University of Twente Student Theses
Reality-centered AI for tighter in-hospital glucose control
Rijnhout, Robin (2024) Reality-centered AI for tighter in-hospital glucose control.
|
PDF
1MB |
| Abstract: | BG prediction aids in decision-support for BG regulation. This study assesses patient-specific ARIMA(X) models in an adaptive identification framework, utilizing available data and considering diabetes's time-variant dynamics. A retrospective cohort study was performed. Quantitative, observational, and secondary CGM, insulin, and carbohydrate data were used from the DIABASE database. Data was extracted from hospitalized patients, including both data from during and right before hospitalization. Model identification was conducted to optimize ARIMA(X) parameters. Predictions were made for PH = {30 min, 60 min, 120 min}. An adaptive identification algorithm was applied for continuously changing model orders. Testing was performed using time series cross-validation. Performance metrics were chosen based on clinical relevance: MAE, RMSE, MAPE, CEGA, amount of training data needed for (accurate) prediction, and computational time. This study contained a group of 15 patients (mean age 56.33 years, n = 10 female), with some patients having multiple admissions. The MAE was, respectively, 0.65 (±0.21), 1.17 (±0.23), and 1.88 (±0.24) mmol/l for PH = 30 min, 60 min, and 120 min. The RMSE was, respectively, 0.87 (±0.27), 1.52 (±0.27), and 2.41 (±0.25) mmol/l for PH = 30 min, 60 min, and 120 min. The models trained and tested with intra-extended data reported totals of 100%, 99.87%, and 99.42% in the A+B zones of the CEG for PH = 30 min, 60 min, and 120 min. Cross-BG data showed the most promising predictions that satisfied the different criteria and benchmarks of clinical relevance across various PHs. The satisfaction of the criteria and benchmarks for clinical relevance depended heavily on the different PHs. For this study to have direct implications, a study could be performed on the use of the model as a decision-support system either only as a 30-minute-ahead alarm or in combination with the bolus algorithm of Pérez et al. under strict expert supervision. |
| Item Type: | Essay (Master) |
| Faculty: | TNW: Science and Technology |
| Subject: | 44 medicine, 50 technical science in general |
| Programme: | Technical Medicine MSc (60033) |
| Link to this item: | https://purl.utwente.nl/essays/104739 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page