Progression prediction model for solid tumors with clinical and immunological parameters.

Abstract

2539 Background: The immune system has well-known relation to tumor progression. Numerous immune-related parameters exist, but only a minor part could be used as biomarkers, especially dynamic ones. We trained a progression prediction model based on clinical features and peripheral immune system assessments. Methods: Patients with immunogenic (melanoma, 295, kidney cancer, 81), non-immunogenic (soft tissue sarcoma, 47, colorectal cancer, 26) and multiple primary tumors (29) with immunologic assessments before treatment (23.5%), on therapy (58.3), and in follow-up after the treatment (18.2%) were randomly divided in 7:3 ratio to the training and test groups. Counts of lymphocytes, T-, B, NK cells, cytotoxic lymphocytes, T-helpers were used as immunologic parameters. Age, sex, disease, stage, therapy, mutational status, last response on treatment, disease and therapy duration, previous treatments were used as clinical ones. The model was trained to predict disease progression in the next three months using “Catboost” gradient boosting. We used ROC AUC to test model performance and Yoden’s index for optimal cutoff calculation. We also studied the influence of model prediction on overall survival (OS) and time to progression (TTP) on the test dataset using the Kaplan-Meyer method and Cox regression. Results: We used 1682 assessments of immune parameters (immune status, IS) done in 354 patients (average 5 per patient) to train the model and 616 IS in 124 patients for validation. All IS of one patient were in the same group. The ROC AUC value of the model was 0.801. The model prediction of progression increased the probability of progressive disease from 37.5 to 62% and decreased the response rate from 37,5% to 8.4% (p = 0.016). The model prediction did not add information over known prognostic factors for OS in the multifactorial model but was an independent prognostic factor for TTP (HR 2.204, p = 0.011). False-positive results separate the group of patients with poor prognosis (OS 16 months, TTP 6 months) among patients with clinical benefit from patients with favorable prognosis (OS 61 months, TTP 18 months, p < 0.001), who had a truly negative model prediction. The possibility of prognosis improvement with therapy change was an essential factor for OS and TTP prediction (р < 0.001). The model was useful in predicting higher OS in patients with disease progression (p = 0.033) and shorter response duration in patients with clinical benefit (р = 0.03). Conclusions: Our progression prediction model provides clinically useful information and can be used for decision making in several clinical situations. Its utility should be tested in a prospective trial.