I am not an expert in neural networks, But in time series, you can of course predict. Typical example: you have several years of observed data (let say 50 data points) and you need to predict few next years. I believe it works even f you have few data points (you mention 3) but confidence interval will be very wide (uncertain prediction).

See also https://www.researchgate.net/post/What_should_be_the_minimum_number_of_observations_for_a_time_series_model

Hello Madhuri.

I think you don't have sufficient data to perform a model and run a time series prediction.

Hanke and Wichern, chapter 3, page 80 recommend a minimum 2xs to 6xs depending on the method (where s is the seasonal period, so s=12 for monthly data). 50 data points would be 50/12 = 4 years of data. But it depends on the regularity of the data. If the seasonal pattern is quite regular, 3 years is OK.

http://rads.stackoverflow.com/amzn/click/0132301202

http://stats.stackexchange.com/questions/119970/whats-the-minimum-sample-size-required-to-do-a-time-series-analysis

http://robjhyndman.com/hyndsight/short-time-series/

http://robjhyndman.com/papers/minimum-sample-size-requirements-for-seasonal-forecasting-models/

Values can be predicted using time series analysis or regression analysis by defining the relationship between variable, time or dependent & independent variables, in your case data  is not sufficient

In my opinion, you need a time series model that is able to give you the reaction of the glucose levels over time after each food intake. With data on glucose levels and food intakes (converted into grams of carbohydrats/fats), you should be able to estimate a vector autoregressive (VAR) model for each particular patient. Two problems:

1) The fact that your variables are not observed every second, or not even every hour may be a problem. The solution would be to use a state-space representation of your VAR model, which allows you to use the Kalman filter for the estimation.

2) You may have a very small dataset, as mentioned earlier, so don't take your results too seriously in that case. However, you may want to try using bayesian inference (e.g. Bayesian VAR, or BVAR for short) in combination with the previous point. Use the "theoretical" reaction function of glucose, which is probably well known in medicine, as prior belief. Then check whether the data for your patients yields a different reaction function that deviates from the theoretical one. In that case, your data may be informative after all!  

Thank You so much for responding me

I am taking the glucose data every 1 hour means i am having 72 glucose samples and i want to predict the glucose level for the next time stamp. i. e 73rd glucose level.

Hi Madhuri,

I have a bit experience with prediction with ANN (construction works execution time) but my mother suffer from diabetes, so she have to inject hersefl with insulin. I Wonder about applying ANN where the input would be:

- number of hours since the previous injection

- suger level in a blood measured last time

- present sugar level

- number of insulin units injected then

- phisical effort from the previous injection (low, middle, high)

- number of calories eaten since the previous injection

As an output - suggested number of insulinę units to be injected

Having 3-4 injection a day, the number of data set will be big enough to apply ANN in two weeks time.

It seems that according to the different age of the person (and her/his pancreas activity) prediction should be done of one person data set. Or... theese two factors should be included in the input set. I'm open for any cooperation (if the civil engineer can help).

Regards

Hubert