A DeepNLME framework for modeling ordinal data to describe disease progression in patients with Alzheimers disease.
Christos Kaikousidis1, Mohamed Tarek1, Roxana Aldea2, Joćo Abrantes2, Marcelo Boareto2, Antoine Soubret2
1Pumas AI, 2Roche Pharma Research and Early Development, Pharmaceutical Sciences
Introduction: Disease modeling in pharmacometrics is essential for understanding and predicting disease progression and optimizing drug development, as it enables the prediction of treatment outcomes and enhances personalized therapy strategies. In the case of Alzheimer's disease (AD), a widely used method for tracking disease progression is by modeling the longitudinal trajectory of the Clinical Dementia Rating (CDR) Scale. The CDR scale evaluates three domains of cognition and three domains of function, assigning patients a discrete score between 0 and 3 for each. Although previous attempts [1,2] have been made to predict CDR scores using linear and non-linear mixed-effects (NLME) models, model development for diverse and multi-dimensional disease score data in the presence of rich covariates and biomarkers remains a challenging and time-consuming task. In this work, we propose a flexible DeepNLME-based ordinal score modelling workflow and model structure for the quick exploration of diverse and multi-dimensional disease progression domain score data, shortlisting promising baseline covariates and biomarkers in the process. DeepNLME[3] is a novel methodology involving a structured use of neural networks in various parts of an existing NLME model, including the dynamics and covariate models. To our knowledge, this is the first attempt to model multi-dimensional ordinal data using DeepNLME. Objectives: •Apply ordinal regression in combination with DeepNLME to model time profiles of CDR domain scores within a mixed-effects framework. •Identify baseline covariates and biomarkers that can predict disease progression in Alzheimer's disease. •Compare disease progression across different CDR domain scores. •Develop a streamlined workflow that automates the entire modeling process, making it applicable to all CDR score types. Methods: The dataset included 326 subjects in the training set and 352 subjects in the test set. Measurements comprised values for 6 CDR scores, with 3 related to cognition and 3 related to functionality with measurements taken over a span of approximately 900 days. Baseline data included demographic information, patient physiological characteristics, the Functional Activities Questionnaire (FAQ), Mini-Mental State Examination (MMSE), and Alzheimer’s Disease Assessment Scale–Cognitive Subscale (ADAS-Cog) scores. Additionally, values for plasma, cerebrospinal fluid (CSF), and PET biomarkers were available. The modeling approach considers CDR scores as ordinal variables. We modeled the logit cumulative probability functions (CDF) for the possible outcomes using a mixed effects Neural-ODE system, with its outputs appropriately transformed to meet the necessary CDF function conditions for ordinal variables. The mixed effects neural ODE employed input random effects to capture the inter-individual variability in the dynamics of the various scores. The model was initially developed using the Memory CDR score and then adapted with minimal adjustments to the other domain scores. Sequential covariate modeling was performed to map baseline covariates to the empirical Bayes estimates of the fitted NLME model. A sensitivity analysis was conducted to identify the most predictive covariates. The proposed model was implemented in the Pumas/DeepPumas software suite. Results: The model accurately described individual profiles, outperforming a linear mixed-effects model used as a reference. Visual predictive checks demonstrated that the model could generate realistic patient profiles. Mean absolute error was low and similar for test and train set (0.29 and 0.27 respectively). Although the structural model discovered by the neural network was developed using only a single domain score, the model was successfully applied to the other scores with minimal adjustments. The models were able to rank disease progression from the most rapidly changing sub-scores (Memory and Orientation) to those that evolved slowest (Personal care) in agreement to the data. The most predictive covariates were the FAQ, baseline CDR, MMSE, and ADAS-Cog scores -leading to improved model predictability and an increase of loglikelihood values between 4% and 13 % on the test set, depending on the sub-score. Conclusion: To the best of our knowledge, this is the first instance in which DeepNLME has been applied to ordinal modeling in disease progression modeling. The process can be generalized easily to several CDR scores to properly fit the data, identify the baseline covariates’ and biomarkers’ predictive power and support model-informed drug development through comprehensive multi-dimensional modeling of domain scores.
Video link:
https://www.youtube.com/watch?v=AEU8RQUu5WI
