The Current and Inevitable Expansion of AI in Oncology

Generally, emailing a colleague or consulting the National Comprehensive Cancer Network guidelines would be a pathway for retrieving information relating to cancer management, but developments of an artificial intelligence (AI)–assistant device may be the new channel for getting questions answered. Millions of individuals across the world consult AI, beginning their questions with “Ok, Google,” or “Alexa,” and health care providers will be no exception. AI is making strides with its ability to process large amounts of data quickly using organizational platforms and tools that make information more accessible than ever.

“AI is everywhere and it’s almost impossible not to be involved in AI at this point,” Christopher E. Mason, PhD, said in an interview with Targeted Therapies in Oncology. Mason is a professor of physiology and biophysics in the Department of Physiology & Biophysics, Weill Cornell Medicine; a professor of computational genomics in computational biomedicine at The HRH Prince Alwaleed Bin Talal Bin Abdulaziz Alsaud Institute for Computational Biomedicine, Weill Cornell Medical College; and a professor of neuroscience at Weill Cornell Medicine Feil Family Brain & Mind Research Institute. He is also director of WorldQuant Initiative for Quantitative Prediction and WorldQuant Foundation Research Scholar, and author of The Next 500 Years: Engineering Life to Reach New Worlds, and The Age of Prediction: Algorithms, AI, and the Shifting Shadows of Risk.

AI-Powered Tools and Devices

Even now there are AI-assisted oncological tools gaining FDA premarket approval such as Tempus One, which is the first generative AI-enabled clinical assistant that provides easy access to patient data. This platform received FDA approval in mid 2023 and the company, Tempus Labs, recently released news of its collaboration with Lens Data Analytics that “will leverage advancements in generative AI to provide…tools to seamlessly analyze Tempus’ de-identified, multimodal data library to glean insights, build patient cohorts, interrogate patient populations, and more,” per a company press release.”^1,2 

“The ability to have a voice assistant augment our efforts in delivering precision oncology allows us to unlock the power of AI for direct clinical benefit for our patients,” Sandip Patel, MD, a professor in the Department of Medicine and an oncologist at UC San Diego Health, said in a news release.³ “Companies [such as] Tempus and Foundation Medicine have come out with AI-assisted tools that connect a lot of research questions into a clinical paradigm, and [a provider] anywhere can have access to all the information from the headquarters of a large oncology company,” Mason explained, who was a consultant for Tempus Labs.

Another AI-assisted oncological tool that gained FDA approval is the wireless, handheld device known as DermaSensor. The point-and-click design uses noninvasive, AI-powered spectroscopy technology to detect cellular and subcellular characteristics of lesions found in melanoma, basal cell carcinoma, and squamous cell carcinoma and gives “an immediate, objective result using an FDA-cleared algorithm.”⁴ It also has a 96% sensitivity for up to 224 different skin cancer types and is decreasing the number of skin cancers missed by half.⁵ 

Current Benefits of AI in Oncology

In areas of basic research, Mason explained that at Weill Cornell Medicine, in New York, New York, they use AI to improve DNA-sequencing data, which involves analyzing tumor samples to identify mutations and determine the best treatment based on those mutations. To achieve this, they rely on algorithms, specifically recurrent neural networks, to process the sequencing data generated by machines. For imaging data, spatial transcriptomics is used, which focuses on studying RNA molecules in tissues, with spatial referring to the location and transcriptome indicating all RNA molecules in a tissue.

For example, when a patient undergoes a tissue biopsy for a breast lump, the sample is cut into small segments for analysis. “Each piece provides billions of data points, and there [is] so much data that it’s impossible for a person to process it all,” Mason said. That’s where AI and machine learning come in to help. These technologies help to interpret the complexities within tumors, such as the tissue biopsies, which use spatial transcriptomics. “This is only made possible by AI and machine learning methods and is one case where it would be…impossible without the algorithm to fully understand what we see inside the tumors,” Mason said.

Additionally, when clinicians are analyzing brain or tumor scans, instead of only looking at one patient’s image, they can now compare it simultaneously with those from millions of other patients. This is something that would be impossible to do manually and it is relatively straightforward with the help of AI tools. “This not only allows us to examine millions of molecules per square centimeter, but also enables us to analyze millions of patients’ data simultaneously, greatly enhancing the scale and types of analyses we can perform,” Mason said.

AI Today vs Yesterday

AI and machine learning have been slowing developing over the years. “Initially,”Mason continued, “our work was limited in scope; we could only analyze hundreds or thousands of datasets at a time. However, in the last 3 to 4 years, there’s been a significant breakthrough and we’re able to handle millions or even billions of data points simultaneously.” 

Much of this recent progress is thanks to the widespread adoption of graphics processing units (GPUs), primarily pioneered by NVIDIA, but also offered by companies such as Advanced Micro Devices, Mason explained. GPUs excel at performing fast mathematical calculations, particularly floating-point operations, which are crucial for machine learning tasks.⁵ This acceleration in processing speed provided by GPUs has revolutionized the field, allowing much larger datasets to be tackled and significantly speeding up computations, Mason said.

AI’s Expertise and Future Challenges

The cancer types that will derive the most benefit from AI assistance are those dealing with imaging data or high-dimensional data, Mason explained. This is“because they necessitate a more in-depth, algorithmic approach to teasing out the sources of signal and noise.” Common cancer types are also particularly well suited for AI applications, he continued, as AI thrives on large amounts of data for training. Rare cancers pose challenges due to the limited availability of training data and mark new thresholds in the application of AI.

There are many AI tools that can separate out whether a patient is likely to respond to a certain treatment, Mason explained. By assessing extensive amounts of data to evaluate factors such as comorbidities that may affect treatment efficacy, multi-omics tools or AI tools go beyond genetic data to consider various factors such as microbiome data and pharmaceutical history to predict if a treatment is more effective based on the patient’s full profile.⁷ Many AI tools can tease this out better than traditional computational methods, he said.

The heterogeneous nature of cancers, especially lymphomas, are difficult to classify, causing individualized care to be a challenge, Ari M. Melnick, MD, explained. Melnick is the Gebroe Family Professor of Hematology/Oncology and a professor of medicine at Weill Cornell Medical College in New York, New York. As classification improves, it still falls short of biological accuracy, and understanding what is driving a particular tumor is essential for tailoring treatments, he continued. With more understanding about how tumors work and considering the complexity of human biology, synthesizing the drivers of each tumor through machine learning algorithms could streamline treatment decisions, enhancing their impact on individual patients, Melnick suggested.

Concerns Associated With AI

Complacency

One concern related to AI implementation is “how we interpret data when we’re dealing with huge amounts of data,” Melnick said. “It’s important to use statistical testing that not only shows if something is statistically significant but also if it’s biologically meaningful.” He explained that many current statistical methods can make variations seem more important than they actually are, and “if we rely too heavily on these methods without thinking critically, we might end up making wrong conclusions.”

Bias

“It could be damaging if you have so much faith in your algorithm that you don’t see where it could be wrong,” Mason said. For instance, he noted the issues with facial recognition software having more accuracy with a Caucasian individual than with a person of color due to the amount of data the technology has gained.⁸ For a long time, the algorithm for tracking organ donations had the wrong calculations for those of African American backgrounds, Mason explained.⁹ “If you think the algorithm is good, but it’s built on assumptions that might be flawed or outdated, that can often be a problem. The overconfidence challenge is something to watch out for,” he said.

Cost

The way in which cost structures are constructed in the US could also be problematic, because medical care is more expensive than it should be, Melnick explained, and other countries may be faster at implementing these advances because they have more control over cost. Mason expects that “in the near term, it might make cost a bit more expensive; however, in the long term, it should make costs less expensive, because you’ll achieve more automation in places that would be more laborious.” For instance, some of the more mundane tasks will be achieved with AI, allowing providers more efficiency and more time for treating patients, he said.

HIPPA Compliance

Some AI algorithms have become extremely accurate at tracking down individuals, Mason stated. That means it will be crucial to ensure privacy is maintained and Health Insurance Portability and Accountability Act (HIPAA) compliance is guaranteed, making implementation more complex and riskier. However, checks are in place such as congressional committees and oversight groups within each scientific discipline, and the FDA has its own review committee, Mason explained.

With the power to analyze large amounts of data––by connecting with organizational tools and devices such as Tempus One to retrieve information and incorporating various categories to predict treatment courses––the health care industry is rapidly changing. Some change has already occurred. “However, I believe that over the next 5 to 10 years, we’ll see a shift,” Mason stated.“We’ll go from a handful of devices to suddenly the majority of devices that providers use…likely [having] some degree of AI embedded into their functionality. It’s already impacted almost every facet of medicine, life, and science,” he said.

REFERENCES:

1. Tempus introduces AI-enabled assistant, Tempus One, into Lens Data Analytics platform to accelerate research. News release. Tempus. February 2, 2024. Accessed February 6, 2024. https://tinyurl.com/448mmbah

2. Tempus Receives US FDA approval for xTCDx, a NGS-based in vitro diagnostic device. News release. Tempus. May 1, 2023. Accessed February 5, 2024. https://tinyurl.com/55t2fn5v

3. Troy J. Tempus Announces Broad Launch of Tempus One. Tempus. Published June 1, 2023. Accessed February 13, 2024. https://tinyurl.com/48bpty8p

4. FDA clearance granted for first AI-powered medical device to detect all 3 common skin cancers (melanoma, basal cell carcinoma, and squamous cell carcinoma). News release. DermaSensor. January 17, 2024. Accessed January 17, 2024. https://tinyurl.com/2ry943sz

5. FDA clearance granted for first AI-powered medical device to detect all three common skin cancers (melanoma, basal cell carcinoma, and squamous cell carcinoma). News release. DermaSensor Inc. January 17, 2024. Accessed January 17, 2024. http://tinyurl.com/jye7euub 

6. Pandey M, Fernandez M, Gentile F, et al. The transformational role of GPU computing and deep learning in drug discovery. Nat Mach Intell. 2022;4:211-221. doi:10.1038/s42256-022-00463-x

7. Subramanian I, Verma S, Kumar S, Jere A, Anamika K. Multi-omics data integration, interpretation, and its application. Bioinform Biol Insights. 2020;14:1177932219899051. doi:10.1177/1177932219899051

8. Gentzel M. Biased face recognition technology used by government: a problem for liberal democracy. Philos Technol. 2021;34(4):1639-1663. doi:10.1007/s13347-021-00478-z

9. Lebret A. Allocating organs through algorithms and equitable access to transplantation–a European human rights law approach. J Law Biosci. 2023;10(1):lsad004. doi:10.1093/jlb/lsad004