摘要
Microvascular complications can adversely impact disease prognosis in adult patients with type 1 diabetes (T1D). Early identification of patients at risk using predictive models through machine learning (ML) can help in T1D management. The objective of current review was to systematically identify and summarize published research on predictive models using ML for microvascular complications (diabetic nephropathy, retinopathy, and neuropathy) in adult T1D patients. Articles were identified from four prior reviews. An additional targeted review of English literature was undertaken in PubMed and Google Scholar from Jan 1, 2016 to May 31, 2019. Following concepts were used in combination in search queries: diabetes, microvascular complication, risk model, and ML. Studies analyzing image data, not developing predictive models, not focusing on an outcome of microvascular complication, not differentiating adult T1D patients, or letters, opinions, and posters were excluded. A total of 6 studies met the eligibility criteria. Four studies developed risk models in T1D patients, whereas two used type of diabetes as a predictor. Diabetic retinopathy, nephropathy and neuropathy were assessed in 3, 3, and 2 studies, respectively. Predictions were based on data from clinical trials (n=2, US:1, Europe:1), EHR (n=3, US:1, Europe:2) and cross-sectional questionnaires (n=1, Iran). Commonly used ML methods included classification and regression tree (CART, including random forest, n=3), support vector machines (n=2), logistic regression (n=2), and neural networks (n=1). Model performance was evaluated by c-statistics (n=3), accuracy (n=2) and confidence intervals (CIs, n=1). Common predictors across complications included age, gender, diabetes duration, BMI, blood pressure, lipid level, and mean or a single A1C value. There is need for developing risk models using ML for microvascular complications, especially neuropathy, in adult T1D patients in the US utilizing contemporary real-world data. Future studies should also evaluate how A1C variability versus a single A1C measure may affect a risk model’s performance.