Unit 7. Conclusion#

This resource was created to help researchers understand and address technical factors that contribute to issues in health equity when conducting public health research. As we’ve seen, these technical factors are often data or method specific. Therefore, this resource focused on potential issues when using different types of data sources, different data characterization and preparation techniques, and common analytic techniques including biostatistics, Machine Learning (ML) methods, Natural Language Processing (NLP), and computer vision.

This conclusion provides final thoughts regarding the following:

  • Expectations for responsible AI in healthcare.

  • Combatting health disparities in public health research.

AI Contributions and Impacts in Healthcare#

Artificial Intelligence (AI) ‘s contributions to healthcare are well established, helping to improve medical imaging, personalized medicine, and tasks such as disease prediction, diagnosis, and treatment. While rule-based systems are still used within Electronic Health Record (EHR) systems, some of the most successful AI applications come from statistically-based machine learning models that help diagnose and guide treatment protocols for patients as well or better than a human clinician [Davenport and Kalakota, 2019]. On a broader level, AI has the ability to improve and automate existing models for transformative patient care, delivery, and practice and the potential to benefit historically underrepresented and marginalized populations with the rise of “big data.”

These success stories have helped influence a cultural shift toward evidence-based medicine and evidence-backed policy in the field of public health. Advanced AI applications are increasingly pervasive in the business and technology sectors as well as healthcare. More recently, the deployment of generative AI applications such as ChatGPT may signal the beginning of a new wave of sophisticated AI systems that can transform many industries. However, generative AI can pose practical limitations and serious risks, safety concerns, and ethical issues within the context of public health. In general, AI still has challenges to overcome in order to be responsibly integrated into the healthcare system and influential in public health policy. Ideally, deployed AI systems must meet several reasonable expectations, such as:

  • Some guaranteed levels of trustworthiness and minimal harm.

  • Ethical integration into current healthcare practices.

  • Honoring and safeguarding patient privacy and safety.

AI impacts in healthcare continues to be an ongoing conversation and concern since regulation for AI in public health is still in its infancy. Moreover, as the research community and general public have noted, when AI applications fall short of these expectations, it can manifest in perpetuating existing bias and producing unfair and inequitable outcomes. For example [Norori et al., 2021]:

  • Patient mortality risk prediction performing differently depending on a race variable, which impacts how hospitals allocate resources to patients.

  • Producing racial disparities in the context of administration of pain medication.

  • Delayed diagnosis of skin cancer melanomas for darker-skinned patients, which lead to worse health outcomes.

  • Biased clinical measures, such as pulse oximeter readings, which overestimates oxygen saturation levels in non-white patients who are then more likely to suffer worse outcomes than white patients.

Understanding Health Disparities in AI#

Mitigating health disparities as a byproduct of AI starts with identifying the source of the bias. This resource discussed health equity challenges and potential sources of bias relevant to each lesson topic and identified bias typically stemming from either human, data, or algorithmic drivers. Taking a cue from machine learning, another categorization can be viewed using the bias-variance trade-off, a way of characterizing a learning algorithm’s performance on unseen data. Categorically, the generalized error can be decomposed into one of the following:

  • Statistical Bias: In statistical terms, this pertains to scenarios where the model is oversimplified, not well-suited for the data or research question, or when there’s insufficient training data to construct a more accurate model.

  • Variance: This can depend on the sample size used to train the model or specifying an overly complex model.

  • Noise: Inherent in the problem. This can be viewed as random variability introduced during data collection and other data quality issues. In addition, systemic inequities and societal bias may also be captured in the noise. It’s important to distinguish between random noise, which affects variance, and non-random noise, which may include systemic inequities and societal bias. Health disparities in AI can arise due to all three sources of error with different mitigation strategies depending on the source of bias. Decisions regarding data preprocessing and model selection, complexity and specification, training, and evaluation can greatly affect model performance across different groups such as race, ethnicity, gender, or those with different socioeconomic backgrounds. These are all steps with potential biases that the individual researcher can directly influence. However, bias arising due to “noise,” including data quality issues and societal or systematic bias, is harder to address during analysis and is not influenced by fine-tuning the model. Therefore, it becomes increasingly important to approach bias mitigation holistically and proactively.

Mitigating Bias in AI#

Systematic bias can often appear in the data itself. Therefore, public health researchers should consider how data quality can impact downstream analysis, especially when it comes to data limitations with race, which accounts for many known health disparities [Thomasian et al., 2021]. Aggregating and using data sets with diverse demographic backgrounds is challenging and sometimes infeasible in practice. However, awareness of the limitations and potential bias within your data set allows for possible strategies that can help mitigate those biases during analysis. For example, data preparation and preprocessing steps can help adjust for more fair representation among subjects, which can help mitigate minority and underrepresented groups from having less predictive power and poorer or unfair model performance. Health data often suffers from representation imbalance for a variety of valid reasons (access, mistrust, etc.) However, efforts should continue to pursue more inclusive, large-scale, data sets in order to help meet expectations of AI in healthcare and to allow AI applications a greater chance to discover more equitable healthcare practices and policies.

Although problems related to data collection are often not within the practitioner’s control, researchers can adopt a mitigation mindset to help combat health inequities throughout their research process. Best practices for researchers may be a thorough, systematic, and thoughtful approach when performing public health research. This resource highlights details such as Health Equity Considerations and Considerations for Project Planning sections that are included in each lesson. These sections were included to help prompt intentional focus on relevant health equity issues given the lesson topic, potential bias that may arise, and questions for the practitioner to ponder when performing a specific type of analysis, using a certain method, or using a certain type of data set.

Proactive Approaches to Bolster Health Equity Science#

In addition to mitigation at the individual level, several steps at the organizational and even societal levels can be very effective such as:

  • Diversifying the research team in order to benefit from different valuable perspectives.

  • Involving community members in the project design process to allow a more inclusive and fair representation of a population under study.

  • Fostering interdisciplinary research collaborations and review panels from a variety of related disciplines such as public health, economics, public policy, and sociology.

  • Prioritizing and promoting high-impact research that advocates for policy changes to promote more health-equitable outcomes for the population.

  • Emphasizing health literacy for both individuals and policymakers to allow greater advocacy for improved health outcomes, especially for more marginalized populations.

  • Requiring clear and transparent communication from researchers and subject matter experts to help the public and policymakers understand all limitations and reasoning within the presented analysis.

  • Continual surveillance for emerging bias in AI implementations in order to help monitor health equity issues.

Bias mitigation strategies in individual analyses will likely always need to occur. However, the future of health equity science will benefit from a proactive mindset in order to enable more trustworthy AI implementations that maximize health equity and minimize harm to public health.