A Comparative Analysis of Unscented Kalman Filter for Smartphone-Based Multi-Parametric Glucose Prediction

Ece Ayfer; Ömer Atılım Koca; Volkan Kılıç

Makale detayları

A Comparative Analysis of Unscented Kalman Filter for Smartphone-Based Multi-Parametric Glucose Prediction

Yazarlar:

Ece Ayfer | Ömer Atılım Koca | Volkan Kılıç

Yayıncı:

Izmir Katip Celebi University

Ana Başlıklar:

Yapay Zeka (Diğer)

Dergi:

Journal of Artificial Intelligence and Data Science Vol.5 Issue 1, pp.1-11

Belge Tipi:

Makale

Yayın Tarihi:

6/2025

Belge Dili:

İngilizce

Document ID (Org No.):

2001351929

pISSN:

2791-8335

Anahtar kelimeler:

Özet:

 Diabetes mellitus, a chronic disease affecting millions of people worldwide, requires monitoring and management of glucose levels to reduce the risks of hyperglycemia and hypoglycemia. Technological advancements have enabled the development of various digital tools, including continuous glucose monitors CGMs for effective management of this disease. However, these tools only provide alerts after glucose levels exceed critical thresholds, which causes delays in taking necessary precautions. To address this issue, various artificial intelligence AI-based models have been developed to predict glucose levels in advance. Traditional AI approaches, however, often rely on standardized datasets, limiting their ability to achieve the accuracy required for individualized treatment. Therefore, it is crucial to develop personalized prediction models that can be trained using the individual data of patients. Here, this paper introduces a personalized glucose prediction approach that employs a three-parameter unscented Kalman filter UKF to predict future glucose levels using CGM data, as well as basal and bolus insulin values. Experiments on OhioT1DM dataset show the advantage of our proposed approach over the baseline KF and UKF for glucose prediction in terms of Root Mean Square Error. Furthermore, the proposed approach is embedded into a custom-designed cross-platform smartphone application, GlucoThinker Advance, capable of providing offline access to the proposed personalized glucose prediction approach to ensure continuous support without requiring an internet connection.

 Diabetes mellitus, a chronic disease affecting millions of people worldwide, requires monitoring and management of glucose levels to reduce the risks of hyperglycemia and hypoglycemia. Technological advancements have enabled the development of various digital tools, including continuous glucose monitors CGMs for effective management of this disease. However, these tools only provide alerts after glucose levels exceed critical thresholds, which causes delays in taking necessary precautions. To address this issue, various artificial intelligence AI-based models have been developed to predict glucose levels in advance. Traditional AI approaches, however, often rely on standardized datasets, limiting their ability to achieve the accuracy required for individualized treatment. Therefore, it is crucial to develop personalized prediction models that can be trained using the individual data of patients. Here, this paper introduces a personalized glucose prediction approach that employs a three-parameter unscented Kalman filter UKF to predict future glucose levels using CGM data, as well as basal and bolus insulin values. Experiments on OhioT1DM dataset show the advantage of our proposed approach over the baseline KF and UKF for glucose prediction in terms of Root Mean Square Error. Furthermore, the proposed approach is embedded into a custom-designed cross-platform smartphone application, GlucoThinker Advance, capable of providing offline access to the proposed personalized glucose prediction approach to ensure continuous support without requiring an internet connection.

 Diabetes mellitus, a chronic disease affecting millions of people worldwide, requires monitoring and management of glucose levels to reduce the risks of hyperglycemia and hypoglycemia. Technological advancements have enabled the development of various digital tools, including continuous glucose monitors CGMs for effective management of this disease. However, these tools only provide alerts after glucose levels exceed critical thresholds, which causes delays in taking necessary precautions. To address this issue, various artificial intelligence AI-based models have been developed to predict glucose levels in advance. Traditional AI approaches, however, often rely on standardized datasets, limiting their ability to achieve the accuracy required for individualized treatment. Therefore, it is crucial to develop personalized prediction models that can be trained using the individual data of patients. Here, this paper introduces a personalized glucose prediction approach that employs a three-parameter unscented Kalman filter UKF to predict future glucose levels using CGM data, as well as basal and bolus insulin values. Experiments on OhioT1DM dataset show the advantage of our proposed approach over the baseline KF and UKF for glucose prediction in terms of Root Mean Square Error. Furthermore, the proposed approach is embedded into a custom-designed cross-platform smartphone application, GlucoThinker Advance, capable of providing offline access to the proposed personalized glucose prediction approach to ensure continuous support without requiring an internet connection.

 Dünya çapında milyonlarca insanı etkileyen kronik bir hastalık olan diabetes mellitus, hiperglisemi ve hipoglisemi risklerini azaltmak için glikoz seviyelerinin izlenmesini ve yönetilmesini gerektirmektedir. Teknolojik gelişmeler, bu hastalığın etkin yönetimi için sürekli glikoz monitörleri CGMler de dahil olmak üzere çeşitli dijital araçların geliştirilmesini sağlamıştır. Ancak bu araçlar yalnızca glikoz seviyeleri kritik eşikleri aştıktan sonra uyarı vermekte, bu da gerekli önlemlerin alınmasında gecikmelere neden olmaktadır. Bu sorunu çözmek amacıyla, glikoz seviyelerini önceden tahmin etmek için çeşitli yapay zeka YZ tabanlı modeller geliştirilmiştir. Bununla birlikte, geleneksel YZ yaklaşımları genellikle standartlaştırılmış veri kümelerine dayanmakta ve bireyselleştirilmiş tedavi için gereken doğruluğu elde etme yeteneklerini sınırlamaktadır. Bu nedenle, hastaların bireysel verileri kullanılarak eğitilebilen kişiselleştirilmiş tahmin modellerinin geliştirilmesi çok önemlidir. Bu makalede, CGM verilerinin yanı sıra bazal ve bolus insülin değerlerini kullanarak gelecekteki glikoz seviyelerini tahmin etmek için üç parametreli kokusuz Kalman filtresi UKF kullanan kişiselleştirilmiş bir glikoz tahmin yaklaşımı tanıtılmaktadır. OhioT1DM veri kümesi üzerinde yapılan deneyler, Kök Ortalama Kare Hatası açısından glikoz tahmini için önerilen yaklaşımımızın temel KF ve UKFye göre avantajını göstermektedir. Ayrıca, önerilen yaklaşım, internet bağlantısı gerektirmeden sürekli destek sağlamak için önerilen kişiselleştirilmiş glikoz tahmin yaklaşımına çevrimdışı erişim sağlayabilen özel tasarlanmış bir çapraz platform akıllı telefon uygulaması olan GlucoThinker Advancea yerleştirilmiştir.

Makale detayları

A Comparative Analysis of Unscented Kalman Filter for Smartphone-Based Multi-Parametric Glucose Prediction

Yazarlar:

Ece Ayfer | Ömer Atılım Koca | Volkan Kılıç

Yayıncı:

Izmir Katip Celebi University

Ana Başlıklar:

Yapay Zeka (Diğer)

Dergi:

Journal of Artificial Intelligence and Data Science Vol.5 Issue 1, pp.1-11

Belge Tipi:

Makale

Yayın Tarihi:

6/2025

Belge Dili:

İngilizce

Document ID (Org No.):

2001351929

pISSN:

2791-8335

Glucose Prediction

Multi-Parameter

Personalized

Smartphone Application

Unscented Kalman FilterGlikoz Tahmini

Çok-Parametre

Kişiselleştirilmiş

Akıllı Telefon Uygulaması

Kokusuz Kalman Filtresi

Sayı

Makale detayları

Yazarlar: Ece Ayfer | Ömer Atılım Koca | Volkan Kılıç

Yayıncı: Izmir Katip Celebi University

Ana Başlıklar: Yapay Zeka (Diğer)

Dergi: Journal of Artificial Intelligence and Data Science Vol.5 Issue 1, pp.1-11

Belge Tipi: Makale

Yayın Tarihi: 6/2025

Belge Dili: İngilizce

Document ID (Org No.): 2001351929

pISSN: 2791-8335

Sayı

Yazarlar:

Ece Ayfer | Ömer Atılım Koca | Volkan Kılıç

Yayıncı:

Izmir Katip Celebi University

Ana Başlıklar:

Yapay Zeka (Diğer)

Dergi:

Journal of Artificial Intelligence and Data Science Vol.5 Issue 1, pp.1-11

Belge Tipi:

Makale

Yayın Tarihi:

6/2025

Belge Dili:

İngilizce

Document ID (Org No.):

2001351929

pISSN:

2791-8335