This figure shows a 16 day time-series matrix from a patient with Ebola sepsis
There are three major (global) clinical transition points along this time matrix and several regional ones.
.
(A major clinical transition state as used here is one of “onset”, “worsening”, “phenotype change” or “recovery”)

It is pivotal to identify global transition states in clinical data especially in time relation to therapy . We have a large cohort with thousands of sepsis cases many having such transition states. Does anyone have advice re: the optimal conventional (non ML or AI) statistical methods to engage complex time matrix data like this and to define the probability of the occurrence of a global transition state.along time series matrix data sets?

So the global transition states are more clear, note object matrix (the figure below) which is derived from the numeric data in the original time matrix

When we convert the same time series matrix data into objects (and then time images) responsive to slope, change magnitude, etc. of the relational time series in the matrix, you can “see” the transition states. Note the term global applies because the transition states develop and evolve across each regional time series matrix responsive each different physiologic system.

Hi llynn, I haven’t done a lot of work in complex time series myself, but I wonder if you have already considered multi-state models? I haven’t had the opportunity to work with them yet, but they essentially fit a series of survival models that define the transition between pre-defined health states. They are essentially an IPD based extension of the (slightly more crude) markov models typically used for these problems in health economics.

Hello.
This is very interesting. Our goal is to maintain the focus throughout on the data by minimizing assumptions so this is something we will explore. Love the quote

“…this multi-state modeling approach builds survival regression models of each of the transitions directly using the individual patient data. It uses the exact time of transitions and therefore is a continuous time state-transition framework, rather than the discrete time framework used in Markov decision-analytic modeling.”