A log-logistic survival model from multimodal data for prediction of Alzheimer's Disease

Pascal Lu (1) , Olivier Colliot (1)

1 ARAMIS - Algorithms, models and methods for images and signals of the human brain = Algorithmes, modèles et méthodes pour les images et les signaux du cerveau humain [ICM Paris]

The early diagnosis of Alzheimer's disease (AD) is important for providing adequate care and is currently the topic of very active research. In particular, a current challenge is to predict the future occurrence of AD in patients with mild cognitive impairment. Multimodal data (such as cogni-tive/clinical, imaging and genetic) can provide complementary information for the prediction. Whereas clinical data, such as cognitive scores, provide an accurate estimate of the current subject's state, genetic variants help to identify whether a subject would develop Alzheimer's disease faster than another. In the state of the art, most papers put clinical and genetic variables on the same level in order to predict the current or future subject's state, although they do not provide the same type of information. In this work, we propose a new survival model based on multimodal data to estimate the conversion date to AD from genetics and clinical data. We chose the log-logistic model which provides a parametric framework where the parameters depends on both clinical and genetic data. The hazard function is unimodal, which seems well-suited to our model. In our proposed formulation, genetic data xG only influences the speed v(xG) at which the conversion would happen, whereas clinical data xC influences the initial state of the subject p(xC). If S denotes the survival function and t_(1/2) the median survival time, we set t_(1/2) = p(xC)/v(xG), and S (t_(1/2)) = −v(xG). By determining v(xG) and p(xC), we are able to determine the associated survival function S. We tested our model on the ADNI-1 dataset (from the Alzheimer's Disease Neuroimaging Initiative), using cognitive scores (such as MMSE, ADAS13, RAVLT) and genetic information (APOE, gender) and compared it to the Cox-regression model using the Kaplan-Meier estimate, and the classical parametric log-logistic model where the parameters depends on the covari-ates using a log-linear model.

Mots clés: Log-logistic model; Multimodal data; Alzheimer's Disease

Domaines: Statistiques [math.ST]; Bio-informatique [q-bio.QM]

Pascal Lu, Olivier Colliot. A log-logistic survival model from multimodal data for prediction of Alzheimer's Disease. SAfJR 2019 - Survival Analysis for Junior Researchers, Apr 2019, Copenhague, Denmark. ⟨hal-02430943⟩