Diagnosis of hypoglycemic episodes using a neural network based rule discovery system
Research highlights
► Low blood glucose is dangerous for Type 1 diabetes mellitus patients (T1DM). ► A neural network based rule discovery system is developed to determine low blood glucose for T1DM. ► Satisfactory results can be obtained. ► Explicit information can be extracted from the neural network based rule discovery system.
Introduction
Episodes of hypoglycemia for Type 1 diabetes mellitus (T1DM) patients, which can result in unconsciousness, seizures or even death, are common and have serious side effect in insulin therapy (DCCT, 1993). Especially at night, at least 50% of all severe episodes of hypoglycemia occur during that time (Pickup, 2000) because usual insulin preparations do not adequately mimic the normal patterns of endogenous insulin secretion (Yale, 2004). However, it is impossible in practice to monitor the episodes of hypoglycemia by measuring the blood glucose levels around the clock. Based on the physiological parameters including heart rate and QT interval of ECG signal, the glucose levels of the T1DM can be determined and thus the episodes of hypoglycemia can be diagnosed (Harris et al., 1996).
The statistical regression method (Seber, 2003) is a common empirical approach to develop such classification models for various medical diagnoses such as diabetic nephropathy (Cho et al., 2008), acute gastrointestinal bleeding (Chu et al., 2008), pancreatic cancer (Chang & Hsu, 2009). However, statistical regression models are accurate over the range of the patients’ data in which they are developed. Therefore it can only be applied if the patients’ data is distributed according to the developed regression model, and the correlation between dependent and independent variables does not exist. If the patients’ data is irregular, the developed regression models have unnaturally too wide possibility range. Genetic programming (Smith et al., 2007, Gray et al., 1998) is another commonly used method to generate classification models for diagnosis purposes for examples in heart disease (Winkler, Affenzeller, & Wagner, 2009), and Parkinson’s disease (Subashini, Ramalingam, & Palanivel, 2009). In this approach, genetic operations are used to generate structures of classification models with nonlinear terms in polynomial forms, and then the least squares algorithm is used to determine the contribution of each nonlinear term of the model classification. However, it is unavoidable that patients’ data involves uncertainty, due to fuzziness of measures. Therefore the genetic programming together with the least square algorithm may not yield the best classification models for diagnosis purposes, since it does not consider the fuzziness of uncertainty in measures. Neural networks (Reggia & Sutton, 1998) have been used to develop classification models for medical diagnosis purposes. The advantages of using neural network approaches in diagnosis are their generalization ability in addressing both the nonlinear and fuzzy nature of the patients’ data. Although neural networks have been applied in building diagnosis models for various diagnoses like gastrointestinal disorders (Aruna, Puviarasan, & Palaniappan, 2007), abdominal pain (Mantzaris, Anastassopoulos, & Gardikis, 2008), urological dysfunctinos (Gil, Johnson, Chamizo, Paya, & Fernandez, 2009), dermatologic disease (Chang & Chen, 2009), breast cancer (Tanaka & Watada, 1998), heart disease (Das, Turkoglu, & Sengur, 2009), these diagnosis models only have the capability to transform the nonlinear or fuzzy patients’ data into simplified black-box structures in which no explicit knowledge or information can observed. Because of the black-box nature of the neural networks, some medical doctors may feel uncomfortable to use neural networks for diagnosis purposes even through the approaches may achieve better accuracy diagnosis than the other explicit modeling methods like classical statistical methods or genetic programming. This could also pose serious issues of one has to provide justification for one’s decision based on the implicit output of the neural network. Therefore this is essential to extract explicit information from the neural networks, so that the decision’s basis is explicit.
Recently neural fuzzy networks have been applied on modeling and classification based on patients’ data for medical diagnosis purposes in breast cancer (Oentaryo et al., 2008, Sim et al., 2006), prostate cancer (Keles, Hasiloglu, Ali, & Aksoy, 2007), heart disease (Kannathal, Lim, Acharya, & Sadasivan, 2006). The fuzzy neural networks model is constructed by distributing input and output relationships to the weights connecting neurons. The error value is limited to a reasonable level via sample training and used for modification of each weight value to acquire the final weight value for connection between neurons. The model of the system is constructed with these weight values. This approach is especially suitable for the construction of highly nonlinear models, and also fuzzy rules which contain certain information of the developed models can be generated (Lin, 2008). However, comparing with traditional network networks, more parameters need to be determined from the fuzzy neural networks, because not only parameters of the weights connecting neurons need to be determined but also the parameters inside the fuzzy rules needed to be determined. Therefore larger memory, more computational time and learning data are required than the ones required to develop neural networks. Even fuzzy rules which represent curtain information from the models can be generated, the domains of inputs and outputs represented by the fuzzy rules are all fuzzy. Medical doctors may find it difficult to make diagnosis decisions based on those fuzzy rules.
In this paper, a neural network based rule discovery system, which consists of a neural network based classification unit and a rule based extraction unit, is proposed to perform diagnosis of hypoglycemic episodes in T1DM patients. The neural network based classification unit is used for determining hypoglycemic episodes in T1DM patients using the specified physiological parameters, and a set of rules, which describe the domains of physiological parameters for which hypoglycemic episodes occur, is extracted from the neural network classification unit by a rule based extraction unit. Based on a set of training data collected from TIDM patients, a neural network based classification unit was developed by a genetic algorithm, which has a multi-objective fitness function with two goals. It maximizes the number of TIDM patients with hypoglycemic episodes diagnosed correctly with hypoglycemic episodes, and the number of TIDM patients in normal conditions diagnosed correctly in normal conditions. It also minimizes the number of TIDM patients with hypoglycemic episodes diagnosed wrongly under normal conditions and the number of TIDM patients under normal conditions wrongly diagnosed with hypoglycemic episodes. The neural network classification unit was validated by a set of testing data, and satisfactory results can be found. After the development of the neural network based classification unit, explicit rules were extracted by a rule discovery unit in which the explicit rules were generated based on a data set generated by the neural network classification unit. The explicit were validated by a set of testing data, and satisfactory results can be also found. The neural network based rule discovery system compensates for the limitation of neural network not providing explicit information.
This paper is organized as follows: Section 2 presents the proposed neural network based rule discovery system, which consists of a neural network based classification unit and a rule based extraction unit. Section 3 describes the nature of the data collected from the TIDM patients, and the presents the evaluation and validation results of the proposed neural network based rule discovery system. Finally a conclusion is given in Section 4.
Section snippets
Diagnosis of hypoglycemic episodes
Diagnosis of hypoglycemic episodes is essential for T1DM patients especially at night, largely because episodes of hypoglycemia are common while usual insulin preparations do not adequately mimic the normal patterns of endogenous insulin secretion (Yale, 2004). Based on the medical doctors’ experiences (Harris et al., 1996), the blood glucose levels of T1DM, y, which indicate whether the patients are hypoglycemia, are significantly related to several physiological parameters of which the three
T1DM’s data
The data is collected from 16 T1DM patients at 14.6 ± 1.5 years of age who volunteered to carry out the 10-h overnight hypoglycemia study at the Princess Margaret Hospital for Children in Perth, Western Australia, Australia. Each T1DM patient was monitored overnight for the natural occurrence of nocturnal hypoglycemia. HypoMon (Hypoglycaemia Monitor from AIMedics Pty, Ltd.) was used to measure the required physiological parameters, while the actual blood glucose levels were measured by a Yellow
Conclusion
In this paper, a neural network based rule discovery system is developed to determine the presence of hypoglycemic episodes based on the TIDM patients’ physiological parameters, rate of change of heart rate, corrected QT interval of electrocardiogram signal and rate of change of corrected QT interval. It was developed based on T1DM patients’ 420 data sets which were collected from 16 T1DM patients by using the genetic algorithm. 320 data sets were used to develop the neural network based rule
Acknowledgment
This works was supported by a grant from Juvenile Diabetes Research Foundation.
References (34)
- et al.
Diagnosis of gastrointestinal disorders using DIAGNET
Expert Systems with Applications
(2007) - et al.
Applying decision tree and neural network to increase quality of dermatologic diagnosis
Expert Systems with Applications
(2009) - et al.
The study that applies artificial intelligence and logistic regression for assistance in differential diagnostic of pancreatic cancer
Expert Systems with Applications
(2009) - et al.
Application of irregular and unbalanced data to predict diabetic nephropathy using visualization and feature selection methods
Artificial Intelligence in Medicine
(2008) - et al.
A decision support system to facilitate management of patients with acute gastrointestinal bleeding
Artificial Intelligence in Medicine
(2008) - et al.
Effective diagnosis of heart disease through neural networks ensembles
Expert Systems with Applications
(2009) - et al.
Application of artificial neural networks in diagnosis of urological dysfunctions
Expert Systems with Applications
(2009) - et al.
Neuro-fuzzy classification of prostate cancer using NEFCLASS-J
Computers in Biology and Medicine
(2007) The effect of knowledge sharing model
Expert Systems with Applications
(2008)- et al.
A non-symbolic implementation of abdominal pain estimation in childhood
Information Sciences
(2008)
GenSoFNN-Yager: A novel brain-inspired generic self-organizing neuro-fuzzy system realizing Yager inference
Expert Systems with Application
Sensitivity glucose sensing in diabetes
Lancet
Breast mass classification based on cytological patterns using RBFNN and SVM
Expert Systems with Applications
A hybrid decision tree/genetic algorithm for coping with the problem of small disjoints in data mining
Proceedings of Conference of Genetic and Evolutionary Computation
Cited by (60)
A Self-Attention Deep Neural Network Regressor for real time blood glucose estimation in paediatric population using physiological signals
2024, Biomedical Signal Processing and ControlEffect of internet of things on manufacturing performance: A hybrid multi-criteria decision-making and neuro-fuzzy approach
2022, TechnovationCitation Excerpt :This technique is used to prioritize and assess the relationship among variables of a system in order to solve problems that emerge from technology and human activity growth (Tseng, 2010; Chen and Chi, 2013). ANFIS is one of the most robust neural network systems (Chan et al., 2011). Petković et al. (2014) states that “ANFIS is about taking an initial fuzzy inference (FIS) system and tuning it with a back-propagation algorithm based on the collection of input and output data.”
Saturation magnetization parameters by adaptive neuro-fuzzy technique
2019, Physica A: Statistical Mechanics and its ApplicationsMobile cloud computing for ECG telemonitoring and real-time coronary heart disease risk detection
2018, Biomedical Signal Processing and ControlCitation Excerpt :In ANFIS, FIS is combined with a back propagation learning algorithm of neural network. In literature, ANFIS has been used in many studies for the extraction of the fuzzy rules from application models [20]. ANFIS has been used for ophthalmic arterial disorders using belief and plausibility functions [21].
Non-invasive hypoglycemia monitoring system using extreme learning machine for Type 1 diabetes
2016, ISA TransactionsCitation Excerpt :However, more parameters were need to be optimized and a large computational time is need for training. Further experiment were conducted by the use of neural network based rule discovery system [19]. The development was based on 420 data sets which were collected from 16 T1DM patients by using the genetic algorithm.