Leveraging Artificial Intelligence/Machine Learning to Reduce Health Disparities

Cancer remains one of the most significant public health challenges in the United States. Each year, an estimated 2 million new cases are diagnosed and 600,000 deaths occur due to cancer. Despite substantial advances in screening, diagnosis, and treatment, notable disparities exist in oncology that impact both clinical outcomes and care delivery. Many factors drive these disparities, including race, socioeconomic status, geographic location, and access to health care resources. These disparities often result in delayed diagnoses, suboptimal treatment, poor outcomes, and even undertreatment of pain in certain racial and ethnic groups.

Eliminating the Pain Gap in Cancer Care

Recent innovations in artificial intelligence and machine learning (AI/ML) offer a promising solution to help mitigate these inequities. By streamlining the diagnostic process, personalizing treatment strategies, reducing clinician bias, and expanding access to care, AI/ML has the potential to transform the landscape of oncology, delivering quality care to all patients regardless of background or circumstances. In July 2024, the National Cancer Institute published a webinar as part of the Events: AI in Cancer Research series. During the session, Emma Pierson, PhD, assistant professor of population health sciences at Weill Cornell Medical College, presented a novel application of AI in health care: using an algorithmic approach to reduce unexplained pain disparities in underserved populations. Pierson explains the pain gap in osteoarthritis of the knee as an illustrative example that can be translated to cancer care.

Studies have demonstrated that health care providers can unconsciously downplay or misinterpret pain symptoms in certain patient populations, particularly people of color, women, and individuals from lower socioeconomic backgrounds. For instance, patients with osteoarthritis of the knee experience an unexplained pain gap, where medically underserved populations report more severe pain scores. This gap persists, even when controlling for disease severity and when blinding physician assessment of X-ray imaging using patient deidentification.

One plausible explanation for this pain gap involves the osteoarthritis disease severity score used in modern clinical practice. This grading system was originally described by Jonas Kellgren, MBBS, MRCP, MSc, FRCS, FRCP, and JS Lawrence, MD, MRCP, using knee X-rays from a primarily White, British, male patient population. Due to the limited diversity in the original cohort, it is conceivable that certain subtle radiographic features correlating to disease severity or pain levels in nonwhite or female patients were overlooked. This oversight may have led to inadequate pain assessment and suboptimal care for these patient populations.

To address this problem, computer scientists defined a clinical question for AI to answer: Are there overlooked physical features in the knee that explain higher pain levels in underserved patients? Traditionally, AI models correlate X-ray features with the physician-reported disease severity scores to replicate clinical judgment. However, in this approach, AI was used to correlate X-ray features directly with patient-reported pain scores, aiming to predict pain levels from X-ray images alone.

The AI algorithm successfully identified osteophytes as the most common radiographic feature associated with pain levels­—a finding consistent with the current osteoarthritis disease severity score. However, the ML model also recognized additional radiographic characteristics that more closely aligned with patient-reported pain scores. These predictions were robust and consistent across all patient demographics.

Acknowledging Limitations

Though reproducible and scalable, one downside of this algorithmic tool is that it is unable to specify which additional radiographic features it identifies when predicting pain scores. This lack of transparency can hinder clinical interoperability, as clinicians desire to trust, but verify AI/ML algorithms. The absence of detailed explanations contributes to a black box bias, where clinicians may hesitate to trust or, conversely, may overly rely on algorithmic outputs without the ability to cross-check results.

The application of AI/ML to bridge health disparities in oncology represents an encouraging development with potential to enhance diagnostic accuracy and personalize care delivery for medically underserved populations. However, limitations in AI transparency and interpretability remain a significant challenge to widespread clinical adoption. Ongoing research and refinement of these technologies are essential to enhance their reliability and address the black box concerns, ensuring that AI/ML tools are robust, high-trust, and ethically aligned with patient needs. As these technologies advance, they hold great promise to reduce disparities, facilitating more equitable health care delivery for all patients.

Nicole A. Colwell, MD, is a senior editor/medical writer for the Association of Cancer Centers (ACCC).