Artificial intelligence models could inform treatment practices in the future

GlobalData Healthcare 3 December 2020 (Last Updated December 3rd, 2020 15:57)

Artificial intelligence models could inform treatment practices in the future
Credit: Shutterstock.

Researchers have used machine learning to develop an artificial intelligence (AI) model that uses histological samples from advanced melanoma patients to predict response to immunotherapy (Johannet et al., 2020). The researchers from the NYU Grossman School of Medicine used over 300 images from melanoma patients who had received either CTLA-4 or PD-1-targeted immune checkpoint inhibitors as either mono- or combination therapies to train their algorithm. They also incorporated information on patient response, as well as several clinical characteristics into their final model. With a reported accuracy of 80%, this tool has the potential to transform clinical practice, but where could such AI models fit in?

Immunotherapy is often the first-line therapy for advanced melanoma patients and is increasingly used in the adjuvant setting to try to prevent recurrence in high-risk patients. While certain patients harbouring BRAF mutations have the option to use targeted BRAF-MEK inhibitors, patients without this mutation are left with very few options should they fail to respond to immunotherapy. Biomarkers that predict response to immunotherapies have been sought in the past, and there was much hope that PD-L1 levels could be utilised in the clinic. However, despite PD-L1 levels correlating with response to immunotherapy, this marker is not widely used to inform treatment decisions. Ultimately, there are patients with low PD-L1 levels who respond to immunotherapy, and importantly, there are currently no available alternatives to immunotherapy for many of these patients. These current constraints mean that until further effective therapies are available, any prognostic test would have to be exceptionally accurate in order to be used to deny potentially life-saving therapy to a patient with advanced melanoma.

However, knowing which patients with BRAF mutations will likely do poorly on immunotherapy could potentially save considerable time and money, by directing those with poor prognoses to take targeted therapy in the first line. Furthermore, in the adjuvant setting, where the goal is to reduce the risk of recurrence, there are different considerations. A small but significant proportion of patients develop irreversible side effects on immunotherapies, including hypothyroidism and diabetes, which require life-long medication. In a patient with a moderate risk for recurrence, the choice about whether to take immunotherapy can be a difficult one, and so anything that could inform this decision would be welcome. If AI models could be utilised in this setting there is a great deal of potential, not just in melanoma but in other indications, to revolutionise treatment decision-making. This could lead to a more personalised approach to treatment and could mean that a more targeted and smaller patient population would be eligible for adjuvant immunotherapy.

It will be important to further validate this tool on larger patient datasets, but this work is an important proof-of-concept study. Histological samples are widely available and with further validation, this tool and others like it could become commonplace in the future.