Identifying and Intervening on Glucose Patterns in Multivariate Data Using Block-Based Recurrence Quantification Analysis

Abstract

Background:

While automated insulin delivery (AIDs) systems have significantly improved glycemic control for individuals with type 1 diabetes (T1D), there remains a need for identifying and acting upon complex physiologic and behavioral patterns which consistently lead to hypo- and hyperglycemia. Prior methods have lacked the ability to automatically identify and extract patterns across mixed-type multidimensional data (eg, insulin, glucose, activity) without instilling bias from stipulations on time-lagged coupling, pattern length, or pre-defining patterns.

Methods:

We introduce a new pattern-detection technique—Block-based Recurrence Quantification Analysis (BlockRQA)—and preliminary results using BlockRQA in an AID on both in silico and in an outpatient feasibility study. We first introduce the BlockRQA algorithm, which extends Recurrence Quantification Analysis for use in categorical and continuous time-series data, while maintaining interpretable patterns in the domain of interest, in contrast to prior state-of-the-art approaches which require embeddings. Next, we demonstrate the feasibility of utilizing these patterns and BlockRQA with an existing AID system (BlockRQA+AID) to identify and dose for patterns leading to hyperglycemia in individuals with T1D.

Results:

We demonstrate how BlockRQA+AID can improve glucose outcomes in patterns leading to hyperglycemia in silico. And we show real-world results using BlockRQA+AID to reduce hyperglycemic events (>250 mg/dL) via an interim safety analysis of a small outpatient pilot study. For all cases, we show BlockRQA efficiently identifies, aggregates, and scores behavioral patterns which can be targeted for clinical intervention.

Conclusions:

The BlockRQA is a powerful pattern recognition tool that may be used to identify glucose outcome patterns to guide AID dosing.