Utilizing incremental learning for the prediction of disease outcomes across distributed clinical data: a framework and a case study

Abstract

In this work, we highlight the need of a supervised learning framework for disease predictive modeling across distributed clinical data to overcome the privacy limitations that are introduced by centralized analysis. Towards this direction, a computational framework is proposed, consisting of six incremental learning algorithms that are based on Stochastic Gradient Descent, Naïve Bayes, and Gradient Boosting Trees, to provide new insight on the construction of supervised learning models across clinical data that are stored in multiple locations. The applicability of the proposed framework is demonstrated through a preliminary case study, where a distributed lymphoma prediction model is constructed across private cloud spaces that consist of clinical data from patients that have been diagnosed with primary Sjögren’s Syndrome (pSS). Our results reveal the dominance of the Gradient Boosting Trees, yielding an average accuracy 91.6% and sensitivity 87.5% towards the correct identification of lymphoma cases.