Machine learning for exploratory data analysis and model diagnosis in oncology

Objectives: In this work, data from five clinical trials investigating Atezolizumab [ 1], an immune checkpoint inhibitor, will be analyzed using a novel machine learning based exploratory data analysis (EDA) and model diagnostic method. In particular, clustering of the subjects’ responses will be performed to achieve the following objectives: more easily visualize the response patterns in a large dataset; automatically identify sub-populations in the response (e.g. responders and non-responders); identify potential outliers; stratify model diagnostics by the response’s cluster to understand which sub-populations are fitted better than others [ 2], aiding the model development process. Methods: We propose a method to perform clustering of longitudinal and heterogeneous data, e.g. from PK-PD studies. Clustering is a well known unsupervised machine learning task with a number of standard algorithms. To perform clustering, a pairwise dissimilarity matrix between the subjects’ responses is first computed using the dynamic time warping (DTW) method. DTW can be used to quantify the dissimilarity between heterogeneous and multi-dimensional longitudinal data from 2 different subjects with a different number of data points per subject. After computing the pairwise dissimilarity matrix, clustering is performed using the k-medoids algorithm to identify k clusters in the response. To demonstrate the value of stratifying model diagnostics by the cluster, a subset of the data from five clinical trials investigating Atezolizumab was analyzed using a nonlinear mixed effects general Bertalanffy [ 3] model. The model was fitted to the data and the clusters were used to produce stratified diagnostics. Results: The clustering procedure was able to automatically reveal a number of response types in the data, aiding with the visualization of a large dataset. The number of clusters k was tuned to produce visually consistent clusters. Clusters with the smallest number of subjects in them were flagged as potential outliers. An NLME Bertalanffy model was fitted to a subset of the data and the visual predictive check (VPC) was stratified by the response’s cluster. The stratification highlighted a weakness of the model in fitting a particular response type that was masked by the full population VPC. Conclusions: Unsupervised machine learning methods such as clustering can be used to enhance EDA and model diagnostics in any data analysis task, revealing non-obvious patterns and insights both in the data and model developed. In this study, a combination of DTW and the k-medoid algorithm was used to demonstrate the feasibility of clustering heterogeneous longitudinal data from clinical trials and to demonstrate its value in an NLME analysis workflow.