Production-Grade Machine Learning Pipelines For Healthcare Predictive Analytics

Abstract

Predictive analytics leverages past occurrences to answer questions about future events, thereby enabling healthcare organizations to identify exceptional cases and optimize strategies and actions. Healthcare predictive analytics applies predictive analytics to clinical prediction, resource allocation, workflow optimization, risk stratification, treatment response prediction, and other tasks with different objectives and stakeholder perspectives. Successful solutions can be instrumental in improving healthcare outcomes, increasing operational efficiency, and achieving better return on investment, thereby stimulating interest among analysts and clinicians. However, the innovation gap in applying machine learning to healthcare is primarily associated with production-grade models rather than algorithmic novelty. Machine learning is maturing into a viable technology for many forms of prediction, but the surround systems and processes needed for real-world adoption remain largely unsolved. Production-grade pipelines fill this need by providing the end-to-end framework connecting clinicians’ predictive ideas to analytically-driven changes in healthcare delivery. They support experimentation by enabling analysts or researchers to quickly build and test pipelines on small-scale data using diverse models without requiring detailed technical expertise. For the clinical effort to yield substantial benefits, however, the pipelines must be production-grade, thereby allowing a solution to work reliably and continuously once analysts uncover an interesting prediction application.