Artificial intelligence (AI) is creating new opportunities to identify the origins of cancer, particularly when a diagnosis does not fit the clinical picture. In this episode of Exploring AI in Oncology, Dr. Vitor Pastorini, a Medical Oncologist at Florida Oncology and Hematology, and Dr. George Sledge, Chief Medical Officer at Caris Life Sciences and former Chief of Oncology at Stanford University, sit down with Dr. Waqas Haque to discuss how AI-enabled genomic profiling can help reassess tumor origin. Through the case of a patient whose presumed metastatic breast cancer was ultimately identified as Hodgkin lymphoma, they explore the potential of genomic prediction tools, the importance of confirming unexpected results with additional testing and expert review, and the role of clinical judgment in ensuring patients receive the right diagnosis and treatment.

When the Clinical Picture Doesn’t Add Up

The story began when 38-year-old Erika Rios presented to Dr. Pastorini’s clinic carrying an outside diagnosis of metastatic triple-negative breast cancer (TNBC). Clinically, she was deteriorating rapidly, having lost over 50 pounds while suffering from debilitating night sweats, severe fatigue, and systemic wasting.

Despite the established TNBC label, several clinical details prompted Dr. Pastorini to pause. A review of her PET imaging revealed an atypical pattern of metastasis: widespread lymph-node involvement above and below the diaphragm without visceral organ metastases—an unusual presentation for aggressive metastatic TNBC. Furthermore, diagnostic breast imaging, including mammography and PET, revealed no primary breast mass. When Dr. Pastorini examined the initial biopsy records, he noted that the tissue was taken from an axillary lymph node rather than breast tissue, and the initial immunohistochemistry (IHC) panel had not included a comprehensive evaluation for hematologic malignancies.

Adhering to a principle he learned early in his training—“trust, but verify”—Dr. Pastorini sent Erika’s tissue for next-generation sequencing (NGS) at Caris Life Sciences, ordering the Genomic Prediction Score AI (GPSai) algorithm to clarify the tissue of origin. Because the patient was rapidly deteriorating and pleading for help, he initiated bridge chemotherapy with carboplatin and paclitaxel while awaiting the results.

The turnaround delivered an unexpected revelation: the molecular profiling and review of the specimen demonstrated that the tumor was not metastatic breast cancer at all, but Hodgkin lymphoma.

“It's not very often that we change a diagnosis from a deadly disease with basically no survival outcome, to an 89% overall survival,” Dr. Pastorini reflected. “It was a blessing in disguise. She was basically getting a second chance in life.”

How GPSai Uncovers the Genetic Fingerprint of Cancers

Dr. Sledge explained the scientific foundation behind GPSai, which leverages deep learning trained on Caris Life Sciences’ massive dataset of comprehensive NGS linked with real-world clinical and pathological outcomes. By evaluating complex genomic signatures across extensive molecular data, the model can identify the distinct “genetic fingerprint” of virtually any cancer in the human body. When benchmarked against well-curated reference cohorts, such as The Cancer Genome Atlas (TCGA), the algorithm demonstrates approximately 95% accuracy in correctly predicting cancer origin.

Dr. Sledge noted that misdiagnoses often stem not from clinician carelessness, but from cognitive bias and standard pattern recognition that trap oncologists and pathologists alike.

“We're fooled by our own pattern recognition frequently,” he said. “If this patient was being presented in tumor board, she would be presented as a 38-year-old woman with an occult primary breast cancer with axillary lymph-node involvement that was ER-, PR-, and HER2-negative, therefore aggressive triple-negative breast cancer, and we would’ve treated it as breast cancer.”

The Hidden Scale: From Squamous Cell Lung Carcinoma to Cancers of Unknown Primary

This diagnostic divergence is not an isolated phenomenon. Dr. Sledge pointed to published and ongoing research demonstrating the frequency with which molecular tools overturn presumed diagnoses:

Squamous cell lung carcinoma: In Caris studies evaluating specimens submitted with a diagnosis of squamous cell carcinoma of the lung, approximately 3% were identified by GPSai as metastatic squamous cell carcinomas arising from other anatomic sites, such as the head and neck, skin, bladder, renal pelvis, or cervix.
Population impact: While a 3% discrepancy rate may seem modest in isolated clinics, across millions of cancer patients nationwide, it represents thousands of individuals receiving therapies targeted to the wrong disease entity.
Cancers of unknown primary (CUP) and dual primaries: Beyond overt misdiagnoses, GPSai provides clarity for patients presenting with true CUP, limited fine-needle aspiration (FNA) tissue where exhaustive IHC panels are unfeasible, or patients with multiple primary tumors experiencing progression at an ambiguous site.

Resolving Discrepancies With Orthogonal Testing

When asked how clinicians should manage situations where an AI read conflicts with traditional pathology, both experts stressed that AI remains an adjunctive decision-support tool, not an absolute arbiter.

Dr. Sledge underscored the importance of running orthogonal validation assays whenever discrepancies arise:

Targeted immunohistochemistry: Rerunning focused IHC stains to verify the lineage suggested by molecular signatures.
Mutational signatures: Evaluating distinct biological footprints, such as ultraviolet (UV) radiation-induced mutational signatures characteristic of cutaneous carcinomas or viral signatures indicative of head and neck cancers.

Dr. Pastorini added that any discordant result should prompt immediate multidisciplinary review, including tumor boards and secondary expert pathology consultations, rather than reflexively following any single test in isolation.

Expanding Access and Advocating for General Medical Oncologists

As comprehensive genomic profiling and machine-learning models mature, Dr. Pastorini argued that every patient diagnosed with metastatic disease should have access to comprehensive NGS and molecular prediction tools. Broad coverage through Medicare, Medicaid, and commercial payers is critical to ensure diagnostic equity across community practices, where the vast majority of cancer patients receive care.

Dr. Sledge concluded by acknowledging the immense complexity faced by community general oncologists, who manage rapidly evolving standards of care and an average of 10 new FDA oncology indications each month.

“The hardest job on the planet right now is being a general medical oncologist,” Dr. Sledge stated. He emphasized that because general oncologists evaluate diverse malignancies across organ systems daily, they are often uniquely positioned to develop the clinical intuition needed to spot when a patient’s disease trajectory does not match the textbook picture.

By combining that vigilant clinical instinct with advanced AI diagnostic tools, oncology teams can ensure that more patients receive the right diagnosis—and the right treatment—from the very beginning.

Key Takeaways

Clinical clues can justify a closer look at the diagnosis. In Erika Rios’s case, the lack of a breast mass and an atypical pattern of lymph-node involvement prompted Dr. Pastorini to question the initial diagnosis of metastatic triple-negative breast cancer.
Genomic AI helped identify a different cancer origin. GPSai analysis, followed by pathology review, pointed to Hodgkin lymphoma rather than breast cancer, changing the treatment pathway.
AI can help counter familiar-pattern bias. Clinicians may understandably interpret a case through the lens of common presentations; molecular profiling can offer another perspective when the clinical picture does not fit.
An AI result should be treated as a signal to investigate, not a final verdict. The physicians emphasized confirming conflicting findings with pathology review, additional tests, and multidisciplinary discussion.

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.

George Sledge, MD, is the Executive Vice President and Chief Medical Officer of Caris Life Sciences, where he oversees medical affairs, research, and medical education. Previously, he was the Chief of the Division of Oncology at Stanford University, and he is a Past President of the American Society of Clinical Oncology (ASCO). Dr. Sledge has devoted his career to understanding the biology and improving the treatment of breast cancer, and he has authored more than 390 scientific publications in laboratory and clinical research.

Vitor Pastorini, MD, is a Medical Oncologist at Florida Oncology and Hematology. Previously, he was Chief Resident at the University of Miami/Jackson Memorial Hospital and held academic and teaching roles at the University of Miami and Duke University. Dr. Pastorini has participated extensively in clinical research and has served as principal investigator of numerous clinical trials in hematology and oncology, including studies focused on critical care outcomes and advances in pancreatic cancer treatment.

How Clinical Intuition and Genomic AI Solved a Cancer Misdiagnosis Mystery