Basal Insulin Titration Using Machine Learning Risk Models

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional self-titration algorithms for basal insulin in diabetes management are inadequate, as they often rely on inaccurate glucose measurements, fail to account for variability in glucose levels, and lack robustness, leading to suboptimal insulin dosing and increased risk of hypoglycemia, especially in patients who struggle to adjust doses between doctor visits.

Innovation Solution

A system and method that utilize timestamped autonomous glucose measurements and insulin records to calculate a glycaemic risk measure and insulin sensitivity factor, employing a supervised machine learning decision model to generate a basal insulin titration schedule and fasting blood glucose profile model, which predicts glucose levels and adjusts the insulin dosing accordingly, reducing the need for frequent glucose measurements and improving safety.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If conventional self-titration algorithms are used, then patients can adjust insulin doses between doctor visits, but the algorithms rely on inaccurate glucose measurements and fail to account for variability, leading to suboptimal dosing and increased hypoglycemia risk

Engineering Contradiction:
Improvepatient ability to adjust insulin dosesVSAvoidaccuracy of insulin dosing
Core Design Contradiction:
Ease of operationVSReliability

Solution Approach 1:

The system implements continuous feedback loops where glucose measurements are constantly monitored and fed back to the machine learning model, which adjusts insulin dosing recommendations in real-time. This closed-loop feedback mechanism ensures that dosing decisions are based on current glucose levels and historical patterns, improving both ease of operation and reliability simultaneously

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system enables patients to self-manage insulin dosing through an automated algorithm that learns from patient-specific data patterns. The machine learning model autonomously analyzes glucose measurements, identifies patterns, and generates dosing recommendations without requiring patient expertise, making the system both easy to operate and reliable

Inventive Principle:
Principle #25Self-service

2Measurement precision

If frequent glucose measurements are taken to improve dosing accuracy, then more data is available for titration decisions, but the burden on patients increases and compliance decreases

Engineering Contradiction:
Improvequality of titration decisionsVSAvoidpatient burden
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system performs preliminary analysis of glucose measurement patterns during an initial learning phase, establishing a personalized baseline before intensive monitoring is required. This preliminary action allows the system to predict glucose trends and reduce the frequency of required measurements while maintaining dosing accuracy

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system implements adaptive measurement frequency where glucose monitoring intensity is adjusted based on current titration needs and patient stability. During stable periods, measurement frequency is reduced to minimize patient burden, while intensifying only when clinically necessary, achieving partial monitoring that maintains precision without excessive burden

Inventive Principle:
Principle #16Partial or excessive action

3Productivity

If aggressive insulin titration is applied to quickly reach target glucose levels, then treatment efficiency improves, but the risk of hypoglycemia increases

Engineering Contradiction:
Improvespeed of achieving glucose targetsVSAvoidhypoglycemia risk
Core Design Contradiction:
ProductivityVSObject-affected harmful factors

Solution Approach 1:

The system employs dynamic titration schedules that adapt the aggressiveness of dose adjustments based on real-time glucose patterns and patient response. The machine learning model continuously learns from patient-specific dynamics, modifying titration speed and magnitude to achieve targets efficiently while automatically reducing aggressiveness when hypoglycemia risk is detected

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system implements safety cushioning mechanisms where conservative dosing recommendations are provided when glucose patterns suggest vulnerability to hypoglycemia. The machine learning model identifies at-risk patterns in advance and applies cushioning adjustments to prevent harmful drops, allowing more aggressive titration during safe periods while maintaining productivity

Inventive Principle:
Principle #11Beforehand cushioning (Prior cushioning)

4Reliability

If manual titration methods are used, then healthcare practitioners maintain control over dosing decisions, but their workload increases and data quality for decision-making deteriorates

Engineering Contradiction:
Improvecontrol over dosing decisionsVSAvoidpractitioner workload
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The system introduces an intelligent intermediary layer between raw glucose data and dosing decisions, where the machine learning model processes data and generates evidence-based recommendations. Healthcare practitioners review and approve these algorithm-generated suggestions, maintaining control while significantly reducing manual workload and improving data quality through automated analysis

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3504652B1Starter kit for basal insulin titration
Publication Date: 2024.02.28 NOVO NORDISK AS
  • EP3504652B1 patent drawingFigure 1
  • EP3504652B1 patent drawingFigure 2
  • EP3504652B1 patent drawingFigure 3A

AI summary

Systems and methods for treating a subject are provided. A first dataset comprising timestamped autonomous glucose measurements of the subject over a first time course is obtained. A second dataset, associated with a standing insulin regimen for the subject over the first time course and comprising insulin medicament records, is also obtained. Each record comprises a timestamped injection event including an amount and type of insulin medicament injected into the subject by an insulin pen. The first and second datasets serve to calculate a glycaemic risk measure and an insulin sensitivity factor of the subject during the first time course, which are used to obtain a basal titration schedule and a fasting blood glucose profile model over a subsequent second time course for the subject. The model predicts the fasting blood glucose level of the subject based upon amounts of basal insulin medicament injected into the subject.