ESTRO 2024 - Abstract Book

S4532

Physics - Machine learning models and clinical applications

ESTRO 2024

Blood tests were collected at various intervals over 5 years to monitor counts of lymphocytes (ALC), neutrophils (ANC), Hemoglobin (Hb), platelets (PLT), red blood cells (RBC) and white blood cells (WBC). The analysis focused on late grade≥2 (G2+) lymphopenia, being ALC at 2 years (or 30 months if missing) after RT conclusion<800/μL, according to CTCAE v4.03.

Patients with complete baseline, RT mid- and/or endpoint, 3 or 6 months, 1 year or 18 months, 2 years or 30 months blood tests were included in this study.

Mann–Whitney U tests selected the fraction of structures receiving xxGy on the DVHs of bones and body that better discriminates patients with/without toxicity (corresponding to the lowest p-value). IDs were examined when no optimal DVH parameters could be determined. Twenty-four clinical/dosimetric covariates were considered for late G2+ lymphopenia prediction among which: ALC, WBC and age at baseline, smoking, craniocaudal PTVLN extent, PTVLN volume, hypofractionation, DVH parameters and IDs. Furthermore, acute grade≥3 (G3+) lymphopenia, defined as ALC at its lowest point between mid and end RT<500/μL, was separately considered as a potential predictor. Preprocessing and training used the information available for the included patients while 1000 bootstrap replicates were generated for internal validation. Univariate logistic regression assessed the effect of each covariate on the binary outcome of toxicity. For those variables with a significance level p<0.3 the association between variables was measured. In case of high-level Spearman correlation (ρ>0.7) between two variables maintained at univariable, the one with the lowest p-value was considered. The training phase tuned a backward stepwise multiple logistic regression, retaining variables with p-value≤0.05. Bootstrap resampling determined the robustness of the model by iteratively resampling the training population with replacement. Median coefficients, p-values, odds ratios, average precision, F1 score, ROC curve AUC, McFadden’s pseudo-R-squared and LLRpvalue across replicates were evaluated.

Results:

Data of 499 patients was available for late HT analysis. Nine-point-two percent of patients experienced Late G2+ lymphopenia.

DVHs were stratified by the presence of Late Grade 2+ Lymphopenia and parameters were checked for significant differences at the Mann–Whitney test. Parameters with the lowest p-values entered a univariate logistic regression.

The variables retained at multivariate analysis were: ALC at baseline (μL -1 ) [HR=0.997, 95%CI 0.996-0.998, p<0.0001], smoke (yes/no) [HR=2.902, 95%CI 1.246-6.761, p=0.013] and the fraction of lumbosacral volume receiving≥24Gy (cc) [HR=1.006, 95%CI 1.002-1.011, p=0.003]. Median and standard errors of performance statistics (average precision: 0.45, 0.07; F1 score: 0.4, 0.04; ROC curve AUC: 0.87, 0.02; McFadden’s pseudo-R-squared: 0.27, 0.04) were obtained aggregating over the replicates for which single covariates and overall model p-values were<0.05. When acute G3+ lymphopenia (yes/no) was retained in addition to the previous three variables [HR=4.517, 95%CI 1.954-10.441, p=0.0004], performance statistics (average precision: 0.52, 0.07; F1 score: 0.42, 0.04; ROC curve AUC: 0.88, 0.02; McFadden’s pseudo-R-squared: 0.32, 0.05) improved.