MSK Research Shows AI Cuts Lung Cancer Tests by 40%

Memorial Sloan Kettering Cancer Center researchers are investigating methods to enhance immunotherapy effectiveness by inducing mismatch repair deficiency in resistant cancers. A clinical trial involving 18 patients with metastatic colorectal cancer showed that while no immediate responses were observed, patients’ tumor DNA developed mutational profiles suggesting the strategy could improve outcomes. Separate retrospective analysis revealed that 17 of 118 nonagenarian patients undergoing surgery for non-small cell lung cancer at MSK achieved a median survival of 43 months, comparable to the age group’s overall life expectancy, with no major complications reported. Additionally, a newly developed cancer-specific large language model, Woollie, demonstrated strong predictive capabilities, achieving scores of 97 and 98 out of 100 when predicting cancer progression from MSK data for all cancers and specifically for pancreatic cancer, respectively.

Engineering Immunotherapy Response

Researchers are investigating methods to enhance immunotherapy efficacy by inducing mutations in resistant cancer cells. Clinical trials have explored combining temozolomide and cisplatin to induce mismatch repair deficiency in colorectal cancer, with analysis revealing that while no immediate responses were observed, the development of compatible mutational profiles correlated with improved outcomes. This suggests the potential for this approach to increase responsiveness to immunotherapy.

A computational tool, EAGLE (EGFR AI Genomic Lung Evaluation), demonstrated effectiveness in assessing EGFR mutations in lung cancer. Analysis of over 8,000 lung cancer slides revealed that AI-assisted analysis reduced the need for molecular tests by over 40% while maintaining existing clinical performance standards, demonstrating real-world clinical utility.

Researchers have developed Woollie, an open-source, cancer-specific large language model (LLM) trained on real-world data to aid in understanding and predicting cancer progression. Validated on data from both MSK and UCSF, Woollie achieved a score of 97 out of 100 for predicting cancer progression from MSK data, and 98 specifically for pancreatic cancer. On UCSF data, the model achieved an overall score of 88, and 95 for lung cancer, indicating a potential for reliable support of oncologists and researchers, and offering opportunities for AI cancer prediction.

Surgery Outcomes for Nonagenarians

Data indicates positive outcomes for individuals in their 90s undergoing lung cancer surgery at MSK. Between 2001 and 2021, the number of nonagenarians seen in Thoracic Oncology clinics increased nine-fold. A retrospective study examined 118 patients in their 90s diagnosed with non-small cell lung cancer, of whom 17 underwent surgery. These surgical patients had a median survival of 43 months, approximating the 53-month life expectancy for individuals in that age group without cancer. No major complications were reported, cognitive functioning remained stable post-surgery, and no patients died within 90 days of the operation. These findings reflect multidisciplinary care and can inform decision-making regarding surgery for patients in their 90s.

Governing AI in Oncology

Responsible governance is critical to the success of AI in oncology. A team from MSK recently published a study reporting on the first year of the center’s comprehensive responsible AI governance model for clinical programs, operations, and research – one of the first such published reports in the field. The study covers the registration and monitoring of 26 AI models (including large language models), two ambient AI pilots, and a review of 33 nomograms. Analysis shows that governance and quality assurance of AI models are feasible at scale, but require key components for success.

The article outlines novel management tools, including for risk assessment and lifecycle management, as well as two case studies illustrating lessons learned. These real-world insights may be useful to others in oncology as these technologies continue to evolve and shape cancer care.

Validating AI-Assisted Biomarker Assessment

A real-world test demonstrated the value of AI-assisted biomarker assessment in lung cancer. Researchers demonstrated the effectiveness of EAGLE (EGFR AI Genomic Lung Evaluation) for assessing EGFR mutations. The researchers assembled a clinical dataset of over 8,000 lung cancer slides, representing the largest international dataset to date. They found that the AI-assisted analysis cut the number of molecular tests needed by more than 40% while maintaining current clinical standards for performance, demonstrating its real-world clinical utility.

Analysis shows that governance and quality assurance of AI models are feasible at scale, but require key components for success. The study covered the registration and monitoring of 26 AI models (including large language models), two ambient AI pilots, and a review of 33 nomograms. The article outlines novel management tools, including for risk assessment and lifecycle management, as well as two case studies illustrating lessons learned. These real-world insights may be useful to others in oncology as these technologies continue to evolve and shape cancer care.

Predictive Value of Cancer-Trained LLMs

The development of Woollie represents an important step toward AI tools that can reliably support oncologists and researchers. Woollie, an open-source, cancer-specific large language model (LLM), was trained on thousands of radiology reports from MSK across five major cancer types: lung, breast, prostate, pancreatic, and colorectal cancers.

Independent validation of Woollie’s performance was conducted using data from UCSF. Analysis showed strong capabilities, achieving a score of 97 out of 100 for predicting cancer progression from MSK data, and 98 specifically for pancreatic cancer. On UCSF data, it achieved an overall score of 88, and 95 for lung cancer. (A score of 50 would be the equivalent of a coin flip, and 100 would indicate perfect prediction.)

Beyond guiding treatment, Woollie could also help uncover broader insights into cancer biology, such as metastatic pathways and patterns of disease progression.