Separating Hype from Reality in AI's Application to Cancer Care

In my ongoing research and analysis of technological advancements in the field of biotechnology, I recently came across the announcement of a particularly ambitious project. The "Stargate" initiative is a $500 billion investment in AI infrastructure that promises to revolutionise healthcare, with a particular focus on the field of oncology. To put this into perspective, the entire annual budget for the National Institutes of Health (NIH) in 2023 was approximately $47.5 billion. While such a commitment to advancing medical technology is, in principle, commendable, experience has taught me to approach such pronouncements with a degree of caution.

The ethical and societal implications of technological progress, a focus of my research, demand a careful separation of hype from demonstrable reality, thus ensuring that optimism is firmly rooted in empirical findings. The "Stargate" project’s bold claims about accelerating AI development to achieve breakthroughs in cancer diagnostics and personalised therapies raise critical questions about technological ambition, financial resources, and the current state of medical science.

The American Cancer Society estimates 2,001,140 new cancer cases and 611,720 cancer deaths in the United States for 2024 alone. Progress in cancer care is undeniably urgent. These statistics showcase the magnitude of the problem and the importance of effective diagnosis and therapy. But a critical look at this potential requires a more nuanced view. How can we ensure that such a powerful tool as AI is developed and deployed responsibly? Does the "Stargate" project represent a truly realistic pathway towards meaningful advancements, or does it fall into the paradox of overpromising and under-delivering?

Let’s dive in and examine the project's core hypotheses and timelines in light of the current state of AI in healthcare, particularly in the complex and challenging field of oncology. The core hypothesis, as announced at a recent press conference, posits that substantial capital investment, combined with the technological expertise of leading technology companies, will accelerate AI development, resulting in much-improved cancer diagnostics and personalised therapies. The proposed timelines are particularly interesting when compared to the average 10-15 years it typically takes to develop a new cancer drug, according to the Pharmaceutical Research and Manufacturers of America (PhRMA). This ambitious vision suggests that AI will soon empower doctors everywhere, including remote regions, with insights comparable to those at top medical institutions and, most remarkably, implies that personalised cancer treatments could be developed and deployed within a mere 48-hour window, or a similarly brief period.

The enthusiasm surrounding AI's role in healthcare and cancer treatment is quite understandable. The proposition that artificial intelligence will revolutionise electronic health records, thereby affording physicians in remote regions the same level of expertise as their counterparts in leading medical centres, is an attractive yet challenging prospect. However, much needs to be done to address data integration across the diverse and fragmented EHR system characteristics that currently exist in the US healthcare environment. For instance, a 2021 report by the Office of the National Coordinator for Health Information Technology (ONC) found that only about 64% of hospitals had adopted a certified EHR system, highlighting the ongoing challenges in achieving universal EHR adoption and interoperability. A study published in the Journal of the American Medical Informatics Association (JAMIA) in 2020 revealed significant variations in data quality and completeness across different EHR systems, posing substantial challenges for AI algorithms that rely on accurate and consistent data.

Accurate and representative data integration is fundamental for effective AI deployment. The precision of artificial intelligence models is significantly influenced by the quality and variety of the training datasets. Current research indicates that AI can indeed improve diagnostic accuracy in specific tasks, provided that technical, regulatory, and ethical considerations are implemented effectively during clinical adoption. For instance, studies have demonstrated that AI can detect pneumonia from chest X-rays with an accuracy comparable to that of human radiologists.

However, this is a relatively narrow application, and further clinical trials and studies are needed.

A system that uses data from various electronic health record systems to improve diagnosis and treatment may not yet be feasible until certain limitations and inadequacies are addressed. The presence of systemic biases exacerbates health disparities. For example, a study published in JAMA Oncology in 2019 highlighted that AI algorithms trained on predominantly white populations performed significantly worse in diagnosing skin cancer in individuals with darker skin tones, demonstrating the critical need for diverse and representative datasets. This raises a crucial question: Who is accountable when an AI's actions cause harm due to the biased and corrupted training practices it has been subjected to?

Several studies have indicated that African Americans have the highest death rate and shortest survival of any racial/ethnic group in the US for most cancers. In fact, the overall cancer death rate for African Americans was 14% higher than for whites in 2019, according to the American Cancer Society. Prostate cancer disproportionately affects black men, who experience a substantially higher incidence and mortality rate than white men do. Specifically, the incidence rate of prostate cancer in Black men is about 70% higher than in white men, and their mortality rate is more than double, as reported by the Centres for Disease Control and Prevention (CDC).

Although black women exhibit lower breast cancer incidence rates than Caucasian women do, their breast cancer mortality rate is 40% higher. Likewise, American Indian and Alaska Native populations exhibit elevated cancer rates compared with non-Hispanic white populations. A comparative analysis revealed that the prevalence of liver cancer is approximately twice as high among AI/AN individuals than among non-Hispanic whites. The five-year survival rate for liver cancer among AI/AN populations is only 12%, compared to 18% for whites, according to data from the National Cancer Institute. Individuals in this group exhibit elevated mortality rates of renal, gastric, and cervical cancers.

It is evident that AI shows promise in cancer detection from medical images, but it is far from replacing traditional human-centric diagnostic methods. I am particularly concerned about the research on personalised cancer therapies utilising genomic profiling, which is a field of growing complexity. The present procedure involves tumour biopsy, genomic sequencing, bioinformatics analysis, multidisciplinary tumour board review, and subsequent drug acquisition. This process typically spans weeks and months. For example, a 2017 study in the New England Journal of Medicine reported that the median time from diagnosis to the initiation of targeted therapy for patients with advanced lung cancer was 23 days, even with expedited genomic testing. The feasibility of compressing this process into a 48-hour window, or any similar timeframe in the near future, lacks scientific support.

The current technological and logistical challenges are significant. Rapid and accurate diagnosis of all cancer types across all stages is not feasible with the available technology and medical research. Comprehensive genomic and molecular profiling, along with the necessary sophisticated analysis, still require days, not hours. Designing personalised, multidrug treatment plans through AI requires an understanding of cancer biology and treatment responses that are currently beyond our understanding. As of 2023, there were only about 50 FDA-approved targeted cancer therapies, and these are effective for only a subset of cancers and patients, highlighting the limitations of current personalised treatment options.

Manufacturing or procuring personalised drug cocktails within hours is impossible at present, given the physical constraints and limitations of pharmaceutical production and supply chains. The concentration of power within a few dominant technology corporations, such as Google, Amazon, and Microsoft controlling vast computational infrastructure and access to global data, both essential for developing advanced AI systems should be discussed further. Their dominance enables them to shape markets and public policy, raising serious concerns about the erosion of national sovereignty and accountability.

The proposed $500 billion investment is substantial, but ample financial resources do not necessitate success, particularly when the projected timeframe is significantly compressed. To make breakthroughs in AI-powered cancer care, we need sustained investment in computational infrastructure, fundamental biological research, new drug discovery, and clinical trials. These processes require extensive infrastructure, including large data centres and a vast network of research institutions, hospitals, and factories.

Although the goal of using AI to improve cancer treatment is commendable, the assertions made in the "Stargate" announcement are overly optimistic based on contemporary medical science and technology. A realistic timeline for major advancements in this domain may extend to several years, requiring ongoing investment, rigorous research, and a collaborative approach involving scientists, clinicians, and technologists.

The "Stargate" project, as presented, appears to be driven more by political aspirations for a rapid technological revolution and a “quick fix approach” than by a reasonable assessment of the challenges and opportunities in the fight against cancer. Such hype or over-optimistic claims could potentially give patients with cancer and their families false hope, which is currently not based on medical evidence. Actual and realistic advancements in this field require considerable time, extensive research, and realistic expectations.