Reynolds, Rebecca

Wake, Debbie

Lindsay, Robert

Manataki, Areti

Kirkwood, Jasmine R.

Medical Research Scotland

MyWay Digital Health

2026-05-26T15:43:58Z

2026-05-26

INTRODUCTION: Gestational diabetes mellitus (GDM) is high blood glucose that is first recognised during pregnancy and is one of the most common pregnancy complications. GDM is managed through lifestyle modification and may require pharmacological therapy to normalise blood glucose levels. Self monitoring blood glucose readings are reviewed at clinical appointments, where treatment may be escalated. Current care in Scotland is time-consuming and does not utilise the potential of risk-stratification and digital tools. This thesis proposes an alternative method of care that does so by risk stratifying women with GDM based on their need for insulin and remotely monitoring them through a digital tool. METHODS & RESULTS: This is an interdisciplinary mixed-methods thesis, the quantitative methods focusing on the development of a need for an insulin treatment prediction model, and the qualitative methods focusing on the design of a digital tool to support both women and clinicians in the management of GDM. To aid with readability, the results chapters’ methods and results are reported together. In the first result chapter I report the findings of a scoping review of 17 studies describing 44 machine learning models to predict the need for pharmacological therapy in GDM from four electronic databases between 1st July 2007 and 31st August 2024. All published literature were binary classifiers with 61.4% (27/44) of models predicting need for any pharmacological therapy and 38.6% (17/44) predicting need for insulin. Models had a median area under the receiver operator curve (AUROC) of 0.75. Common clinical variables were found to be predictors, such as history of GDM, gestational week at GDM diagnosis, pregestational body mass index (BMI), maternal age, HbA1c, fasting – and 1-hr-glucose from a 75g oral glucose tolerance test (OGTT), with logistic regression being a popular algorithm. There was a lack of external validation and clinical implementation. In the following two results chapters, I present my quantitative analyses. From a database of 30,666 pregnancy episodes in Greater Glasgow and Clyde, I selected 10,694 pregnancy episodes that had a booking date and received care between 1st April 2022 to 31st December 2023. The selected data were cleaned, and a cohort of singleton pregnancies that were complicated by GDM (10.4%, 1,109) was identified alongside a complementary non-GDM cohort (89.6%, 9,585). The maternal characteristics and pregnancy outcomes were described. In the descriptive analysis of 10,694 singleton pregnancies in Glasgow, the rate of GDM was high (10.4%, 1,109) despite only screening women with a BMI ≥ 35kg/m2. Women with pregnancies complicated by GDM were older (GDM: mean[SD] 32 [5.3], non-GDM: 31 [5.4] years, p<0.001), had a higher BMI (32 [7.4], 27 [5.8] kg/m2, p<0.001), and more likely to live in the most deprived areas (Scottish Index of Multiple Deprivation quantile 1, 45.4%(504), 39.5% (3,789), compared to those without GDM. Among women with GDM, those managing it through diet were younger than those who required pharmacological treatment (Diet: 32 [5.4], Metformin: 33 [4.8], Insulin: 33 [5.3] years, p = 0.001). Women requiring insulin had a higher BMI (32 [7.2], 32 [7.6], 35 [7.5]kg/m2, p<0.001). HbA1c at booking was higher in the insulin-treated GDM (35 [3.3] mmol/L, 35 [3.1] mmol/L, 36 [3.8] mmol/L, p=0.038) as were fasting OGTT results (5.2 [0.5] , 5.3 [0.46] , 5.6[0.57] mmol/L, p<0.001) which was taken earlier (26 [6], 24 [5.6] ,23 [5.6] gestational weeks, p<0.001) . Caesarean birth rates were higher in GDM pregnancies (53.5%(593) vs. 40.0%(3,834), p < 0.001), with insulin-treated GDM showing increased elective Caesarean rates (29.8%(239), 35.9%(79), 41.4%(36), p=0.033). Large for gestational age (LGA) was more common in GDM-complicated pregnancies (15.0%(166), 9.0%(841), p<0.001). Neonatal unit admissions were higher in GDM pregnancies compared to non-GDM (41.2%(158), 9.4%(899), p<0.001). Although admissions were more frequent in GDM managed with insulin, the difference was not statistically significant (13.8%(111), 14.1%(31), 18.4%(16), p=0.513). The penultimate results chapter uses the results of the scoping review and statistical analysis to direct the development of machine learning models for predicting insulin for GDM. Two algorithms (logistic regression and Classification and Regression Tree (CART)) using univariate feature selection and least absolute shrinkage and selection operation (LASSO) were compared. Synthetic Minority Oversampling Technique (SMOTE) was used to address the unbalanced data. Using a complete case analysis, 996 GDM complicated pregnancy episodes were included in the dataset which split into 70% for training and 30% for testing. The training performance was assessed using 10-fold cross-validation, and the final model performance was validated on the unseen test data. Both models generalised well, and overall were sensitive but not specific. The logistic regression had a mean 10-fold cross validation AUROC in train data of 0.79 [0.07] and AUROC of 0.71 on the unseen data. The CART model had mean 10-fold cross validation AUROC in train data of 0.78 [0.02] and AUROC of 0.71 on the unseen data. In the final results Chapter, a user-centred digital tool, ‘MyGDM’, was designed from ethnographic observation and 11 semi-structured interviews with end-users (6 healthcare professionals and 5 women with GDM). The initial design was evaluated in feedback sessions with 31 participants (17 healthcare professionals, 14 researchers) and 13 questionnaires with women with GDM. MyGDM has a clinical dashboard linked to a patient-facing app, aiming to enhance clinical workflows, identify off-target blood glucose levels and provide GDM-specific information. The initial design drew upon existing literature and insights from 11 semi-structured interviews conducted with HCPs and women with GDM. In a survey of 13 women with GDM, every participant reported that the tool would fit well into their lifestyles and aid in managing their GDM. Educational resources, along with the ‘request a call’ feature, were particularly well received, with 61.5% (8 out of 13) and 69.2% (9 out of 13) indicating they were very likely or likely to utilise these options, respectively. End-user evaluations of the interactive design were favourable, confirming that it effectively met their needs. CONCLUSION: GDM care could be personalised through risk-stratification and digital tools. There is an opportunity to translate the key concepts explored in this thesis into a clinically implementable model of care for GDM.

https://era.ed.ac.uk/handle/1842/44750

https://doi.org/10.7488/era/7265

en

KIRKWOOD, J. R., DICKSON, J., MANATAKI, A., LINDSAY, R. S., WAKE, D. J. & REYNOLDS, R. M. 2025a. User Needs Assessment for Gestational Diabetes Care and Management Digital Tool: A Qualitative Study. Intelligent Health Systems – From Technology to Data and Knowledge. Medical Informatics Europe

KIRKWOOD, J. R., DICKSON, J., STEVENS, M., MANATAKI, A., LINDSAY, R. S., WAKE, D. J. & REYNOLDS, R. M. 2024. The User-Centered Design of a Clinical Dashboard and Patient-Facing App for Gestational Diabetes. J Diabetes Sci Technol, 0(0)

KIRKWOOD, J. R., GALLOWAY, N., LINDSAY, R. S., MANATAKI, A., WAKE, D. J. & REYNOLDS, R. M. 2025b. The use of machine learning to predict pharmacological therapy in gestational diabetes: A scoping review. Diabetic Medicine

gestational diabetes mellitus

GDM

managing blood sugar levels

remote monitoring

medication risk prediction model

machine learning models

body mass index

logistic regression

personalised approach

Towards a data-driven personalised approach to gestational diabetes care

Thesis

Doctoral

PhD Doctor of Philosophy