets for training, vague definitions of algorithmic fairness, and bias in society [2, 5, 8].

The impact of algorithmic bias is apparent in the medical field. For example, Winterlight Labs built an AI tool to identify people with Alzheimer’s disease, but the technology ended up functioning properly only for native English speakers. Nonnative speakers’ mispronunciations and pauses were interpreted as indicators of Alzheimer’s by the model and led to incorrect diagnoses. There are also various clinical algorithms with “race corrections” justified by developers due to the analyses of data. One such algorithm is an American Heart Association algorithm that “scores individuals for heart-failure risk”, and in its process, assigns three additional points to “nonblack” patients [2]. This effectively categorizes Black patients as lower risk of death from heart failure, although this is not necessarily true [2]. The potential of false positives and false negatives in the medical field as a result of biased AI can be very dangerous and must be addressed.

Existing Inaccuracy

Related to bias is the issue of inaccuracy. The results from artificial intelligence machines are often interpreted as factual and applied as such in the medical field. This is due to the numerical nature and objective methods of AI models. However, AI is still very subjective and the quantified values do not indicate an objective truth [5].

In reality, it is the researchers who interpret the numerical output so there is potential for bias [5]. Additionally, there may be errors in the dataset itself, errors such as gaps in data, unreliable sourcing of data, and unrepresentative data [5]. These factors can also take away from the validity of AI, and in healthcare, this can cause enormous issues as these machines are being used in life-changing situations.

Another cause for inaccuracy is the lack of human supervision in the models. AI models are often trained and then put to work, there is little understanding of how the models make decisions or what factors they are giving importance to. This is more commonly known as the “black box” problem, algorithms’ inner workings are opaque to human observers [4]. This prevents healthcare professionals from intervening in the AI process and sometimes even from implementing AI properly because they are unsure of its exact features [4].

The lack of internal monitoring also raises the issue of apophenia, seeing patterns where they do not exist [5]. With such large datasets, it is common for models to mistake data correlation for causation. Leinweber demonstrated this through a study in which data mining showed a strong correlation between changes in a stock index and butter production [5]. In the health field, this could mean creating a strong connection between features such as demographics and immunity from diseases, which would prevent proper treatment from being given to patients. Inaccuracy in a field such as healthcare must be mitigated as patients’ lives are being put on the line.

Recommendations

Confidentiality

Legal regulations such as the Health Insurance Portability and Accountability Act (HIPPA) already require the protection of personally identifiable information (PII) of patients. But the nature of machine learning and other AI methods can add complexities to this task, making it more difficult to control the flow of information, and to detect potential leaks.

Ensuring confidentiality of patient information now requires control over more than direct, explicit disclosures. Additionally, AI analysis and training typically involves a much larger volume of data than many medical service providers may use day-to-day, and therefore concerns more than specific patients.

To address these challenges and maintain compliance, we recommend two main categories of policies that any developer or researcher seeking to process medical data should establish. Medical providers should then expect these policies in third parties before any collaboration.

Thorough de-identification

De-identification, put simply, is the reprocessing of data in order to remove its association with individuals’ identities.

1. Collectors of medical data must remove any and all PII from datasets before any public distribution. Especially in the field of machine learning, it is common for medical data to be made available to other teams for scientific study. This is a beneficial practice, but a shift from controlled individual access to general availability naturally carries high risk.

Caution: A simple first step is scrubbing direct identifiers such as patients names and ID numbers. However, this is often insufficient. One study on a dataset for physical activity research found that over a majority of patients could still be reidentified, despite efforts to anonymize it [3]. We address this surprising outcome below.

1. Careful study of collected data features should be performed to find correlations with patient identity. Even with direct identifiers removed, it may still be possible to deduce the patient from the remaining information. For example, the age, occupation, and hospital location can drastically narrow down the possibilities. As machine learning itself is a powerful tool for discovery of correlations, this analysis could be incorporated into the training process.

Minimal data, minimal access

The storage and access of data should be limited to what is necessary for the task at hand.

1. Collect all features with purpose, and delete unneeded data. Although interesting predictors could be found in unexpected places, data features should not be collected indiscriminately. There should be some reason to believe that a piece of information could be relevant, as every new addition brings additional risk. If it is determined that a dataset or specific feature is no longer necessary, it should be disposed of in a timely manner.
2. Access to identifiable data should be strictly limited to a specified whitelist of parties, and explicit permission must always be obtained before any transmission. For some purposes, it is infeasible to meet the previous goals of de-identification while leaving enough information. In such cases, the usual precautions regarding disclosure of medical history should apply, and the patient is entitled to control over its use.
3. Decentralized storage and training methods should be considered. One such emerging approach is federated learning, where multiple client nodes (such as a patient’s personal device, or one clinic location) collaborate on training a model, consolidating the training results, but keeping the data to themselves locally [4]. This is a shift from the centralized storage of large datasets under one authority that is common today. A decentralized paradigm means that the benefits of learning from patient data can be shared without pooling the data itself.

Caution: Implementing methods such as federated learning can be challenging. These systems remove certain assurances about the data and training quality, such as the assumption that samples are independent and identically distributed (IID) [4]. Validation and balancing algorithms can be employed to combat issues relating to accuracy and fairness. A hybrid, gradual transition could also be used to detect shortcomings.

A possible federated learning setup with a central orchestrator

Bias

It is clear to see why inherent bias in AI models is an important risk factor to address. Bias can lead to unfair and inaccurate outcomes or actions that can act as a detriment and disadvantage certain groups of people. Thus it is necessary to acknowledge that bias would be greatly prevalent in a system that does not account and engineer against bias. Therefore, it is critical for there to be some transparency between the AI development process and the consumers/patients on how the model was created, and with what precautions to mitigate bias.

Bias in Artificial Intelligence can stem from multiple places in its development cycle. Because of this, it is hard to introduce a one size fits all approach to address these issues. Especially because different AI models can differ greatly between one another.

To account for these changes, we propose two types of policies to reduce and monitor bias. The first type addresses bias during the design stages by creating clear transparency to proactively incorporate countering bias into development. The second type focuses on policies that affect AI models post-training retroactively.

Proactive Bias Policies

There are several things to consider when trying to minimize bias. The earliest time where bias can accrue during the development of an AI model is during data collection which is used to train the model. If the sample space that the AI learned from is biased, bias is all that the model will ever know. Therefore it is important to collect a wide and diverse sample of input data. By capturing a fair amount of data for all types of groups and cases, it can allow the AI model to properly identify solutions for a diverse population.

Another problem point is the training algorithm the AI is using itself. Essentially, if the AI’s learning mechanism has inherent and unintended bias, even if the data inputs are very diverse, the AI will construe the information in a skewed way which inturn produces biased results. This can be due to societal bias found in the developers or simply unregulated algorithms that do not take bias into account. Because of this, we propose the following policies which aim to create a stronger level of transparency between these stages of AI development, and users/patients.

1. Publicize data collection methods and demographics of diversity. Increased transparency can help users understand how the data was obtained and processed which can help identify potential concerns for bias. Users and the public can look at the data pool and address poor data quality and diversity. The inclusion of diversity metrics also promotes a fair and nondiscriminatory training environment for the AI model. By creating a demographic of the input data, developers can have a much better idea of where their input data is lacking, which could be a possible cause for a biased AI model. Then they can alleviate this by collecting more data to close the gap between the population sample data.

Caution: There can be an issue when developers are collecting data to fill the diversity gap for the sake of the metrics at the expense of quality data. To prevent this from happening, it may be beneficial for the data collection process to be monitored to make sure quality data is being collected in conjunction with diversity.

1. Publicize algorithmic efforts to counter bias. Similar to data collection, actually creating and using the algorithm is another point where bias can stem from. Transparency will allow users a better understanding of what the developers are doing to address this issue. Publicizing algorithm details with regards to bias during the design process promotes thinking about and the utilization of various methods to mitigate bias. With this knowledge becoming public, potential AI consumers can be careful and diligent about which AI algorithms they want to use to produce fair and just outcomes.

Reactive Bias Policies

Even after an AI model has been fully implemented and in use, there can still be some actions done to help reduce bias in a current model and also in the future. Even after a model is fully trained, its final outcome can be slightly altered by developers to skew its outcome. This is important especially when the data used to train an AI model become slightly outdated and the current data for the population itself becomes slightly skewed. By skewing the outcome of the AI model it can better match the more up to date information. It can also be important to hold companies and developers accountable retroactively when their AI model is in use and there exists a prevalent discrimination due to an oversight.

1. Create a post-processing algorithm to skew outcomes away from bias. The term post-processing is used for any algorithm that sits on top of an AI model and slightly adjusts its outcomes based on the design of the algorithm. This can be used to try to shift an already existing AI model away from biased results. If it is clear that an AI model is showing slight bias that can be corrected by one of these algorithms, it should be required by the developers to create a post-processing algorithm to utilize it.

Caution: If there is a large amount of bias, it can be unreasonable to try to accommodate it using this type of algorithm, especially if the root cause is due to poor data collection or biased training. In these cases it would be necessary to completely retrain the AI. Constantly forcing post-processing algorithms in this case is a short sighted solution that will lead to less inaccuracies.

1. Hold AI companies/distributors accountable. In the case where an AI model is unjust, it can harm or cause a mismanagement of resources toward certain groups of people (which is a real concern in the healthcare industry). With this policy, there would be a greater initiative for research potential places where bias can start from and then promptly solving the issues. It would let there be some justice to the people who have been greatly affected by the discrimination from the use of AI models.

Inaccuracy

Inaccuracy has been a provocative question in AI's domain, and many great talents are eager to prevent it. However, no matter how the model reaches perfection or how bias is controlled, the situation of having an exception out of the cluster is inevitable. In general, we can continuously evolve the algorithm by feeding accumulated data sets and improving the practice while on the fly. But such a strategy will fail in the realm of medical treatments where every flawed physical practice will cost great time and energy from patients and may even suffer from potential death.

We neither have the confidence to eliminate inaccuracy in AI nor should we neglect ongoing malfunction using simple explanation by probability. In order to fully incarnate the ability of AI in healthcare, and promote assurance from both specialists and patients, we would like to propose neutral responsibility agreements for all to follow. The agreements will be from the aspects of patients, physicians, and AI providers:

Patients:

1. Patients should be fully educated on AI. They should be aware of AI’s role, its benefits, and limitations, and especially the potential risks of incorrect results. Every patient has the right to consent or decline the use of AI in their treatment. When provided with AI suggested diagnosis, the patient has the right to request real-person assistance, and should be made aware of this.
2. A structured process should be established to gather patients' insights and experiences. Patient feedback is invaluable in refining AI applications, ensuring they are user-friendly and in meeting patients' needs. Meanwhile, patients’ data is safeguarded against unauthorized access, ensuring their privacy and confidentiality are uncompromised.

Physicians:

1. Physicians should receive comprehensive training to adeptly integrate AI insights into clinical decision-making processes. This policy will ensure AI acts as a complementary tool rather than largely replacing physicians. The physician must balance AI recommendations with their professional judgment, ensuring that technology enhances, not overrides, personalized patient care.
2. Physicians are responsible for communicating AI’s role, benefits, and potential risks transparently to the patients. Fostering an environment of trust and informed consent is important in the healthcare field. Having open communication about the AI between physician and patient will help create this environment. Without pre-warning patients, the treatments are substantially vulnerable and the patients reserve the right to ask forfeit from the practice.

AI Providers:

1. AI Providers must offer clear information about the capabilities, limitations, and updates of AI systems. This is crucial to instill confidence among users and stakeholders. The providers should be accountable for the performance of AI systems, adhering to established ethical, legal, and professional standards.
2. AI must be continuously monitored and refined to reduce inaccuracies, and biases, and enhance performance, aligning with the evolving healthcare landscape. For each unwanted medical failure, the provider must take responsibility for future refinements and provide complements for affected users.

This tripartite agreement fosters a collaborative ecosystem. Patients are empowered and protected, physicians are equipped and supported, and AI providers are accountable and responsive. Every party is both a contributor and beneficiary, ensuring that AI’s integration into healthcare is balanced, ethical, and beneficial.

One example is the company called Airdoc, which is dedicated to using AI to analyze patient’s retinal images. The retina is an important sign for detecting the early stage of disability, chronic cardiac disease, and brain malfunctions. Traditional physicians may take years of training to understand the signal from each image and seldom delve into it since changes are trivial. The AI is apposite to such a trivial problem and by collecting millions of pictures from patients’ retinas, it is able to detect as well as those who have decades of experience in the field.

What we can learn from this company is its commitment to both patients and users. Instead of asking users to take their own risk, the company encourages people to take retinal exams first and keep track of their health status. The aim of AI wasn’t to present the full procedure for patients to follow, but rather to point out implicit ailments such that the user can choose further and more professional diagnosis. The company has successfully introduced this program to many hospitals as a supplementary tool for physicians rather than dictating while earning profits.

With this lively example, we believe it’s worthwhile for all to follow the standard that will benefit all parties. This holistic, collaborative approach ensures that AI in healthcare is not just technically proficient but is ethically sound, user-friendly, and publicly accepted and trusted. We are committed to fostering this balanced ecosystem where technology, ethics, and humans coalesce seamlessly, optimizing healthcare outcomes for all.

Budget & Personnel

We propose funding a team of professionals to connect the bridge between the healthcare industry and ethical AI practices. By hiring a team of professionals to work with the medical companies and AI companies, they would be able to assess the efficacy of the AI model throughout all of its stages of development and confirm that they adhere to the recommendations that we have laid out in this proposal. The team would measure the efficacy of different AI’s to help promote AI companies to follow better practices and guide healthcare companies to understand and choose the best AI model for their and their patients’ needs.

Efficacy

To evaluate measures relating to confidentiality: A basic scan for the presence of directly identifiable data records in widely accessible databases can be conducted before and after the trial period. A more in-depth search for inferences conducted by third party analysts can also be arranged. If successful, we should see significantly fewer instances of these issues.

Inventories of datasets currently stored locally and remotely can be compared. A decrease in the footprint, and a higher percentage of datasets with clearly documented descriptions and histories would suggest that the measures were effective.

The specific baselines will depend on the quality of previous management.

Efficacy should also take into account effort to mitigate Bias. This includes proper measures taken to proactively reduce the possibility of bias by publicizing relevant information on how the AI input data are cultivated and processed. If needed, reactionary efforts to reduce bias should also be taken into account such as skewing post-processed AI data and holding developers and companies accountable.

Efficacy should also be established in correcting inaccuracies. The nature of significant potential risk caused by single flaws propels us to construct a robust gateway for releasing and using AI. We should proactively examine the practice of AI in each treatment using the standard by collaborating with hospitals and educating patients with legitimate information. By working collaboratively, we can ensure that these standards are not just formulated but are also consistently implemented across the board, ensuring the safety and effectiveness of AI applications in medical treatments.

Personnel

To satisfy the above goals, the team would require the following:

Medical AI Specialist (Lead): Provides overall leadership and ensures that AI solutions align with medical needs.

Clinical Experts: Medical doctors with expertise in digital health and physical experiences. Represent the opinions of first-hand users and in charge of evaluating the accuracy and efficacy of the final models.

Data Scientists: Analyze medical data and outputs for patterns of interest. Knowledgeable on AI algorithms and development, information theory, statistics.

Medical Data Engineers: Handle and process medical data, and develop secure software for the transfer, storage, and access of data.

Patient Advocates: Ensure patient perspectives are considered. Hold both medical and AI companies accountable for their effects on the patients.

External Auditors: Review and evaluate the project's effectiveness, as described above, as well as compliance, and ensure new actions and policies align with medical ethics and regulations.

The total estimated yearly salaries for all the roles combined is $1,947,500 (Data collected from Glassdoor).

Conclusion

Our solution is intricately designed to align with the present challenges associated with AI in healthcare. It stands as a robust response to issues of data privacy, inherent biases, and inaccuracies, aiming to transform vulnerabilities into strengths. We are confident in the potency of our approach to not only address the prevailing concerns but to also foster an environment where AI is synonymous with precision, safety, and trust.We have meticulously planned each element of our proposal to ensure that it is both comprehensive and adaptable. The multidimensional strategy encompasses rigorous standardization, enhanced ethical frameworks, and fortified data privacy protocols. Collaborating with your esteemed agency will expedite the translation of this plan into tangible, enduring benefits for the healthcare sector and the public at large.

References

[1]       S. Cruz Rivera, X. Liu, A.-W. Chan, A. K. Denniston, and M. J. Calvert, “Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension,” Nature Medicine, vol. 26, no. 9, pp. 1351–1363, Sep. 2020, doi: 10.1038/s41591-020-1037-7.

This article offers guidelines for reporting AI intervention trials and how it should be manipulated and observed in medical conditions. These guidelines promote transparency of complex AI models and aid understanding of AI design and bias risk. By generalizing these conceptions among all physicians and editors, the danger of the unsupervised learning process and the handling of input and output data will be minimized. The article stressed that this is a rapidly evolving field reiterated the limitation of the current guidelines in clinical involvement, and expressed the importance of continuously updating the guidelines.

In our proposal, we are able to introduce guidelines from this article and extract common standards in the existing industry. All the mentioned SPIRIT guidelines are the consensus from the majority of physicians and editors and have proven to be constructive in the past few years with multiple refinements.

[2]       F. McKay, B. J. Williams, G. Prestwich, D. Bansal, N. Hallowell, and D. Treanor, “The ethical challenges of artificial intelligence‐driven digital pathology,” The Journal of Pathology: Clinical Research, Feb. 2022, doi: 10.1002/cjp2.263.

This paper discusses the ethical challenges associated with AI-driven digital pathology, focusing on privacy, choice, equity, and trust. As digital pathology involves scanning, storing, and electronically sharing pathology slides, concerns about patient data privacy are significant. The authors emphasize the need for robust de-identification methods to maintain privacy. They also explore the public’s right to choose how their data is used, the equity issues related to the value of data and algorithmic bias, and the trustworthiness of AI systems, especially when commercial entities are involved. The paper suggests a continuous examination of these ethical dimensions and underscores the importance of public engagement to address these concerns comprehensively.

The article strongly supports our proposal, elaborating on the concepts in detail and extending our proposal’s applicability beyond health-related issues. It will serve as a comprehensive generalization that underscores the broader relevance and applicability of our proposal.

[3]       B. Murdoch, “Privacy and artificial intelligence: challenges for protecting health information in a new era,” BMC Medical Ethics, vol. 22, no. 1, Sep. 2021, doi: 10.1186/s12910-021-00687-3.

This article discusses new challenges in ensuring patient confidentiality and privacy raised by AI technologies, which is already an important responsibility for medical providers and researchers. In particular, it focuses on the issue of access to and control of patient data, especially with new types of partnerships and data custodians. For example, it notes concerns with access to medical data by private entities. It also touches on emerging risks of privacy breaches such as reidentification, where patient identity and information can be recovered from seemingly de-identified data, and often algorithmically. This source will help us motivate the need for newer procedures for handling patient data, justify our suggestion for more demanding standards, and discuss some major roadblocks that can be expected in addressing these issues.

[4]       E. by: P. Kairouz and H. B. McMahan, “Advances and Open Problems in Federated Learning,” Foundations and Trends® in Machine Learning, vol. 14, no. 1, 2021, doi: 10.1561/2200000083.

This article offers an introduction to federated learning, an emerging technique for machine learning that has gained significant attention in research. Federated learning is the concept of having multiple endpoint clients (such as a user’s personal device, or an organization’s network) collaborate on training a model under the guidance of a central authority, consolidating the training results but keeping the data to themselves locally, and with loose cooperation between clients. The article also discusses challenges raised by using this method.

This is one proposed paradigm shift in ML that could be effective in mitigating issues relating to confidentiality, integrity, and the overall risk of centralized handling of large sets of sensitive data, which has seen some use in real applications. But beyond this, many of the problems covered in the article can be generalized to the challenges of processing data with limited access to the complete, raw dataset, which also makes this a good basis for discussing the general considerations to make for solutions to our issue.

[5]       D. Boyd and K. Crawford, “Critical questions for Big Data,” Information, Communication & Society, vol. 15, no. 5, pp. 662–679, May 2012, doi: 10.1080/1369118x.2012.678878.

This paper discusses ethical issues regarding big data such as false claims of accuracy and misuse of private data. Often researchers assume that because the data is quantified, it is the objective truth, however this is inaccurate as the interpretation of the data defines its meaning. When researchers interpret data, they introduce bias, and because big data is not self-explanatory, this interpretation and bias is inevitable. Further, this paper discusses the permissions for the data used in creation of large datasets. Big data is often created from openly available data, so individuals are not consulted for permission and this introduces the potential for harm. This source comments on general privacy, accuracy, and bias concerns related to big data which then become concerns of AI since they make use of big data. Thus this paper is relevant to our proposal in discussing the current issues of AI in the health field.

[6]       S. Barocas and A. D. Selbst, “Big Data’s Disparate Impact,” California Law Review, vol. 104, no. 3, pp. 671–732, 2016, Available: https://www.jstor.org/stable/24758720

This source highlights the impact of biased data in terms of discrimination and discusses potential remedies along with their limits. AI models are trained on large sets of data in an effort to be largely applicable. However, training models on current societies leads to biased models and discriminatory AI because there is bias in our communities. To truly create unbiased models, the training data needs to be constantly evaluated with an unbiased perspective. Additionally, the government needs to lay clear laws relating to discrimination. These are two very difficult requirements to meet, however they would lead to progress for society as a whole. The discussion in this paper lays out various reasons for bias in AI and proposes a few solutions to it. Based on this, we can target discrimination as a reason for bias in AI in the medical field as well as propose some ways to mitigate it.

[7]       S. Hoffman, “The Emerging Hazard of AI‐Related Health Care Discrimination,” Hastings Center Report, vol. 51, no. 1, pp. 8–9, Dec. 2020, doi: 10.1002/hast.1203.

The premise of this article is to bring to light the effects of the inherent bias within AI use in healthcare systems. It mentions how the algorithms used show a clear racial preference which can deter proper treatment and care to people in need. They even mention how many researchers are raising awareness about these AIs, and the harmful consequences that could come about if they were more widely adopted. It mentions some specific output data points from some algorithms, the difference in risk evaluation based on color. The end of the article also mentions ways to combat this bias by governmental action from Congress and the FDA. Our proposal would use this article as an example of how current AI systems introduce bias into the healthcare industry creating an ineffective and unfair system. It also provides some action items to help reduce this which our proposal could draw from for reducing the bias in AI.

[8]      A. S. Tejani, T. A. Retson, L. Moy, and T. S. Cook, “Detecting Common Sources of AI Bias: Questions to Ask When Procuring an AI Solution,” Radiology, vol. 307, no. 3, pp. e230580–e230580, Mar. 2023, doi: 10.1148/RADIOL.230580.

The article talks about some of the big possible sources and reason bias can be present in an AI model algorithm. It mentions seven main sources generally involving how the data is created, organized, categorized, used in the algorithm model, and how it has been affected by social bias. It then goes into depth about why these sources can create bias within a model. For some of the examples, the article also addressed how the effect of bias can be mitigated to some degree. We can use this article as it provides specific practices that lead to the biases that lead to the issues that we want to address in our proposal. We can use this to show where some of the bias in current AI systems used in the healthcare industry originates from. By giving us some exact sources of the problem, we can address these issues directly to help reduce the effect of bias presently in AI in the medical field.