Big Data/Machine Learning/AI
Personalised dynamic prediction in dialysis using a novel super learning framework Arthur Chatton* Arthur Chatton Michèle Bally Renée Lévesque Ivana Malenica Robert W Platt Mireille E Schnitzer
Obtaining continuously updated predictions is a major challenge for personalized medicine. In end-stage kidney diseases, a major cause of morbidity and mortality worldwide, dialysis is the standard therapy. However, achieving high blood-filtered volumes time after time and across patient populations requires clinical skills and readily accessible information and data. Nephrologists and nurses must continually re-assess multiple parameters refreshed with each HDF session and consider time-varying clinical status changes, which is daunting in busy dialysis centres.
Dynamic prediction models provide predicted outcome values that can be updated over time for an individual as new measurements become available. Previous approaches to prediction were mainly based on parametric models, but there is a current trend towards using more flexible machine learning approaches. Ensemble methods leverage combinations of parametric regressions and machine learning approaches into one final prediction.
We extend an ensemble method called super learner for (i) dynamically predicting a repeated continuous outcome and (ii) optimizing the prediction for the patients the clinician faces up by combining approaches trained on the personal history of the patient or on an external (i.e., “historical”) cohort. We also propose a new way to validate such personalized prediction models. We illustrate its performance by predicting the convection volume of patients undergoing hemodiafiltration, a specific dialysis technique, in Montréal, Canada.
The personalized dynamic super learner outperformed its candidate learners with respect to median absolute error, calibration-in-the-large, discrimination, and net benefit. We finally discuss the choices and challenges underlying its use and implementation.