IDNetwork: A deep illness-death network based on multi-state event history process for disease prognostication.

Multi-state models can capture the different patterns of disease evolution. In particular, the illness-death model is used to follow disease progression from a healthy state to an intermediate state of the disease and to a death-related final state. We aim to use those models in order to adapt treatment decisions according to the evolution of the disease. In state-of-the art methods, the risks of transition between the states are modeled via (semi-) Markov processes and transition-specific Cox proportional hazard (P.H.) models. The Cox P.H. model assumes that each variable makes a linear contribution to the model, but the relationship between covariates and risks can be more complex in clinical situations. To address this challenge, we propose a neural network architecture called illness-death network (IDNetwork) that relaxes the linear Cox P.H. assumption within an illness-death process. IDNetwork employs a multi-task architecture and uses a set of fully connected subnetworks in order to learn the probabilities of transition. Through simulations, we explore different configurations of the architecture and demonstrate the added value of our model. IDNetwork significantly improves the predictive performance compared to state-of-the-art methods on a simulated data set, on two clinical trials for patients with colon cancer and on a real-world data set in breast cancer.

MS-CPFI: A model-agnostic Counterfactual Perturbation Feature Importance algorithm for interpreting black-box Multi-State models.

Multi-state processes (Webster, 2019) are commonly used to model the complex clinical evolution of diseases where patients progress through different states. In recent years, machine learning and deep learning algorithms have been proposed to improve the accuracy of these models' predictions (Wang et al., 2019). However, acceptability by patients and clinicians, as well as for regulatory compliance, require interpretability of these algorithms's predictions. Existing methods, such as the Permutation Feature Importance algorithm, have been adapted for interpreting predictions in black-box models for 2-state processes (corresponding to survival analysis). For generalizing these methods to multi-state models, we introduce a novel model-agnostic interpretability algorithm called Multi-State Counterfactual Perturbation Feature Importance (MS-CPFI) that computes feature importance scores for each transition of a general multi-state model, including survival, competing-risks, and illness-death models. MS-CPFI uses a new counterfactual perturbation method that allows interpreting feature effects while capturing the non-linear effects and potentially capturing time-dependent effects. Experimental results on simulations show that MS-CPFI increases model interpretability in the case of non-linear effects. Additionally, results on a real-world dataset for patients with breast cancer confirm that MS-CPFI can detect clinically important features and provide information on the disease progression by displaying features that are protective factors versus features that are risk factors for each stage of the disease. Overall, MS-CPFI is a promising model-agnostic interpretability algorithm for multi-state models, which can improve the interpretability of machine learning and deep learning algorithms in healthcare.