Causal Inference for Personalized Treatment Effect Estimation for given Machine Learning Models

We propose a causal machine learning inference pipeline that combines a given predictive machine learning model with analytical estimations of average treatment effects. It enables to utilize any predictive model for causal inference, which makes it easy to adapt the approach to existing systems. By first estimating the average treatment effect of an intervention on predictors instead of the outcome variable, a causal relationship between an intervention and a wide range of variables is determined. Next, artificial samples are created that are evaluated using the given predictive model to link interventions and outcomes and also allows inferring measurements of uncertainty. Finally, simulations using again the given predictive model are performed to compute measurements of confidence and that allow to compare – according to the given predictive model – the effect of specific treatments. We furthermore demonstrate how this inference engine can be adapted to a privacy-preserving federated learning environment where training data is horizontally distributed across multiple datasets without compromising on our approach’s accuracy. The approach has been evaluated on a use case with a predictive model for the quality of life score of cancer patients, to determine medical interventions to improve their individual quality of life score.