Electronic disease surveillance system based on inputs from people with diabetes: an early outbreak detection mechanism.

Abstract

Objective: Generally, the purpose of this thesis project is to develop an effective electronic disease surveillance system, which is capable of detecting a cluster of people with elevated blood glucose (BG) levels within a specific region and timeframe by analyzing diabetes data. Specifically, we mainly focus on developing an early outbreak detection algorithm that can analyze BG data and detect individuals with an elevated BG level (aberrant patterns) using continues blood glucose measurement (CGM) and the mobile phone- based diabetes patients’ historical data – the diabetes diary. Material: This thesis project was conducted using data from two individuals with type-1 diabetes. The Dexcom continuous glucose monitoring device (CGM)) was used for the data collection. The collected data were CGM (in 5 minutes’ intervals) for a period of one month. We used these datasets to train and validate the developed system. After training and validating the system, for its goodness of fit to the individual BG dynamics, in the non-infection status of the two subjects using normal BG values, we tested our system with artificially simulated datasets, which resemble the individual BG evolution during infections. The simulated datasets were consisted of elevated or high BG values of varies size, duration and shape through a course of time, i.e. a week or more. It was simulated so as to resemble the elevated BG after one is infected, by considering various increments per minutes (∆BG/ (minutes (t))) and various durations of elevated BG. The system was developed using Matlab version R2015b. Method: We presented a system that is consisted of four modules: the data collection module, the blood glucose prediction, the outbreak detection, and the information dissemination and reporting module. There are two types of early outbreak detection approaches incorporated in the system, a type of statistical control (prediction interval-based) algorithm and a moving window based z-score process. The first approach, the prediction interval-based algorithm combined a novel mechanism for BG prediction, which is an interval prediction based on a set of autoregressive models and predicts the expected BG intervals for an individual with diabetes. The actual BG value is compared against the predicted intervals, which is generated using auto-regressive (AR) and Autoregressive moving average (ARMA) methods. We evaluated and compared the performance of these methods using the mean square errors (MSE) and root mean square errors (RMSE) functions. The second approach, the moving window based z-score process calculates a running mean and standard deviation based on a specific window size. The running mean and standard deviation are used to check the agreement of the current BG reading with the previous trend in the window. The performance of the process was evaluated based on the accuracy of detecting the specific surveillance case definition, i.e. sensitivity, specificity and positive predictive value (PPV). Result: Both the prediction interval-based algorithm and the moving window based z-score process were tested against the artificially simulated datasets and were capable of detecting statistically significant BG deviation of various sizes and durations. The prediction methods were capable of predicting the single step - ahead BG values with a reasonable accuracy, which were tested against validation datasets (unseen datasets during training). All the methods, autoregressive (AR), autoregressive (AR) with ratio of consecutive data as inputs, and autoregressive moving average (ARMA) have attained minimum root mean square errors (RMSE) for both subjects. However, the second methods predict well attaining the lowest RMSE for both subjects, which demonstrates the advantage gained through the use of ratio of consecutive data points rather than the raw blood glucose data. Moreover, we accurately monitored the BG fluctuations of both individuals with a significance level of α =0.01. However, there are difference in window size and RMSE attained by these subjects for a comparable interval width, where the first subject attained smaller than the second subject. In addition, for comparable detection capability, the size of the moving window used to calculate the z-score for the first subject is less than the second subject. This clearly shows the effect of personal behavior towards diabetes management on the detection capability, which is mainly due to the significant fluctuations in BG readings as a result of poor personal behavior in managing his/her diabetes. Conclusion: Generally, both of our early outbreak detection approaches have produced optimal detection results and were capable of detecting statistically significant BG deviation of various size and duration. However, considering flexibility, simplicity, computational time, and needs of computational power the moving window based z-score process is better than the prediction interval-based algorithm. Moreover, both the approaches are found to be affected by the quality of personal behavior towards diabetes management and this needs to be taken into account during large scale implementations. Besides, these results have clearly shown the effectiveness of the proposed approaches for detecting a cluster of people with similar patterns. . Consequently, after validating these approaches on a large scale basis, this promising results will hopefully lead the way for the development of the early outbreak detection system (prototype) based on inputs from people with diabetes, which is considered to be the next generation electronic disease surveillance system.