ESTRO 2024 - Abstract Book

S4551

Physics - Machine learning models and clinical applications

ESTRO 2024

explained (i) the mathematical association between [Hb] variations and the dose, (ii) the impact of the haemoglobin concentration at baseline ([Hb]_bas), while the random effects summarized the inter-patient factors that we still cannot describe (while being aware of their importance in modulation the single-patient response), such as genetics, vascular health status and comorbidities.

Before modelling, we conducted a preliminary analysis evaluating the relationship between the simulated (by TPS) and delivered dose (GAFchromic film dosimeters) on the patient's skin surface.

Results:

The LMM maximum likelihood estimation highlighted the association between variation in haemoglobin concentration and the square root of the dose. We also found a strong impact of the [Hb]_bas, suggesting the hypothesis that patients with good perfusion are less prone to consistent alterations of the Hb level. For a graphical representation, we defined three models according to the [Hb]_bas quartiles; we found that the slope of the dose response relationship is highly modulated, with steeper slopes for lower [Hb]_bas (see Figure 1, Q2 and Q3 showed similar slopes and were merged).

The LMM model predicts an increase in [Hb] of 70% after 20 Gy and 130% after 40 Gy for patients in the 1st quartile. The [Hb] increase reduces to 45% after 20 Gy and 80% after 40 Gy for patients in the 2nd and 3rd quartiles and a 10%