AI in oncology: Improving tumor board workflows

• Integration of artificial intelligence (AI) into clinical workflows is expanding, with notable projected market growth, requiring laboratory and clinical leaders to evaluate how such tools fit into established practices
• AI applications in oncology include support for tumor board workflows, particularly in generating structured case summaries
• A pilot study at Northside Hospital examined the use of AI-powered clinical summarization, reporting improvements in efficiency while also identifying areas for further evaluation

Artificial intelligence (AI) is increasingly applied in healthcare to enhance patient care, streamline operations, and reduce administrative burden. Forecasts suggest that the global AI healthcare market may grow at an annual rate of nearly 40%, reaching more than US$187 billion by 2030, compared with approximately US$27 billion in 2024.¹ This expansion reflects AI’s widening role across diagnostics, treatment planning, and clinical workflows.

In oncology, AI tools are being tested to support multidisciplinary workflows, including tumor boards. One area of interest is automated clinical summarization, which aims to reduce the time required to synthesize complex patient information across genomics, imaging, pathology, laboratory tests, and clinical notes.²˒³ A recent pilot at Northside Hospital in Georgia, U.S., illustrates how this technology may function in practice, as well as its limitations.⁴

Tumor boards are multidisciplinary meetings where providers, including physicians, nurse clinicians, clinical coordinators, and managers, review and plan treatment for complex cancer cases. These meetings are essential for fostering coordinated care and ensuring timely, well-informed clinical decision-making.⁵ Effective preparation for tumor board meetings requires aggregating and synthesizing comprehensive patient data from multiple sources such as genomics, imaging, pathology, lab reports, clinical notes.5 Traditionally, this preparation is performed manually, making it a time-intensive process that places considerable strain on clinical workflows.^2,3

At Northside Hospital, preparing clinical summaries for tumor board meetings is a labor-intensive process, with providers spending an average of 20 to 30 minutes per patient case. With approximately 60 to 85 cases reviewed weekly across 8 to 10 tumor boards, this preparation equated to nearly 25-37 hours of staff time each week.⁶

A major source of inefficiency stems from the fragmented nature of clinical data. The care team must gather information from various healthcare information sources, often pulling together disparate data types such as clinical notes, genomic test results, imaging, pathology reports, and laboratory values.

This manual, decentralized approach not only consumes valuable clinical time but also introduces variability and increases the risk of errors, especially in the absence of standardized workflows or quality control mechanisms. 

To address these inefficiencies, Northside Hospital piloted an AI-enabled summarization tool built on large language models (LLMs) to support tumor board case preparation. These models can transform unstructured EHR data into coherent, clinically relevant summaries. In oncology, where summarization requires integrating diverse data types such as genomics, imaging, pathology, and laboratory results, LLMs may help reduce variability, improve workflow standardization, and decrease preparation time. Automating this process enables clinicians to devote more attention to higher-order decision-making; however, real-world validation of AI-generated summaries remains critical to ensure safety, accuracy, and clinical reliability.⁷

The Oncology Quality Department conducted a qualitative evaluation in routine workflow. Over two months, clinicians used the tool to summarize 100 real patient cases, and semi-structured interviews were conducted with nurse clinicians, a clinical coordinator, and a clinical manager.⁴ Key evaluation domains included time efficiency, user experience, and perceived accuracy, completeness, and conciseness.⁴

By adopting this AI-enabled summary capability, Northside Hospital’s oncology department reduced summarization time by 40-50%, cutting it from 20-30 minutes per case to just 10-15 minutes.⁴ This time savings translates to approximately 15-21 hours per week, allowing physicians to shift their focus to more complex clinical tasks.

Clinicians found that the AI-generated summaries were highly concise, structured, and easy to review, often outperforming manual summaries in clarity and efficiency. The content was succinct, minimally redundant, and well-organized, enabling faster clinical review and decision-making. Coherence was generally strong, with most summaries presenting information in a logical and easy-to-follow format. While overall accuracy, completeness, and consistency were positive, clinicians identified occasional issues such as misinterpreted staging, rare hallucinations, or missing clinical details. Even with these minor challenges, the AI-generated clinical summaries were viewed as a valuable tool that significantly supported tumor board preparation.⁴

Clinicians reported that the AI-enabled summarization tool was easy to use and provided meaningful support in their workflow. While they emphasized that it does not replace clinical expertise, they found it to be a useful tool that complements their work, particularly in reducing tumor boards preparation time. Based on their experience, clinicians viewed the tool as suitable for broader use in other departments and recognized its potential to improve efficiency in clinical summarization across various clinical contexts.⁴

This single-site pilot suggests that AI-based summarization may reduce workload and improve consistency in tumor board preparation, with reported time savings and favorable user feedback.⁴ At the same time, observed inaccuracies underscore the necessity of human oversight and further evaluation to ensure clinical reliability.³˒⁷ Future work should include prospective studies across diverse settings, standardized quality controls, and assessment of downstream clinical impact.³˒⁷