Model based Hypoglycaemia Detection in Patients with Type 1 Diabetes using Electrocardiogram monitoring
Translated title
Modelbaseret hypoglykæmi detektering af patienter med Type 1 diabetes ved hjælp af elektrokardiografi-monitorering
Author
Osmolovski, Juri
Term
4. term
Publication year
2013
Submitted on
2013-05-13
Abstract
Effektiv behandling af type 1-diabetes kræver insulin, men øger risikoen for hypoglykæmi. Dette projekt undersøger, om ændringer i elektrokardiogrammet (EKG) kan bruges til at opdage hypoglykæmi, og udvikler en model, der kan advare om lavt blodsukker. Datasættet omfattede 8 personer med type 1-diabetes, hver med to besøg med insulininduceret hypoglykæmi og to kontrolbesøg med samme insulindosis samt tilførsel af glukose for at forhindre fald i blodglukose; EKG og blodglukose blev målt ved alle besøg. EKG-data blev forbehandlet (bl.a. splineinterpolation, filtrering af EKG-features og normalisering), og en indledende analyse anvendte Mood’s medianetest til at undersøge forskelle. To patientspecifikke logistiske regressionsmodeller blev udviklet: model 1 til at detektere insulininduceret hypoglykæmi og model 2 til at skelne mellem reel hypoglykæmi og blot insulineffekt; den endelige løsning kombinerede modellerne i serie. Modellens ydeevne varierede afhængigt af, hvilke besøg der blev brugt til træning og test, hvilket tyder på overfitting; flere eksperiment- og kontrolbesøg bør inkluderes for at opnå klinisk relevans. Som proof-of-concept blev der desuden udviklet en applikation, der fungerer som hypoglykæmialarm og indeholder handlingsguides ved svær hypoglykæmi. Samlet set peger arbejdet på potentialet ved EKG-baseret hypoglykæmidetektion, men understreger behovet for større datasæt og yderligere validering.
Insulin is essential to manage type 1 diabetes but increases the risk of hypoglycaemia. This project investigates whether changes in the electrocardiogram (ECG) can be used to detect hypoglycaemia and develops a model to warn about low blood glucose. The dataset included 8 people with type 1 diabetes, each with two visits featuring insulin-induced hypoglycaemia and two control visits with the same insulin dose plus glucose to prevent blood glucose from falling; ECG and blood glucose were recorded at all visits. ECG data were preprocessed (including spline interpolation, ECG feature filtering, and normalization), and an initial analysis used Mood’s median test to examine differences. Two patient-specific logistic regression models were built: model 1 to detect insulin-induced hypoglycaemia and model 2 to distinguish true hypoglycaemia from insulin effects; the final approach combined the models in series. Performance varied depending on which visits were used for training and testing, indicating overfitting; additional experiment and control visits are needed before clinical relevance can be achieved. As a proof of concept, an application was also developed to act as a hypoglycaemia alarm and to provide action guides in severe events. Overall, the work shows promise for ECG-based hypoglycaemia detection but highlights the need for larger datasets and further validation.
[This summary has been generated with the help of AI directly from the project (PDF)]
Documents