A mixed-effect modeling framework to personalize treatment of low-grade glioma patients
P. Mazzocco(1), C. Barthélemy(2), M. Lavielle(2), F. Ducray(3,4,5), B. Ribba(1)
(1) Inria, Project-team NUMED, Ecole Normale Supérieure de Lyon, 46 allée d’Italie, 69007 Lyon Cedex 07, France ; (2) Inria Saclay, Project-team POPIX; (3) Hospices Civils de Lyon, Hôpital Neurologique, Neuro-oncologie,Lyon, 69003 France; (4) Inserm U1028 ; CNRS UMR5292 ; LyonNeuroscience Research Center, Neuro-oncology and Neuro- inflammation team, Lyon,F-69000, France; (5) University Lyon 1, Lyon, F-69000, France.
Objective: We aim to develop a mixed-effect modeling framework to predict the effect and the duration of temozolomide (TMZ) treatment in low-grade glioma (LGG) patients.
Method: We analyze a dataset containing mean tumor diameters (MTD) in 77 LGG patients treated with TMZ (952 total observations). Among these 77 patients, 45 (58%) experienced tumor progression after 14.3 months in median from treatment onset, this progression occurring during treatment for 28 patients (36%).
We propose a mixed-effect model to describe the observed MTD, accounting for possible tumor progression during treatment. Model parameters are estimated with Monolix (Lixoft). Genetic statuses (1p19q codeletion and p53 mutation) are included as covariates with a stepwise forward/backward analysis .
We investigate the ability of the model to predict 2 clinically-relevant metrics:
1- the time to tumor growth (TTG)
2- the minimal tumor size (MTS).
We only consider observations before the 3rd month of treatment to compute the empirical Bayes estimates (EBE) of the individual parameters, using a MAP algorithm implemented in Matlab (Mathworks) and MLXTRAN. EBEs are used to simulate the model and compute the 2 metrics in the 45 patients with observed progression.
Results: With model simulations we find that beyond 18 months of treatment, p53-mutated patients have a smaller TTG (less effective treatment) compared to the others and more than 50% of them experience tumor progression during treatment. The MTS is also always worse for p53-mutated patients.
Regarding individual prediction capability, the prediction of TTG is correct until 24 months following treatment onset. Beyond that, the model tends to under-estimate the effect of TMZ. For 85% of patients (38 patients), the MTS is correctly predicted, i.e the error on the prediction is less than 20% relatively to the tumor size at treatment onset, which represents an error of less than 1cm.
Conclusions: Our results indicate that knowing whether a LGG patient is p53-mutated or not is the first step to personalize his/her therapy, 18 months of TMZ appearing to be a maximum for p53-mutated patients.
Extracting information through the MAP algorithm only based on observations before the 3rd month of treatment is sufficient to predict the amplitude of the response for 85% of patients and its duration for 2 years.
As a perspective, this modeling framework can be used by clinicians to personalize the duration of TMZ treatment in LGG patients.
 Wählby U, Jonsson EN, Karlsson MO. Comparison of stepwise covariate model building strategies in population pharmacokinetic-pharmacodynamic analysis. AAPS PharmSci. 2002;4(4):E27. doi:10.1208/ps040427.