Using machine learning to predict severe hypoglycaemia in hospital

Capillary Ketone Level and Future Ketoacidosis Risk in Patients With Type 1 Diabetes Using Sodium-Glucose Cotransporter Inhibitors

OBJECTIVE

We aimed to determine if routine capillary blood ketone testing on well days predicts future diabetic ketoacidosis (DKA) in type 1 diabetes (T1D) using sodium-glucose cotransporter inhibitors (SGLTi).

RESEARCH DESIGN AND METHODS

We examined previously collected data from empagliflozin-assigned participants in a T1D trial that included weekly fasted ketone levels. Over 6-12 months, ketone levels were subdivided into 28-day periods, and the outcome was subsequent adjudicated DKA or severe ketosis.

RESULTS

Among 1,194 participants, 325 had 49 DKA and 568 severe ketosis events. On-treatment maximum ketone levels were higher in the 28 days before an outcome compared with levels in those without an outcome, with area under receiver operating characteristic curve of 0.76 (95% CI 0.71-0.82). Maximum ketone level ≥0.8 mmol/L had sensitivity of 66.0%, specificity of 79.6%, and diagnostic odds ratio of 7.6.

CONCLUSIONS

Routine surveillance of capillary ketone levels in T1D using SGLTi may represent a DKA mitigation strategy and implies a potential threshold for continuous ketone monitoring.

Enhancing severe hypoglycemia prediction in type 2 diabetes mellitus through multi-view co-training machine learning model for imbalanced dataset

Patients with type 2 diabetes mellitus (T2DM) who have severe hypoglycemia (SH) poses a considerable risk of long-term death, especially among the elderly, demanding urgent medical attention. Accurate prediction of SH remains challenging due to its multifaced nature, contributed from factors such as medications, lifestyle choices, and metabolic measurements. In this study, we propose a systematic approach to improve the robustness and accuracy of SH predictions using machine learning models, guided by clinical feature selection. Our focus is on developing long-term SH prediction models using both semi-supervised learning and supervised learning algorithms. Using the action to control cardiovascular risk in diabetes trial, which includes electronic health records for over 10,000 individuals, we focus on studying adults with T2DM. Our results indicate that the application of a multi-view co-training method, incorporating the random forest algorithm, improves the specificity of SH prediction, while the same setup with Naive Bayes replacing random forest demonstrates better sensitivity. Our framework also provides interpretability of machine learning models by identifying key predictors for hypoglycemia, including fasting plasma glucose, hemoglobin A1c, general diabetes education, and NPH or L insulins. The integration of data routinely available in electronic health records significantly enhances our model's capability to predict SH events, showcasing its potential to transform clinical practice by facilitating early interventions and optimizing patient management. By enhancing prediction accuracy and identifying crucial predictive features, our study contributes to advancing the understanding and management of hypoglycemia in this population.

Demographics, medication use, and admission characteristics of patients hospitalized with diabetes in Ontario, Canada: A retrospective cohort study

BACKGROUND

In Canada, one in seven adults has diabetes (i.e., 2.3 million) and the lifetime risk of developing diabetes is approximately 30% by age 65. Although 30% of patients admitted to the hospital have diabetes, data from inpatient hospitalizations for patients with diabetes are lacking, both in Canada and globally.

OBJECTIVE

To validate International Classification of Diseases 10th edition Canadian version (ICD-10-CA) codes for the identification of patients with diabetes, to create a multicenter database of patients with diabetes hospitalized under internal medicine in Ontario, and to determine their baseline characteristics, medication use, and admission characteristics.

STUDY DESIGN

We created a database of people who had diabetes and were hospitalized between 2010 and 2020 at 8 hospitals in Ontario that were part of the General Medicine Inpatient Initiative (GEMINI) hospital data-sharing network. Patients who had diabetes were identified using chart review, based upon either (i) a previous physician diagnosis of diabetes, (ii) a recorded hemoglobin A1c ≥ 6.5% or (iii) outpatient prescription of a diabetes medication preceding the hospitalization. The test characteristics of ICD-10-CA codes for diabetes were evaluated. We compared baseline demographics, medication use and hospitalization details among patients with and without diabetes. For hospitalization details, we collected information on the admission diagnosis, comorbidity index, length of stay, receipt of ICU-level care, and inpatient mortality.

RESULTS

There were 384,588 admissions within the total study cohort, of which 118,987 (30.9%) had an ICD-10-CA diagnosis code of diabetes (E10.x, E11.x, E13.x, E14.x). The sensitivity and specificity of ICD-10-CA diagnostic codes was 95.9% and 98.8%, respectively. Most patients with an ICD-10-CA code for diabetes had a code for type 2 diabetes (93.9%) and a code for type 1 diabetes was rare (6.1%). The mean age was 66.4 years for patients without diabetes and 71.3 years for those with an ICD-10-CA diagnosis code for diabetes. Patients with diabetes had a higher prevalence of hypertension (64% vs. 37.9%), coronary artery disease (28.7% vs. 15.3%), heart failure (24.5% vs. 12.1%) and renal failure (33.8% vs. 17.3%) in comparison to those without diabetes. The most prevalent diabetes medications received in hospital were metformin (43%), DPP4 inhibitors (22.7%) and sulfonylureas (18.8%). The most common reason for admission among patients with diabetes was heart failure (9.0%), and among patients without diabetes was pneumonia (7.8%). Median length of stay was longer for patients with diabetes (5.5 vs. 4.5 days) and in-hospital mortality was similar between groups (6.8% with diabetes vs. 6.5% without diabetes).

IMPORTANCE

Diabetes is one of the most prevalent chronic medical conditions, affecting roughly one third of all patients hospitalized on an internal medicine ward and is associated with other comorbidities and longer hospital stays. ICD-10-CA codes were highly accurate in identifying patients with diabetes. The development of an inpatient cohort will allow for further study of in-hospital practices and outcomes among patients with diabetes.

A machine learning radiomics model based on bpMRI to predict bone metastasis in newly diagnosed prostate cancer patients

OBJECTIVES

To develop and evaluate a machine learning radiomics model based on biparametric magnetic resonance imaging MRI (bpMRI) to predict bone metastasis (BM) status in newly diagnosed prostate cancer (PCa) patients.

METHODS

We retrospectively analyzed bpMRI scans of PCa patients from multiple centers between January 2016 and October 2021. 348 PCa patients were recruited from two institutions for this study. The first institution contributed 284 patients, stratified and randomly divided into training and internal validation cohorts at a 7:3 ratio. The remaining 64 patients were sourced from the second institution and comprised the external validation cohort. Radiomics features were extracted from axial T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI) tumor regions. We developed the radiomics prediction model for BM in the training cohort and validated it in the internal and external validation cohorts. As a benchmark, we trained the logistic regression model with lasso feature reduction (LFR-LRM) in the training cohort and further compared it with Naive Bayes, eXtreme Gradient Boosting (XGboost), Random Forest (RF), GBDT, SVM, Adaboost, and KNN algorithms and validated in both the internal and external cohorts. The performance of several predictive models was assessed by receiver operating characteristic (ROC).

RESULTS

The LFR-LRM model achieved an area under the receiver operating characteristic curve (AUC) of 0.89 (95% CI: 0.822-0.974) and an accuracy of 0.828 (95% CI: 0.713-0.911). The AUC and accuracy in external validation were 0.866 (95% CI: 0.784-0.948) and 0.769 (95% CI: 0.648-0.864), respectively. The RF and XGBoost models outperformed the LFR-LRM, with AUCs of 0.907 (95% CI: 0.863-0.949) and 0.928 (95% CI: 0.882-0.974) and accuracies of 0.831 (95% CI: 0.727-0.907) and 0.884 (95% CI: 0.792-0.946). External validation for these models yielded AUCs and accuracies of 0.911 (95% CI: 0.861-0.966), 0.921 (95% CI: 0.889-0.953), and 0.846 (95% CI: 0.735-0.923) and 0.876 (95% CI: 0.771-0.945), respectively.

CONCLUSIONS

The XGboost machine learning model is more accurate than LFR-LRM for predicting BM in patients with newly confirmed PCa.

Development and Validation of Inpatient Hypoglycemia Models Centered Around the Insulin Ordering Process

BACKGROUND

The insulin ordering process is an opportunity to provide clinicians with hypoglycemia risk predictions, but few hypoglycemia models centered around the insulin ordering process exist.

METHODS

We used data on adult patients, admitted in 2019 to non-ICU floors of a large teaching hospital, who had orders for subcutaneous insulin. Our outcome was hypoglycemia, defined as a blood glucose (BG) <70 mg/dL within 24 hours after ordering insulin. We trained and evaluated models to predict hypoglycemia at the time of placing an insulin order, using logistic regression, random forest, and extreme gradient boosting (XGBoost). We compared performance using area under the receiver operating characteristic curve (AUCs) and precision-recall curves. We determined recall at our goal precision of 0.30.

RESULTS

Of 21 052 included insulin orders, 1839 (9%) were followed by a hypoglycemic event within 24 hours. Logistic regression, random forest, and XGBoost models had AUCs of 0.81, 0.80, and 0.79, and recall of 0.44, 0.49, and 0.32, respectively. The most significant predictor was the lowest BG value in the 24 hours preceding the order. Predictors related to the insulin order being placed at the time of the prediction were useful to the model but less important than the patient's history of BG values over time.

CONCLUSIONS

Hypoglycemia within the next 24 hours can be predicted at the time an insulin order is placed, providing an opportunity to integrate decision support into the medication ordering process to make insulin therapy safer.

Multicomponent prediction of 2-year mortality and amputation in patients with diabetic foot using a random survival forest model: Uric acid, alanine transaminase, urine protein and platelet as important predictors

The current methods for the prediction of mortality and amputation for inpatients with diabetic foot (DF) use only conventional, simple variables, which limits their performance. Here, we used a random survival forest (RSF) model and multicomponent variables to improve the prediction of mortality and amputation for these patients. We performed a retrospective cohort study of 175 inpatients with DF who were recruited between 2014 and 2021. Thirty-one predictors in six categories were considered as potential covariates. Seventy percent (n = 122) of the participants were randomly selected to constitute a training set, and 30% (n = 53) were assigned to a testing set. The RSF model was used to screen appropriate variables for their value as predictors of 2-year all-cause mortality and amputation, and a multicomponent prediction model was established. Model performance was evaluated using the area under the curve (AUC) and the Hosmer-Lemeshow test. The AUCs were compared using the Delong test. Seventeen variables were selected to predict mortality and 23 were selected to predict amputation. Uric acid and alanine transaminase were the top two most useful variables for the prediction of mortality, whereas urine protein and platelet were the top variables for the prediction of amputation. The AUCs were 0.913 and 0.851 for the prediction of mortality for the training and testing sets, respectively; and the equivalent AUCs were 0.963 and 0.893 for the prediction of amputation. There were no significant differences between the AUCs for the training and testing sets for both the mortality and amputation models. These models showed a good degree of fit. Thus, the RSF model can predict mortality and amputation in inpatients with DF. This multicomponent prediction model could help clinicians consider predictors of different dimensions to effectively prevent DF from clinical outcomes .

EMERGE: Evaluating the value of Measuring Random Plasma Glucose Values for Managing Hyperglycemia in the Inpatient Setting

IMPORTANCE

A diagnosis of diabetes is considered when a patient has hyperglycemia with a random plasma glucose ≥200 mg/dL. However, in the inpatient setting, hyperglycemia is frequently non-specific, especially among patients who are acutely unwell. As a result, patients with transient hyperglycemia may be incorrectly labeled as having diabetes, leading to unnecessary treatment, and potential harm.

DESIGN, SETTING, AND PARTICIPANTS

We conducted a multicenter cohort study of patients hospitalized at six hospitals in Ontario, Canada, and identified those with a glucose value ≥200 mg/dL (including standing measurements and randomly drawn). We validated a definition for diabetes using manual chart review that included physician notes, pharmacy notes, home medications, and hemoglobin A1C. Among patients with a glucose value ≥200 mg/dL (11.1 mmol/L), we identified patients without diabetes who received a diabetes medication, and the number who experienced hypoglycemia during the same admission.

MAIN OUTCOMES AND MEASURES

To determine the diagnostic value of using random blood glucose to diagnose diabetes in the inpatient setting, and its impact on patient outcomes.

RESULTS

We identified 328,786 hospitalizations from hospital between 2010 and 2020. A blood glucose value of ≥200 mg/dL (11.1 mmol/L) had a positive predictive value of 68% and a negative predictive value of 90% for a diagnosis of diabetes. Of the 76,967 patients with an elevated glucose value reported, 16,787 (21.8%) did not have diabetes, and of these, 5375 (32%) received a diabetes medication. Hypoglycemia was frequently reported among the 5375 patients that received a diabetes medication, with 1406 (26.2%) experiencing hypoglycemia and 405 (7.5%) experiencing severe hypoglycemia.

CONCLUSIONS AND RELEVANCE

Hyperglycemia in hospital is common but does not necessarily indicate a patient has diabetes. Furthermore, it can lead to treatment with diabetes medications with potential harm. Our findings highlight that clinicians should be cautious when responding to elevated random plasma glucose tests in the inpatient setting.

Natural Language Processing of Nursing Notes: An Integrative Review

Natural language processing includes a variety of techniques that help to extract meaning from narrative data. In healthcare, medical natural language processing has been a growing field of study; however, little is known about its use in nursing. We searched PubMed, EMBASE, and CINAHL and found 689 studies, narrowed to 43 eligible studies using natural language processing in nursing notes. Data related to the study purpose, patient population, methodology, performance evaluation metrics, and quality indicators were extracted for each study. The majority (86%) of the studies were conducted from 2015 to 2021. Most of the studies (58%) used inpatient data. One of four studies used data from open-source databases. The most common standard terminologies used were the Unified Medical Language System and Systematized Nomenclature of Medicine, whereas nursing-specific standard terminologies were used only in eight studies. Full system performance metrics (eg, F score) were reported for 61% of applicable studies. The overall number of nursing natural language processing publications remains relatively small compared with the other medical literature. Future studies should evaluate and report appropriate performance metrics and use existing standard nursing terminologies to enable future scalability of the methods and findings.

Artificial Intelligence for Predicting and Diagnosing Complications of Diabetes

Artificial intelligence can use real-world data to create models capable of making predictions and medical diagnosis for diabetes and its complications. The aim of this commentary article is to provide a general perspective and present recent advances on how artificial intelligence can be applied to improve the prediction and diagnosis of six significant complications of diabetes including (1) gestational diabetes, (2) hypoglycemia in the hospital, (3) diabetic retinopathy, (4) diabetic foot ulcers, (5) diabetic peripheral neuropathy, and (6) diabetic nephropathy.

Hospital Diabetes Meeting 2022

The annual Virtual Hospital Diabetes Meeting was hosted by Diabetes Technology Society on April 1 and April 2, 2022. This meeting brought together experts in diabetes technology to discuss various new developments in the field of managing diabetes in hospitalized patients. Meeting topics included (1) digital health and the hospital, (2) blood glucose targets, (3) software for inpatient diabetes, (4) surgery, (5) transitions, (6) coronavirus disease and diabetes in the hospital, (7) drugs for diabetes, (8) continuous glucose monitoring, (9) quality improvement, (10) diabetes care and educatinon, and (11) uniting people, process, and technology to achieve optimal glycemic management. This meeting covered new technology that will enable better care of people with diabetes if they are hospitalized.

A machine learning-based risk warning platform for potentially inappropriate prescriptions for elderly patients with cardiovascular disease

Potentially inappropriate prescribing (PIP), including potentially inappropriate medications (PIMs) and potential prescribing omissions (PPOs), is a major risk factor for adverse drug reactions (ADRs). Establishing a risk warning model for PIP to screen high-risk patients and implementing targeted interventions would significantly reduce the occurrence of PIP and adverse drug events. Elderly patients with cardiovascular disease hospitalized at the Sichuan Provincial People’s Hospital were included in the study. Information about PIP, PIM, and PPO was obtained by reviewing patient prescriptions according to the STOPP/START criteria (2nd edition). Data were divided into a training set and test set at a ratio of 8:2. Five sampling methods, three feature screening methods, and eighteen machine learning algorithms were used to handle data and establish risk warning models. A 10-fold cross-validation method was employed for internal validation in the training set, and the bootstrap method was used for external validation in the test set. The performances were assessed by area under the receiver operating characteristic curve (AUC), and the risk warning platform was developed based on the best models. The contributions of features were interpreted using SHapley Additive ExPlanation (SHAP). A total of 404 patients were included in the study (318 [78.7%] with PIP; 112 [27.7%] with PIM; and 273 [67.6%] with PPO). After data sampling and feature selection, 15 datasets were obtained and 270 risk warning models were built based on them to predict PIP, PPO, and PIM, respectively. External validation showed that the AUCs of the best model for PIP, PPO, and PIM were 0.8341, 0.7007, and 0.7061, respectively. The results suggested that angina, number of medications, number of diseases, and age were the key factors in the PIP risk warning model. The risk warning platform was established to predict PIP, PIM, and PPO, which has acceptable accuracy, prediction performance, and potential clinical application perspective.

Machine Learning Models for Nocturnal Hypoglycemia Prediction in Hospitalized Patients with Type 1 Diabetes

Nocturnal hypoglycemia (NH) is a dangerous complication of insulin therapy that often goes undetected. In this study, we aimed to generate machine learning (ML)-based models for short-term NH prediction in hospitalized patients with type 1 diabetes (T1D). The models were trained on continuous glucose monitoring (CGM) data obtained from 406 adult patients admitted to a tertiary referral hospital. Eight CGM-derived metrics of glycemic control and glucose variability were included in the models. Combinations of CGM and clinical data (23 parameters) were also assessed. Random Forest (RF), Logistic Linear Regression with Lasso regularization, and Artificial Neuron Networks algorithms were applied. In our models, RF provided the best prediction accuracy with 15 min and 30 min prediction horizons. The addition of clinical parameters slightly improved the prediction accuracy of most models, whereas oversampling and undersampling procedures did not have significant effects. The areas under the curve of the best models based on CGM and clinical data with 15 min and 30 min prediction horizons were 0.97 and 0.942, respectively. Basal insulin dose, diabetes duration, proteinuria, and HbA1c were the most important clinical predictors of NH assessed by RF. In conclusion, ML is a promising approach to personalized prediction of NH in hospitalized patients with T1D.

Machine learning models for diabetes management in acute care using electronic medical records: A systematic review

BACKGROUND

Machine learning (ML) is a subset of Artificial Intelligence (AI) that is used to predict and potentially prevent adverse patient outcomes. There is increasing interest in the application of these models in digital hospitals to improve clinical decision-making and chronic disease management, particularly for patients with diabetes. The potential of ML models using electronic medical records (EMR) to improve the clinical care of hospitalised patients with diabetes is currently unknown.

OBJECTIVE

The aim was to systematically identify and critically review the published literature examining the development and validation of ML models using EMR data for improving the care of hospitalised adult patients with diabetes.

METHODS

The Preferred Reporting Items for Systematic Reviews and Meta Analyses (PRISMA) guidelines were followed. Four databases were searched (Embase, PubMed, IEEE and Web of Science) for studies published between January 2010 to January 2022. The reference lists of the eligible articles were manually searched. Articles that examined adults and both developed and validated ML models using EMR data were included. Studies conducted in primary care and community care settings were excluded. Studies were independently screened and data was extracted using Covidence® systematic review software. For data extraction and critical appraisal, the Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS) was followed. Risk of bias was assessed using the Prediction model Risk Of Bias Assessment Tool (PROBAST). Quality of reporting was assessed by adherence to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guideline. The IJMEDI checklist was followed to assess quality of ML models and the reproducibility of their outcomes. The external validation methodology of the studies was appraised.

RESULTS

Of the 1317 studies screened, twelve met inclusion criteria. Eight studies developed ML models to predict disglycaemic episodes for hospitalized patients with diabetes, one study developed a ML model to predict total insulin dosage, two studies predicted risk of readmission, and one study improved the prediction of hospital readmission for inpatients with diabetes. All included studies were heterogeneous with regard to ML types, cohort, input predictors, sample size, performance and validation metrics and clinical outcomes. Two studies adhered to the TRIPOD guideline. The methodological reporting of all the studies was evaluated to be at high risk of bias. The quality of ML models in all studies was assessed as poor. Robust external validation was not performed on any of the studies. No models were implemented or evaluated in routine clinical care.

CONCLUSIONS

This review identified a limited number of ML models which were developed to improve inpatient management of diabetes. No ML models were implemented in real hospital settings. Future research needs to enhance the development, reporting and validation steps to enable ML models for integration into routine clinical care.

A machine learning-based risk warning platform for potentially inappropriate prescriptions for elderly patients with cardiovascular disease

Potentially inappropriate prescribing (PIP), including potentially inappropriate medications (PIMs) and potential prescribing omissions (PPOs), is a major risk factor for adverse drug reactions (ADRs). Establishing a risk warning model for PIP to screen high-risk patients and implementing targeted interventions would significantly reduce the occurrence of PIP and adverse drug events. Elderly patients with cardiovascular disease hospitalized at the Sichuan Provincial People's Hospital were included in the study. Information about PIP, PIM, and PPO was obtained by reviewing patient prescriptions according to the STOPP/START criteria (2nd edition). Data were divided into a training set and test set at a ratio of 8:2. Five sampling methods, three feature screening methods, and eighteen machine learning algorithms were used to handle data and establish risk warning models. A 10-fold cross-validation method was employed for internal validation in the training set, and the bootstrap method was used for external validation in the test set. The performances were assessed by area under the receiver operating characteristic curve (AUC), and the risk warning platform was developed based on the best models. The contributions of features were interpreted using SHapley Additive ExPlanation (SHAP). A total of 404 patients were included in the study (318 [78.7%] with PIP; 112 [27.7%] with PIM; and 273 [67.6%] with PPO). After data sampling and feature selection, 15 datasets were obtained and 270 risk warning models were built based on them to predict PIP, PPO, and PIM, respectively. External validation showed that the AUCs of the best model for PIP, PPO, and PIM were 0.8341, 0.7007, and 0.7061, respectively. The results suggested that angina, number of medications, number of diseases, and age were the key factors in the PIP risk warning model. The risk warning platform was established to predict PIP, PIM, and PPO, which has acceptable accuracy, prediction performance, and potential clinical application perspective.