Increasingly, and with exponential pace, AI algorithms are finding their way into the oncology clinic.
A 65-year-old man with a remote smoking history is automatically referred to your clinic after an artificial intelligence (AI) algorithm-enhanced routine CT screening exam detects a nodule with 90% certainty of malignancy. A 45-year-old woman with metastatic breast cancer arrives at your office and her AI-integrated electronic health record (EHR) algorithm alerts you that she has a 60% probability of a major cardiac event within 5 years if given an anthracycline.
You hold an uninterrupted, computer screen–free discussion with your patient about their newly diagnosed pancreatic cancer; meanwhile, an AI-powered device processes the conversation and produces all documentation and billing information.
Only a few years ago, these imagined scenarios may have seemed far-fetched, but no longer. Increasingly, and with exponential pace, AI algorithms are finding their way into the oncology clinic. Here, I will review some of the promising ways in which these AI algorithms are poised to improve clinical cancer care for providers and patients.
AI is a field of research that began in the mid-20th century with the goal of developing machines that could perform tasks that previously only a human could do. Under this umbrella, “machine learning” was born several decades later, defined as the ability of algorithms to learn from data without explicit programming. Traditional machine learning algorithms, such as logistic regression, support vector machines, and random forest, generally take structured data inputs—variables such as age, cancer stage, smoking status, etc—and output predictions based on these features.
Recently, major progress in medical AI applications has stemmed from “deep learning,” an advanced form of machine learning that uses neural networks to ingest, synthesize, and make predictions with large quantities of raw, unstructured data. These neural networks draw on, among others, imaging, EHRs, and genomics.1 When it comes to unstructured data analysis, deep learning approaches have been found to outperform most traditional methods allowing for the modeling of data types that were not previously feasible.
Over the past decade, exponentially increasing quantities of digital patient data, the complexity of cancer care, advances in AI research, and the ubiquity of advanced computers have led to tremendous interest and opportunity for AI to make a large impact in oncology.
The past decade has seen an explosion in the number of algorithms published for a variety of applications in oncology, including over 5500 in 2020 alone (an average of 15 per day, according to a simple PubMed search). Despite this, there are relatively few applications that have found real-world clinical adoption. Some reasons for this delay include the nascent state of the field, trust issues with algorithm results, and a lack of demonstration of value proposition with clinically meaningful end points. Regarding this latter point, although there are over 340 FDA-cleared medical AI applications, there have been very few randomized controlled trials (RCT) of AI applications in oncology. Notably, this number is also rapidly increasing, with more RCTs published in 2021 than all previous years combined.
For these algorithms to truly make a difference in the clinic, they must cross the “translational gap,” which involves proving that these models can generalize to different demographic populations and treatment settings outside the setting in which they were developed.2 It also means showing clinically meaningful utility— improved resource utilization, patient quality of life, disease control, or survival—not just accuracy improvements. Finally, these algorithms need to be designed to fit into the daily oncology clinic workflow in such a way that they are user-friendly and easy to navigate. These tasks are often underappreciated during AI development, yet require a tremendous amount of attention for the potential promise of AI in cancer care to be fulfilled.
As the number of algorithms and trials to assess their use in oncology increases, there are 3 major areas of oncology practice that AI is already affecting and improving.
Tasks that are simple yet redundant are low-hanging fruit for AI solutions. In radiation oncology, for example, contouring of organs that are at risk in radiotherapy planning is a relatively straightforward task that takes hours of time to complete per week. Now, there are multiple open-source and commercial applications based on AI that provide organ at risk contouring that is as accurate as human experts and quicker. Furthermore, some institutions have begun adopting end-to-end automated radiotherapy treatment planning for certain malignancies, like prostate cancer.3
AI is also being used for automated diagnosis of malignancy on histopathology, to support pathologists’ workflow and enable focus on more complex cases.4 Additionally, AI is a tool that can address the well-documented challenge of enrolling patients with cancer on to clinical trials. AI-driven software that combs and digests an ever-ballooning patient EHR has been implemented as a clinical trial autoscreening tool and shown to improve the identification of eligible trial patients.5
Improved Screening and Diagnosis There are now several FDA-approved platforms for CT-based lung cancer nodule detection that have been found to reduce false positive and negative results, compared with expert radiologists’ results.6,7 There are AI-assisted endoscopy devices that improve adenomatous polyp detection for colonoscopy screening, with several randomized trials showing improvements in detection.8,9 Additionally, automated mammography systems are now finding their way into clinical adoption, having shown improved sensitivity over experts10 and the ability to perform well in different patient groups from around the world.11
Traditional means of cancer prognostication, such as American Joint Committee on Cancer staging and clinical risk prediction models, although powerful, have limits in their ability to predict personalized outcomes. Machine learning has enabled a class of EHR-based platforms able to risk-stratify patients more effectively and guide treatment recommendations.
For instance, 1 model predicts patients receiving chemoradiotherapy who are at high risk of hospitalization, and in a clinical trial found that, for those patients identified as high risk, more frequent routine clinic visits prevented hospitalization and emergency department visits.12 Another AI model that identified cancer patients with short life expectancy was found to increase the rate of serious illness conversations via physician nudges—cues and reminders to influence a behavior in a patient.13
While AI has seen adoption in several cancer care path scenarios, most applications continue to trickle through the development stage. Slowly, these applications are demonstrating clinical utility via prospective testing and increasingly common randomized controlled trials that may lead to adoption over the next few years.
For example, there is a tremendous amount of interest in using AI to predict treatment response. Immunotherapy response has been difficult to predict with traditional biomarkers and there has been significant progress using various machine learning algorithms to improve predictive performance. Imagingbased approaches, relying on quantitative imaging features [ie, “radiomics”],14 and genomic-based models have each demonstrated promise,15 but further validation is needed. Algorithms such as these may be able to guide and support treatment recommendations throughout the patient care path.
AI has also demonstrated the ability to “see” things that the naked eye cannot. Through improved digital pathology analysis, AI has shown the ability to instantly deduce cancer mutational status simply from traditional tissue slides, which may replace the often weeks-long turnaround time for running genetic panels on tumor specimens to allow for quicker decision-making and treatment initiation.16 Our group has shown the ability for AI to accurately detect extranodal extension in head and neck lymph nodes on computed tomography, a difficult task for humans that is extremely important in choosing a surgical vs nonoperative management approach.17
Finally, anti-cancer drug development is an area where AI is making an impact. Al can accurately predict the mechanism of action of certain cancer metabolites,18 as well as predict combinations of drugs that may work synergistically, which may help guide preclinical and clinical study design.19
The early successes of AI in accomplishing the most mundane of tasks at the level of clinicians should be celebrated as a means to free clinicians from redundant, low-skill exercises to reduce “click culture,” unchain us from our computers, and focus on what matters: patient care. A recent study found that clinicians now spend almost as much of their time using the EHR as they do providing direct patient care.20 Much of this time is spent combing through dozens, if not hundreds, of parameters, lab values, notes, and images.
Adoption of more sophisticated EHR systems has come with the unintended consequence of more onerous and time-consuming documentation. AI-driven systems are uniquely positioned to streamline these tasks. There are several ongoing efforts to develop “digital scribes” for automated documentation, which would free up significant time to focus on patient care.21
Utilizing natural language processing, there are also AI systems under development to automatically synthesize complex EHR patient oncologic histories into streamlined patient profiles, saving a significant amount of administrative work.22,23 In addition, genomic data are being produced at an increasingly faster pace, exacerbating the problem. The DeepPhe software enables automated extraction of detailed phenotype information from electronic medical records of cancer patients. The system implements advanced Natural Language Processing and knowledge engineering methods within a flexible modular architecture, and was evaluated using a manually annotated dataset of the University of Pittsburgh Medical Center breast cancer patients. The resulting platform provides critical and missing computational methods for computational phenotyping. Working in tandem with advanced analysis of high-throughput sequencing, these approaches will further accelerate the transition to precision cancer treatment. Off-loading the burden of record-keeping and charting to AI comes with significant challenges but has the potential to allow for providers to shift more focus to the direct care of patients.
In summary, we remain in the very early stages and pitfalls remain, there is tremendous potential for AI to have a transformative impact in oncology, helping make clinical cancer care more effective, efficient, and patient centric.
1. Esteva A, Robicquet A, Ramsundar B, et al. A guide to deep learning in healthcare. Nat Med. 2019;25(1):24-29. doi:10.1038/s41591- 018-0316-z
2. Kann BH, Hosny A, Aerts HJWL. Artificial intelligence for clinical oncology. Cancer Cell. 2021;39(7):916-927. doi:10.1016/j. ccell.2021.04.002
3. McIntosh C, Conroy L, Tjong MC, et al. Clinical integration of machine learning for curative-intent radiation treatment of patients with prostate cancer. Nat Med. 2021;27(6):999-1005. doi:10.1038/ s41591-021-01359-w
4. Campanella G, Hanna MG, Geneslaw L, et al. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nat Med. 2019;25(8):1301-1309. doi:10.1038/ s41591-019-0508-1
5. Calaprice-Whitty D, Galil K, Salloum W, Zariv A, Jimenez B. Improving clinical trial participant prescreening with artificial intelligence (AI): a comparison of the results of AI-assisted vs standard methods in 3 oncology trials. Ther Innov Regul Sci. 2020;54(1):69-74. doi:10.1177/2168479018815454
6. Venkadesh KV, Setio AAA, Schreuder A, et al. Deep learning for malignancy risk estimation of pulmonary nodules detected at lowdose screening CT. Radiology. 2021;300(2):438-447. doi:10.1148/ radiol.2021204433
7. Ardila D, Kiraly AP, Bharadwaj S, et al. End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography. Nat Med. 2019;25(6):954-961. doi:10.1038/s41591-019-0447-x
8. Wu L, Shang R, Sharma P, et al. Effect of a deep learning-based system on the miss rate of gastric neoplasms during upper gastrointestinal endoscopy: a single-centre, tandem, randomised controlled trial. Lancet Gastroenterol Hepatol. 2021;6(9):700-708. doi:10.1016/ S2468-1253(21)00216-8
9. Kamba S, Tamai N, Saitoh I, et al. Reducing adenoma miss rate of colonoscopy assisted by artificial intelligence: a multicenter randomized controlled trial. J Gastroenterol. 2021;56(8):746-757. doi:10.1007/s00535-021-01808-w
10. Lotter W, Diab AR, Haslam B, et al. Robust breast cancer detection in mammography and digital breast tomosynthesis using an annotation-efficient deep learning approach. Nat Med. 2021;27(2):244-249. doi:10.1038/s41591-020-01174-9
11. McKinney SM, Sieniek M, Godbole V, et al. International evaluation of an AI system for breast cancer screening. Nature. 2020;577(7788):89-94. doi:10.1038/s41586-019-1799-6
12. Hong JC, Eclov NCW, Dalal NH, et al. System for high-intensity evaluation during radiation therapy(SHIELD-RT): a prospective randomized study of machine learning–directed clinical evaluations during radiation and chemoradiation. J Clin Oncol. 2020;38(31):3652- 3661. doi:10.1200/JCO.20.01688
13. Manz CR, Parikh RB, Small DS, et al. Effect of integrating machine learning mortality estimates with behavioral nudges to clinicians on serious illness conversations among patients with cancer: a stepped-wedge cluster randomized clinical trial. JAMA Oncol. 2020;6(12):e204759. doi:10.1001/jamaoncol.2020.4759
14. Sun R, Limkin EJ, Vakalopoulou M, et al. A radiomics approach to assess tumour-infiltrating CD8 cells and response to anti-PD-1 or anti-PD-L1 immunotherapy: an imaging biomarker, retrospective multicohort study. Lancet Oncol. 2018;19(9): 1180-1191. doi:10.1016/ S1470-2045(18)30413-3
15. Chowell D, Yoo SK, Valero C, et al. Improved prediction of immune checkpoint blockade efficacy across multiple cancer types. Nat Biotechnol. doi:10.1038/s41587-021-01070-8
16. Coudray N, Ocampo PS, Sakellaropoulos T, et al. Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning. Nat Med. 2018;24(10):1559-1567. doi:10.1038/s41591-018-0177-5
17. Kann BH, Hicks DF, Payabvash S, et al. Multi-institutional validation of deep learning for pretreatment identification of extranodal extension in head and neck squamous cell carcinoma. J Clin Oncol. 2020;38(12):1304-1311. doi:10.1200/JCO.19.02031
18. Madhukar NS, Khade PK, Huang L, et al. A Bayesian machine learning approach for drug target identification using diverse data types. Nat Commun. 2019;10(1):5221. doi:10.1038/s41467-019-12928-6
19. Gayvert KM, Aly O, Platt J, Bosenberg MW, Stern DF, Elemento O. A computational approach for identifying synergistic drug combinations. PLoS Comput Biol. 2017;13(1):e1005308. doi:10.1371/ journal.pcbi.1005308
20. Melnick ER, Fong A, Nath B, et al. Analysis of electronic health record use and clinical productivity and their association with physician turnover. JAMA Netw Open. 2021;4(10):e2128790. doi:10.1001/ jamanetworkopen.2021.28790
21. Quiroz JC, Laranjo L, Kocaballi AB, Berkovsky S, Rezazadegan D, Coiera E. Challenges of developing a digital scribe to reduce clinical documentation burden. NPJ Digit Med. 2019;2:114. doi:10.1038/ s41746-019-0190-1
22. Savova GK, Tseytlin E, Finan S, et al. DeepPhe: a natural language processing system for extracting cancer phenotypes from clinical records. Cancer Res. 2017;77(21):e115-e118. doi:10.1158/0008- 5472.CAN-17-0615
23. Morin O, Vallières M, Braunstein S, et al. An artificial intelligence framework integrating longitudinal electronic health records with real-world data enables continuous pan-cancer prognostication. Nat Cancer. 2021;2(7):709-722. doi:10.1038/s43018-021-00236-2