UB and Roswell Park develop AI tool to assess lung cancer surgery complication risk

BUFFALO, N.Y. – Researchers from the University at Buffalo and Roswell Park Comprehensive Cancer Center are developing an artificial intelligence tool to help surgeons better identify lung cancer patients at risk for postoperative complications. The work builds on a longstanding collaboration between the institutions, bringing together UB’s strengths in AI and Roswell Park’s expertise in thoracic oncology.

The system, called MIRACLE (Multimodal Integrated Radiomics and Clinical Language-based Explanation), is believed to be the first to combine clinical data, CT imaging and large language model (LLM)-generated explanations to provide personalized risk estimates for patients who may be candidates for lung cancer surgery. It also produces a summary that surgeons can review and refine to reflect their own clinical insight.

“For more than five decades, UB has been a leader in artificial intelligence for public good,” said Venu Govindaraju, PhD, senior vice president for research, innovation and economic development, and co-author of the study whose leadership in AI laid the foundation for this breakthrough. “Today, this expertise is transforming cancer research, using AI to detect disease earlier and opening new pathways to prevention and cure.”

Lung cancer is the leading cause of cancer-related death worldwide, and surgery remains one of the most effective treatment options. Because many patients are medically complex, postoperative complications can affect up to 40% of cases, making accurate risk assessment critical. However, existing risk calculators often rely on population-level data and subjective clinician judgment, which can limit their effectiveness for individual patients.

According to Kenneth Patrick Seastedt, MD, a thoracic surgeon at Roswell Park and study co-author, this can lead to difficult decisions as some patients who could have safely undergone surgery are not recommended for it, while others at higher risk proceed without sufficient preoperative and postoperative planning.

“One of the biggest challenges for thoracic surgeons is that preoperative risk calculators are very generic and don’t capture the real complexity of the patients we see every day. Many lung cancer patients are older, have multiple comorbidities or present with subtle imaging findings that generic models miss,” said Seastedt, an assistant professor in thoracic surgery at Roswell Park who also serves as an assistant professor of surgery in the Jacobs School of Medicine and Biomedical Sciences at UB. “Our goal was to improve how surgeons assess risk, plan perioperative care and counsel patients by bringing together clinical nuance, imaging detail and physician insight in a way that directly supports more personalized care.”

The team designed MIRACLE to integrate three forms of preoperative data:

• Clinical information such as age, smoking history, lung function and other health conditions
• Measurable characteristics from CT scans that capture important details about the patient’s anatomy
• Plain-language explanations generated by an LLM grounded in surgical oncology literature and clinical guidelines.

MIRACLE combines these inputs to estimate a patient’s risk of complications. The model updates the estimate by incorporating prior knowledge and patient‑specific data.

“The strength of this multimodal framework is that it brings together what clinicians know, what imaging shows and what AI can help explain,” said Srirangaraj (Ranga) Setlur, co-author of the study, and associate director for UB’s Institute for Artificial Intelligence and Data Science. “Each piece captures a different aspect of the patient’s condition, helping us assess postoperative risk in a way that’s tailored to the individual.”

A core feature of MIRACLE is its ability to generate a summary of a patient’s risk factors that surgeons can edit. For instance, surgeons can add details not captured in structured data — such as frailty or functional limitations — and the model recalculates the risk estimate accordingly.

“Allowing surgeons to account for factors outside of the data helps align the tool with real-world clinical decision-making. That’s essential for earning clinicians’ trust and acceptance,” Seastedt said.

To keep these edits grounded in clinical data, the team built in safeguards to prevent the LLM from introducing unsupported information or hallucinations. Key clinical factors, including age, gender and body mass index, are locked and cannot be modified by the model.

In addition, MIRACLE processes clinical inputs only for the duration of a session and deletes them immediately afterward, ensuring patient data is neither retained nor transmitted externally. This approach supports HIPAA compliance and makes the tool easier to implement in practice.

Researchers evaluated MIRACLE using a dataset of 3,094 lung cancer patients who underwent surgery at Roswell Park between 2009 and 2023. The model outperformed five machine learning methods, three open‑source LLMs, and practicing thoracic surgeons, who correctly identified complications about 45% of the time, compared with the model’s 75–80% sensitivity.

The top‑performing version achieved approximately 81% accuracy in distinguishing higher‑ from lower‑risk patients and maintained low false‑positive rates.

A panel of thoracic surgeons also reviewed the model’s explanations and found that many aligned with their clinical reasoning. At the same time, they noted that the model could overestimate risk or miss subtle interactions among comorbidities, reinforcing the importance of clinician judgment in interpreting and refining its output.

“Our results demonstrate that a multimodal approach generates more accurate and clinically coherent predictions than single‑source models. By making the explanations editable, we’ve moved from a static prediction model to an interactive and transparent decision-support tool,” Setlur said.

The research team recently presented their study, titled “LLM Augmented Intervenable Multimodal Adaptor for Postoperative Complication Prediction in Lung Cancer Surgery,” last month at the IEEE/CVF Winter Conference on Applications of Computer Vision. Additional contributors include UB alumnus Bhavin Jawade, PhD; and Shubham Pandey, a current UB PhD student, both working under the mentorship of Govindaraju and Setlur

“This collaboration between UB and Roswell Park is what makes projects like this possible,” Seastedt said. “You need teams that combine deep clinical insight with real technical expertise to build tools that truly serve patients. Otherwise, industry will do it for us, and its priorities aren’t always aligned with patient care. When clinicians and scientists work side by side, we can design solutions that put patients’ best interests first. That’s why this kind of partnership matters.”

The team’s next steps include prospectively validating MIRACLE by testing it in real time at Roswell Park to confirm it can reliably identify high-risk patients. If successful, they hope to adapt the model for use across other surgical specialties.