Prediction and prevention of hypoglycaemic events in type-1 diabetic patients using machine learning

Related work

A recent report on artificial intelligence applications for diabetes management ¹² points out an increasing interest in studies dealing with the anticipation of BG prediction and the personalization of decision support systems (DSSs) for the management of diabetes. Although most research efforts have been focused on the prediction of BG values, they have immediate application in avoiding adverse glycaemic events, such as hypoglycaemia. ¹³ These proposals include, among others, studies based on artificial neural networks (ANNs), ¹⁴ support vector machines (SVMs), ¹⁵ and evolutionary algorithms. ¹⁶

Next, we review some of the relevant studies focused on the use of machine learning techniques to detect future hypoglycaemic events. A study presented by Zhang ¹⁷ used data from 3116 patients in the intensive care unit to build a classification tree. The study used information of glucose measurement along with insulin infusion rate to predict future hypoglycaemic events in a 1-h window. Results showed a potential tool to provide early warnings of hypoglycaemia events. Similar methodological approach but using CGM data was used by Eljil et al. ¹⁸ They reported accurate results for prediction horizons of 30 min. Later, Georga et al. ¹⁹ presented a glucose model based on SVM for the prediction of hypoglycaemic events of 10 T1D patients under free-living conditions. The input features were the recent glucose profile, meals, insulin intake, and physical activities. This study provided hypoglycaemia predictions using 30- and 60-min prediction horizons. Dubosson et al. ²⁰ studied postprandial hypoglycaemia prediction for T1D patients on MDI therapy. The study presented poor postprandial hypoglycaemia prediction results for two different datasets: (1) a cohort of 10 real patients (CGM data, insulin injections, carbohydrate and alcohol consumption, and physical activities) and (2) CGM data from 10 virtual patients from the UVa/Padova T1D simulator along with insulin and meal information. They concluded that the features glucose, insulin dose, and carbohydrate estimations are insufficient to offer reliable predictions.

A DSS for diabetes management has been evaluated in real patients recently. ²¹ The system is composed by different modules intended to reduce glycaemic variability in patients with T1D. The proposed strategy improves BG control by advising patients regarding some activities (meals and exercise), and also tunes individuals parameters related with insulin treatment based on hypo- and hyperglycaemic risks. Results obtained showed that the incorporation of the DSS in insulin therapy improved both glycaemic control and also patients’ safety, by reducing the time spent in the hypoglycaemic range.

The study presented here proposes a system based on various methods of artificial intelligence for the prediction and prevention of hypoglycaemic events in combination with data mining algorithms for the classification of glycecaemic control profiles for patients with T1D. Details of the techniques and datasets considered for the analysis are presented on the following sections, as well as the results obtained for each approach.

Overall system description

The application of machine learning algorithms to predictive modelling encompasses countless methodologies that have arisen to deal with multiple applications and computational requirements. There is no general purpose, universal machine learning approach for every predictive modelling problem. The fittest approach for each problem depends on the specific requirements, data, and focus of the problem being addressed. This study proposes the application of a set of machine learning techniques to tackle patient safety in diabetes management.

Applying these machine learning methods to predictive modelling requires various preprocessing stages. As a first step, the targeted experimental datasets, consisting of the information from MDI or CSII-CGM therapy and the data from a fitness tracker band, was subjected to an exploratory analysis. The analysis provided basic information about the available data types, the data distribution, the number of observations and features, and the correlations between variables. Next, data cleaning was applied to fix structural errors in pump data values such as those coming from different pump firmware, clean suspicious outliers indicating erroneous data, and drop observations that contained missing values. Finally, we performed a feature engineering phase, which encompasses tasks to provide additional value to the dataset. The most representative transformations involve the following physiological models:

d C 1 ( t ) d t = u ( t ) − K D I A C 1 ( t ) d C 2 ( t ) d t = K D I A ( C 1 ( t ) − C 2 ( t ) ) I O B ( t ) = C 1 ( t ) + C 2 ( t )

(1)

where t is the time instant, the compartments C 1 and C 2 have initial values of C 1 ( 0 ) = C 2 ( 0 ) = 0 , u(t) is the insulin dose, and K D I A = 0 . 0195 is a constant related to the duration of insulin action (DIA), which characterises the patient’s insulin activity dynamics.

R a ( t ) = C i n C b i o t e ( − t / t m a x , G ) t max , G 2

(2)

where t it the time instant, C i n is the amount of carbohydrates ingested in a meal (in grammes), C b i o = 0 . 8 is the carbohydrate bioavailability, and t m a x , G = 50 (min) denotes the time of the maximum appearance rate of glucose in the accessible glucose compartment.

A O B ( t ) = steps ( t ) e ( − k s t )

(3)

where t is the time instant, steps ( t ) is the total number of steps, and k s = 0 . 0115 is a constant related to the duration of the effects of physical activity on BG control. ²⁴ One AOB curve is obtained for each time instant, and the final value of the AOB represents the superposition of all the curves.

Once we have conducted an effective feature engineering phase and tapped into the domain knowledge, we present the core system, which introduces four powerful mechanisms to manage possible adverse events. Table 1 depict the four modules used in this work and the following sections will detail each of the approaches.

OPEN IN VIEWER

Table 1.

Summary of objectives, databases, and algorithms of the models that compose the system.

Objective	Window	Methodology	Database
Continuous prediction	1 h	Grammatical evolution	1, 2, and 3
Postprandial prediction	4 h	Support vector machine	1
Nocturnal prediction	6 h	Artificial neural network	2
Daily profiles classification	24 h	Data mining	1 and 3

Grammatical evolution for mid-term continuous prediction

The ability to provide an early warning of ineffective or poor insulin treatment, which could lead to adverse glycaemic levels, is of considerable interest. Short- to mid-term continuous predictions of BG levels are possible, but challenging due to various factors, including variability, delays of meal and insulin absorption, and lagging BG measurements. Methods appropriately addressing such delays can provide accurate forecasts, which in turn could offer sufficient time to anticipate BG measurements and act to prevent adverse glycaemic events.

Continuous prediction is managed by a module based on grammatical evolution (GE). The evolutionary methodology aims to identify the best model within a search space of candidate models, where the best model is the one that optimises a fitness function for the problem of BG prediction. The GE module generates personalised models aiming to prevent any adverse glycaemic event. Thus, the continuous predictions provide a trade-off between the detection of hypoglycaemia (BG ⩽ 70) and hyperglycaemia (BG ⩾ 180) events. ²⁵

GE implements a modular design based on a problem-specific free-context grammar and a fitness function. The GE grammar consists of a set of rules that incorporate the information relating to the addressed problem and defines the structure of the candidate models. The basic flow of the algorithm is an iterative process, which updates a set of solutions named the population. In each iteration, all members of the population are mapped by the grammar and assessed according the fitness function. A new population is generated, selecting the best adapted individuals in the present population. Some of these individuals remain unaltered in the next generation (elitism), and the remainder participate in the generation of the new offspring, applying transformative operations to the solutions (crossover and mutation).

The GE module was based on a version of a java evolutionary computation library (JECO). ²⁶ It implements a fitness function that consists of a glucose-specific mean squared error. ²⁷ The adopted grammar involves solutions that combine four sub-models associated with the BG, carbohydrates, insulin, and circadian rhythm. The grammar and the remaining hyper-parameters are detailed in Contreras et al. ²⁸ Finally, patient data were divided into a dataset for training (66.6%) and one for validation (33.3%).

The system generates continuous prediction models using a prediction horizon of 60 min which provide punctual BG predictions based on the following semi-continuous inputs from the preprocessing stage:

Support vector classifier for postprandial risk assessment

The prediction of hypoglycaemic events when a patient announces a meal allows the assessment of the impact of the insulin bolus on the postprandial response and makes it possible to optimise the bolus to achieve safer dosages. The prediction architecture presented in this work is specialised to postprandial hypoglycaemia, which enables bolus re-estimation for insulin pump and MDI users. Previous approaches have been restricted mainly to nocturnal hypoglycaemia and 30–45 min prediction horizons. ²⁹ We developed models for the prediction of hypoglycaemia within 240 min after a meal, given that the patients had a fairly regular meal intake.

Postprandial risk can be assessed using the result of a class prediction. This class can be determined by a support vector classifier (SVC) using relevant features extracted from the available inputs. SVC is a supervised learning algorithm, which has been widely employed in many classification and regression problems. Given a set of N training examples with a feature vector x and labels vector y ∈ − 1 , 1 , the parameters ω , b represent the model of the separating hyperplane. ³⁰ This model is obtained by searching for the hyperplane equation that maximises the margin between two linearly separable classes. When the classes are not linearly separable, a mapping function is employed from the input space to a k-dimensional space where the classes are linearly separable. This mapping function is also known as the kernel. Typical kernels used in SVC are polynomial function, radial basis function (RBF), and hyperbolic tangent function. There is no consensus on which method is best for selecting the most appropriate kernel function for a given problem. We picked the RBF kernel, which provided the best results compared with the other two. A Python implementation based on the Scikit-learn package was used in this module.

In this work, we extracted an initial set of features from retrospective data, specifically the CGM signal and pump information. The selected features for postprandial hypoglycaemia were as follows:

G R = a b s ( B G [ k − 12 ] − B G [ k ] )

where B G [ k ] is the last CGM value and BG[k – 12] is the CGM value one hour before.

Several scenarios were employed as models for each patient, and after testing, these were compared to select the best personalised model according to the Matthews correlation coefficient (MCC). After feature extraction and target labelling, the SVC hyper-parameters were tuned using a grid search with fivefold cross-validation. Finally, we trained the tuned models using the complete training set (80% of the data) and tested the final models using the remaining 20 per cent of the data, which was purposely excluded from the cross-validation process.

ANNs to predict nocturnal hypoglycaemia

The occurrence of hypoglycaemic events during the night is a critical hazard in T1D management. During this period, patients are sleeping and may not be able to recognise falling glucose levels, which constitutes the main defence against severe hypoglycaemia. This recognition is based on a complex hormonal system, which alerts patients to the falling levels. ³¹ Nevertheless, subjects that suffer from impaired awareness of hypoglycaemia are unable to detect falling glucose levels and suitably take a correction action to initiate self-treatment, increasing the risk of severe episodes. ³²

ANNs simulate the human brain, by acquiring and maintaining knowledge based on information through an extensive parallel distributed process performed by single processing units, referred to as neurons. A multilayer perceptron network is one of the most widely employed ANNs for supervised learning and contains at least one layer that is hidden from both the input and output layers. Each layer is composed of a finite number of neurons, and each neuron is connected to every neuron in the next and previous layer. The output signal of a neuron given by the weighted sum of the inputs applied to a nonlinear activation function.^33,34

The nocturnal hypoglycaemia predictor is driven by an ANN-based module. The aim of this module is to alert patients concerning the occurrences of future hypoglycaemic events in the upcoming hours while patients are preparing to sleep. Patients must announce to the system that they are going to bed, and based on historical data, the system will return a message informing the patient if there is a high risk of low glucose levels during the next 6-h overnight period. Considering this information, patients are able to conduct self-treatment before the occurrence of hypoglycaemia, such as by reducing basal insulin delivery rates and/or consuming a snack without bolusing, to elevate glucose levels before sleep.

Input features were obtained from historical data on glucose, insulin, meals, and physical activity considering the time instant at which the patient announced that they were preparing to sleep (sleep announcement). Inputs related to glucose data were the CGM value at sleep announcement, hourly average of CGM readings over the last 6 h before sleep announcement, hourly area under the curve below 70 mg/dL of CGM readings over the last 6 h before sleep announcement, rate of change (ROC) of CGM readings during the previous 30 min before sleep announcement, low blood glucose index (LBGI) and high blood glucose index (HBGI) ³⁵ for the last 6 h before sleep announcement, and finally the BG difference between 60 and 30 min with respect with BG levels at sleep announcement. In addition, the values of the I O B , A O B , and remaining R a for meal consumption at sleep announcement were also considered as inputs for the ANN. The 6-h period following sleep announcement was used to assign the class of the respective day. During this period, Class 1 was assigned in case of Level 1 or Level 2 hypoglycaemia. Otherwise, Class 0 was assigned.

Individualised models were obtained for each patient. Exhaustive search was considered to find the best input features for each patient. As the dataset exhibits an unequal distribution between its classes, the adaptive synthetic sampling (ADA SYN) algorithm ³⁶ was employed to provide a balanced distribution during the training process. Due to the small amount of instances obtained from the dataset, fivefold cross-validation was applied in the entire dataset. This procedure was repeated 100 times to produce different splits in each repetition and reduce the influence of sample selection bias in the outcomes. Individualised results were obtained by averaging the outcomes over the 100 runs. The software MATLAB^® R2016b was considered to develop this methodology.

Patient condition assessment

The high variability between glycaemic profiles of patients makes tight glycaemic control a complex task. T1D treatments are governed by an inter-patient variability, which leads to a lack of models that can respond to the particularities of each patient. In addition, the so-called intra-patient variability makes it difficult to generalise models able to capture the glucose dynamics of a singular patient. Although some systems (see for instance Medtronic CareLink Pro/Personal^®) ³⁷ have developed some functionalities to analyse BG dynamics depending on events occurring at specific periods, the majority of scenarios with a high impact on BG values are not yet detectable. The implementation of a personalised DSS with an ability to identify scenarios such as exercise, menstrual periods, sick days, seasons, diet disturbances, or habits, among others, would allow the generation of tailored T1D prediction models that can assist in the control of BG levels.

A data mining approach based on the normalised compression distance (NCD) ³⁸ was applied to identify multiple scenarios for T1D patients. This module was implemented using a custom Java software based on the Lempel Ziv Markov compression algorithm for detecting redundancies and patterns in the content of a time series. First, we perform hierarchical clustering, in which the scenario instances are automatically and blindly grouped into a tree structure. Therefore, we consider a set of N time series to be grouped, and a similarity matrix (N × N measurements). The employed hierarchical clustering process constructs a binary tree from individual elements by progressively merging the clusters containing the two closest elements. This process is guided by the complete linkage method, ³⁹ which was selected for its compromise between simplicity and the potential for quality solutions. The initial clustering of the training dataset, which establishes the initial classes, involves 66 per cent of the dataset. Then, an NCD-based classifier is employed to identify the underlying profile class of the validation set (66%). The approach identifies unseen daily time series by computing the minimum distance between the pre-classified classes from the initial dataset and the new instance. The features selected to profile T1D scenarios consisted of the BG measurements and insulin doses.

Databases

Three different databases were used to assess the methods individually. The datasets were labelled using several thresholds according to the definitions of hypoglycaemia ² and hyperglycaemia. ²⁵ In both cases, an adverse glycaemic event was predicted whenever three consecutive BG measurements (15 min) were detected above or below their corresponding levels. ⁴⁰ The participants of the clinical studies related to the datasets 1 ⁴¹ and dataset 2 ⁴² provided written informed consent on data utilisation.

Assessment metrics

Each method was evaluated individually. Final classification results are presented using sensitivity, specificity, and accuracy (Table 2). In addition, MCC ⁴⁵ values and the Clarke error grid (CEG) percentages are reported. The MCC (Table 2) is a metric that merges the confusion matrix into a single value and is generally regarded as a balanced measure, even for imbalanced classification problems. The MCC index returns a value in the range of [ − 1 , 1 ] , where 1 is a perfect prediction and –1 is an erroneous classification. The CEG metric ⁴⁶ evaluates the clinical significance of the deviation of the estimated BG values. The CEG uses a Cartesian diagram divided in five regions. Predictions within regions A and B are clinically acceptable, while pairs in regions C, D, and E are potentially dangerous and are considered significant clinical errors.

OPEN IN VIEWER

Table 2.

Metrics used to assess each of the methodologies using metrics of a positive example classified as positive (TP), negative example classified as negative (TN), negative example classified as positive (FP), and positive example classified as negative (FN).

Performance metrics
Accuracy	T P + T N T P + F N + T N + F P
Sensitivity	T P T P + F N
Specificity	T N T N + F P
MCC	T P × T N − F P × F N ( T P + F P ) ( T P + F N ) ( T N + F P ) ( T N + F N )

MCC: Matthews correlation coefficient.

Mid-term continuous prediction

Table 3 presents results for all the databases considered in this article, presenting more than 96 per cent of the predictions inside regions A and B for the test data, which implies that the majority of the predictions were safe from a therapeutic point of view. Regarding the general distribution of the clinical harmfulness of the deviations, Table 3 presents that most of the errors out of zones A + B are concentrated in zone D. Despite always being under 4 per cent, estimations located in region D are highly undesirable because they mean that the predictions missed severe hypoglycaemia or hyperglycaemia events.

OPEN IN VIEWER

Table 3.

CEG analysis for the mid-term continuous prediction of BG excursions (PH = 60 min).

CEG zones (%)	A	B	C	D	E
Database 1	62.1	34.7	0.3	2.8	<0.1
Database 2	73.1	24.7	0.2	1.9	<0.1
Database 3	79.1	18.6	0.1	2.1	<0.1

CEG: Clarke error grid; BG: blood glucose; PH: prediction horizon.

This study assesses the GE method, by computing the performance metrics (defined in Table 2) based on the number of predicted Level 1 or Level 2 hypoglycaemic events. The average results for Database 3 (20 runs) are presented in Table 4.

OPEN IN VIEWER

Table 4.

Population outcomes for the mid-term continuous prediction module to predict hypoglycaemia.

Metric	Mean	Min	Max
Accuracy	86.1%	60.3%	97.9%
Sensitivity	48.5%	0%	98.2%
Specificity	93%	46.2%	100%
MCC	0.51	–0.18	0.86

MCC: Matthews correlation coefficient.

Postprandial risk assessment

The averaged results of the prediction models are displayed in Table 5. The results of the postprandial prediction strategy show the best MCC from the 10 patients corresponding to the Database. A sensitivity and specificity of 69 per cent and 80 per cent, respectively, for Level 1 hypoglycaemia, and 75 per cent and 81 per cent, respectively, for Level 2 hypoglycaemia.

OPEN IN VIEWER

Table 5.

Population outcomes considering the most favourable model of each patient for hypoglycaemia Level 1 and Level 2.

	Metric	Mean	Min	Max
Level 1	Sensitivity	69%	42%	82%
	Specificity	80%	73%	97%
	MCC	0.24	0.20	0.59
Level 2	Sensitivity	75%	60%	85%
	Specificity	81%	60%	98%
	MCC	0.24	0.20	0.68

MCC: Matthews correlation coefficient.

These results were obtained for the postprandial period, using a prediction window of 4 h after meal consumption and assuming a bi-class classification scheme.

Nocturnal hypoglycaemic events

To assess the capability of the ANN to predict nocturnal hypoglycaemic events, Database 2 was considered. This database was chosen because it is the only one in which physical activity signals and sleep information are available. Population results are presented in Table 6, showing the outcomes for the entire cohort.

OPEN IN VIEWER

Table 6.

Population outcomes for the nocturnal hypoglycaemic events predictor module.

Metric	Mean	Min	Max
Accuracy	80.1%	70.3%	94.9%
Sensitivity	44.0%	14.5%	67.5%
Specificity	85.9%	81.6%	96.3%

Patient condition assessment

Figure 1 presents the results of the last experiment performed in this study. The experiment uses the complete set of profiles which comprises a total of 280 daily time series from Database 3. The figure presents the three stronger connections of each node according to the NCD (discarding similarity values with themselves). The figure represents each time series with a circular node and, each profile with a green, blue, or grey shape (colours are just used to improve readability of the figure). The links between the time series of same profiles are coloured in black, while the connections between different profiles are coloured in red. The figure presents a total of 280 nodes with 840 connections. The 94 per cent of the edges are connecting nodes between same profiles while just 6 per cent of the links comprise an edge between different profiles. The graph shows that it is able to cluster majority of profiles, even isolating some of them completely. Although most of the profiles included in this graph are strongly connected with themselves, it also shows examples of more disperse clusters such as profile 8. Results indicate that the similarities between the days generated with different virtual models are much less frequent. By contrast, the approach is able to find a high level of similarity between the time series generated with the same virtual model. The classifier outcomes calculated for this study and Database 3 show a specificity of 92 per cent and sensitivity of 99 per cent.

OPEN IN VIEWER

Figure 1.

Clustering of glucose profiles. Each cluster represents a simulated T1D profile, each node a historical time series data from a type 1 diabetic pump.

Once the identification features of the NCD distance were demonstrated, we validate the approach with a classifier tool based on the NCD distance. A set of virtual profiles are used as source of knowledge to classify unseen time series. The classifier system uses 10 time series per profile as database and 4 time series aimed to validate the system. A K-fold cross-validation (K = 20) was used to select the items comprised in the sets. The classifier outcomes calculated for this study and Database 3 show a specificity of 92 per cent and sensitivity of 99 per cent.