Development of Machine Learning Models for the Identification of Elevated Ketone Bodies During Hyperglycemia in Patients with Type 1 Diabetes.

Aims: Diabetic ketoacidosis (DKA) is a serious life-threatening condition caused by a lack of insulin, which leads to elevated plasma glucose and metabolic acidosis. Early identification of developing DKA is important to start treatment and minimize complications and risk of death. The aim of the present study is to develop and test prediction model(s) that gives an alarm about their risk of developing elevated ketone bodies during hyperglycemia. Methods: We analyzed data from 138 type 1 diabetes patients with measurements of ketone bodies and continuous glucose monitoring (CGM) data from over 30,000 days of wear time. We utilized a supervised binary classification machine learning approach to identify elevated levels of ketone bodies (≥0.6 mmol/L). Data material was randomly divided at patient level in 70%/30% (training/test) dataset. Logistic regression (LR) and random forest (RF) classifier were compared. Results: Among included patients, 913 ketone samples were eligible for modeling, including 273 event samples with ketone levels ≥0.6 mmol/L. An area under the receiver operating characteristic curve from the RF classifier was 0.836 (confidence interval [CI] 90%, 0.783-0.886) and 0.710 (CI 90%, 0.646-0.77) for the LR classifier. Conclusions: The novel approach for identifying elevated ketone levels in patients with type 1 diabetes utilized in this study indicates that CGM could be a valuable resource for the early prediction of patients at risk of developing DKA. Future studies are needed to validate the results.

Identification of individuals with diabetes who are eligible for continuous glucose monitoring forecasting.

BACKGROUND AND OBJECTIVES: Predicting glucose levels in individuals with diabetes offers potential improvements in glucose control. However, not all patients exhibit predictable glucose dynamics, which may lead to ineffective treatment strategies. We sought to investigate the efficacy of a 7-day blinded screening test in identifying diabetes patients suitable for glucose forecasting. METHODS: Participants with type 1 diabetes (T1D) were stratified into high and low initial error groups based on screening results (eligible and non-eligible). Long-term glucose predictions (30/60 min lead time) were evaluated among 334 individuals who underwent continuous glucose monitoring (CGM) over a total of 64,460,560 min. RESULTS: A strong correlation was observed between screening accuracy and long-term mean absolute relative difference (MARD) (0.661-0.736; p < 0.001), suggesting significant predictability between screening and long-term errors. Group analysis revealed a notable reduction in predictions falling within zone D of the Clark Error Grid by a factor of three and in zone C by a factor of two. CONCLUSIONS: The identification of eligible patients for glucose prediction through screening represents a practical and effective strategy. Implementation of this approach could lead to a decrease in adverse glucose predictions.

Development and Validation of a Machine Learning Model to Predict Weekly Risk of Hypoglycemia in Patients with Type 1 Diabetes Based on Continuous Glucose Monitoring.

Aim: The aim of this study was to develop and validate a prediction model based on continuous glucose monitoring (CGM) data to identify a week-to-week risk profile of excessive hypoglycemia. Methods: We analyzed, trained, and internally tested two prediction models using CGM data from 205 type 1 diabetes patients with long-term CGM monitoring. A binary classification approach (XGBoost) combined with feature engineering deployed on the CGM signals was utilized to predict excessive hypoglycemia risk defined by two targets (time below range [TBR] >4% and the upper TBR 90th percentile limit) of TBR the following week. The models were validated in two independent cohorts with a total of 253 additional patients. Results: A total of 61,470 weeks of CGM data were included in the analysis. The XGBoost models had an area under the receiver operating characteristic curve (ROC-AUC) of 0.83-0.87 (95% confidence interval; 0.83-0.88) in the test dataset. The external validation showed ROC-AUCs of 0.81-0.90. The most discriminative features included the low blood glucose index, the glycemic risk assessment diabetes equation (GRADE), hypoglycemia, the TBR, waveform length, the coefficient of variation and mean glucose during the previous week. This highlights that the pattern of hypoglycemia combined with glucose variability during the past week contains information on the risk of future hypoglycemia. Conclusion: Prediction models based on real-world CGM data can be used to predict the risk of hypoglycemia in the forthcoming week. The models showed good performance in both the internal and external validation cohorts.

Prediction of pancreatic cancer risk in patients with new-onset diabetes using a machine learning approach based on routine biochemical parameters.

OBJECTIVE: To develop a machine-learning model that can predict the risk of pancreatic ductal adenocarcinoma (PDAC) in people with new-onset diabetes (NOD). METHODS: From a population-based sample of individuals with NOD aged >50 years, patients with pancreatic cancer-related diabetes (PCRD), defined as NOD followed by a PDAC diagnosis within 3 years, were included (n = 716). These PCRD patients were randomly matched in a 1:1 ratio with individuals having NOD. Data from Danish national health registries were used to develop a random forest model to distinguish PCRD from Type 2 diabetes. The model was based on age, gender, and parameters derived from feature engineering on trajectories of routine biochemical variables. Model performance was evaluated using receiver operating characteristic curves (ROC) and relative risk scores. RESULTS: The most discriminative model included 20 features and achieved a ROC-AUC of 0.78 (CI:0.75-0.83). Compared to the general NOD population, the relative risk for PCRD was 20-fold increase for the 1 % of patients predicted by the model to have the highest cancer risk (3-year cancer risk of 12 % and sensitivity of 20 %). Age was the most discriminative single feature, followed by the rate of change in haemoglobin A1c and the latest plasma triglyceride level. When the prediction model was restricted to patients with PDAC diagnosed six months after diabetes diagnosis, the ROC-AUC was 0.74 (CI:0.69-0.79). CONCLUSION: In a population-based setting, a machine-learning model utilising information on age, sex and trajectories of routine biochemical variables demonstrated good discriminative ability between PCRD and Type 2 diabetes.

Risk of pancreatic cancer in people with new-onset diabetes: A Danish nationwide population-based cohort study.

BACKGROUND: New onset diabetes (NOD) in people 50 years or older may indicate underlying pancreatic ductal adenocarcinoma (PDAC). The cumulative incidence of PDAC among people with NOD remains uncertain on a population-based level. METHODS: This was a nationwide population-based retrospective cohort study based on the Danish national health registries. We investigated the 3-year cumulative incidence of PDAC in people 50 years or older with NOD. We further characterised people with pancreatic cancer-related diabetes (PCRD) in relation to demographic and clinical characteristics, including trajectories of routine biochemical parameters, using people with type 2 diabetes (T2D) as a comparator group. RESULTS: During a 21-year observation period, we identified 353,970 people with NOD. Among them, 2105 people were subsequently diagnosed with pancreatic cancer within 3 years (0.59%, 95% CI [0.57-0.62%]). People with PCRD were older than people with T2D at diabetes diagnosis (median age 70.9 vs. 66.0 years (P < 0.001) and had a higher burden of comorbidities (P = 0.007) and more prescriptions of medications used to treat cardiovascular diseases (all P < 0.001). Distinct trajectories of HbA1c and plasma triglycerides were observed in PCRD vs. T2D, with group differences observed for up to three years prior to NOD diagnosis for HbA1c and up to two years for plasma triglyceride levels. CONCLUSIONS: The 3-year cumulative incidence of PDAC is approximately 0.6% among people 50 years or older with NOD in a nationwide population-based setting. Compared to T2D, people with PCRD are characterised by distinct demographic and clinical profiles, including distinctive trajectories of plasma HbA1c and triglyceride levels.

Using random forest machine learning on data from a large, representative cohort of the general population improves clinical spirometry references.

INTRODUCTION: Spirometry is associated with several diagnostic difficulties, and as a result, misdiagnosis of chronic obstructive pulmonary disease (COPD) occurs. This study aims to investigate how random forest (RF) can be used to improve the existing clinical FVC and FEV1 reference values in a large and representative cohort of the general population of the US without known lung disease. MATERIALS AND METHODS: FVC, FEV1, body measures, and demographic data from 23 433 people were extracted from NHANES. RF was used to develop different prediction models. The accuracy of RF was compared with the existing Danish clinical references, an improved multiple linear regression (MLR) model, and a model from the literature. RESULTS: The correlation between actual and predicted FVC and FEV1 and the 95% confidence interval for RF were found to be FVC = 0.85 (0.85; 0.86) (p < 0.001), FEV1 = 0.92 (0.92; 0.93) (p < 0.001), and existing clinical references were FVC = 0.66 (0.64; 0.68) (p < 0.001) and FEV1 = 0.69 (0.67; 0.70) (p < 0.001). Slope and intercept for the RF models predicting FVC and FEV1 were FVC 1.06 and -238.04 (mL), FEV1: 0.86 and 455.36 (mL), and for the MLR models, slope and intercept were FVC: 0.99 and 38.56 39 (mL), and FEV1: 1.01 and -56.57-57 (mL). CONCLUSIONS: The results point toward machine learning models such as RF have the potential to improve the prediction of estimated lung function for individual patients. These predictions are used as reference values and are an important part of assessing spirometry measurements in clinical practice. Further work is necessary in order to reduce the size of the intercepts obtained through these results.

Using machine learning to design a short test from a full-length test of functional health literacy in adults-The development of a short form of the Danish TOFHLA.

INTRODUCTION: Patients are compelled to become more involved in shared decision making with healthcare professionals in the self-management of chronic disease and general adherence to treatment. Therefore, it is valuable to be able to identify patients with low functional health literacy so they can be given special instructions about the management of chronic disease and medications. However, time spent by both patients and clinicians is a concern when introducing a screening instrument in the clinical setting, which raises the need for short instruments for assessing health literacy that can be used by patients without the involvement of healthcare personnel. This paper describes the development of a short version of the full-length Danish TOFHLA (DS-TOFHLA) that is easily applicable in the clinical context and where the use does not require a trained interviewer. MATERIALS AND METHODS: Data were collected as a part of a large-scale telehomecare project (TeleCare North), which was a randomized controlled trial that included 1225 patients with chronic obstructive pulmonary disease. The DS-TOFHLA was developed solely using an algorithm-based selection of variables and multiple linear regression. A multiple linear regression model was developed using an exhaustive search strategy. RESULTS: The exhaustive search showed that the number of items in the full-length TOFHLA could be reduced from 17 numeracy items and 50 reading comprehension items to 20 reading comprehension items while maintaining a correlation of r = 0.90 between the scores from full-length and short versions. A generic model-based approach was developed, which is suitable for development of short versions of the TOFHLA in other languages, including the original American version. CONCLUSIONS: This study demonstrated how a generic model-based approach could be applied in the development of a short version of the TOFHLA, thereby reducing the 67 items to 20 items in the short version. Furthermore, this study showed that the inclusion of numeracy items was not necessary. The development of the DS-TOFHLA presents an opportunity to reliably identify patients with inadequate functional health literacy in approximately 5 minutes without involvement of healthcare personnel. The approach may be used in the development of short versions of any scaling questionnaire.

Optimal Data Collection Period for Continuous Glucose Monitoring to Assess Long-Term Glycemic Control: Revisited.

BACKGROUND AND OBJECTIVE: It is not clear how the short-term continuous glucose monitoring (CGM) sampling time could influence the bias in estimating long-term glycemic control. A large bias could, in the worst case, lead to incorrect classification of patients achieving glycemic targets, nonoptimal treatment, and false conclusions about the effect of new treatments. This study sought to investigate the relation between sampling time and bias in the estimates. METHODS: We included a total of 329 type 1 patients (age 14-86 years) with long-term CGM (90 days) data from three studies. The analysis calculated the bias from estimating long-term glycemic control based on short-term sampling. Time in range (TIR), time above range (TAR), time below range (TBR), correlation, and glycemic target classification accuracy were assessed. RESULTS: A sampling time of ten days is associated with a high bias of 10% to 47%, which can be reduced to 4.9% to 26.4% if a sampling time of 30 days is used (P < .001). Correct classification of patients archiving glycemic targets can also be improved from 81.5% to 91.9 to 90% to 95.2%. CONCLUSIONS: Our results suggest that the proposed 10-14 day CGM sampling time may be associated with a high correlation with three-month CGM. However, these estimates are subject to large intersubject bias, which is clinically relevant. Clinicians and researchers should consider using assessments of longer durations of CGM data if possible, especially when assessing time in hypoglycemia or while testing a new treatment.

Body Composition Prediction-BOMP: A New Tool for Assessing Fat and Lean Body Mass.

BACKGROUND: Pragmatic and easy-to-use alternatives to estimating body composition, such as lean body mass and fat mass, could be valuable tools for assessing the risk of diabetes or other metabolic diseases. Previous work has shown how demographic and anthropometric data could be used in a neural network to estimate body composition with high precision. However, there is still a need for a publicly available and user-friendly format before these results can have clinical impact. METHODS: We used data from 18 430 NHANES participants and stepwise linear regression with inclusion of linear, interactions, and quadratic terms to model lean body and fat mass. HTML and Javascript was used to develop a webapp as a frontend of the model. RESULTS: The models had a correlation cofficent R = 0.99-0.98 (P < .001) withstandard error of estimate [SEE] = 2.07-2.05. CONCLUSIONS: The results indicate that it is possible to develop a "white-box" model with high precision.The proof of concept webapp is available as open source under the MIT license.

Long-term glucose-lowering effect of intermittently scanned continuous glucose monitoring for type 1 diabetes patients in poor glycaemic control from Region North Denmark: An observational real-world cohort study.

BACKGROUND: Lowering glucose levels is a complex task for patients with type 1 diabetes, and they often lack contact with health care professionals. Intermittently scanned continuous glucose monitoring (isCGM) has the potential to aid them with blood glucose management at home. The aim of this study was to investigate the long-term effect of isCGM on HbA1c in type 1 diabetes patients with poor glycaemic control in a region-wide real-world setting. METHODS: All patients with type 1 diabetes receiving an isCGM due to poor glycaemic control (≥70 mmol/mol [≥8.6%]) in the period of 2020-21 in Region North Denmark ("T1D-CGM") were compared with all type 1 diabetes patients without isCGM ("T1D-NOCGM") in the same period. A multiple linear regression model adjusted for age, sex, diabetes duration and use of continuous subcutaneous insulin infusion was constructed to estimate the difference in change from baseline HbA1c between the two groups and within subgroups of T1D-CGM. RESULTS: A total of 2,527 patients (T1D-CGM: 897; T1D-NOCGM: 1,630) were included in the study. The estimated adjusted difference in change from baseline HbA1c between T1D-CGM vs T1D-NOCGM was -5.68 mmol/mol (95% CI: (-6.69 to -4.67 mmol/mol; p<0.0001)). Older patients using isCGM dropped less in HbA1c. CONCLUSIONS: Our results indicate that patients with type 1 diabetes in poor glycaemic control from Region North Denmark in general benefit from using isCGM with a sustained 24-month improvement in HbA1c, but the effect on HbA1c may be less pronounced for older patients.

Is predicted body-composition and relative fat mass an alternative to body-mass index and waist circumference for disease risk estimation?

BACKGROUND AND AIMS: New methods to estimate body-composition have recently been proposed, but their relation to diseases, such as diabetes and coronary heart disease, needs further investigation. The purpose of this study was to investigate the association between proposed prediction of body-composition (PBC); Relative Fat Mass (RFM), Body Mass Index (BMI), Waist Circumference (WC) and disease. METHODS: In a cross-sectional cohort (NHANES) the association between the four body measures and diabetes, high blood pressure, coronary heart disease, cancer, arthritis, and hospitalization were assessed. A total of 13,348 people was included in this study. Receiver operating characteristic (ROC), Area Under Curve (AUC) and statistical testing were used to evaluate the differences. RESULTS: PBC/RFM had significant higher AUC than BMI or WC for diabetes, high blood pressure, hospitalization, and arthritis. PBC had a significant higher AUC than RFM, BMI, WC for Cancer and coronary heart disease. CONCLUSIONS: RFM and PBC could be a better indicator to distinguish amongst people with a risk of diseases compared to traditional measures such as BMI and WC. However, future studies need to investigate the longitudinal association between RFM, PBC and the risk of disease development to assess if these measures are better suited for risk-stratification.

A novel approach for predicting acute hospitalizations among elderly recipients of home care? A model development study.

BACKGROUND: Frail elderly are at high risk of hospitalizations and have a complex pattern of risk factors that makes it hard to foresee potential needs for additional treatment and care. Machine learning algorithms are potentially well-suited to discover hidden patterns in registrations that are routinely made across sectors. OBJECTIVE: To investigate predictors and performance of machine learning algorithms designed to predict acute hospitalizations in elderly recipients of home care services. MATERIALS AND METHODS: A development study based on a retrospective social sector cohort with 1,282 participants was designed. Included subjects were aged 65 or older and received home care services in Aalborg Municipality at least once a week from 1/1-2016 to 31/12-2017. Data were collected from a newly developed triage tool in combination with administrative and clinical data routinely collected in the Danish healthcare and social care sector. 857 predictors were tested and evaluated based on the area under the precision-recall curve (PR-AUC). The data was divided into a 70/30 training and test split with 5-fold cross-validation. A sliding window approach combining random under-sampling with a boosting algorithm (RUSBoost) was applied with a standard logistic regression included for comparison. RESULTS: The logistic regression achieved a PR-AUC of 0.01 (CI 0.00; 0.01) while the PR-AUC was 0.71 (CI 0.56; 0.76) for the RUSBoost algorithm. Four of the five most important citizen-level features used to accurately predict an acute hospitalization was the total number of services provided by the municipality to the citizen, the number of personal care registrations as well as number of medication handlings and nutritional registrations. A final important predictor was the number of physical complaints derived from the triage tool. CONCLUSIONS: Municipalities routinely collect valuable administrative and clinical data that can be used for early prediction of acute hospitalizations. However, future studies are needed to validate the results.

Hypoglycemia event prediction from CGM using ensemble learning.

This work sought to explore the potential of using standalone continuous glucose monitor (CGM) data for the prediction of hypoglycemia utilizing a large cohort of type 1 diabetes patients during free-living. We trained and tested an algorithm for the prediction of hypoglycemia within 40 minutes on 3.7 million CGM measurements from 225 patients using ensemble learning. The algorithm was also validated using 11.5 million synthetic CGM data. The results yielded a receiver operating characteristic area under the curve (ROC AUC) of 0.988 and a precision-recall area under the curve (PR AUC) of 0.767. In an event-based analysis for predicting hypoglycemic events, the algorithm had a sensitivity of 90%, a lead-time of 17.5 minutes and a false-positive rate of 38%. In conclusion, this work demonstrates the potential of using ensemble learning to predict hypoglycemia, using only CGM data. This could help alarm patients of a future hypoglycemic event so countermeasures can be initiated.

Using Case-Based Reasoning in a Learning System: A Prototype of a Pedagogical Nurse Tool for Evidence-Based Diabetic Foot Ulcer Care.

BACKGROUND: Currently, evidence-based learning systems to increase knowledge and evidence level of wound care are unavailable to wound care nurses in Denmark, which means that they need to learn about diabetic foot ulcers from experience and peer-to-peer training, or by asking experienced colleagues. Interactive evidence-based learning systems built on case-based reasoning (CBR) have the potential to increase wound care nurses' diabetic foot ulcer knowledge and evidence levels. METHOD: A prototype of a CBR-interactive, evidence-based algorithm-operated learning system calculates a dissimilarity score (DS) that gives a quantitative measure of similarity between a new case and cases stored in a case base in relation to six variables: necrosis, wound size, granulation, fibrin, dry skin, and age. Based on the DS, cases are selected by matching the six variables with the best predictive power and by weighing the impact of each variable according to its contribution to the prediction. The cases are ranked, and the six cases with the lowest DS are visualized in the system. RESULTS: Conventional education, that is, evidence-based learning material such as books and lectures, may be less motivating and pedagogical than peer-to-peer training, which is, however, often less evidence-based. The CBR interactive learning systems presented in this study may bridge the two approaches. Showing wound care nurses how individual variables affect outcomes may help them achieve greater insights into pathophysiological processes. CONCLUSION: A prototype of a CBR-interactive, evidence-based learning system that is centered on diabetic foot ulcers and related treatments bridges the gap between traditional evidence-based learning and more motivating and interactive learning approaches.