RESUMO
This article proposes two ensemble neural network-based models for blood glucose prediction at three different prediction horizons-30, 60, and 120 min-and compares their performance with ten recently proposed neural networks. The twelve models' performances are evaluated under the same OhioT1DM Dataset, preprocessing workflow, and tools at the three prediction horizons using the most common metrics in blood glucose prediction, and we rank the best-performing ones using three methods devised for the statistical comparison of the performance of multiple algorithms: scmamp, model confidence set, and superior predictive ability. Our analysis provides a comparison of the state-of-the-art neural networks for blood glucose prediction, estimating the model's error, highlighting those with the highest probability of being the best predictors, and providing a guide for their use in clinical practice.
Assuntos
Glicemia , Diabetes Mellitus Tipo 1 , Algoritmos , Automonitorização da Glicemia , Humanos , Redes Neurais de ComputaçãoRESUMO
Predicting glucose values on the basis of insulin and food intakes is a difficult task that people with diabetes need to do daily. This is necessary as it is important to maintain glucose levels at appropriate values to avoid not only short-term, but also long-term complications of the illness. Artificial intelligence in general and machine learning techniques in particular have already lead to promising results in modeling and predicting glucose concentrations. In this work, several machine learning techniques are used for the modeling and prediction of glucose concentrations using as inputs the values measured by a continuous monitoring glucose system as well as also previous and estimated future carbohydrate intakes and insulin injections. In particular, we use the following four techniques: genetic programming, random forests, k-nearest neighbors, and grammatical evolution. We propose two new enhanced modeling algorithms for glucose prediction, namely (i) a variant of grammatical evolution which uses an optimized grammar, and (ii) a variant of tree-based genetic programming which uses a three-compartment model for carbohydrate and insulin dynamics. The predictors were trained and tested using data of ten patients from a public hospital in Spain. We analyze our experimental results using the Clarke error grid metric and see that 90% of the forecasts are correct (i.e., Clarke error categories A and B), but still even the best methods produce 5 to 10% of serious errors (category D) and approximately 0.5% of very serious errors (category E). We also propose an enhanced genetic programming algorithm that incorporates a three-compartment model into symbolic regression models to create smoothed time series of the original carbohydrate and insulin time series.
Assuntos
Glicemia/análise , Algoritmos , Inteligência Artificial , Diabetes Mellitus Tipo 1 , Humanos , Insulina , EspanhaRESUMO
Effective blood glucose management is crucial for people with diabetes to avoid acute complications. Predicting extreme values accurately and in a timely manner is of vital importance to them. People with diabetes are particularly concerned about suffering a hypoglycemia (low value) event and, moreover, that the event will be prolonged in time. It is crucial to predict hyperglycemia (high value) and hypoglycemia events that may cause health damages in the short term and potential permanent damages in the long term. This paper describes our research on predicting hypoglycemia events at 30, 60, 90, and 120 minutes using machine learning methods. We propose using structured Grammatical Evolution and dynamic structured Grammatical Evolution to produce interpretable mathematical expressions that predict a hypoglycemia event. Our proposal generates white-box models induced by a grammar based on if-then-else conditions using blood glucose, heart rate, number of steps, and burned calories as the inputs for the machine learning technique. We apply these techniques to create three types of models: individualized, cluster, and population-based. They all are then compared with the predictions of eleven machine learning techniques. We apply these techniques to a dataset of 24 real patients of the Hospital Universitario Principe de Asturias, Madrid, Spain. The resulting models, presented as if-then-else statements that incorporate numeric, relational, and logical operations between variables and constants, are inherently interpretable. The True Positive Rate and True Negative Rate metrics are above 0.90 for 30-minute predictions, 0.80 for 60 min, and 0.70 for 90 min and 120 min for the three types of models. Individualized models exhibit the best metrics, while cluster and population-based models perform similarly. Structured and dynamic structured grammatical evolution techniques perform similarly for all forecasting horizons. Regarding the comparison of different machine learning techniques, on the shorter forecasting horizons, our proposals have a high probability of winning, a probability that diminishes on the longer time horizons. Structured grammatical evolution provides advanced forecasting models that facilitate model explanation, modification, and retesting, offering flexibility for refining solutions post-creation and a deeper understanding of blood glucose behavior. These models have been integrated into the glUCModel application, designed to serve people with diabetes.
Assuntos
Glicemia , Hipoglicemia , Aprendizado de Máquina , Humanos , Glicemia/metabolismo , Diabetes Mellitus , Modelos Teóricos , AlgoritmosRESUMO
This dataset provides a collection of Continuous Glucose Monitoring (CGM) data, insulin dose administration, meal ingestion counted in carbohydrate grams, steps, calories burned, heart rate, and sleep quality and quantity assessment ac- quired from 25 people with type 1 diabetes mellitus (T1DM). CGM data was acquired by FreeStyle Libre 2 CGMs, and Fitbit Ionic smartwatches were used to obtain steps, calories, heart rate, and sleep data for at least 14 days. This dataset could be utilized to obtain glucose prediction models, hypoglycemia and hyperglycemia prediction models, and research on the relationships among sleep, CGM values, and the rest of the mentioned variables. This dataset could be used directly from the preprocessed version or customized from raw data. The data set has been used previously with different machine learning algorithms to predict glucose values, hypo, and hyperglycemia and to analyze influences among the features and the quality and quantity of sleep in people with T1DM.
RESUMO
This article proposes the first hardware implemen-tation of a low-power LSTM neural network targeting a wearable medical device designed to predict blood glucose at a 30-minute horizon. This work aims to reduce energy consumption by propos-ing new activation functions that target hardware implementation. On top of this proposal, we also prove there is room for improve-ment in energy consumption by applying neural network optimiza-tions at the algorithmic, such as quantization, and architecture level, LSTM hyperparameters, that consider the target hardware. To validate our proposal, we devise an optimized version of the neural network aimed to be wearable and, therefore, to reduce its energy consumption while preserving its accuracy as much as possible. The hardware is implemented on a Xilinx Virtex-7 FPGA VC707 Evaluation Kit. It is compared with (i) a faithful design of the original neural network implemented on the same evaluation kit, (ii) three state-of-the-art LSTM-based FPGA implementations, and (iii) software implementations running in cutting-edge smartphones:OnePlus NordTM and an Apple iPhone 13 ProTM with artificial in-telligence hardware accelerators. Our proposal consumes between ×1020 and ×7 less energy than the software implementations, being the most efficient system compared to the smartphones. On the other hand, its energy efficiency, measured in GFLOP/J, is between ×2.84 and ×7.82 greater than other state-of-the-art LSTM implementations, proving to be the most suitable implementation for a wearable system for blood glucose prediction.
RESUMO
Background: Sleep quality disturbances are frequent in adults with type 1 diabetes. However, the possible influence of sleep problems on glycemic variability has yet to be studied in depth. This study aims to assess the influence of sleep quality on glycemic control. Materials and methods: An observational study of 25 adults with type 1 diabetes, with simultaneous recording, for 14 days, of continuous glucose monitoring (Abbott FreeStyle Libre system) and a sleep study by wrist actigraphy (Fitbit Ionic device). The study analyzes, using artificial intelligence techniques, the relationship between the quality and structure of sleep with time in normo-, hypo-, and hyperglycemia ranges and with glycemic variability. The patients were also studied as a group, comparing patients with good and poor sleep quality. Results: A total of 243 days/nights were analyzed, of which 77% (n = 189) were categorized as poor quality and 33% (n = 54) as good quality. Linear regression methods were used to find a correlation (r =0.8) between the variability of sleep efficiency and the variability of mean blood glucose. With clustering techniques, patients were grouped according to their sleep structure (characterizing this structure by the number of transitions between the different sleep phases). These clusters showed a relationship between time in range and sleep structure. Conclusions: This study suggests that poor sleep quality is associated with lower time in range and greater glycemic variability, so improving sleep quality in patients with type 1 diabetes could improve their glycemic control.
Assuntos
Diabetes Mellitus Tipo 1 , Transtornos do Sono-Vigília , Humanos , Adulto , Diabetes Mellitus Tipo 1/complicações , Glicemia , Qualidade do Sono , Automonitorização da Glicemia , Inteligência Artificial , Controle GlicêmicoRESUMO
Background: The bacteraemia prediction is relevant because sepsis is one of the most important causes of morbidity and mortality. Bacteraemia prognosis primarily depends on a rapid diagnosis. The bacteraemia prediction would shorten up to 6 days the diagnosis, and, in conjunction with individual patient variables, should be considered to start the early administration of personalised antibiotic treatment and medical services, the election of specific diagnostic techniques and the determination of additional treatments, such as surgery, that would prevent subsequent complications. Machine learning techniques could help physicians make these informed decisions by predicting bacteraemia using the data already available in electronic hospital records. Objective: This study presents the application of machine learning techniques to these records to predict the blood culture's outcome, which would reduce the lag in starting a personalised antibiotic treatment and the medical costs associated with erroneous treatments due to conservative assumptions about blood culture outcomes. Methods: Six supervised classifiers were created using three machine learning techniques, Support Vector Machine, Random Forest and K-Nearest Neighbours, on the electronic health records of hospital patients. The best approach to handle missing data was chosen and, for each machine learning technique, two classification models were created: the first uses the features known at the time of blood extraction, whereas the second uses four extra features revealed during the blood culture. Results: The six classifiers were trained and tested using a dataset of 4357 patients with 117 features per patient. The models obtain predictions that, for the best case, are up to a state-of-the-art accuracy of 85.9%, a sensitivity of 87.4% and an AUC of 0.93. Conclusions: Our results provide cutting-edge metrics of interest in predictive medical models with values that exceed the medical practice threshold and previous results in the literature using classical modelling techniques in specific types of bacteraemia. Additionally, the consistency of results is reasserted because the three classifiers' importance ranking shows similar features that coincide with those that physicians use in their manual heuristics. Therefore, the efficacy of these machine learning techniques confirms their viability to assist in the aims of predictive and personalised medicine once the disease presents bacteraemia-compatible symptoms and to assist in improving the healthcare economy.