AutoML: Powering the New Human-Machine Learning Ecosystem

1 minute read


Title: AutoML: Powering the New Human-Machine Learning Ecosystem

Date: 2020-09-30

Series: Machine Learning in Medicine Seminar

Speaker: Mihaela van der Schaar

Summary and thoughts:

Dr. van der Schaar from the University of Cambridge presented the talk on automated machine learning (AutoML) in healthcare. AutoML is the process of automating the development of machine learning models that can be applied to real-world problems. Her lab focuses on various challenges in machine learning which make development of models for healthcare difficult. While AutoML has been successfully applied to other areas, the healthcare sector presents new challenges and needs that must be addressed to use machine learning to revolutionaize healthcare. These challenges are two-fold: data and models. Health data tends to be biased, missing, unstructured, multimodal, with high dimensionality and often inaccessible or unavailable. Moreover, the models developed for personalized medicine and clinical decision support must be interpretable, explainable, trustworthy (for clinicians), reproducible and with known uncertainty estimates. Due to this, there is no “one size fits all” model available for all the problems. Often different metrics of performance must be evaluated based on the problem and these are often compared against the best known clinical scores.

AutoML attempts to develop machine learning pipelines that can do the crafting (instead of brute force selection of the model and parameters). Dr. van der Schaar presented four challenges that her lab has addressed using AutoML in healthcare: (1.) automating risk prediction modeling, (2.) survival analysis, (3.) individualized treatment recommendations, and (4.) time series forecasting. The methods use complex Bayesian optimization, structured kernel learning, ensemble methods to generate pipelines, survival quilts, and automated causal inference techniques to develop automated machine learning. Her final slides focused on automated interpretabe and explainable models through the method INVASE which is comparable to other methods in the area such as SHAP and LIME. Her lab also hosts engagement sessions for students interested in machine learning in healthcare.