Researchers at Karolinska Institutet have investigated how well different AI models can predict the prognosis of triple-negative breast cancer by analysing certain immune cells inside the tumour. The study, published in the journal eClinicalMedicine, is an important step towards using AI in cancer care to improve patient health.
Tumour-infiltrating lymphocytes are a type of immune cell that plays an important role in fighting cancer. When they are present in a tumour, it means that the immune system is trying to attack and destroy the cancer cells.
These immune cells can be important in predicting how a patient with so-called triple-negative breast cancer will respond to treatment and how the disease will progress. But when pathologists assess the immune cells, the results can vary. Artificial intelligence (AI) can help standardise and automate this process, but it has been difficult to demonstrate that AI works well enough to be used in healthcare.
Compared ten AI models
The researchers tested ten different AI models and compared their ability to analyse tumour-infiltrating lymphocytes in triple-negative breast cancer tissue samples.
The results showed that the AI models varied in their analytical performance. Despite these differences, eight of the ten models showed good prognostic ability, meaning they were able to predict patients’ future health in a similar way.
“Even models trained on fewer samples showed good prognostic ability, suggesting that tumour-infiltrating lymphocytes are a robust biomarker,” says Balazs Acs , researcher at the Department of Oncology-Pathology , Karolinska Institutet.
Independent studies needed
The study shows that large datasets are needed to compare different AI tools and ensure that they work well before they can be used in healthcare. While the results are promising, more validation is needed.
“Our research highlights the importance of independent studies that mimic real clinical practice,” says Balazs Acs. “Only through such testing can we ensure that AI tools are reliable and effective for clinical use.”
Information on funders and potential conflicts of interest can be found in the scientific article.
Publication
“The analytical and clinical validity of AI algorithms to score TILs in TNBC: Can we use different machine learning models interchangeably?” , Joan Martínez Vidal, Nikos Tsiknakis, Johan Staaf, Ana Bosch, Anna Ehinger, Emma Nimeus, Roberto Salgado, Yalai Bai, David L. Rimm, Johan Hartman, Balazs Acs. eClinicalMedicine, online 15 November 2024, doi: 10.1016/j.eclinm.2024.102928.
