ESTRO 2024 - Abstract Book

S4468

Physics - Machine learning models and clinical applications

ESTRO 2024

(3) Definition of the subgroups

The cutoff value for grouping the patients with low or high TIL density was determined as 270.33, which is the value that maximizes Youden's index in the receiver operating characteristic (ROC) analysis targeting the DFS. The optimization of the cutoff values was done by ten-fold cross-validation using the data of a hold-out sub-cohort (N=174). The remaining cohort (N=240) was used for analysing the impact of the pre-treatment TIL density ( TIL 0 ). The post-treatment TIL density ( TIL 1 ) was quantified for a sub-cohort with available resection WSIs (N=195). We notated the cohorts with low or high TIL density in the pre/post-treatment specimens as TIL 0/1 + and TIL 0/1 − . TIL ++ , TIL −− , TIL +− and TIL −+ represent the groups of patients whose TIL density remained high, remained low, changed from high to low, or changed from low to high after the treatment, respectively.

(4) Statistical Analysis

Kaplan-Meier (KM) estimation was used for analysing the DFS and OS. Log-rank testing was performed to test the separation of survival curves in different subgroups. Hazard Ratios (HR) with a 95% confidence interval (CI) were given by the Cox proportional hazard model.

Results:

The TIL 0 + patients had significantly better DFS (HR = 0.59, 95% CI: 0.39 - 0.90, p = 0.013) and OS (HR = 0.42, 95% CI: 0.24 - 0.73, p < 0.005), compared to the TIL 0 − patients. In this cohort, 89 cases showed re-stratification after the treatment and a significantly shorter DFS was observed for the 31 TIL −− patients ( Figure 2 ). The TIL −+ patients appeared to have significantly improved DFS (HR = 0.70, 95% CI: 0.50 – 0.97, p = 0.028), compared to the TIL −− patients.