Artificial intelligence (AI) is revolutionizing healthcare, offering powerful tools for cancer detection and diagnosis. However, the growing reliance on AI models brings an urgent concern: what happens when these algorithms fail certain populations? Recent studies have highlighted the serious risks associated with AI bias in cancer diagnostics, particularly for underrepresented populations. The consequences can be life-altering, as misdiagnosis leads to delayed treatments, unnecessary procedures, and increased mortality rates.

Overfitting to Western-Centric Datasets

A study analyzing AI models used for melanoma detection found that the algorithms performed well for Caucasian patients but were significantly less accurate in detecting melanoma in individuals with darker skin tones. The primary issue? Training datasets overwhelmingly consisted of images from white patients, limiting the model’s effectiveness for diverse populations. This leads to delayed or missed diagnoses, worsening health outcomes for patients of African, Asian, and Hispanic descent [1].

Failure to Detect Cancer in Diverse Populations

Bias in AI-driven diagnostics extends beyond dermatology. Studies have found that machine learning models for colorectal and breast cancer can misclassify tumors in minority populations due to inadequate training data. For example, an AI tool designed to detect breast cancer was found to underperform in Black women compared to white women, potentially contributing to disparities in early detection and treatment. This discrepancy exacerbates existing healthcare inequalities, where minority groups already face delays in diagnosis and limited access to high-quality care [1].

Algorithmic Blind Spots in Colorectal and Lung Cancer Diagnosis

AI-driven diagnostics for colorectal and lung cancer have also shown significant biases due to inadequate data diversity. Machine learning models trained primarily on Western datasets have struggled to accurately classify tumors in Asian and African patients. In one case, a deep learning model for lung cancer subtyping misclassified tumors in patients from different racial backgrounds, leading to potential misdiagnoses and inappropriate treatment recommendations [2].

The Bias in AI-Based Pathology and Genomic Analysis

Histopathology and genomic sequencing are key areas where AI has the potential to enhance precision medicine. However, bias in training datasets can have serious repercussions. For example, a deep learning algorithm designed to detect lymph node metastasis in breast cancer was found to outperform human pathologists in some cases—but its effectiveness varied widely depending on the demographic distribution of the training data. When tested on diverse patient cohorts, the model’s performance declined for racial minorities [2].

Furthermore, genomic AI models trained on data from predominantly European ancestry populations have failed to capture key genetic mutations found in African and Asian populations. This has critical implications for precision oncology, as targeted treatments based on biased models may not be as effective for non-white patients [1].

Lack of Transparency: A Major Roadblock

Many AI models in oncology are developed with minimal transparency regarding dataset composition and validation processes. This lack of accountability means that biases often go undetected until they manifest in real-world clinical settings. A review of AI-driven cancer studies found that only 22% of the models provided full transparency on data sources, making it difficult to assess their reliability across different populations [2].

PAICON’s View on Fair and Inclusive AI

To mitigate these biases, researchers and healthcare professionals must take proactive steps to ensure fairness and accuracy in AI models:

• Diverse and Representative Datasets: AI models must be trained on datasets that include patients from a broad range of ethnicities, ages, and genetic backgrounds.
• Bias Auditing and Transparency: AI developers should publicly disclose dataset composition and performance metrics across different demographic groups.
• Human-AI Collaboration: AI should augment, not replace, medical professionals. Ensuring human oversight in AI-driven diagnostics can help catch potential errors and misdiagnoses.
• Regulatory Standards: Policymakers must establish guidelines requiring AI models to be tested across diverse patient populations before approval for clinical use.

At PAICON, we recognize the challenges posed by algorithmic bias in cancer diagnostics and are actively working to create more equitable solutions. Our AI-driven platform is built on a genetically and technologically diverse and harmonized cancer datalake, ensuring that our models are trained on a broad and representative dataset. By using diverse data sources, we enhance the reliability of our AI models across different populations, reducing the risk of biased outcomes.

Algorithmic bias in cancer diagnosis is a pressing issue that requires urgent attention. While AI holds immense potential to revolutionize healthcare, it must be developed responsibly. At PAICON, we are leading the way in ensuring that AI-driven cancer diagnostics serve all patients regardless of their background.

Want to learn more about how PAICON is driving fair and transparent AI in cancer care? Visit our website and explore our latest innovations in digital pathology and AI-driven diagnostics.

References

1. Dankwa-Mullan I, Weeraratne D. Artificial intelligence and machine learning technologies in cancer care: Addressing disparities, bias, and data diversity. Cancer Discov. 2022;12(6):1423-7. doi:10.1158/2159-8290.CD-22-0373.
2. Smiley A, Reategui-Rivera CM, Villarreal-Zegarra D, Escobar-Agreda S, Finkelstein J. Exploring artificial intelligence biases in predictive models for cancer diagnosis. Cancers. 2025;17(3):407. doi:10.3390/cancers17030407.