The SAS Virtual Health Artificial Intelligence (AI) Summit on Cancer Research1 was held this year to share best practices, ongoing challenges, and future opportunities for advancing cancer treatment through analytics. Innovations in applying computer vision to medical images and using machine learning (ML) to build predictive models may help clinicians assess therapeutic results more efficiently, thereby enhancing personalized approaches to cancer treatment.
AI is the application of digital devices and computers to enhance human intelligence.2 In this article, we focus on the use of AI to develop ML and deep learning (DL) models. Whereas ML is the subfield of AI using mathematical and statistical approaches to derive models from data, DL is a specific class of ML that leverages complex networks in its learning process (Figure 1).3
Four Key Themes from the Summit
1. Applying Response Evaluation Criteria for Solid Tumors (RECIST 1.1) criteria to solid tumors involves measuring the largest diameter of a tumor, but tumor volume and morphology give a more comprehensive assessment of treatment response,4,5 so there is an opportunity to improve RECIST 1.1 with AI, ML, and DL.6 AI can determine volumetric changes in the three-dimensional morphology of cancers that are not simple spherical or elliptical structures, while eliminating subjectivity and observer variability7-10 and reducing time assessing tumor response. The combination of clinical data features, such as AI-assisted interpretation of test reports and longitudinal patient level data, can train DL and ML models to improve the diagnostic accuracy of radiographic studies.11,12
AI may also be used to give objective histopathological results, as in a recent study of patients with pancreatic cancer. Digitized, segmented images of tumors were used to segment residual tumor burden after chemotherapy to aid follow-up therapy.13
2. AI and ML enable personalized medicine through disease detection and risk assessment models. For early detection of cancer, AI was used to examine biochemical parameters such as serum albumin and platelets. Data presented showed that low albumin levels were associated with higher cancer risk than high albumin.14 Elevated platelet count may also be associated with high risk for lung or colorectal cancer.15
In another model, ML was used to develop a risk score for severe and febrile neutropenia in patients receiving chemotherapy. This would aid clinicians in deciding whether to prescribe filgrastim and health care systems in constructing clinical pathways to guide use of such drugs. Point-of-care electronic medical record data were used to train and validate a variety of ML models (Figure 2). Of six ML models studied, the preferred one requires only 20 clinical features; the model offers interpretability and a low data extraction burden, addressing two common barriers to adoption.16
3. Health disparities arising from various demographic factors have been well-documented, the 111 percent higher risk of Black men dying from prostate cancer compared with White men being just one example.17 Contributing factors include socioeconomic status, access to health care and treatment, culture, genetic variants, and molecular differences in tumors.17,18 AI can integrate the impact of social determinants on cancer rates and treatment outcomes.19 These data can help address gaps in health care through policy, preventative care programs, and target clinical intervention.
Often the introduction of a new technological intervention, e.g., mammography, may inadvertently lead to disparities. Currently, Black women are 40 percent more likely to die from breast cancer than White women.20 Care must be taken so that AI is trained on unbiased datasets to correct for possible unequal access to diagnostics, so that the resulting clinical decision technology is generalizable to the entire population.21 Thus, AI requires strategies from the start to address equitable use, so that the health of the many will be improved, instead of the few.
4. AI, ML, and DL have implications on patient data privacy and equitable care delivery, concerns that must be addressed before broader adoption. Privacy is particularly relevant when considering the growing role of companies outside of the traditional health ecosystem.22 The UK has already experienced breaches of patient privacy due to lax procedures by a technology company.22 Such incidents demonstrate the urgent need for regulation and security standards, but these generally advance more slowly than the development of AI tools. Currently the European Medicines Agency is drafting guidance,23 while the FDA is creating an action plan for AI and ML.24,25 ML researchers are also working to solve these problems, such as using federated learning—a method by which models are trained without sharing private data—which could offer a privacy-preserving solution.26
AI, ML, and DL applied to high-quality datasets, which ideally would be large and from diverse groups of individuals, will be increasingly used to interpret medical imaging, automate analyses, build predictive models, transform written text into coded data, and improve population health.
Questions remain about integrating AI into health care systems. As new AI tools are developed, what are the best ways to prospectively validate models, deploy AI in the clinic, and modify RECIST criteria in clinical trials? Who is responsible for the governance of analytic models used in clinical practice, and will reimbursement policies be barriers to adopting new solutions that leverage AI to improve healthcare outcomes?
PETER YU, MD, is with the Hartford HealthCare Cancer Institute, Hartford Hospital, Hartford, CT. GEERT KAZEMIER, MD, PHD, is with the Amsterdam University Medical Centers, Vrije Universiteit Amsterdam, Department of Surgery, Cancer Center Amsterdam. IVAN BRANDSLUND, MD, DMSC, is at the University of Southern Denmark. ASBA TASNEEM, PHD, works at Project Data Sphere, Morrisville, N.C. HOLLY WIBERG, BS, is at Operations Research Center of MIT. NINA J. WESDORP, MD, is with the Amsterdam University Medical Centers, Vrije Universiteit Amsterdam, Department of Surgery, Cancer Center Amsterdam. MARK LAMBRECHT, PHD, is at the SAS Institute, Inc. HAI HU, PHD, works at the Chan Soon-Shioing Institute of Molecular Medicine at Windber. ROBERT A. WINN, MD, is at the VCU Massey Cancer Center, Richmond, Va. STEVE KEARNEY, PHARMD, is at the SAS Institute, Inc. JOOST HUISKENS, MD, PHD, is at the SAS Institute, Inc. ELDER GRANGER, MD, is at at The 5 Ps LLC in Centennial, CO.
1. SAS Health. Virtual Health Artificial Intelligence (AI) Summit on Cancer Research. 2021.
2. Chua IS, Gaziel-Yablowitz M, Korach ZT, et al. Artificial intelligence in oncology: Path to implementation. Cancer Medicine
3. Shimizu H, Nakayama KI. Artificial intelligence in oncology. Cancer Sci
4. Chun YS, Vauthey JN, Boonsirikamchai P, et al. Association of computed tomography morphologic criteria with pathologic response and survival in patients treated with bevacizumab for colorectal liver metastases. JAMA
5. Wesdorp NJ, Bolhuis K, Roor J, et al. The Prognostic Value of Total Tumor Volume Response Compared With RECIST1.1 in Patients With Initially Unresectable Colorectal Liver Metastases Undergoing Systemic Treatment. Annals of Surgery Open
6. Eisenhauer EA, Therasse P, Bogaerts J, et al. New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1). Eur J Cancer
7. Hayes SA, Pietanza MC, O'Driscoll D, et al. Comparison of CT volumetric measurement with RECIST response in patients with lung cancer. Eur J Radiol
8. Schiavon G, Ruggiero A, Bekers DJ, et al. The effect of baseline morphology and its change during treatment on the accuracy of Response Evaluation Criteria in Solid Tumours in assessment of liver metastases. Eur J Cancer
9. Rothe JH, Grieser C, Lehmkuhl L, et al. Size determination and response assessment of liver metastases with computed tomography-comparison of RECIST and volumetric algorithms. Eur J Radiol
10. Yoon SH, Kim KW, Goo JM, Kim DW, Hahn S. Observer variability in RECIST-based tumour burden measurements: a meta-analysis. Eur J Cancer
11. Arbour K, Tuan L, Rizvi H, Yala A, Hellmann M, Barzilay R. ml-RECIST: Machine learning to estimate RECIST in patients with NSCLC treated with PD-(L)1 blockade. J Clinical Oncol
12. Arbour KC, Luu AT, Luo J, et al. Deep Learning to Estimate RECIST in Patients with NSCLC Treated with PD-1 Blockade. Cancer Discov
13. Janssen B, Theijse R, van Roessel S, et al. Artificial Intelligence-Based Segmentation of Residual Tumor in Histopathology of Pancreatic Cancer after Neoadjuvant Treatment. Cancers
14. Soerensen PD, Christensen H, Gray Worsoe Laursen S, Hardahl C, Brandslund I, Madsen JS. Using artificial intelligence in a primary care setting to identify patients at risk for cancer: a risk prediction model based on routine laboratory tests. Clin Chem Lab Med
15. Bailey SE, Ukoumunne OC, Shephard EA, Hamilton W. Clinical relevance of thrombocytosis in primary care: a prospective cohort study of cancer incidence using English electronic medical records and cancer registry data. Br J Gen Pract
16. Wiberg H, Yu P, Montanaro P, et al. Prediction of Neutropenic Events in Chemotherapy Patients: A Machine Learning Approach. JCO Clin Cancer Inform
17. American Association for Cancer Research. AACR Cancer Disparities Progress Report. https://cancerprogressreport.aacr.org/disparities/
. Published 2020. Accessed October 21, 2020.
18. Huo D, Hu H, Rhie SK, et al. Comparison of Breast Cancer Molecular Features and Survival by African and European Ancestry in The Cancer Genome Atlas. JAMA Oncol
19. Freeman VL, Boylan EE, Pugach O, et al. A geographic information system-based method for estimating cancer rates in non-census defined geographical areas. Cancer Causes Control
20. DeSantis CE, Ma J, Goding Sauer A, Newman LA, Jemal A. Breast cancer statistics, 2017, racial disparity in mortality by state. CA: A Cancer Journal for Clinician
21. Peek ME, Han JH. Disparities in screening mammography. Current status, interventions and implications. J Gen Intern Med
22. Murdoch B. Privacy and artificial intelligence: challenges for protecting health information in a new era. BMC Medical Ethics
23. European Medicines Agency. Artificial intelligence in medicine regulation. https://www.ema.europa.eu/en/news/artificial-intelligence-medicine-regulation
. Published 2021. Accessed October 21, 2021.
24. Food & Drug Administration. Software as a medical device. Department of Health & Human Services. https://www.fda.gov/medical-devices/digital-health-center-excellence/software-medical-device-samd
. Published 2018. Accessed October 4, 2021.
25. Food & Drug Administration. Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan In: Services HH, ed. Bethesda MD2021.
26. Hao K. A little-known AI method can train on your health data without threatening your privacy. In. MIT Technology Review
. Cambridge MA: MIT Press; 2019.
27. Bertsimas D, Orfanoudaki A, Pawlowski C. Imputation of clinical covariates in time series. Machine Learning