Prefer compositions to features in ML-Ops

I had an interesting conversation with a colleague this week. He has built a model that predicts physiological stability for inpatients on intravenous antibiotics using a recent history of vital signs. The purpose is to identify patients who could be switched to oral antibiotics and therefore discharged sooner.

I had also considered that the underlying physiological stability, or its inverse, was a common component of many acute hospital prediction problems. But I wonder if we could take this further when considering how we build features under the ML-Ops paradigm.

That is the features that we seek to build could be framed as concepts and then the tasks built up from those concepts. This is more readily explained by example. Take physiological stability and frailty as concepts. The first would be well summarised by some composition of vital signs. The latter by a composition of co-morbidities and physical fitness. But they could both be reused as components of several different downstream prediction tasks from readiness for discharge, to IV to PO switch, to patient deterioration, to bed demand etc.

So a team managing and maintaining models for a variety of tasks in an acute hospital setting might want to solve those tasks by combining these first order compositions. This would be a more manageable work flow.

And because the first order composition is biologically and clinically grounded then it would be more readily developed independent of the downstream tasks.