Can AI Improve the Safety of Immune Checkpoint Inhibitors?

In a scoping review published in JCO Clinical Cancer Informatics, researchers found that AI is being applied to improve the safety of immune checkpoint inhibitors across 40 studies involving nearly 46,000 patients, with most applications focused on risk prediction.

“The findings support the potential of AI to advance real-world pharmacovigilance, enable earlier recognition of toxicities, and inform personalized monitoring strategies,” the researchers, including corresponding author Christine Y. Lu, PhD, of the University of Sydney School of Pharmacy and Kolling Institute, Faculty of Medicine and Health, The University of Sydney and the Northern Sydney Local Health District, Australia, commented. However, they noted, “Prospective validation within real-world clinical workflows remains rare but is essential to determine whether AI-based tools improve patient outcomes.”

Review Methods

The researchers searched MEDLINE (Ovid), Embase, and Scopus for studies published between January 1, 2015, and August 24, 2025. Eligible studies applied at least one AI method, such as machine learning or natural language processing, to investigate immune-related adverse events. Studies were grouped into risk prediction, identification and detection, and clinical information and decision support, with data synthesized narratively and mapped descriptively.

From an initial identification of 4,070 records, 40 studies met the inclusion criteria, encompassing 45,897 patients with cancer who were treated with immune checkpoint inhibitors.

AI in Treatment Safety

Among the analyzed studies, AI applications were predominantly directed toward risk prediction (n = 27), followed by identification and detection (n = 10) and decision support (n = 3). Machine/deep learning and natural language processing was predominantly used for risk prediction and adverse event prediction, while large language model approaches were used more commonly for decision support tasks. 

According to the researchers, AI approaches demonstrated the potential to detect immune-related adverse events using structured and unstructured data and real-world sources, as well as to stratify patient-level risk and inform clinical decision-making. For example, in one study, a large language model identified severe immune-related adverse events from clinical notes with higher sensitivity than diagnostic code–based methods in both internal (0.95 vs 0.69) and external (0.98 vs 0.79) validation, and reduced chart review time from approximately 15 minutes to under 10 seconds per case.

In risk prediction, several models were reported to achieve strong (area under the curve [AUC] ≥ 0.8) or acceptable (AUC = 0.7–0.8) discriminative ability.

In terms of decision support, large language models supplied accurate and clinically relevant information, with two of the studies reporting greater accuracy with ChatGPT-4 over other models. The researchers suggested that these models could be used as adjunct reference tools for clinicians. 

Despite these “promising” findings, the researchers cautioned that methodologic limitations were common, with most studies based on retrospective data and lacking external validation—factors that may constrain clinical applicability.

They concluded, “Applications of AI for immune-related adverse events are evolving rapidly, demonstrating credible potential to enhance early detection of toxicities, enable patient-level risk stratification, and support scalable clinical decision-making.”

“To translate these advances into routine practice, future research should prioritize methodologic standardization, rigorous external and prospective validation, real-world impact evaluation, and the development of robust clinical governance frameworks,” they continued. “With these elements in place, AI can move beyond proof of concept and become a reliable, evidence-based tool in the ongoing effort to optimize immunotherapy safety and patient care.”

DISCLOSURES: Dr. Yiu reported no conflicts of interest. For full disclosures of the other study authors, visit ascopubs.org.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.