Clinical Validation of a Deep Learning Model for Immunotherapy Response Prediction Using Longitudinal Blood Test Dynamics

In recent years, immunotherapy—a broad cancer treatment category that boosts the body's immune system to fight cancer—has become an important treatment option for many types of cancer. Immune checkpoint blockade (ICB) is a specific type of immunotherapy that works by blocking signals that normally stop the body's immune system from attacking cancer. The protein PD-1 (Programmed Death-1), found on T-cells in the immune system, deliberately binds together with the protein PD-L1 (Programmed Death Ligand-1), found on cancer cells, and acts as an "off switch" for the immune system, preventing T-cells from attacking the cancer cells. Checkpoint inhibitors "take the brakes off" by blocking this interaction, allowing immune cells called T cells to better recognize and destroy tumor cells.

This treatment has transformed outcomes for some patients. For example, before 2010, the chance of surviving three years with advanced melanoma (a serious type of skin cancer) was less than 15% with chemotherapy. With the use of checkpoint inhibitor drugs called programmed death-1 (PD-1) and programmed death-ligand 1 (PD-L1) inhibitors, survival has improved to around 40–60% in clinical trials.
Despite this progress, there are two major challenges. First, immunotherapy is very expensive. Second, only a small number of patients—usually 10–20% depending on the cancer type—actually benefit from the treatment. This means many patients undergo side effects and high costs without clear benefit. Finding reliable ways to predict who will respond is therefore a key priority.

Traditionally, doctors have relied on baseline information—such as the size of the tumor or patient health characteristics—collected before treatment begins. However, this only provides a single snapshot in time. We believe that changes in routine blood tests taken during the first few weeks of treatment may be more informative. Early shifts in blood cell counts or inflammation markers may reveal whether the immune system is being activated or if resistance to treatment is starting to develop. These signals often appear weeks before scans or other tests can confirm whether the therapy is working.

In this study, we will test and validate a computer-based deep learning model (a type of advanced machine learning method that can detect complex patterns in data). This model will combine early blood test results with clinical information to predict a patient's likelihood of benefiting from immunotherapy. By confirming the accuracy of this tool, we hope to help doctors identify non-responders sooner, avoid unnecessary treatment costs and side effects, and adjust treatment strategies in time to improve patient outcomes.