Artificial Intelligence Proves Its Worth in Cancer Treatment, Drug Development, and Other Real-World Applications

Artificial Intelligence
in Cancer Treatment
| Image Credit: © Crystal light
– stock.adobe.com

As artificial intelligence (AI) seeps more and more into the daily lives of professionals in every industry, cancer care has proven to be no different. In fact, AI-based approaches are already having a dramatic effect on multiple areas of oncology, including treatment selection, the lowering of administrative burden, and drug development.

“One of the most timely and pressing [concerns] in modern oncology is how AI is shaping the clinical care environment, and whether it is truly improving efficiency, communication, and outcomes,” Enrique Velazquez Villarreal, MD, PhD, MPH, MS, assistant professor, Department of Integrative Translational Sciences at City of Hope in Duarte, California, said in an interview with OncLive®. “Real-world applications of AI are already entering the clinic [in forms such as] ambient listening tools, patient messaging chat bots, and inbox management systems, all of which aim to reduce administrative burden and enhance patient-[clinician] communication. [We are] at the point when we are starting to use the available models, [which] is the first step in clinicians exploring the [existing] options.”

AI Is Already Making a Splash in the Clinic

AI-based approaches, including natural language processing and machine learning, are already having a dramatic effect on the ways that patients with cancer can be diagnosed and treated, as well as showing potential in drug development.

One example is a machine learning model developed by a team of investigators from the European Society for Blood and Marrow Transplantation (EBMT) which can predict overall survival for patients with myelofibrosis following allogeneic hematopoietic cell transplantation (allo-HCT).1 The model accounts for patient age,comorbidity index, performance status, blood blasts, hemoglobin, leukocytes, platelets, donor type, conditioning intensity, and graft-vs-host disease prophylaxis.2 In March 2025, the open-access tool, which features an interactive web-based calculator that visualizes the risk score based on inputs of the 10 variables, was launched online and it is free to use.1

Data from 5,183 patients across 288 EBMT-registered centers were used to train the machine learning model.2 Findings from a study published in Blood demonstrated that the model identified 25% of patients as being at high risk for a poor outcome with allo-HCT, defined as 1-year non-relapse and overall mortality rates of approximately 35% and 40%, respectively. Comparatively, traditional methods using Cox-derived and Center for International Blood and Marrow Transplant Research scores assigned patients to the high-risk group at respective rates of 10.1% and 8.2%.

“We currently have good [transplant risk projection] scoring systems, but you [also] have to consider the toxicity and the results of transplant, and we don’t have very good tools at this time [for that],” Juan Carlos Hernández-Boluda, MD, PhD, associated physician, Hematology and Medical Oncology Service, Hospital Clínico of Valencia, Spain said in an interview with OncLive. “With this [machine learning] tool, we will have information on the disease and transplantation risk, and we can share this information with the patient. We also have a web application that is visual, and we can share this image with the patient if they are interested. [Using this tool], we can make a more informed decision, which is critical because transplantation is a curative treatment, but also it poses a lot of danger [in terms of] the life of the patient. Combining what we had before with this tool will help us to make that decision.”

AI has also already shown its value in reducing administrative burdens by mining electronic health records (EHRs) for valuable patient data much faster and more accurately compared with human capabilities.3 In a study published in Nature Communications, investigators developed a natural language processing model that can extract structured cancer outcome variables from EHRs for creation of multi-site clinico-genomic datasets.

The ‘teacher’ portion of the model was trained using EHR data from Dana-Farber Cancer Institute (DFCI) to label imaging reports and discharge summaries. The ‘student’ portion of the model then used this information to predict the labels that were assigned by the teacher component and sent the information to Memorial Sloan Kettering Cancer Center (MSKCC) for evaluation.

The DFCI cohort included 37,274 annotated imaging reports from 3,213 patients and 39,191 annotated oncologist notes from 3,588 patients, across 7 cancer types. The MSKCC cohort included 24,472 annotated imaging reports from 2,672 patients and 40,701 annotated oncologist notes from 3,617 patients, across 5 cancer types. Findings from the study showed that the model’s approach was feasible in terms of extracting clinical variables from the EHR across cancer centers. Importantly, the model did not expose protected health information to investigators during its training. The study authors concluded that this approach could significantly expand the scale of clinical datasets for use in precision oncology research.

Additionally, AI-based approaches have shown great potential in terms of reducing research costs and increasing the speed of drug development.4 For example, AI has shown the ability to predict drug-target interaction performance, achieve target prediction accuracy of approximately 90% on over 2,000 small molecules, and the capability of predicting the druggability of anticancer drug targets. It can also identify drugs that could be repurposed based on drug-target interaction and assist in the accurate prediction of reactions to anticancer agents.

Addressing the Current Limitations of AI in Cancer Care

Despite the great promise that AI has already displayed in cancer care, there are concerns among experts that some of the limitations of the technology still need to be improved before it is truly ready for universal adoption. One issue that has concerned investigators is the presence of biases that can be introduced during the training of AI models, rendering them ineffective for some groups of patients.

“We don’t want to introduce biases or errors, particularly if they are going to impact patient care, so it’s a work in progress,” Mack Roach III, MD, professor of radiation oncology, medical oncology and urology at the University of California San Francisco, explained in an interview with OncLive. “There have been missteps along the way. For example, a model can be created in one population that’s made up of individuals with fair skin, then you try to to look at the retina to see [whether] a patient is at risk for retinal disease. If it is applied in a population where individuals have more pigmentation [compared with those it was trained with], the AI model isn’t going to work effectively.”

In a study published in JCO Clinical Cancer Informatics, Roach and his coauthors developed a novel multimodal AI deep learning system that incorporated digital histopathology and clinical data.5 The study authors aimed to assess whether the tool could be applied across risk-stratified African American patients with prostate cancer without algorithmic bias compared with non–African American patients.

The study included data from the randomized phase 3 NRG/RTOG 9202, 9408 (NCT00002597), 9413 (NCT00769548), 9910 (NCT00005044), and 0126 (NCT00033631) studies, in which all patients were treated with definitive external radiotherapy with or without androgen deprivation therapy. The study included high-quality digital histopathology data from 5,708 patients from the 5 studies; 948 were African American and 4,731 were non–African American.

Findings from the study showed that the model performed similarly in terms of showing a strong prognostic signal for distant metastasis in both the African American (subdistribution HR, 1.2; 95% CI, 1.0-1.3; P = .007) and non–African American (subdistribution HR, 1.4; 95% CI, 1.3-1.5; P < .001) groups. Additionally, the multimodal AI model for prostate cancer–specific mortality displayed a strong prognostic signal in both groups, with subdistribution HRs of 1.3 (95% CI, 1.1-1.5; P = .001) and 1.5 (95% CI, 1.4-1.6; P < .001), respectively.

“These men were treated on prospective phase 3 randomized trials, so we feel that this is the most reliable information that’s available,” Roach said. “[Our findings] demonstrate algorithms can be created based on AI that improve our ability to predict prognosis and guide treatment without introducing bias.”

“AI algorithms are now being applied in terms of deciding [which patients] should receive hormone therapy with radiation and who should not, as well as who should receive long-term vs short-term hormone therapy. Having established as a baseline that the algorithm did not introduce bias, we would deduce that if we’re deciding how long a patient should get hormone therapy and whether they should get hormone therapy or not, that we don’t have to worry about introducing bias. [These decisions] have a major impact not only on the patient’s survival but also quality of life, because patients that don’t need [therapy] won’t be receiving it. The downstream effects in terms of what to do for an individual patient are a real benefit.”

Investigators Anticipate Upcoming AI Research

During the upcoming 2025 ASCO Annual Meeting, several planned sessions will explore the role of AI in areas including drug development, community practice, disease screening, disease assessment, imaging, and clinical decision support. Abstracts will also be presented during oral sessions on a multimodal AI model used to identify benefit from second-generation androgen receptor inhibitors in patients with high-risk nonmetastatic prostate cancer and an AI-based platform designed for dynamic visualization of the National Comprehensive Cancer Network guidelines for patients with breast cancer.

“The question [regarding the future of AI in cancer care] should almost be, ‘Where will it not be useful?’” Roach said. “We need to have good data from studies available, then patients will be able to independently get a second opinion from both their doctors [and] the computer. The only risk there is that we need to provide accurate information; we need to make sure that we don’t have any biases. AI may also [prove to] be useful in evaluating the quality of publications [by] identifying false information included in studies. [It will be able to] identify fraudulent manuscripts [as well as] evaluate the quality of conclusions, providing transparency. There is tremendous potential [overall].”

References

1. Machine learning program enhances transplant risk assessment in myelofibrosis patients better than current models. News release. American Society of Hematology. March 27, 2025. Accessed May 20, 2025. https://www.hematology.org/newsroom/press-releases/2025/machine-learning-program-enhances-transplant-risk-assessment-in-myelofibrosis-patients
2. Hernandez-Boluda JC, Mosquera Orgueira A, Gras L, et al. Use of machine learning techniques to predict poor survival after hematopoietic cell transplantation for myelofibrosis. Blood. Published online March 27, 2025. doi:10.1182/blood.2024027287
3. Kehl KL, Jee J, Pichotta K, et al. Shareable artificial intelligence to extract cancer outcomes from electronic health records for precision oncology research. Nat Commun. 2024;15(1):9787. doi:10.1038/s41467-024-54071-x
4. Wang L, Song Y, Wang H, et al. Advances of artificial intelligence in anti-cancer drug design: a review of the past decade. Pharmaceuticals (Basel). 2023;16(2):253.doi:10.3390/ph16020253
5. Roach M 3rd, Zhang J, Mohamad O, et al. Assessing algorithmic fairness with a multimodal artificial intelligence model in men of African and non-African origin on NRG oncology prostate cancer phase III trials. JCO Clin Cancer Inform. Published online May 9, 2025. doi:10.1200/CCI-24-00284