Personalizing Surgical Risk Assessment in Lung Cancer With AI

For many patients with lung cancer, surgery offers the best chance for cure and long-term survival, but that opportunity can come with meaningful risk—particularly for older patients with smoking histories, reduced lung function, frailty, or other medical conditions. In this episode of Exploring AI in Oncology, Waqas Haque, MD, MPH, Hematology/Oncology Fellow at the University of Chicago, speaks with Dr. Kenneth Patrick Seastedt, Thoracic Surgeon at Roswell Park Comprehensive Cancer Center, about MIRACLE, an AI-enabled platform designed to help predict postoperative pulmonary complications after lung cancer surgery. By combining clinical data, radiomic information extracted from CT imaging, and large language model (LLM)–generated risk explanations, the platform creates a more personalized assessment of surgical risk—all while keeping surgeons involved in every stage of interpretation and decision making. 

Moving Beyond the Eyeball Test

Assessing whether a patient can safely tolerate lung cancer surgery is a familiar challenge in thoracic oncology. Patients may have comparable pulmonary function tests, similar ages, or overlapping comorbidities, yet their ability to recover from surgery can differ substantially. Standard risk calculators and pulmonary function tests remain valuable tools, but they may not fully account for the nuances of a patient’s overall health and functional status.

Dr. Seastedt described the practical judgment surgeons often make in clinic as an eyeball test.

“Sometimes you’ll see a patient who looks like they could tolerate it really well, and they end up not tolerating it very well,” he said. “Or there’s a patient who you think is going to be a disaster, and they end up doing really well after the surgery.” That uncertainty was a central motivation for developing MIRACLE. The goal, Dr. Seastedt explained, was to use AI to “help personalize that risk calculation ahead of surgery” and give clinicians and patients “a more informed decision about what they want to do with their treatment.”

Why a Personalized Risk Model Matters

Traditional surgical risk calculators generally estimate outcomes by comparing a patient with a larger population of people with similar characteristics—such as age, pulmonary function, or comorbidities. But patients are not averages. “A lot of our patients are in the same age range, 70 to 75, and have similar pulmonary function tests,” Dr. Seastedt said. “You can get very generic outputs from risk calculators as to what the morbidity and mortality is going to be, but we want much more.” He added, “All of our patients are very different.”

MIRACLE aims to reflect that variation by integrating structured clinical data with radiomic features derived from each patient’s preoperative CT scan. This multimodal approach is intended to produce a risk estimate tailored to the individual patient rather than a broad population-level prediction. The tool focuses specifically on postoperative pulmonary complications, which Dr. Seastedt identified as the most common and often most serious complications following lung cancer surgery.

Comparing MIRACLE With Clinical Judgment

The model was developed using data from approximately 3,100 patients at Roswell Park. In the study, MIRACLE achieved an area under the receiver operating characteristic curve of approximately 0.81 for predicting postoperative pulmonary complications. As Dr. Seastedt noted, that figure should not be interpreted as simple accuracy. Rather, it reflects how well the model distinguishes patients at higher versus lower risk for a complication.

The study also compared MIRACLE’s performance with that of experienced thoracic surgeons. Both the model and surgeons received the same clinical information and CT imaging for a 200-patient test set. At a comparable false-positive rate of approximately 20%, MIRACLE showed sensitivity of roughly 77%, compared with about 45% for surgeons. The surgeon group had a median of approximately 17 years of clinical experience.

Dr. Seastedt emphasized that this comparison is not intended to criticize surgeons or suggest that AI should supplant their expertise. “It’s not like we’re trying to replace the surgeons or replace their judgment,” he said. “It just helps in their clinical decision making.” The results point to the potential for AI to identify relationships across data that may be difficult to recognize through conventional clinical review alone. “There are patterns across the data that we can’t necessarily see with our eyes, and this could help us make better decisions for us and our patients,” he said.

Keeping the Clinician in the Loop 

A key feature of MIRACLE is that it does not simply generate a risk score and leave the clinician with an unexplained output. The platform provides a plain-language summary describing the factors contributing to the patient’s risk. It highlights unfavorable factors separately from favorable ones, allowing the clinician to review why the model produced its assessment. Dr. Seastedt intentionally designed the tool so that surgeons and other clinicians can edit this summary and incorporate additional context before recalculating the risk estimate. “I wanted to have surgeons and clinicians be in the loop of the model and be in control of it,” he said. 

For example, a surgeon might feel that a patient appears more frail in the clinic than is reflected by a structured ECOG performance status. A clinician might also want to account for a recent treatment history—such as neoadjuvant chemotherapy or immunotherapy—that may not yet be fully represented in the model’s training data. These clinician-added details can affect the estimate, but early use suggests that the resulting changes are generally modest, often in the range of 0% to 5%. This is an intentional balance: clinical expertise can meaningfully inform the assessment without allowing subjective additions to radically alter underlying risk calculation.

Building Guardrails for LLM Use 

Because MIRACLE incorporates an LLM to generate and support the editable explanation, the team recognized the need for strong safeguards. Dr. Seastedt noted that an early version of the tool could generate incorrect factual values, such as an inaccurate body mass index. For a clinical decision support tool, that was unacceptable. “We didn’t want the model hallucinating any ground truth data points,” he said. 

The platform now locks all verified, structured variables used in the risk model. The LLM cannot modify factual patient information such as age, body mass index, pulmonary function test values, or radiomic features extracted from CT imaging. Instead, the LLM’s role is limited to the narrative clinical summary. This design preserves the integrity of the patient data while allowing clinicians to provide additional context that may be difficult to capture in a conventional data field.

From Retrospective Results to Real-World Clinical Use

Although the study included more than 3,100 patients, the model was developed at a single institution. That means the next critical question is whether it will perform similarly in other settings, with different patient populations, clinical workflows, and surgical practices. Dr. Seastedt and his colleagues are planning a prospective pilot at Roswell Park.

The study will evaluate whether surgeons find the tool useful, how it fits into clinical workflow, and whether it changes management decisions. Ultimately, the team hopes to evaluate the platform across additional institutions. The tool was designed around data commonly available in thoracic surgery: standard demographic and clinical information, pulmonary function testing, and preoperative CT imaging. However, Dr. Seastedt acknowledged that a model trained with data from one institution may still carry local biases and will need external validation before broader use.

Why Clinician-Led AI Development Is Essential

MIRACLE also represents a broader shift in how clinicians can participate in AI development.

Dr. Seastedt believes that physicians should be actively involved in defining the problems AI tools are intended to solve, helping ensure that technology fits real clinical needs rather than being driven solely by commercial or technical priorities. “A clinician really has to be the one setting the terms of what the clinical problem is that we’re trying to solve,” he said.

Industry partners can play an important role in supporting and scaling healthcare technology, but clinicians bring firsthand knowledge of the patient journey, real-world workflow constraints, and the practical consequences of an inaccurate or poorly designed tool. 

Dr. Seastedt developed MIRACLE in collaboration with the University at Buffalo’s Department of Computer Science and Engineering. The model was built from scratch and designed to be inspectable, allowing the team to identify and correct issues rather than relying on an opaque proprietary system. “I’m all for clinicians getting their hands dirty and trying to use some of these newer technologies to build their own products,” he said. To him, the underlying purpose of that work remains patient-centered: “Our ultimate goals are usually what’s best for us and what’s best for the patients.”

Key Takeaways

• Even experienced thoracic surgeons can struggle to predict which patients will thrive after surgery and which will suffer complications.

• In a head-to-head comparison with surgeons who had a median of 17 years of experience, the MIRACLE model showed significantly higher sensitivity (77%) for predicting pulmonary complications compared to surgeons (45%), at the same false-positive rate.

• Unlike traditional calculators that treat patients as population averages, the MIRACLE platform fuses three distinct data sources: structured clinical data, radiomics, and LLMs for explanations.

• By analyzing patterns in CT imaging that are invisible to the human eye, the model can identify subtle indicators of frailty or lung health that standard pulmonary function tests might miss.

• Clinicians must lead the development of AI tools to ensure they solve real clinical problems and integrate seamlessly into hospital workflows, rather than being driven solely by profit or generalizability.

• Next steps involve prospective pilot studies and testing at other institutions to ensure the algorithm’s accuracy across different patient populations and surgical practices. 

About the Speakers

Waqas Haque, MD, MPH, is a second-year Hematology/Oncology Fellow at the University of Chicago. He recently completed his Internal Medicine Residency at New York University (NYU) in a Clinical Investigator Track. Dr. Haque has balanced his patient care work with a variety of research projects. During his fellowship training at the University of Chicago, he plans to further his work in innovative clinical trial design, value-based care delivery to cancer patients, and clinical investigation. 

Kenneth Patrick Seastedt, MD, MBA, FACS, is an Assistant Professor of Oncology in the Department of Thoracic Surgery at Roswell Park Comprehensive Cancer Center. He specializes in minimally invasive thoracic surgical oncology and is certified in robotic surgery. Dr. Seastedt’s research focuses on medical applications of AI to thoracic cancers, improving perioperative outcomes and long-term survival for patients with lung and esophageal cancers, advancing minimally invasive surgical techniques, and developing real-time predictive tools for precision risk prediction.