Variational Gaussian Mixture Models with robust Dirichlet concentration priors for virtual population generation in hypertrophic cardiomyopathy: a comparison study

Abstract

Nowadays, there is an emerging need for the development of computationally efficient virtual population generators for large-scale clinical trials. In this work, we utilize Gaussian Mixture Models (GMM) with variational Bayesian inference (BGMM) using robust estimations of Dirichlet concentration priors for the generation of virtual populations. The estimations were based on an exponential transformation of the number of Gaussian components. The proposed method was compared against state-of-the-art virtual data generators, such as, the Bayesian networks, the supervised tree ensembles (STE), the unsupervised tree ensembles (UTE), and the artificial neural networks (ANN) towards the generation of 20000 virtual patients in hypertrophic cardiomyopathy (HCM) clinical trials. Our results suggest that the proposed BGMM can yield virtual distributions with small inter-and intra-correlation difference (0.013 and 0.012), in less execution time (0.432 sec) than the STE which achieved the second-best performance.

Bayesian Inference-Based Gaussian Mixture Models With Optimal Components Estimation Towards Large-Scale Synthetic Data Generation for In Silico Clinical Trials

“…Similar to previous studies [11,16], the BGMM-OCE places particular emphasis on the quality of the input data since lack of data quality reduces the statistical power of the outcomes. Thus, the quality of the real data is reflected on the synthetic data.…”

“…The UTE, STE, and ANN are unable to capture the inter- and intra- correlation differences. As far as the GMM algorithm is concerned, although it is more computationally efficient, but it requires multiple hyperparameters which are arbitrarily defined [16] and thus they introduce biases. However, the precise definition of components and the estimation of the weight concentration parameter is a technical challenge.…”

“…Since GMM maximizes only the data likelihood based on the expectation maximization (EM) approach, it might yield specific structures that might not apply to the data. This can be solved by variational inference (VI) [14] – [16] which is more efficient than EM and reduces the computational complexity. Other attempts [17] , [18] focused on the automated adjustment of Gaussian distributions for background modeling.…”

Advancing treatment of retinal disease through in silico trials

“…For the warfarin trial, Bayesian Network Models were trained to learn and constraint the patient parameters with Monte Carlo simulation being used to sample to generate the VP [111]. Pezoulas et al also used tree-based methods (supervised and unsupervised) to generate VPs for cardiomyopathy drug development [227] but found that Gaussian Mixture Models with Variational Bayesian Inference outperformed supervised tree ensembles when comparing the VPs to the data [228].…”