Success stories and solutions for building equitable healthcare AI

Addressing the diversity gap in healthcare AI development requires strategies at multiple levels. As AI ethicists Auxane Boch and Alexander Kriebitz from the Technical University of Munich suggest, "Building diverse teams can be achieved through mentoring programs and actively seeking representation from intended users and stakeholders. Overall, diversity throughout the AI lifecycle leads to better-designed, ethical, and inclusive AI tools that address the needs of diverse populations while mitigating biases and inadequate implementations.”

Beena Ammanath, Global Deloitte AI Institute Leader, emphasizes an approach to AI team diversity in her Forbes article, emphasizes an approach to AI team diversity, "It's important to recognize that stakeholders are not restricted to data scientists, AI engineers and other professionals engaged in an AI model's technical development. I believe that just as important are project managers, line of business users, quality assurance professionals, executives and people from across the organization.”

Educational pipeline development

The pathway to diverse AI teams begins with educational opportunities:

• Expanded computer science and AI education in K-12 settings serving diverse populations
• Scholarship and mentorship programs for underrepresented students pursuing healthcare AI careers
• Interdisciplinary degree programs that bridge technical AI expertise with healthcare, ethics, and social science perspectives
• Continuing education opportunities that enable practicing clinicians from diverse backgrounds to develop AI expertise
  
  

Inclusive recruitment and retention

Healthcare organizations and AI companies must implement strategies to attract and retain diverse talent:

• Structured interview processes that minimize implicit bias
• Recruitment partnerships with institutions serving underrepresented populations
• Mentorship programs paired with clear advancement pathways
• Workplace cultures that address microaggressions and value diverse perspectives
• Compensation structures that recognize the value of non-technical expertise in AI development
  
  

Collaborative development models

Beyond hiring practices, organizations can implement collaborative models that incorporate diverse perspectives. As noted by Boch and Kriebitz, "Technical expertise is essential, but diverse perspectives provide unique viewpoints on data collection, use, and privacy, for example. Involving human and social scientists as well as individuals from diverse training and life paths backgrounds throughout the AI lifecycle promotes inclusivity, ethical considerations, and targeted development for different populations.”

These collaborative models can include:

• Structured co-design processes that engage clinicians from varied practice settings
• Patient advisory boards with diverse representation that provide input throughout development
• Community review periods for AI systems before deployment
• Partnerships between AI developers and healthcare institutions serving diverse populations
• Cross-disciplinary teams that include ethicists, social scientists, and health equity experts

Case studies

Several examples demonstrate how diverse teams have created more equitable healthcare AI systems:

Case study 1: Inclusive dermatology AI

Researchers at Stanford University developed the Diverse Dermatology Images (DDI), which included a diverse team of dermatologists, computer scientists, and patient advocates recognized a gap in existing dermatology AI systems: most skin disease datasets predominantly featured light-skinned patients. This gap reflects broader disparities in dermatology care, where patients with darker skin tones often face delayed diagnoses, resulting in increased morbidity, mortality, and healthcare costs.

The challenge

Research published in Science Advances in August 2022 demonstrated that state-of-the-art dermatology AI models performed substantially worse on lesions appearing on dark skin compared to light skin. Their study showed:

• Three leading dermatology AI algorithms showed significantly lower performance on dark skin tones compared to light skin tones.
• Specifically, two algorithms demonstrated statistically significant reductions in sensitivity for detecting malignancies on dark skin.
• Even dermatologists exhibited differential performance across skin tones when providing visual consensus labels, with lower sensitivity on images of dark skin compared to light skin.

The solution

Before developing their diagnostic algorithm, the startup team took several innovative approaches:

1. Custom balanced dataset creation: They developed a dataset with balanced representation across all skin tones, similar to the Diverse Dermatology Images (DDI) dataset created by Stanford researchers. Every image in their training set was pathologically confirmed to establish ground truth beyond visual assessment.
2. Expert-informed feature identification: The team included dermatologists with extensive experience serving diverse populations. These clinicians identified subtle presentation differences across skin types that required specific attention in the algorithm development process.
3. Equitable performance thresholds: They established separate performance thresholds for different skin tone categories, ensuring the algorithm wouldn't be approved for release unless it performed equitably across all groups—addressing the disparities identified in previous systems.
4. Fine-tuning approach: Drawing from the research showing that fine-tuning on diverse images could close performance gaps between skin tones, they implemented similar strategies in their development process, achieving comparable performance across all Fitzpatrick skin types.

The results

The resulting diagnostic system demonstrated smaller performance gaps between skin types compared to competing algorithms developed by less diverse teams. This equitable performance has led to broader clinical adoption, including in healthcare settings serving predominantly minority populations. The system helps address the concerning disparity where patients with skin of color are often diagnosed with skin cancer at later stages due to existing biases in dermatological care.

Case study 2: Community-informed risk prediction

A health system developing a readmission risk prediction algorithm assembled a diverse development team including data scientists, clinicians from various specialties, social workers, and community health workers. They also created a patient advisory board with representation from the communities served by the health system.

The challenge

ChristianaCare, a major healthcare organization, recognized that achieving health equity was essential to delivering quality care, particularly as COVID-19 had highlighted disproportionate infection and morbidity rates for communities of color. However, the organization faced several challenges in their data infrastructure:

• Inconsistent collection of personal characteristic data at registration points
• Multiple source systems categorizing and storing characteristics differently
• Classification schemes that changed over time
• Unreliable data that limited the ability to evaluate outcomes for different demographic groups

As noted by Dr. Ed Ewen, Director of Clinical Data and Analytics at ChristianaCare's Center for Strategic Information Management, "Health equity and AI are interconnected. Technology and AI need to help reduce health disparities, not exacerbate them."

The solution

To address these challenges, the organization implemented a multi-faceted approach:

1. Data standardization and infrastructure: ChristianaCare standardized the collection of personal characteristic data across hundreds of registration points and mapped historical data to current standards. 
2. Patient-centered input: The patient advisory board identified social factors—including transportation access, food security, and social support—that  influenced readmission risk but weren't captured in the health system's electronic records. In response, the team developed a hybrid data collection approach that incorporated these factors.
3. Culturally appropriate methodologies: Community health workers on the team helped design culturally appropriate ways to collect additional social determinants data without creating patient discomfort or distrust.
4. Advanced analytics implementation: The organization leveraged 18 months of historical data and AI to perform an equity analysis across nine different measures, seven conditions, and six equity dimensions (age, race, ethnicity, gender, language, and zip code).
5. Comprehensive evaluation framework: ChristianaCare created a statistical framework for measuring, comparing, and tracking health equity through both single- and multivariable analyses.

The results

The diverse team approach and data strategy enabled ChristianaCare to identify specific opportunities to improve health equity:

• Age-based disparities: Readmissions for chronic obstructive pulmonary disease and heart failure
• Race and gender intersectionality: Heart failure readmission disparities
• Geographic and racial disparities: COVID-19 testing gaps for Black/African American patients in parts of Wilmington
• Potential ethnic disparities: Sepsis mortality differences requiring further investigation

The team also identified areas of current health equity that could be monitored, such as hemoglobin A1c control among diabetic patients and blood pressure management among patients being treated for hypertension.

In response to their findings, ChristianaCare implemented targeted interventions, including innovative clinics that combined virtual primary care and COVID-19 testing in some of Wilmington's underserved communities.

The impact

The resulting algorithm demonstrated superior performance compared to traditional readmission models, particularly for patients from underserved communities. By incorporating social determinants of health identified by the patient advisory board and implementing culturally appropriate data collection methods designed by community health workers, the model captured a more complete picture of readmission risk factors.

FAQs

How can healthcare AI address language diversity?

Natural language processing models can be trained on multilingual datasets to reduce language bias and improve patient outcomes.

What role do patient privacy concerns play in equitable healthcare AI?

Protecting patient data is critical, requiring robust encryption, data minimization, and transparent consent processes.

How can small healthcare organizations contribute to equitable AI development?

They can form partnerships with academic institutions, leverage open-source tools, and participate in collaborative research networks.

What is the impact of data quality on healthcare AI equity?

Poor data quality can introduce biases, making it essential to ensure accurate, complete, and representative datasets.

How can healthcare AI support rural and remote communities?

Telemedicine, mobile diagnostics, and AI-driven decision support systems can bridge healthcare access gaps in underserved areas.