ESTRO 2024 - Abstract Book

S4554

Physics - Machine learning models and clinical applications

ESTRO 2024

Low values of circulating absolute lymphocyte counts (ALCs), leading to lymphopenia, are associated with worse overall survival (OS)[1] and reduce the efficacy of adjuvant immunotherapy in stage III non-small cell lung cancer (NSCLC)[2]. While radiation-induced lymphopenia is usually reported towards the end of radiation therapy, lymphocyte loss can already be observed early during concurrent chemoradiotherapy (CCRT). We aim to model the ALCs dynamics during CCRT, using dose to blood-containing organs[3–5], and dose to lymphocyte-proliferating organs. We then investigate the association of the modeled ALCs with OS, and validate it in an independent cohort.

Material/Methods:

The derivation cohort comprised 101 stage III NSCLC patients treated between 2019-2021. ALCs were routinely measured in blood twice weekly during CCRT. The validation cohort comprised 370 stage III NSCLC patients treated between 2009-2017, without ALCs collection. CCRT was mostly 24x2.75 Gy to the primary tumor and 24x2.42Gy or 2.75Gy to the involved lymph nodes with daily low dose Cisplatin (6 mg/m²).

The ALCs model was constructed in two-steps: (1) a mathematical description of the lymphocyte dynamics similarly as described in[6], (2) the estimation of the fraction dose to blood (D blood ) and to proliferating organs (D proliferation ).

Lymphocyte counts (L) as a function of time (t) without radiation was expressed as equation (1) in Figure 1. Daily treatment at time t R was modeled as an instantaneous response as equations (2) and (3) in Figure 1.Simultaneous administration of chemotherapy and RT was assumed.

Dose distributions per fraction were computed as the planned dose divided by 24 followed by EQD2 conversion (α/β=10Gy). We estimated the daily dose to blood (D blood ) as [5]:

D blood = 0.12·MLD+0.08·MHD+(0.45+0.35)·IBD

with MLD the mean lung dose, MHD the mean heart dose, and IBD the integral body dose per fraction, with the preceding coefficients representing the percentage of blood contained in these compartments. The IBD was derived with the dose within a body contour that excluded lung, heart, and bone segmentations. Similarly, D proliferation was computed as integral dose generating first another body contour excluding lung, heart and primary GTV. The model was implemented to numerically calculate the ALCs for each patient for a given set of parameter values. The model parameters were optimized based on derivation cohort ALCs dynamics using LMFIT package (Python). ALCs were calculated using the fitted model. For the validation cohort, the average modeled intercept L ’ 0 from derivation cohort was used as input. Pearson’s correlations R of modeled ALCs against actuals were computed per derivation cohort patient.

Nadir was identified for each patient in the modeled and actual ALCs. Two univariate Cox regressions with modeled and actual nadirs were computed for the derivation cohort.

For the validation cohort, the association of modeled nadir with OS was assessed in a multivariable Cox regression with the following clinical variables: age, BMI, sex, and total GTV. Since correlation between GTV and nadir is known [7] a linear regression of nadir against ln(GTV) was computed first, and the linear model nadir residuals were included.