Tissue-of-Origin AI Model Identifies Cases of Misdiagnosed Lung Metastases

A deep learning tissue-of-origin model integrated into routine molecular profiling in combination with orthogonal evidence identified several cases from a database that were misdiagnosed as lung squamous cell carcinoma, but that were found to actually be metastases from other primary tumors.

“Distinguishing primary lung squamous cell carcinoma from squamous metastases to the lung is a clinical challenge due to histopathologic similarities,” the investigators, including lead study author Mark G. Evans, MD, of Caris Life Sciences in Phoenix, Arizona, wrote in the findings of their cross-sectional study published in JAMA Network Open. “Accurate diagnosis is essential to guide treatment decisions.”

The investigators also commented: “These findings suggest the importance of an AI-assisted approach to distinguishing tissues of origin in patients with presumed primary lung squamous cell carcinoma, thus avoiding misdiagnosis and associated impacts on prognosis and therapy selection.”

Study and Model Methods

This study leveraged GPSai, a tissue-of-origin AI model automatically applied to each sample submitted for molecular profiling, to identify potential misdiagnoses among research-eligible cases initially classified as lung squamous cell carcinoma. Molecularly profiled cases from the Caris Life Sciences clinicogenomic database, spanning January 2024 to January 2025, were analyzed, and all cases underwent review by board-certified pathologists.

The deep learning GPSai model was trained on gene expression, DNA variant, and binary sex data from 201,612 cases. The model achieved a 95% accuracy in identifying tumors with a known diagnosis and achieved 84% accuracy for identifying the correct tissue of origin in carcinomas of unknown primary upon retrospective validation.

When the model claimed that the diagnosis was not consistent with the submitted diagnosis (with a high confidence score of ≥ 90% for another lineage and a low confidence score for lung origin), then a pathologist evaluated clinical history, imaging, and next-generation sequencing data, with additional diagnostic immunohistochemistry ordered if additional tissue was available.

The primary outcome was the misdiagnosis rate among presumed lung squamous cell carcinomas, confirmed through pathologist review and orthogonal evidence. This included clinical history and findings, GATA3 and uroplakin II immunohistochemistry for urothelial carcinoma, ultraviolet variant signature for cutaneous squamous cell carcinoma, CD5 and CD117 (c-KIT) immunohistochemistry for thymic carcinoma, and human papillomavirus positivity for orogenital squamous cell carcinoma (eg, head and neck, cervical).

Key Findings

Using a combination of AI and orthogonal evidence, 123 of 3,958 cases (3.1%) initially diagnosed as presumed lung squamous cell carcinoma were confirmed as misdiagnoses, with affected patients having a median age of 71 years and 76.4% being male.

The cohort comprised 50 cutaneous squamous cell carcinomas (40.7%), 33 orogenital squamous cell carcinomas (26.8%)—including 25 in the head and neck (75.8%)—20 urothelial carcinomas (16.3%), 15 thymic carcinomas (12.2%), 4 nuclear protein in testis carcinomas (3.3%), and 1 prostate squamous cell carcinoma (0.8%).

Clinical history or findings consistent with the revised diagnosis were present in 92 of 123 patients (74.8%). In 88 cases (71.5%), first-line systemic therapy recommendations per guidelines changed following reclassification.

“This cross-sectional study of patients diagnosed with lung squamous cell carcinoma found that an AI-assisted approach integrated into the routine molecular profiling workflow identified a meaningful number of misdiagnoses,” the investigators concluded. “Comprehensive evaluation of orthogonal evidence supported these diagnosis changes, which had important implications for prognosis and therapy selection.”

Added Perspective

In an invited commentary, Stephanie E. Siegmund, MD, PhD, of the Department of Pathology at UCSF, explained how the study highlights the potential of AI-assisted prediction models to enhance and assist with diagnosis. She also noted that the study demonstrates the importance of cross-disciplinary communication.

However, Dr. Siegmund also stressed that “caution is needed when implementing novel algorithms in clinical practice. Algorithms may flag or rediagnose cases in which no diagnostic update is needed, prompting potential patient harm through unnecessary additional workup and testing and potential systems-wide harm through wasteful allocation of resources. Thus, before using diagnostic algorithms in patient care, treating clinicians should be familiar with the model’s characteristics and limitations.”

DISCLOSURES: The study was funded by Caris Life Sciences. For full disclosures of all study authors, visit jamanetwork.com.

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.